The Springer Series on Demographic Methods

and Population Analysis 38

Gordon A. Carmichael

Fundamentals of
Demographic Analysis:
Concepts, Measures
and Methods
The Springer Series on Demographic Methods
and Population Analysis

Volume 38

Series Editor
Kenneth C. Land, Duke University
In recent decades, there has been a rapid development of demographic models and
methods and an explosive growth in the range of applications of population analysis.
This series seeks to provide a publication outlet both for high-quality textual and
expository books on modern techniques of demographic analysis and for works that
present exemplary applications of such techniques to various aspects of population
analysis.
Topics appropriate for the series include:
• General demographic methods
• Techniques of standardization
• Life table models and methods
• Multistate and multiregional life tables, analyses and projections
• Demographic aspects of biostatistics and epidemiology
• Stable population theory and its extensions
• Methods of indirect estimation
• Stochastic population models
• Event history analysis, duration analysis, and hazard regression models
• Demographic projection methods and population forecasts
• Techniques of applied demographic analysis, regional and local population
estimates and projections
• Methods of estimation and projection for business and health care applications
• Methods and estimates for unique populations such as schools and students
Volumes in the series are of interest to researchers, professionals, and students
in demography, sociology, economics, statistics, geography and regional science,
public health and health care management, epidemiology, biostatistics, actuarial
science, business, and related fields.

More information about this series at http://www.springer.com/series/6449

Gordon A. Carmichael

Fundamentals
of Demographic Analysis:
Concepts, Measures
and Methods

123
Gordon A. Carmichael
School of Demography
Australian National University
Canberra, ACT
Australia

ISSN 1389-6784 ISSN 2215-1990 (electronic)

The Springer Series on Demographic Methods and Population Analysis
ISBN 978-3-319-23254-6 ISBN 978-3-319-23255-3 (eBook)
DOI 10.1007/978-3-319-23255-3

Library of Congress Control Number: 2015949814

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media (www.

springer.com)
Preface

This book has its origins in lectures prepared while teaching the course Principles of
Population Analysis (PPA) in the Demography Program of the Australian National
University (ANU) between 1989 and 1998. After a 1-year break I also taught
the course for a further 2 years in 2000 and 2001 after moving to the ANU’s
National Centre for Epidemiology and Population Health (NCEPH) in early 1999.
There it was taught in conjunction with a heavily overlapping NCEPH course titled
Population Analysis for Health Research. It seemed to me absurd that two courses
with such substantial common content had been taught side by side in the same
university, each to limited numbers of students, virtually throughout the 1990s, and
my argument that common sense recommended in the future teaching them together
was accepted by both my previous (Program) and my new (Centre) directors.
By the middle stages of my time employed by the Demography Program, what
began as lecture notes delivered to students orally and via a whiteboard, overhead
projector and occasional handouts had morphed into draft chapters of a book that
were distributed one by one as the course progressed. August 2001, however,
saw the arrival of a new director at NCEPH, and armed with an external review
recommending such action, he soon decreed that teaching a coursework master’s
degree that was regularly attracting low single-figure numbers of students to most
of its courses was an inefficient use of resources that could more profitably be
redeployed boosting the Centre’s research productivity.
My classroom teaching consequently ended, and PPA reverted to being the
responsibility of the Demography Program, subsequently to become first the
Demography and Sociology Program then the Australian Demographic and Social
Research Institute (ADSRI). My successor teaching PPA was a former student, Dr
Rebecca Kippen, who sought my permission to continue using my draft chapters as
core teaching material (she actually had them spiral bound for distribution as an in-
house ‘textbook’), and would thereafter encourage me towards formal publication
whenever she saw me – with the addition of a chapter on population projections
(excluded because projections were taught in a separate dedicated course). Rebecca
continued using my ‘book’ until she left the ANU in early 2010 to take up a future

v
vi Preface

fellowship at the University of Melbourne, but I was unable to justify spending

the time needed to bring it to publication while in the employ of NCEPH – a
demographic methods text was not core business for a Centre with a public health
focus.
Retirement in late 2011 put the project back on my agenda as I moved back to
ADSRI as an adjunct (i.e., honorary) associate professor. It was not my first priority
given the freedom I now enjoyed to do my own thing – I had a series of papers that
had also laid dormant because they were not public health-focused that I wanted to
write first. But in late 2013 ‘the book’ rose to the top of the ‘to do’ list, and during
2014 I have extensively reviewed and revised what I initially wrote many years ago
and have added the recommended chapter on projections.
The book chapters that developed out of my lecture notes were written in
the context of teaching groups of generally around 20 students, most of whom
were international students from a variety of developing countries. The master’s
programme in demography towards which my teaching was directed was at that
time domiciled in the ANU’s Development Studies Centre, which relied heavily
on international student fee income to fund the salaries of academics it directly
employed. An overriding characteristic of those classes was an enormous range
of quantitative aptitudes, with students ranging from a few with first degrees in
statistics to larger numbers with only the most rudimentary high school maths
training. The task was to pitch my teaching at a level that would cater to this
enormous range and in particular not lose those at its lower end. I also needed
to contend with the reality that most of my students had only just arrived in
Australia, and many had come from cultures where the teacher-student relationship
was distinctly hierarchical and were not attuned to questioning and interacting as
equals with a lecturer. I wanted them to ask questions and let me know if I was not
getting through to them.
A number of strategies were pursued to these ends. One was to take my own
tutorials, through which I could ensure via worked examples that key issues and
principles were reinforced and could personally field student questions. Another
was to offer optional extra tutorials for students who were having difficulty, at which
they could identify where their problems lay and set the tutorial agenda accordingly
and we could work through issues at their pace. But the major strategies in these
pre-‘PowerPoint’ days were to (i) develop teaching material that would obviate the
need for students to spend lectures furiously scribbling notes and enable them to
instead focus on understanding lecture content and (ii) present that teaching material
in a way that was as comprehensible as I could make it for those less at ease with
quantitative processes and concepts.
This book is the ultimate product of my efforts to make demographic analysis
more accessible to students with limited quantitative backgrounds, without being too
pedestrian for those with greater mathematical aptitude. Others will be the judges of
how successful I have been. It is intended to be an introductory text, leaving ample
scope for resort to additional texts and other sources pitched at higher levels with
students for whom such material is appropriate. The book has benefitted from the
time I spent teaching two courses as one, in particular through the addition of much
Preface vii

of the material in the section ‘Further Issues in the Analysis of Mortality’ at the end
of Chap. 4. This was introduced in more embryonic form to my teaching in 2000–
2001 because of its relevance to students of public health, but has been substantially
expanded in the process of revision and updating engaged in this year. Some of the
examples presented in the book may seem a little dated. This is a consequence of
their often having been compiled during the 1990s when I originally wrote much of
this material. Where it has seemed updating was essential (e.g., where trends were
being traced), I have taken the trouble to make the necessary revisions, but if an
example was a good illustration of a particular technique or principle, I have not
bothered replacing it with a more recent one merely to be more ‘up to date’. That
examples also frequently draw on Australian data reflects my having taught in an
Australian university; I no more apologize for this than previous authors of similar
texts I’ve read have for drawing on British or American examples with which they
were familiar.
Some of the material in this book has origins in teaching material prepared some
years ago by one-time colleagues Professor Peter McDonald (who taught me as a
PhD student in 1978) and the late Dr Alan Gray. I gratefully acknowledge their
inspiration.

Canberra, ACT, Australia Gordon A. Carmichael

Contents

1 Basic Sources, Concepts, Definitions and Types of Measures . . . . . . . . . . . 1

Demography, Demographic Analysis and Population Studies . . . . . . . . . . . . . . 1
Sources of Demographic Data .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
Population Censuses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
Sample Surveys . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 13
Vital Registration Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 15
Administrative Systems.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 17
Data Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 20
Absolute Measures: The Population Balancing Equation . . . . . . . . . . . . . . . . . . . 21
Relative Measures: Rates and Probabilities .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 23
Demographic Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 24
Demographic Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 28
A Digression: Concepts of Age, Incidence and Prevalence . . . . . . . . . . . . . . . . . 30
Refinement of Demographic Rates and the Effect of Age Structure .. . . . . . . 31
The Concept ‘Person-Years Exposed to Risk’ . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 36
The Principle of Correspondence .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
Other Types of Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 46
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 48
2 Comparison: Standardization and Decomposition . . .. . . . . . . . . . . . . . . . . . . . 49
Population Composition and Comparison of Summary Measures .. . . . . . . . . 49
Making More Meaningful Comparisons of Summary Measures . . . . . . . . . . . 54
Standardization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 55
Direct Standardization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 56
Indirect Standardization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 68
Some Additional Points About Standardization .. . . . .. . . . . . . . . . . . . . . . . . . . 72
Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
Decomposition by Reverse Subtraction . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 78
Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 84

ix
x Contents

3 The Cohort and Period Approaches to Demographic Analysis .. . . . . . . . 85

Conceptualizing Problems in Demographic Analysis: Lexis Diagrams . . . . 85
Time and Its Straight Line Representation .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 85
The Lexis Diagram.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 88
The Lexis Diagram in Operation . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 96
Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 96
Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 97
Example 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 100
Postscript to Example 3 .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 107
Issues Pertaining to Cohort Demographic Processes . . . .. . . . . . . . . . . . . . . . . . . . 109
The Tempo and Intensity of Cohort Processes . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
Renewable and Non-renewable Events . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
Attrition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 112
Period Analysis and Synthetic Cohorts . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 115
The Tendency of Synthetic Cohort Measures to Exaggerate Change . . . 118
Other Issues in Period Analysis . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 125
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 127
4 Analysis of Mortality: The Life Table and Survival. .. . . . . . . . . . . . . . . . . . . . 129
Introductory Matters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 129
The Life Table: A General Perspective .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 130
The Single-Year-of-Age Life Table . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 132
Obtaining Age-Sex-Specific Death Rates to Construct a Life Table . . . . 133
Components of a Life Table . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 134
Generating a Single-Year-of-Age Life Table: The 1 qx Column.. . . . . . . . . 135
Obtaining 1 q0 : Separation Factors. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 139
Generating a Single-Year-of-Age Life Table: The lx , 1 dx
and 1 px Columns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 145
Generating a Single-Year-of-Age Life Table: The 1 Lx , Tx
and eo x Columns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 148
Generating a Single-Year-of-Age Life Table: Other Issues . . . . . . . . . . . . . . 151
Recipe for Constructing a Single-Year-of-Age Life Table .. . . . . . . . . . . . . . 157
Abridged Life Tables .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 158
Calculating Abridged Life Tables from Single-Year-of-Age
Life Tables.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 161
Generating an Abridged Life Table: Estimating Values of n qx . . . . . . . . . . 161
Generating an Abridged Life Table: Estimating Other Life
Table Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 172
The Shapes of Life Table Functions . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 173
Using Life Tables: Survival .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 176
Population Survival Ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 176
Survival Ratios from Birth. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 178
Life Table Survival Ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 180
Relationship Between Population and Life Table Survival Ratios . . . . . . 182
Estimation of Births. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 184
Contents xi

Estimation of Intercensal Migration . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 188

Survival Between Birthdays (Exact Ages): Joint Survival .. . . . . . . . . . . . . . 191
Further Issues in the Analysis of Mortality . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 192
Causes of Death .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 192
Analysis of Differential Mortality .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 195
Perinatal Mortality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 195
More Sophisticated Life Tables . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 196
Years of Potential Life Lost (YPLL) .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 200
Focusing on the Quality of Life: Health Expectancies .. . . . . . . . . . . . . . . . . . 201
Disability-Adjusted Life Years (DALYs) and the Burden of Disease . . . 203
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 210
5 Marriage, Marital Status and Relationships . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 213
Why Study Marriage? .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 213
Defining Marriage .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 214
Some Other Features of Marriage Systems . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 218
Marital Status .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 219
General Measures of the Marriage Process . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 221
First Marriage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 222
Measures of the Intensity of the First Marriage Process . . . . . . . . . . . . . . . . . 222
Measures of the Timing of the First Marriage Process .. . . . . . . . . . . . . . . . . . 226
Nuptiality Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 230
Gross Nuptiality Tables .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 231
Net Nuptiality Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 234
The Singulate Mean Age at Marriage . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 240
Consensual Partnering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 243
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 245
6 Analysis of Fertility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 247
The Importance of Fertility Analysis in Demography .. .. . . . . . . . . . . . . . . . . . . . 247
Some Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 248
Basic Fertility Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 249
Fertility Measures from Censuses and Surveys .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . 254
Measures Based on Children Ever Born . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 254
‘Current Fertility’ Measures .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 256
The Child-Woman Ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 256
Measures of Reproduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 257
The Gross and Net Reproduction Rates. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 257
Mean Length of Generation . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 260
Describing the Age Pattern of Fertility . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 261
Graphs of Age-Specific Fertility Rates . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 262
Measures of the Average Age of Fertility .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 266
Natural Fertility and Associated Fertility Models . . . . . . .. . . . . . . . . . . . . . . . . . . . 267
The Coale-Trussell Fertility Model . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 268
The Coale (or Princeton) Fertility Indices . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 270
xii Contents

Other Fertility Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 274

The Brass Relational Gompertz Model . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 274
The Bongaarts Fertility Model . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 278
Analysis of Birth Intervals .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 284
Parity Progression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 287
Biological Aspects of Fertility .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 293
Fecundability .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 293
Post-partum Amenorrhoea.. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 295
Sterility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 296
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 297
7 Population Distribution, Urbanization and Migration . . . . . . . . . . . . . . . . . . 299
Demography and Geography . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 299
Population Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 299
Levels at Which Data Are Tabulated .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 300
Urban and Rural Populations .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 301
Methods of Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 303
Urbanization .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 308
Methods of Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 309
Counterurbanization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 313
Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 314
Mobility and Migration .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 314
Data and Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 315
Demographic Perspectives on Migration . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 318
Measuring International Migration.. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 318
Measuring Internal Migration . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 327
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 342
8 Stable Population Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 343
Stationary and Stable Populations .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 343
Generating Stable Population Measures . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 344
The Intrinsic Rate of Natural Increase . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 344
The Intrinsic Birth and Death Rates. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 346
The Stable Age Distribution .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 347
Applications of Stable Population Theory . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 350
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 351
9 Population Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 353
The Bread and Butter of Demography . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 353
Population Projections and Population Forecasts . . . . . . . .. . . . . . . . . . . . . . . . . . . . 354
Some Other Features of Population Projections . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 355
Approaches to Population Projection.. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 358
Trend Extrapolation Projections .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 361
Simple Extrapolation Methods . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 361
Complex Extrapolation Methods .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 362
Ratio Extrapolation Methods .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 364
Contents xiii

The Cohort-Component Method of Population Projection .. . . . . . . . . . . . . . . . . 365

The Projection Process: A Simplified Model .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 367
Introducing Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 374
Projection Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 379
MortPak for Windows (Version 4.3) . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 379
DemProj . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 380
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 385

Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 387
Chapter 1
Basic Sources, Concepts, Definitions
and Types of Measures

Demography, Demographic Analysis and Population Studies

What is ‘demography’? A short answer, sufficient perhaps to quickly satisfy the

lay inquirer, is ‘the study of human populations’. More elaborate definitions are,
however, to be found in the literature. Several are reproduced in Table 1.1, where
page references make it clear that it is common for books like this one to commence
by addressing the question just posed. Some definitions are more restrictive than
others. The Belgian Achille Guillard is credited with having first coined and defined
the term ‘demography’. Of the more recent definitions cited, that attributed to the
IUSSP advances us a little beyond ‘the study of human populations’, but remains
general and non-specific. Hauser and Duncan provide more detail, introducing
the notion of ‘territorial distribution’ to provide a touchpoint with geography and
itemizing four processes through which populations change over time. Bogue asserts
the ‘statistical and mathematical’ nature of demography, adds ‘marriage’ as a fifth
process generating change and claims theory-building as a long-run disciplinary
goal. Shryock, Siegel and Associates in their encyclopaedic two-volume survey of
demographic techniques offer that demography may be defined either ‘narrowly or
broadly’, whilst Wunsch and Termote explicitly acknowledge only three processes
of change, although their second sentence broadens the agenda á la Shryock et al.
Weeks’s definition is a real catch-all, that of Hinde highlights future prediction as
an important element of demography, Preston, Heuveline and Guillot are succinct
(perhaps too succinct to be very helpful), and Weinstein and Pillai emphasize that
demographers deal with aggregates of living individuals. Siegel and Swanson’s
definition revises that of Shryock, Siegel and Associates in the earlier edition of
The Methods and Materials of Demography.
Clearly the boundaries of demography are fuzzy. There is, however, general
agreement that the discipline has a core, variously referred to as demographic
analysis or formal demography, which is concerned with the measurement of popu-

© Springer International Publishing Switzerland 2016 1

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_1
2 1 Basic Sources, Concepts, Definitions and Types of Measures

Table 1.1 Some definitions of demography

Guillard (1855) cited in Shryock, Siegel and Associates (1973: 2)
[Demography is] the natural and social history of the human species or the mathematical
knowledge of populations, of their general changes, and of their physical, civil, intellectual, and
moral condition.
International Union for the Scientific Study of Population (1958: 3)
Demography is the scientific study of human populations, primarily with respect to their size,
their structure and their development.
Hauser and Duncan (1959: 31)
Demography is the study of the size, territorial distribution, and the composition of population,
changes therein, and the components of such changes, which may be identified as natality,
mortality, territorial movement (migration), and social mobility (change of status).
Bogue (1969: 1–2)
Demography is the statistical and mathematical study of the size, composition, and spatial
distribution of human populations, and of changes in these aspects through the operation of the
five processes of fertility, mortality, marriage, migration and social mobility. : : : [Its] long-run
goal is to develop a body of theory to explain the events that it charts and compares.
Shryock, Siegel and Associates (1973: 2)
Demography is the science of population [and] may be defined narrowly or broadly. The
narrowest sense : : : [deals] with the size, distribution, structure, and change of populations. : : :
A broader sense includes : : : ethnic, social, and economic characteristics. : : : The widest sense
: : : extends to : : : problems : : : related to demographic processes: the pressure of populations
upon resources, depopulation, family limitation, eugenics, the assimilation of immigrants,
urban problems, legislative apportionment, manpower, and the maldistribution of income.
Wunsch and Termote (1978: 1)
Demography is the study of population, its increase through births and immigration, and its
decrease through deaths and emigration. : : : demography is also the study of the various
determinants of population change and of the impact of population on the world around us.
Weeks (1994: 4)
Demography is concerned with virtually everything that influences, or can be influenced by
population size, distribution, processes, structure, or characteristics.
Hinde (1998: 1)
Demography is the study of population structure and change. With the increasing complexity of
modern society, it is becoming ever more important : : : to make estimates of : : : the future
size and composition of the population. Demographers : : : carry out this task.
Preston et al. (2001: Preface)
[D]emography [is] the study of population processes. [It studies] the behavior of human
populations.
Weinstein and Pillai (2001: 5)
Demography [is] the science of population : : : its chief concern is with the characteristics of
aggregates of living individuals. : : : They are born, they age, they move from place to place,
they reproduce, they become ill, and they die.
Siegel and Swanson (2004: 1)
Demography is the scientific study of human population, including its size, distribution,
composition, and the factors that determine changes [therein]. [It] focuses on five aspects of
human population: (1) size, (2) distribution, (3) composition, (4) population dynamics, and (5)
socioeconomic determinants and consequences of population change.
Sources of Demographic Data 3

lation phenomena, and a broader interpretative component, often termed population

studies, which seeks to understand and explain the patterns, differentials and trends
revealed by demographic analysis and to assess their implications and linkages
to other phenomena. Population studies tends to focus on relationships between
demographic and non-demographic variables, and to the extent that demographers’
views as to the scope of their discipline vary, they vary largely in their perceptions
of the breadth of the field of population studies. You may also encounter the terms
mathematical demography, social demography and anthropological demography.
The former refers to the application of formal mathematics to the analysis of
demographic change and relationships; the latter two terms to study of the social
and cultural underpinnings of demographic levels, trends and differentials.
Population studies is unambiguously interdisciplinary, overlapping with fields
like sociology, anthropology, psychology, economics, geography, history, epidemi-
ology and medicine. Indeed many specialists in population studies develop their
interests following initial trainings in one of these other fields, and claim allegiance
to that field in addition to their allegiance to population studies. Many academic
demographers are in fact employed by university Schools or Departments of
Sociology or Geography. But demographic analysis is also interdisciplinary, having
obvious overlap with mathematics and statistics and sharing analytic techniques
with the likes of biology, epidemiology and medicine. The strongly interdisciplinary
nature of both population studies and demographic analysis makes demography,
as the discipline that embraces both of them, a particularly interesting, fertile and
useful (i.e., policy-relevant) area of academic and intellectual endeavour.
This is a book about demographic analysis; about the technical core of demogra-
phy upon which population studies depends. A prerequisite to demographic analysis
is demographic data: that is, counts of the populations of geographic areas; data on
people’s demographic characteristics (age, sex, marital status, labour force status,
etc.); and counts of demographic events (births, deaths, marriages, migratory moves,
etc.). We will now examine the major sources of such data, and the strengths and
drawbacks of each.

Sources of Demographic Data

There are four major sources of demographic data: population censuses, sample
surveys, vital (sometimes termed ‘civil’) registration systems which in a few
countries are embedded in more extensive population registers, and a variety
of administrative systems which yield demographic data as by-products of the
administrative functions they serve.
4 1 Basic Sources, Concepts, Definitions and Types of Measures

Population Censuses

Population censuses, usually although not always also incorporating housing cen-
suses, are conducted with varying degrees of regularity in most countries. According
to the United Nations Statistics Division (2013), countries and territories that failed
to conduct a population census during the decade 1995–2004 were Angola, Burundi,
Cameroon, Chad, DR Congo, Djibouti, Eritrea, Ethiopia, Guinea-Bissau, Liberia,
Madagascar, Mayotte, Nigeria, Somalia, Sudan, Togo, Western Sahara, El Salvador,
Colombia, Peru, Afghanistan, Bhutan, DPR Korea, Lebanon, Myanmar, Uzbekistan
and Bosnia-Herzegovina. As of mid-2013 seven countries had yet to set a census
date for the decade 2005–2014 – Eritrea, Somalia, Western Sahara, Iraq, Lebanon,
Pakistan and Uzbekistan. Another 17 had set only a vague date (e.g., 2013) that sug-
gested they were in grave danger of not holding a census, or had passed a nominated
date without a census having occurred – Benin, Central African Republic, Comoros
Islands, DR Congo, Equatorial Guinea, Gabon, Gambia, Guinea, Madagascar,
Sierra Leone, Guatemala, Haiti, Honduras, Afghanistan, Syria, FYR Macedonia
and Ukraine. From these lists it is easy to appreciate that among the forces that
discourage census taking are smallness, poverty and internal political strife. Small
countries sometimes have alternative sources of demographic data, or are so sparsely
populated that there is limited demand for such data. Governments of poor countries
may simply have other priorities for expenditure of limited economic resources and
may lack the administrative infrastructure to support a census. Strife-torn areas tend
to present insurmountable practical obstacles to census taking, in addition to which
the population may be abnormally mobile, substantially geographically displaced
and have sections intent on not being located by government representatives. The
government in turn may have other priorities in terms of expenditure and occupying
the time of civil servants. There is, in addition to countries like those listed above,
a range of other, mostly European, countries that also do not conduct full censuses
because they maintain population registers that either on their own, or in conjunction
with other sources such as sample surveys, yield equivalent data. These types of
sources are dealt with later in this chapter.
A census aims at complete coverage of a population, although sometimes an
element of sampling intrudes, with some respondents being asked a broader range
of questions than others. Information is sought about each individual present within
a defined geographic area at a defined point in time, with some attempt perhaps also
being made to obtain data on persons ‘usually resident’ in the area but temporarily
absent from it. Population censuses have several significant disadvantages. They
are expensive, which along with the long lead-time needed to plan them and the
time required to process and disseminate data after collection limits their frequency.
Australia is one of a minority of countries (Canada, New Zealand, Japan, South
Korea, Hong Kong SAR, and several small Pacific nations are others) to conduct
censuses as often as every 5 years, although even in these countries the unexpected
can intervene. The scheduled 2011 New Zealand Census, for example, was deferred
for 2 years because the Christchurch earthquake occurred a fortnight before census
Sources of Demographic Data 5

night, devastating the offices from which it was being administered. The more usual
intercensal interval is of the order of a decade, which means that the timeliness of
census data is frequently less than ideal. Some countries attempt to address this
problem by conducting intercensal surveys, but while these can pinpoint major
changes at national and regional levels they tend not to meet the needs of local
area planners and those interested in small minority groups within the population.
The expense of population censuses contributes as well to their characteristically
limited content. The huge number of respondents means that the number of
questions must be kept in check so as to contain costs of stationery, printing,
coding, data capture and, if the census is not self-enumerated, interviewing. Modern
computer technologies, such as the Intelligent Forms Processing (IFP) software used
for the 2001, 2006 and 2011 Australian Censuses, can facilitate major savings on
data processing in countries equipped to use them (Australia also made savings
in 2011 through 28 % online submission of census forms, up from 9 % in 2006,
the first year with an ‘e-census’ option), but the case for restricting census content
relies only partly, and secondarily, on cost. Censuses are often self-enumerated; the
questionnaire is dropped off, the respondent or a household representative on behalf
of all household members answers the questions, and a fieldworker collects the
completed form. Given this system, gathering data of acceptable quality necessitates
limiting respondent burden, so as not to give rise to unacceptable levels of non-
response or inaccurate response. An important element of respondent burden is the
length of the questionnaire. If perceived to be too long or too time-consuming
respondents may avoid answering it altogether, stop answering it part way through,
or become careless or even frivolous with their answers. A second important element
is the complexity of questions asked. Especially when designing a self-completion
census it is essential to appreciate the diverse intellectual capabilities to be catered
to, and to keep questions simple so that they can be coped with at the lower end
of this intellectual range. Inability to understand a question obviously adversely
affects response to that question, but may also contribute to a rapid drop in a
person’s enthusiasm for cooperating with the census in a more general sense. Thus,
demographers and other users have to accept that a self-enumerated census is not a
vehicle through which to ask questions of the complexity and depth one might ask
in an interviewer-administered sample survey. Nor is it a vehicle through which
to ask sensitive questions. Again, data quality at the levels of both individual topic
and the census as a whole can be undermined by questions which are perceived to
be offensive, intrusive, or to rekindle unpleasant memories (e.g., a question on the
number of children ever born to women aged 15 or older often asked in Australian
censuses has always explicitly targeted live births only, in recognition that to ask
respondents to include stillbirths, or imply that they should be included, could be
distressing).
As a result of the considerations just discussed, population censuses typically
ask a limited range of core questions that enable the basic structure and com-
position of national, regional and other geographically defined populations to be
established. Attributes covered in most censuses include age, sex, marital status,
birthplace and/or ethnic group, level of education and occupation. Other questions
6 1 Basic Sources, Concepts, Definitions and Types of Measures

frequently aim at capturing household composition and/or family type, childbearing

experience, religion, income, and any recent change in place of usual residence
(typically within the last year and/or the last 5 years). In countries without reliable
vital registration systems, considerable emphasis may be placed on censuses gaining
insight into fertility and mortality levels and trends. In general, they have been
less effective in providing data on mortality than on fertility. Death is a sensitive
topic, and for that reason alone difficult to collect direct information on. It also
not infrequently leads to the breakup of a household, survivors dispersing to other
locations, including other countries. This may lead to a death being attributed to
the wrong (regional) population, to its not being reported at all, or to multiple
reporting where the responsibility for reporting is unclear. In addition there may be
no survivor to report a death; or defective recall may create uncertainty over whether
such an event occurred within a nominated reference period (e.g., the preceding
12 months). There are, however, less direct census-based approaches to estimating
mortality levels available. One involves comparing the reported age-sex structure of
the population with those of what are known as model stable populations, while
estimates of child and adult mortality can respectively be made by questioning
mothers about the survival of their children, and people about the survival of their
parents or first spouses.
The tenor of this discussion has been rather negative, stressing the limitations
of population censuses. Obviously they also have positive attributes, otherwise they
would never take place. Above all, population censuses are head counts, and as
such have administrative utilities that in politicians’ eyes normally outweigh their
importance as data sources for demographers. In many democracies the census
is the basis upon which electoral boundaries are regularly revised, and simple
population counts may play a major role in the distribution of resources among
competing geographic areas. In Australia, for instance, constitutional and legislative
requirements concerning electoral redistributions and the allocation of Federal tax
revenues (in particular GST consumption tax revenues) to State and Territory
Governments are major reasons for conducting population censuses, especially as
regularly as 5 years apart.
The complete coverage aimed at by population censuses means that, provided
they come close to achieving that objective, reliable data can be obtained for
small geographical areas and small subgroups of the population. Small-area data
are indispensable to effective local area planning (determining where to locate
or close down schools, where the demand for various community services is
concentrated, etc.), and provide crucial input to the location decisions and marketing
strategies of businesses. Indeed, data at small-area and higher subnational levels
of geographic aggregation are the ‘bread and butter’ of applied demography, a
phrase coined to encompass analyses of demographic data that are intended first
and foremost to inform decisionmaking. The ability to focus on small subgroups
(e.g., small immigrant and ethnic groups) is a feature of population censuses that
is very attractive to academic demographers as well as to planners. These are
groups that in any general sample survey of modest size would not be sufficiently
well represented for reliable conclusions about them to be drawn, and whose
Sources of Demographic Data 7

whereabouts often cannot otherwise be adequately pinpointed for more specialized

survey work. Potentially, an attractive feature of censuses is their capacity to
provide sampling frames for such survey work, but in practice the need to strictly
enforce confidentiality assurances in order not to jeopardize the overall integrity of
the census tends to make authorities extremely reluctant, if not totally unwilling,
to permit this sort of application of census data, especially by non-government
agencies.
Complete coverage also means that detailed cross-tabulations can be prepared
which are not subject to sampling error. Moreover, in the event that some questions
in a census are not asked universally, the size of the sample typically is so large that
relevant data items are subject to very low sampling errors.
Population censuses are the primary source of the denominators required by
a host of demographic rates and ratios. Countries with reliable vital registration
systems could not optimally exploit them without regular censuses, many demo-
graphic indices combining a ‘number of events’ numerator from a vital register with
a ‘population at risk’ denominator from, or based upon, a census. Often these indices
can be estimated using survey data as well, the survey yielding both numerator and
denominator. But the ideal is reliable registration data combined with appropriate
census data. For census years the census may directly yield denominators, perhaps
with a small adjustment depending on the timing of the census during the year,
since denominators usually need to be mid-year populations at risk. For other years
denominators normally have a prior or subsequent census base, adjusted for vital
and other (especially migration) events that modified that base in the interim.
Finally, successive censuses are invaluable for the historical record of a pop-
ulation they build up, the capacity they provide for focusing on demographic
change, and the insights into the nature of demographic processes which studies
of intercensal change can provide. Censuses as a historical record obviously have
broad academic interest, as Fig. 1.1, showing census-based age-sex pyramids for
Australia between 1861 and 2011, attests (those before 1911 amalgamate colonial
census data for New South Wales, Victoria, Queensland, South Australia, Western
Australia and Tasmania). Age-sex pyramids are a graphical device in which males
are conventionally represented on the left and females on the right, with age groups
(typically, but not necessarily, 5-year age groups) ranging from the youngest at the
pyramid base to the oldest at its apex. Pyramids shown in Fig. 1.1 have male and
female surpluses in age groups distinctively shaded for emphasis. This assemblage
of age-sex pyramids captures the development of a population that initially bears
the imprint of male-dominated gold rush immigration during the 1850s through a
structure still displaying remnants of that era but also, at younger ages, evidence
of a more settled, but high fertility, population in 1881 to one that in 1891 had
experienced a new influx of immigrants aged in their late twenties and early
thirties. Undercutting of the base of the 1901 pyramid highlights depression-induced
fertility decline through the 1890s, while by 1911 the base is broadening again,
signifying recovery. The 1933 pyramid then reflects the effect on fertility of rampant
unemployment during the Great Depression following the Wall Street stockmarket
crash of October 1929. By this time the male excesses of population at most ages
8 1 Basic Sources, Concepts, Definitions and Types of Measures

Fig. 1.1 Age-sex pyramids for Australia: selected censuses 1861–2011

Sources of Demographic Data 9

Fig. 1.1 (continued)

10 1 Basic Sources, Concepts, Definitions and Types of Measures

Fig. 1.1 (continued)

Sources of Demographic Data 11

Fig. 1.1 (continued)

12 1 Basic Sources, Concepts, Definitions and Types of Measures

have also become much more muted, the heavily male-dominated immigrant cohorts
of the nineteenth century having largely died out. Age-sex pyramids for 1947 and
1961 are then notable for three features: development of the post-war Baby Boom
bulge; the re-emergence due to immigration of a male surplus at young adult ages;
and the appearance of a marked excess of females at older ages as female longevity
increased more rapidly than did that of males. By 1971 the first hint of renewed
fertility decline appears at the pyramid base, oral contraception having by then been
available for a decade, and this has become more pronounced by 1981. Note also
that male excesses, reflecting the biological reality of a male-dominated sex ratio at
birth, continue to prevail at childhood ages. By 1996 the Baby Boomers are in their
thirties and forties, the pyramid has a narrower base, and female population excess
at the oldest ages has become more pronounced. Ageing of the population is even
more marked by 2011, by which time female population excesses are evident at all
ages 30–34 years and above. There is also the hint of a new fertility resurgence,
possibly temporary and reflecting the short-term impact of pronatalist government
policies dating from 2004.
The historical record censuses provide may also have very practical significance,
for example in allowing the demographic evolution of small geographic areas to
be traced with a view to predicting future change and planning accordingly. Some
suburbs within cities, for example, are initially rapidly settled by young families
and age through subsequent censuses with those families. Others maintain more
stable demographic profiles through constant turnover of population in well defined
categories (single young adults; young families occupying rental accommodation
prior to acquiring their own homes; etc.). Still others suddenly change character
through gentrification, invasion by minority groups, etc. Each of these scenarios has
distinctive planning implications.
As data from a new census become available a priority of many users is to
establish change since the previous census. Often there is strong suspicion of change
in a particular direction, the census being relied upon to confirm that suspicion and
measure the extent of the change. The use of census data to study change ideally
requires that at each date censuses be based on the same or equivalent questions,
coding schemes and field procedures. For those who design censuses the lesson
here is to carefully evaluate the impact on data comparability of making alterations
in these areas. If an adverse impact is envisaged, advantages expected in terms of
better quality data on a topic, cost savings, etc. should be weighed against the loss of
temporal comparability. For users of census data the lesson is to always check for
intercensal changes to question format, coding conventions and field procedures.
Otherwise it is easy to find oneself inventing fantastic explanations for spurious
trends.
Sources of Demographic Data 13

Sample Surveys

A sample survey involves questioning a small, ideally randomly selected, sample

of a population, although samples may also be stratified to overrepresent groups
(strata) that otherwise would not have sufficient numbers for separate analysis.
Biases introduced to the sample by stratification can be offset by weighting survey
data. Surveys may be cross-sectional (i.e., relate to a set point in time), longitudinal
(reinterview the same respondents at several points in time), retrospective (collect
data about past events – life histories), or prospective (record information about
events as they happen over time). They often focus on only parts of a population
(e.g., persons aged 65 and over; women of reproductive age (15–49)), and this
capacity to focus on population subgroups of special interest is a major attraction.
It allows questionnaire design to concentrate on probing a subgroup’s character,
behaviour and attitudes in-depth. Another attraction of surveys relative to censuses
is their comparative cheapness, making them affordable for agencies besides those
of national governments.
Unless mounted by government agencies empowered to compel response,
sample surveys can be biased by non-response. Among those who do respond,
however, data quality is easier to control in surveys than in censuses. Higher
data quality is not, though, automatic; indeed survey data quality is extremely
variable, and it is vital to always check and evaluate quality control procedures
followed before analysing a survey dataset. But whereas censuses frequently rely
on self-enumeration, surveys, while they can also use self-enumeration, including
these days online enumeration, more often than not use face-to-face or telephone
interviews. If interviewers are well trained (and assessing this is part of an evaluation
of quality control procedures), they should be able to establish a rapport which
maximises the chance of a respondent treating the questionnaire seriously, guide
respondents to correct interpretations of questions, probe for accurate answers to
questions respondents are unsure of the answers to, recognize and seek resolution
of inconsistencies in answers to related questions, and so on. Because of this quality
control it is common for surveys to in fact be used to check the accuracy of census
and even vital registration data. Thus census agencies often run ‘post-enumeration’
surveys with this objective in mind, and indeed may use surveys to test proposed
new questions before deciding whether or not to include them in a census.
As a corollary largely of the quality control trained interviewers bring to a sample
survey, but also of the capacity to define one’s target group and topic more narrowly,
surveys allow much more detailed data to be gathered than do censuses. Two
of the better known demographic surveys (or, more strictly, groups of surveys)
have been the World Fertility Surveys (WFS) and the Demographic and Health
Surveys (DHS). The former (Cleland and Scott 1987) saw comparable sample
surveys of women aged 15–49 carried out in 66 countries between 1973 and 1984.
Most (44) were developing countries, but several large countries (India, China,
Brazil and the Soviet Union) declined to participate. Very detailed information was
collected on pregnancy histories, the survival of children, knowledge and practice
14 1 Basic Sources, Concepts, Definitions and Types of Measures

of contraception, and potentially related background social and economic variables.

The DHS program, initiated in 1984 and continuing to this day, has sponsored
around 300 surveys in over 90 developing countries to collect data on fertility,
reproductive and maternal health and mortality, child health, survival and immuniza-
tion, HIV/AIDS, malaria and nutrition. Most have targeted women or ever married
women aged 15–49, although a few, chiefly in Africa, have also targeted husbands.
With both the WFS and the DHS considerable resources have been devoted to
developing, evaluating and improving survey methodologies and instruments, with a
view to producing high quality data. Another survey program focused on providing
comparative data on demographic change, this time in developed countries since
the mid-1960s (the ‘second demographic transition’ as it is often referred to), was
the series of Fertility and Family Surveys (FFS) sponsored by the United Nations
Economic Commission for Europe. Designed to facilitate analysis of recently
emergent demographic trends using techniques known as ‘event history analysis’,
these surveys were carried out in 24 chiefly European countries during 1988–1999
(Canada, New Zealand and the USA also participated) (Festy and Prioux 2002).
Individual countries (especially more developed countries) also have their own
regular surveys that gather demographic data (often in conjunction with labour force
data – e.g., the monthly Current Population Survey run since the early 1940s by the
US Bureau of the Census for the Bureau of Labor Statistics), and over time there
have been myriad one-off or longitudinal demographic surveys, or surveys with
content of interest to demographers, mounted by university researchers, specialized
research agencies, government departments, etc.
An increasingly important variant of the sample survey in demography over
recent decades has been the ‘micro’ approach. It is generally applied in small
geographic areas (often a village) over an extended period of time, combining
the formalities of localized census-taking and demographic surveying with detailed
semi-structured or unstructured anthropological or sociological investigations (so-
called ‘qualitative’ research). The ‘micro’ approach is more a tool of population
studies than of formal demography; a means of shedding light on the complex social
and cultural determinants of demographic variables and change. In one sense it
can be a troublesome approach. Whereas with conventional survey research there
is typically a questionnaire and a description of sampling and field procedures that
can be scrutinized and critically evaluated, ‘micro’ studies do not inherently offer
opportunity for such checks. Considerable faith must be placed in the integrity of
the fieldworker, who while perhaps wishing to test hypotheses should not have
embarked on the project intent on verifying preconceived ideas. With a semi-
structured or unstructured survey instrument and the capacity to quote informants
selectively, it is relatively easy to tell, and difficult to refute, any story one cares
to tell. Nevertheless, in the hands of an honest, skilled practitioner the ‘micro’
approach is a very valuable tool of demographic explanation. It is, incidentally,
useful to distinguish between ‘respondents’ (those who answer structured survey
and census questionnaires where question formats remain unchanged from person
to person) and ‘informants’ (those who respond to more conversational semi-
structured or unstructured questioning in in-depth interviews, which tend to be
Sources of Demographic Data 15

topic-driven and sufficiently flexible to allow questioning to digress to follow up

leads as they present during the interview).

Vital Registration Systems

Vital registration systems, sometimes termed ‘civil’ registration systems, exist more
often and tend to be better developed in more developed countries. They derive, as
their name implies, from requirements that designated members of the population
register vital events (births, deaths and marriages), usually within some prescribed
period after their occurrence. In Australia, responsibility for registration rests with
a parent in the case of births and the marriage celebrant in the case of marriages.
Registration of deaths is more complex, being ‘based on information supplied by
a relative or another person acquainted with the deceased, or an official of the
institution where the death occurred and on information supplied by a medical
practitioner or a coroner as to the cause of death’ (Australian Bureau of Statistics
2007: 82).
While demographers might like to imagine otherwise, vital registers, especially
in developed countries, generally exist primarily for legal and administrative
reasons rather than to facilitate demographic research. They establish certain facts
about individuals, and are used for purposes such as providing documentary proof
of age, settling deceased estates and inheritance entitlements, and establishing
obligations to support children and entitlement to marry. Their utility as data
sources is considerable, since they normally record not merely the occurrence
of a demographic event but certain attributes of the event and of the person(s)
experiencing it. However, that utility is apt to be viewed by those who maintain
vital registers as of secondary importance to their legal and administrative utility. In
Australia the situation is complicated by registration being the responsibility of six
State and two Territory governments, each with its own registration forms, and there
has, for example, been less than universal enthusiasm for amending birth registration
forms to eliminate a serious defect in data on mothers’ parities (i.e., the numbers of
live births they have had) (Carmichael 1986; Corr and Kippen 2006). This defect
arises from forms historically requesting details only of previous children ‘of the
current marriage’, and more recently only of previous children ‘of the current
relationship’, whence women with children who repartner and have further children
can in doing so revert to parities previously attained but since passed beyond (e.g., a
woman with two children who divorces, remarries, then has a third child statistically
has a second ‘first’ birth, not a third birth). From 2007 the Australian Bureau of
Statistics began publishing parity (or ‘previous issue’) data based on children of all
relationships for those States and Territories now collecting such data, but Victoria
and Queensland, between them accounting for around 45 % of national population,
still refuse to do so. The central issue is one of sensitivity – whether a woman
should be required to disclose the existence of a previous child of which her current
husband or partner may be unaware. In the 1960s when non-marital births often
16 1 Basic Sources, Concepts, Definitions and Types of Measures

led to placement for adoption this was a real issue, but with such births these days
typically occurring to cohabiting couples and placements for adoption by strangers
uncommon, the sensitivity argument has worn thin.
Like censuses, vital registration systems are costly, and this expense is a major
deterrent to their establishment in low-income countries. In this case the cost has two
dimensions: the need to maintain a permanent infrastructure for an ongoing data
collection operation; and the need for a widely dispersed network of registration
points readily accessible to all members of a population. Cost is not, however,
the only impediment to establishing viable vital registration systems in developing
countries. Even where these do exist, coverage is often incomplete. Several factors
may be responsible for this: lack of capacity to enforce a legal obligation to register
events (it is difficult to expect compliance if failure to comply is rarely detected
and/or attracts no sanction); lack of a popular tradition of registration (in many
developed countries registration at the behest of the state followed lengthy periods
when the Church performed a similar function); and the absence of a comprehensive
system of state-funded entitlements, establishing eligibility for which acts as an
incentive to register vital events in many developed countries.
While in many countries the problems of cost and achieving reasonable coverage
have meant that no attempt has been made to establish vital registration systems,
others have chosen to tackle the task gradually by setting up systems province
by province or district by district. India, on the other hand, has set up systems
in a series of sample areas, and uses the data generated as the basis for national
fertility and mortality estimates. It is not necessary for registration of vital events to
be complete before it provides usable fertility and mortality estimates; techniques
have been developed for adjusting incomplete data (provided they are not too
incomplete) to, for example, yield estimates of adult mortality. It is, however,
important to always assess the completeness of data from a vital registration system
before proceeding with analysis. Blindly assuming complete registration when it is
seriously incomplete obviously can lead to totally erroneous results and conclusions.
Vital registration data typically are used in conjunction with census data. They
provide the numerator for a demographic rate or ratio while a census provides
the denominator. As already noted, though, the latter may have had to be updated
for the period between the census and the date to which the calculation pertains,
and this updating of population estimates following a census is in itself another
common application of vital registration data. Other data may also feature in such
exercises, most commonly migration data. These may derive from border control
documentation in the case of national population estimates, with data on internal
migration being needed in the case of subnational population estimates. For further
detail on vital registration systems see United Nations (2001).
Sources of Demographic Data 17

Administrative Systems

Data sources falling under this heading exist primarily for administrative purposes
but have demographic utility as a by-product of those purposes. It could be argued
that this description applies to vital registers, but they are such a fundamental source
of demographic data as to be regarded as a separate category. The administrative
priority which attaches to what might be termed ‘miscellaneous’ administrative
sources can sometimes mean that access to them is hard to obtain or carries
restrictions, and hence their demographic potential is difficult to realize. It can also
mean that items of considerable demographic interest either are not collected or are
vulnerable to being discarded when data collection procedures and instruments are
reviewed.
Data on divorce are in some jurisdictions another type of vital registration data.
Divorce is certainly a vital event, but it is not one that normally is registered in the
sense that births, marriages and deaths are. Rather, divorce data tend to emanate
from the legal process through which those seeking to formally end marriages
pass. They are a by-product of the legal system, courts rather than registrars of
vital events furnishing statistical agencies with the returns from which statistics are
compiled, and so are considered here to be one of the more important miscellaneous
administrative sources. In Australia they provide a good example of the vulnerability
of such sources, from demographers’ perspective, to administrative whim. In 1995
a revision of the form on which Australian divorce data are transferred to the
Australian Bureau of Statistics deleted, as a cost-saving measure, the two items
giving parties’ marital statuses at marriage (whether never married, divorced or
widowed). In making this decision the review eliminated researchers’ capacity
to distinguish dissolutions of first marriages from those of remarriages following
divorce and widowhood, a fundamental distinction to be made in any thorough
analysis of divorce. Studies invariably show divorce to be more common in
marriages where one or both parties has previously been divorced than in those
that are first marriages for both parties.
Briefly mentioned in passing earlier as obviating the need for censuses in
some countries, universal population registers are another administrative source of
demographic data (Verhoef and van de Kaa 1987). They are essentially extensions
of, and incorporate, vital registers, although these exist as separate entities for each
type of vital event whereas in a population register they are brought together and
amalgamated. ‘The population register is a mechanism for the continuous recording
of selected information pertaining to each member of the resident population of a
country or area, making it possible to determine up-to-date information about the
size and characteristics of the population at selected points in time’ (United Nations
2001: 75). Like vital registration systems, population registers have legislated
underpinnings to create and enshrine the administrative systems that maintain them,
require citizens and others to provide information they rely on in a timely way,
and perhaps allow certain information to be imported into them from other more
specialized databases. They start from essentially a census base – an inventory
18 1 Basic Sources, Concepts, Definitions and Types of Measures

of inhabitants of an area and their demographic characteristics, including their

residential locations – then update individual records as subsequent events and
changes of status are experienced, add new ones to cover births and new immigrants,
and terminate or suspend records of deceased persons and emigrants. This process
includes recording non-temporary changes of residential address, so one important
advantage of population registers over vital registers is their capacity to provide data
on migration. Changes of occupation may also be recorded, allowing occupational
mobility and labour force composition to be monitored.
Because they entail setting up a record with a unique personal identification
number for each individual at creation of the register, birth or on joining the
population as an immigrant, population registers are bothersome at one level in
representing a major, ongoing bureaucratic intrusion into the lives of citizens;
in having considerable potential for data to be used for purposes not initially
advised or intended; and in being potentially tools of repression. Indeed, these
sorts of issues go a long way to explaining why they are not more widespread. The
Australian Government, for example, proposed in the mid-1980s the establishment
of a national identity card system that would have had the potential to become a
population register, but was ultimately dissuaded from proceeding by strong oppo-
sition to its ‘big brother’ overtones. From the demographer’s viewpoint, however,
provided the data are accessible, population registers offer attractive potential for
both up-to-date cross-sectional and longitudinal analysis. They are well developed
in Nordic countries – especially Sweden, Denmark, Norway and Finland – but can
raise interesting issues when attempts are made to interrogate them to generate
census-type data. In Norway, for example, considerable difficulty was experienced
in this context in trying to define ‘households’, because notified addresses in multi-
unit apartment blocks often failed to identify the particular apartment a person lived
in. This necessitated creation of a whole new numbering system for apartments in
such complexes (Trude Lappegárd, personal communication, July 2013).
The precise content of population registers varies from country to country,
and these differences dictate the capacity to produce census-type data exclusively
from interrogation of the register. Other countries use a mixture of data from
a register and other sources, including sample surveys, to produce census-like
statistical snapshots of their populations. Indeed, providing sampling frames for
more targeted surveys is another useful attribute of population registers. According
to the United Nations, countries producing ‘census’ data exclusively from ‘admin-
istrative registers’ during the decade 2005–2014 were Andorra, Austria, Denmark,
Finland, the Holy See, Monaco, Norway, Slovenia, Sweden, Bahrain and Singapore.
Those using a combination of such registers and other sources (chiefly sample
surveys) were Belgium, the Czech Republic, Estonia, Germany, Iceland, Italy,
Latvia, Lichtenstein, Lithuania, the Netherlands, Poland, Spain, Switzerland and
Turkey. Population registers are therefore primarily a European data source.
A third commonly used administrative data source is records of international
movements of population derived from border control procedures. Most countries
have laws governing who is and isn’t eligible to enter the country and under what
conditions, and some also restrict departures of their citizens. In implementing
Sources of Demographic Data 19

these laws information is gathered from persons arriving at, and departing from,
recognized points of entry and exit. The coverage of these data varies from
country to country as a function of (i) the level of incentive for illegal migration,
(ii) governments’ capacities to patrol national borders, and (iii) the existence of
populations whose traditional lands straddle national boundaries. For countries like
Australia and New Zealand, surrounded by water, with international travel occurring
through small numbers of airports and seaports, and with coastal surveillance
adequate to intercept most attempts at illegal entry elsewhere in the country,
coverage is extremely good. But for a country like the USA, with a long common
border with a neighbouring country (Mexico) whose residents it mostly seeks to
exclude, yet whose much lower standard of living creates a strong incentive to
migrate, coverage may be much poorer.
Not that completeness of coverage is the only consideration in assessing the
quality of border control data. As a comparison of Australian and New Zealand
data on population flows between the two countries has shown (Carmichael 1993),
differences in questions asked and in classification concepts and data processing
conventions used can have non-trivial effects on estimates of the sizes of migrant,
as distinct from visitor, flows between pairs of countries. A phenomenon known as
‘category jumping’, in which individuals, through deliberate deception or because
expectations on arrival or departure change, are allocated to movement categories
which ultimately prove to be incorrect, may also have a major impact on data quality.
Accurately separating migrants from visitors in border control data is especially
important these days, given the hugely increased volumes of international visitor
movement since the advent of international air travel. Border control databases have
in fact in recent decades often been overwhelmed by volumes of visitor movement,
resulting in Australia and New Zealand, for example, deciding to process data for
only samples of those deemed to be involved in such movement.
Records associated with the provision of medical and welfare services are
another useful category of administrative data. Their more obvious potential is for
informing studies of fertility, mortality, morbidity, the aged, etc., but if recording
changes of address they may also be a source of data on internal migration. In
Australia data on internal migration have been gathered through the quinquen-
nial censuses since 1971, but between censuses, quarterly estimates of interstate
migration were until 1986 based on changes of address notified to the Department
of Social Security by recipients of the then universal Family Allowance, while
since that time they have been based on changes of address advised to the Health
Insurance Commission as administrator of the national health insurance system,
Medicare. Other administrative systems which record changes of address on a fairly
universal basis, such as those operated by telephone or electricity companies, are
also, if accessible, potential sources of data on internal migration in the absence of
other sources, although the trend from landline to mobile phones with a wider array
of providers constantly urging customers to switch allegiance has undermined the
use of telephone records for this purpose.
20 1 Basic Sources, Concepts, Definitions and Types of Measures

Data Quality

Demographic data are frequently defective. They may be incomplete, as when there
is significant non-response to a census or survey, underregistration of vital events,
or an administrative source is only partial in its coverage. They may be out of
date, whence they should not be discussed in a manner that pretends otherwise
and, if the aim is to assess the current situation, explicit attention should be paid
to possible changes since the date to which the data pertain. They may also contain
measurement errors, which may arise at any stage of data collection and processing;
from faulty questionnaire design, defective sampling, inadequate training and/or
performance of field staff, lack of knowledge or defective recall on the part of
respondents, substandard coding, errors in computer processing, etc.
It is important that a demographer always be alert to the possibility that data
may be defective in some way, particularly (though by no means exclusively) when
the population concerned is a developing country population. Before undertaking
analysis of any dataset you should evaluate it with a view to identifying any
shortcomings or idiosyncracies that might have implications for your intended
analysis. Some types of problem have in the past proved so common that checking
for their existence should be routine – age misstatement and underregistration of
vital events are two examples (see, for example, United Nations (1955)). Age
misstatement may in particular be an issue if you are intent on studying age in
single-year-of-age detail. You should also, in the course of analysis, be awake to
unanticipated results, the first reaction to which should be to ponder whether any
problem with your data or their processing could be responsible. Certainly do not
embark on dreaming up fantastic explanations for unexpected results until you are
absolutely satisfied the results are genuine. And remember, the fact that data exist in
a computerized form, or are neatly printed in an official statistical publication, does
not render them automatically correct. In evaluating data it is always healthy to be
slightly suspicious rather than blindly trusting.
Just because data are discovered to be defective does not make them auto-
matically unusable. In extreme cases a decision to jettison an entire dataset may
be justified, but there exist a battery of techniques for handling deficient data.
The well known United Nations Manual X, for example, is described online as
‘an aid to demographers and population experts to carry out the best possible
evaluation and exploitation of data sources, especially those that are incomplete
or deficient’ (United Nations 1983). Moreover, defective data problems often can
be dealt with adequately by incorporating appropriate caveats into discussions of
results of data analyses, or by demonstrating that a realistic tolerance for data error
has no substantive effect on your conclusions.
Absolute Measures: The Population Balancing Equation 21

Absolute Measures: The Population Balancing Equation

Demographers deal with two broad types of measure – absolute measures and
relative measures. Absolute measures in demography are counts of people, or of
events that occur to people. The population balancing equation (also known as
the ‘basic demographic equation’), a simple accounting equation that expresses
population change through time, makes use of such measures. It takes one of two
forms, depending on whether the population in question is ‘closed’ or ‘open’.
A closed population experiences no migration, either inward or outward. Its
size therefore changes as a function of births and deaths alone. With interplanetary
travel not yet an option the obvious example is the world’s population, although
populations defined by attributes that exist at birth and thereafter are unalterable
(e.g., the Australia-born population) are also closed (albeit that if geographically
scattered they may be difficult to obtain data on), and some national populations
(e.g., until the late 1980s, those of many Soviet bloc countries, and these days that
of a country like North Korea) can be considered to be virtually closed. An open
population, by contrast, is one whose size is affected by migration as well as by
births and deaths. Most of the populations demographers deal with are, of course,
in this category.
For a closed population the population balancing equation is:

P2 D P1 C B  D (1.1)

Where P2 D population at time 2; P1 D population at time 1; B D births between

time 1 and time 2; D D deaths between time 1 and time 2.
The change in population between time 1 and time 2 is then given by:

dP D P2  P1 D B  D (1.2)

For an open population the population balancing equation is:

P2 D P1 C B  D C I  O (1.3)

Where P2 , P1 , B and D have the same meanings as in Eq. 1.1; I D in-migrants

arriving between time 1 and time 2; O D out-migrants departing between time
1 and time 2.
The change in population between time 1 and time 2 is then given by:

dP D P2  P1 D B  D C I  O (1.4)

The righthand side of Eq. 1.4 adds together two components of population
change which have special names. The component (B – D) is known as natural
increase, and the component (I – O) as net migration. The sizes of open populations
22 1 Basic Sources, Concepts, Definitions and Types of Measures

are subject to change through both natural increase (which may, of course, be
negative – i.e., natural decrease – if deaths over the period in question outnumber
births) and net migration; the sizes of closed populations are altered by natural
increase alone.
The population balancing equation can in theory be rearranged to make any
unknown element its subject, so as to enable that element to be calculated. Its most
common application is, however, in the estimation of net migration, especially for
subnational geographic units (for which usually there are no migration data deriving
from border control procedures) over periods between censuses. It thus often
underpins studies of the contribution of migration to intercensal population change
at the state/regional/district/suburban level. This type of study produces estimates of
net migration which measure the combined impact of international (across national
boundaries) and internal (across subnational boundaries) migration. If we denote
the (I – O) element in Eq. 1.3 by M (standing for ‘net migration’) we can rewrite the
equation in the form:

M D .P2  P1 /  .B  D/ (1.5)

In other words, net migration between time 1 and time 2 is total population
change, less natural increase. In applying the population balancing equation, and
in particular this variant of it, it is important to be aware of whether the population
counts P1 and P2 are de facto or de jure counts.
A de facto population count counts all persons present in the area of interest
(country, region, etc.) at the time the count is made, and counts them at the
geographic location where they are found at that time. It excludes usual residents
of the area who are temporarily absent, but includes visitors who are not usual
residents. A de facto census thus enumerates everybody present at the time of the
census according to their geographic location at that time. A de jure population
count, on the other hand, counts persons usually resident in the area of interest, and
counts them according to their geographic location of usual residence. It includes
usual residents who are temporarily absent (to the extent that they can be traced) but
excludes visitors who are not usual residents. A de jure census thus enumerates the
usually resident population according to individuals’ usual places of residence.
It is not always clearcut what constitutes being a ‘usual resident’ or a ‘visitor’.
Arbitrary definitions based on duration of residence are sometimes imposed, but it
is more common to rely on self-classification by respondents – they are allocated to
whichever category, ‘usual resident’ or ‘visitor’, they perceive themselves to belong
to, and usual residents are also left to indicate whether their address on census
night is their ‘usual’ address. Note, too, that in distinguishing between de facto and
de jure censuses we are making a distinction at two levels about any individual:
first, whether (s)he should be counted at all; and second, assuming (s)he should be
counted, where, geographically, (s)he should be counted.
De facto censuses and surveys can cause problems if one’s interest is household
composition or family type. Temporary absentees or visitors can distort household
and family structures. At the 1981 Australian Census, for example, use of a
Relative Measures: Rates and Probabilities 23

strictly de facto count resulted in families with one parent temporarily absent on
business etc. on census night being classified as ‘single parent’ families. Addition
of a question on persons temporarily absent from the household at the 1986
Census introduced a de jure element to counter this ludicrous misclassification.
This was an example of a data problem being identified because of increased
focus on a particular population subgroup. Single parent families became of much
greater interest to Australian researchers during the 1970s after (i) the divorce
rate rose substantially following mid-decade liberalization of divorce laws and
(ii) the introduction of a Supporting Mother’s Benefit in 1973 made choosing to
retain children born outside marriage rather than place them for adoption more
economically feasible.
We have strayed from the population balancing equation, but the point is this.
When applying Eq. 1.3 in respect of P1 and P2 values which are de facto population
counts, I and O refer to all arrivals and departures respectively. When applying it
in respect of de jure population counts, I and O refer to permanent arrivals and
departures only (i.e., to those which add people to, or remove them from, the ‘usual
resident’ population; the word ‘permanent’ is an overstatement for some individuals,
but is intended to encompass all who, in crossing the relevant geographic boundary,
change their place of usual residence). Similarly, if using Eq. 1.5 to compute net
migration, whether the result gives net total or net permanent migration depends on
whether P1 and P2 are de facto or de jure population counts respectively. Never mix
the two!

Relative Measures: Rates and Probabilities

Absolute measures such as the population balancing equation both produces as

output and requires as input are of limited use to demographers, except as input
into the calculation of more sophisticated measures. Suppose, for example, we are
told that in Great Britain there were 606,000 deaths in 1871 and 560,000 deaths in
1921. If we subtract the later figure from the earlier one we discover that there were
46,000 fewer deaths in 1921 than in 1871, and 46,000 is 7.6 % of 606,000. But can
we conclude that mortality in Great Britain declined by 7.6 % during 1871–1921?
The answer is an emphatic ‘No!’ Why? Because the populations at risk of dying
were not the same in 1871 and 1921.
In 1871 the population of Great Britain was 27.4 million; by 1921 it had increased
to 44.0 million. Suppose this population had not changed, and had still been 27.4
million in 1921. What effect would we expect this to have had on the number
of deaths in 1921? Obviously we would expect considerably fewer than 560,000
deaths. If a population of 44.0 million produced 560,000 deaths, then a much
smaller population (27.4 million) experiencing the same mortality conditions would
be expected to yield many fewer deaths, widening the gap between the numbers
of deaths in 1871 and 1921. Thus we can conclude that the absolute decline in the
number of deaths understates the extent of the mortality decline. Had the population
24 1 Basic Sources, Concepts, Definitions and Types of Measures

of Great Britain remained constant at 27.4 million a much larger decline in the
number of deaths would have occurred.
To make comparisons between populations (whether two or more populations
at one point in time or, as in the example above, one population at two or more
points in time) we need measures of the relative occurrence of demographic events;
measures which allow (or control) for differences in the sizes of the populations
producing those events. Always be sceptical of analyses of demographic data that
rely exclusively on absolute numbers. Journalists seeking to quickly write up the
latest press release from the national statistical agency or to localize the focus of
demographic data to the area their publication serves are serial offenders in this
matter.

Demographic Rates

Two major categories of measures of relative occurrence used by demographers are

rates and probabilities. The same two categories are sometimes referred to as m-
type rates and q-type rates (Hinde 1998), the ‘m’ and ‘q’ being standard life table
notations (see Chap. 4). Common use of the word ‘rates’ in this dichotomy is an
unhelpful source of confusion, but in this book the concept ‘rate’ will be rather
strictly defined, and the distinction between it and the concept ‘probability’ will be
observed.
Both rates and probabilities are examples of ratios, the name given to any
expression of the form ‘a divided by b’, in which the ‘a’ quantity is known as the
numerator and the ‘b’ quantity as the denominator. Even aside from the confusion
just discussed, the term ‘rate’ is used very loosely in the demographic literature, and
you should be aware that some measures that conventionally are called ‘rates’ do not
conform with the strict definition to be presented here. Two prominent examples are
the infant mortality rate (IMR), which is not a ‘rate’, but an estimate of a probability,
and the total fertility rate (TFR), which is not a single rate but the sum of a series of
age-specific rates. In some quarters the practice has developed of calling the TFR the
total fertility ratio, but even that is not an accurate description, leading the prominent
British demographic journal Population Studies to insist that authors refer simply
to ‘total fertility’. Some indices that are no more than simple percentages are also
conventionally called ‘rates’; for example, the ‘literacy rate’ – the percentage of a
population which is literate.
A demographic rate can be defined as the number of occurrences of a given type
of demographic event (birth, first birth, death, first marriage, etc.) in a given period
of time (usually a calendar year, or a 1-year period of cohort experience) divided
by the average size of the population at risk of experiencing the event during
the given time period. Strictly speaking, the denominators of rates should be the
number of person-years of exposure to risk. This concept is discussed later in this
chapter, where it is shown that the average, or mean population at risk is ordinarily a
satisfactory approximation of the number of person-years of exposure to risk. As the
Relative Measures: Rates and Probabilities 25

former is invariably far more easily calculated it tends to be the denominator used
for practical purposes. There are various ways of calculating the mean population at
risk; for example, averaging the populations at risk at the beginning and end of the
reference year (i.e., the year for which the rate is to be calculated), or averaging the
populations at risk at the beginning of that year and at quarterly intervals thereafter.
It is also common to use the mid-period (mid-year) population at risk as an estimate
of the mean population at risk, mid-year population estimates often being routinely
produced by national statistical agencies. This can lead to slight variations in results
if events modifying the population at risk over time are abnormally clustered in one
part of the reference year. Having said that, however, use of a mid-year, compared
to a mean, population at risk normally has very little impact on a rate’s value, even
under conditions of quite marked clustering.
Suppose, for example, we have the following data for a closed population in
which a disease epidemic occurred during the final months of 2001. We want to
calculate a crude death rate, being the ratio of deaths in 2001 to the mean population
at risk of dying during that year.

Population at 31 December 2000 1,050,850

Deaths in first quarter 2001 5,400
Deaths in second quarter 2001 6,200
Deaths in third quarter 2001 5,800
Deaths in fourth quarter 2001 15,200
Live births in first quarter 2001 10,200
Live births in second quarter 2001 10,400
Live births in third quarter 2001 10,600
Live births in fourth quarter 2001 10,300

We have a marked clustering of deaths in the final quarter of 2001, which might
lead us to suspect that it would be inadvisable to use the mid-year population at
risk as an estimate of the mean population at risk. Since we are dealing with a
closed population we can obtain populations at the end of each quarter by taking the
31 December 2000 population and successively adding live births and subtracting
deaths for each quarter. This process yields the following populations:

Population at 31 December 2000 1,050,850

Population at 31 March, 2001 1,055,650
Population at 30 June 2001 (mid-year) 1,059,850
Population at 30 September 2001 1,064,650
Population at 31 December 2001 1,059,750

The mean population at risk can be estimated by averaging these figures, since
they are equally spaced through the reference year. The result (1,058,150) indeed
26 1 Basic Sources, Concepts, Definitions and Types of Measures

differs from the mid-year population at risk (which is 1,059,850). However, if we

divide total deaths during 2001 (32,600) by each of these figures and multiply by
1,000 (to obtain a rate per 1,000 at risk) we get answers of 30.8 per 1,000 and
30.8 per 1,000. In other words, even with a marked clustering of events modifying
the population at risk, use of the mid-year population at risk to approximate the
mean population at risk had no impact on the value of a crude death rate calculated
to one decimal place. Thus, we can almost always use a mid-year population at
risk as the denominator of a demographic rate. As a general rule, demographic
rates are far more sensitive to changes in their numerators than to changes in
their denominators. The latter can tolerate some approximation, hence inaccuracy,
without results being significantly affected. On the other hand, changes in rates over
short time periods are normally attributable mainly to changes in their numerators.
We will return now to our example of mortality in Great Britain in 1871 and 1921.
In place of the absolute numbers of deaths with which we began we will compute
relative measures by calculating crude death rates. The crude death rate (CDR) is
given by:

CDR D .D=P/ :1;000 (1.6)

Where D D deaths during year y; P D mean (or mid-year) total population in year y
Substituting the data for Great Britain in this equation (assuming the populations
given to be mid-year populations) we have:

CDR1871 D .606;000=27;400;000/ 1;000 D 22:1 deaths per 1;000 total population

CDR1921 D .560;000=44;000;000/ 1;000 D 12:7 deaths per 1;000 total population

This establishes a decline of 22.1  12.7 D 9.4 deaths per 1,000 total population,
or 42.5 %, between 1871 and 1921, a far more spectacular decline than the 7.6 %
decline in total deaths established earlier.
In this example, measures of absolute and relative occurrence do at least both
show declines. It is, however, possible for them to change in opposite directions.
Suppose, for example, there had been 660,000 deaths in Great Britain in 1921
instead of 560,000. This would have been 54,000 more deaths than in 1871, but
would have yielded a CDR of 15.0 per 1,000 total population, still appreciably below
the 1871 figure. Thus, despite the absolute incidence of mortality having increased,
the relative incidence would have declined, the latter being the more meaningful
finding because of its controlling for change in the population at risk.
The CDR is an example of a true rate. Its denominator accurately captures the
population at risk of experiencing the event (death) that is the focus of its numerator.
There are other so-called ‘crude’ rates that represent looser uses of the term ‘rate’.
While they also employ the mean (or mid-year) total population as denominator, not
all persons in this purported population at risk are actually at risk of experiencing
the event focused on in the numerator. Examples are the crude birth rate (CBR) and
crude marriage rate (CMR), which are calculated as follows:
Relative Measures: Rates and Probabilities 27

CBR D .B=P/ :1;000 (1.7)

CMR D .M=P/ :1;000 (1.8)

Where B D live births during year y; M D marriages during year y; P D mean (or
mid-year) total population in year y.
While all persons in a population are at risk of dying over any given period, not
all are at risk of giving birth or getting married. Males, women yet to reach menarche
(the onset of first menstruation, and hence the biological ability to bear children) and
women past menopause (the cessation of menstruation, and the biological ability to
bear children) are not at risk of giving birth. Likewise, persons who are legally too
young to marry or are currently married are not (in a monogamous society) at risk of
getting married. Obviously neither the CBR nor the CMR is a true rate, because in
neither case does the denominator conform with the definition of a rate given above.
It is not hard to see, either, why these ‘rates’ are called ‘crude’.
But what about the CDR? As a ‘true’ rate, why is it also called a ‘crude’ rate?
It is crude for another reason; it does not allow for extreme variability in the risk
of dying at different ages or for the concentration of high risk at one end (the older
end) of the age distribution. This means that CDRs can differ, or change, simply
because the age structures of the relevant populations differ or change, without
there necessarily being any difference/change in the underlying risks of dying at
different ages. In practice, of course, differences/changes in CDRs typically reflect
differences/changes in both age structure and underlying risks of dying; but the
CDRs themselves tell us nothing of the relative strengths of these forces or, indeed,
whether both are operating in the same direction. As an example to be presented
shortly will show, it is not uncommon for one CDR to be higher than another, but
for the differential in underlying age-specific risks of dying to run in the opposite
direction. In such a situation, differences in age structure have both compensated
for, and outweighed, those in underlying risks of dying, creating the false impression
that mortality is higher in the population in which, age for age, it is actually lower.
Those other crude rates that are not ‘true’ rates are also vulnerable to differ-
ences/changes in age structure (and other dimensions of population composition),
and are therefore in a sense doubly crude. But so extreme are age differentials
in the risk of dying, and so concentrated is high risk at the oldest ages, that the
vulnerability of the CDR is acute. Whereas a ranking of national populations by
crude birth rate usually conforms quite well with one based on a more refined
measure of fertility, a ranking by crude death rate, despite the CDR being a ‘true’
rate, may deviate appreciably from one based on a mortality index which controls
for differences in age structure.
28 1 Basic Sources, Concepts, Definitions and Types of Measures

Demographic Probabilities

Differentiating probabilities from rates gets confusing when common usage can see
one type of measure called the other (the example of the infant mortality ‘rate’
was mentioned above). Basically, a demographic probability measures the risk of
experiencing a specified type of demographic event during a given stage of a
person’s life, or a given life cycle phase. While the use of estimation procedures
may mean that in practice there is not always strict conformity with the statement,
a probability measures risk in the context of one’s being a member of a particular
cohort; a group of individuals who experienced some previous demographic event
(most commonly birth – whence we speak of a ‘birth cohort’) during a specified
period of time (most commonly a particular calendar year). Rates may also be
computed for cohorts, but a cohort setting is not conceptually as integral.
Usually ‘stage of life’ or ‘life cycle phase’ is defined in terms of time elapsed
since the event conferring membership of the cohort was experienced. If that event is
birth, time elapsed since birth is a person’s age, and we calculate probabilities of, for
example, dying between specified birthdays (or exact ages). But the previous event
could also be something like marriage or divorce, whence the life cycle focused
on is one spent in the married or divorced state (as compared to the ‘alive’ state);
time elapsed since marriage or divorce is duration of marriage or divorce; and
we calculate, for example, probabilities of divorcing between specified wedding
anniversaries (exact durations of marriage), or of remarrying between specified
anniversaries of divorce (exact durations of divorce). It is also possible for ‘stage
of life’ or ‘life cycle phase’ to be defined in terms of calendar dates (e.g., the
probability of a person born in year x dying during – i.e., between the beginning
and the end of – year y) or even a combination of a calendar date and an exact age
or duration. However, most of the probabilities that interest demographers pertain
to life cycle phases that begin and end at nominated exact ages or durations.
Demographic probabilities are distinguished from demographic rates primar-
ily by the nature of their denominators. Both types of measure have numerators that
are counts of demographic events during a reference period, or a reference phase of
the life cycle, but whereas a rate relates those events to a measure of the average
(or mid-period) size of the population at risk, a probability can be defined as the
number of occurrences of a given type of demographic event during a specified
life cycle phase divided by the population at risk at the BEGINNING of that
life cycle phase. Strictly speaking, this definition conforms to a common approach
to estimating probabilities. For refined work adjustments to allow for extraneous
demographic processes (migration is a common one) removing people from risk
or adding people ‘at risk’ during the reference phase may be called for. But the
definition does capture common practice, just as the earlier definition of a rate did
in referring to ‘the average size of the population at risk’ rather than to the more
technically correct ‘number of person-years of exposure to risk’.
Consider as an example of a demographic probability the infamous infant
mortality ‘rate’. The IMR is given by:
Relative Measures: Rates and Probabilities 29

IMR D 1 q0 D D0 =B (1.9)

Where D0 D deaths at age 0 during year y; B D live births during year y; 1 q0 is life
table notation for the probability of dying at age 0 (this notation is explained in
Chap. 3, but is introduced here to reinforce the point that 1 q0 and the IMR are to
all intents and purposes one and the same).
Infant mortality is by definition mortality during the first year of life; that is,
mortality between birth (exact age 0) and a child’s first birthday (exact age 1). In
Eq. 1.9, B is a measure of those entering, or reaching the beginning of, the period of
their lives when they would be at risk of infant death during year y. D0 is an estimate
of the deaths during infancy experienced by this group of newborn babies. Thus we
have an estimate of the probability that a child born in year y dies before reaching
its first birthday.
Why is D0 , and hence the IMR, only an estimate? The answer is that, since it
takes a year for any child to move from birth to its first birthday, the period of risk
of infant death for any child invariably straddles two calendar years (unless the
child happens to be born at midnight on 31st December). Hence, (i) not all of the
infant deaths making up our numerator D0 involved children from our denominator
B; some were deaths of children born not in year y, but in year y  1. Similarly,
(ii) some of the infant deaths of children in our denominator B will have occurred
not during year y, but during year y C 1, in which case they are not included in
our numerator D0 . Our estimation procedure assumes that these two elements of
imprecision cancel out; that infant deaths during year y of children who were born
during year y  1 are a good approximation of infant deaths of children born during
year y that in fact occur during year y C 1.
Assumptions of this type are often made when calculating demographic proba-
bilities. It is common practice to estimate probabilities by dividing the number of
people experiencing the reference event (in the above example, death) at a specified
life cycle stage (infancy) during a calendar year by the population who entered
that stage in the same year. The major reason for this practice, which produces
only estimates because numerator and denominator do not relate to exactly the
same group of individuals, is the reality that statistical agencies tabulate event
counts by age/duration and calendar year rather than by age/duration and cohort,
the form more suited to calculating probabilities. Using the principles of a graphical
device known as the Lexis diagram (see Chap. 3), demographers frequently convert
data from the former to the latter type. But as this, too, is a process of estimation
involving assumptions, it is sometimes barely worth the effort. And in the particular
case of the IMR, as we shall see, the assumptions underlying Lexis diagram
principles are simply untenable.
A straightforward example will illustrate the difference between a rate and a
probability in demography. Consider the data over page:
Using Eq. 1.9 we have IMR1990 D (D0 / B) 1,000 D (9,760 / 97,600) 1,000 D 100.0
deaths per 1,000 live births (we have multiplied by 1,000 to obtain an answer in
a form comparable to an age-specific death rate). The death rate at age 0 is given
30 1 Basic Sources, Concepts, Definitions and Types of Measures

Births in 2005 97,600

Infant deaths in 2005 9,760
Population aged 0 at 30 June 2005 91,800

by (D0 / P0 ) 1,000 (where P0 is the mid-year population aged 0) D (9,760 / 91,800)

1,000 D 106.3 deaths per 1,000 aged 0. The two measures have identical numerators
(albeit that for the IMR this numerator is an estimate of the true numerator), but
both conceptually and numerically quite distinct denominators.

A Digression: Concepts of Age, Incidence and Prevalence

You may have noted, in the foregoing few paragraphs, reference to two different
concepts of age. Demographers distinguish between exact ages and conventional
ages, otherwise termed ages in completed years, or ages last birthday. For the
individual, exact age is an instantaneous experience; no sooner has one attained
a nominated exact age than one is older, and no longer that exact age. One is exact
age 0 at birth, exact age 1 precisely a year later, and so on. While in countries like
Australia birthdays conventionally are celebrated over a full day, in reality one is
only ever exactly x years old for an instant. Moments beforehand one is still younger
than x, and moments after one is older (x plus a fraction of a second). By contrast a
conventional age, or age in completed years, or age last birthday is an attribute that
remains with a person for a full year. One is aged 0 (completed years) for the entire
year between birth and one’s first birthday; 10 (completed years) for the entire year
between one’s tenth and eleventh birthdays; and so on.
Demographers often are interested in (i) populations which attained a nominated
exact age (reached a particular birthday) during a specified period of time (often a
particular calendar year), and (ii) populations which at a nominated point in time
(say the date of a census, or the mid-point of a calendar year) shared in common a
specified age, or series of consecutive ages, in completed years (i.e., were members
of a particular age group). Similarly, and by direct analogy, they may be interested
in (i) populations which attained a nominated exact duration (of marriage, divorce,
etc.) during a specified period of time, and (ii) populations which at a nominated
point in time shared in common a specified duration, or series of consecutive
durations (of marriage, divorce, etc.), in completed years. The distinction between
these two types of populations should become clearer once Lexis diagrams have
been covered in Chap. 3, but it can be used to provide another perspective on the
difference between demographic rates and probabilities. A demographic rate has as
its denominator a population quantity based on ages (or durations) in completed
years. A demographic probability usually has as its denominator a population
quantity based on exact ages (or durations).
Refinement of Demographic Rates and the Effect of Age Structure 31

While strictly an epidemiological rather than a demographic distinction, it is

worth also noting at this juncture the difference between measures of incidence and
prevalence. This distinction is made in the context of producing relative measures
of the pervasiveness of particular diseases or medical conditions in a population. Its
relevance to demographers lies in their interest in understanding the disease patterns
and trends that underpin mortality patterns and trends. Incidence measures capture
the number of new occurrences of a disease or medical condition in a population
over a defined reference period (typically a calendar year), relative to the size of
the population at risk (typically the mid-year population at risk). Such measures
are rates, strictly defined, and thus appropriately referred to as incidence rates.
Their numerators are counts of events (new occurrences, or diagnoses, of diseases
or medical conditions).
Prevalence measures, in contrast, focus not on a reference period, but on a
point in time (a date), and capture the number of cases of a disease or medical
condition that exist in a population at that reference date, relative to the size of
the population at risk. Strictly speaking, such a measure is a measure of point
prevalence. Particularly when dealing with medical conditions of typically short
duration (e.g., the common cold), you may also encounter measures of period
prevalence. In this case the numerator is a count not of people with a disease
or condition at a point in time, but of those who had that condition at any time
over a defined period (perhaps a month). It is, of course, important not to compare
measures of point and period prevalence, since the latter have an inherent upward
bias, the degree of which depends on the length of the reference period. In more
general terms, though, measures of prevalence are neither rates nor probabilities.
They are, in fact, simple proportions. Their numerators are not counts of events,
but of persons in a particular category at a point in time (or at some stage during
a defined period). When the category was entered is irrelevant (it could have been
years previously). On the other hand, persons who would feature in the numerator
of an incidence rate for the year in which the reference date/period lies, but who
at that date/period had already left the category (through death, out-migration or
recovery), or were yet to enter it (i.e., would enter it after the reference date/ period),
are excluded. Whether incidence, or prevalence measures for a disease/condition
tend to be higher depends on its average duration from onset to death or recovery.
If disease/condition episodes on average last longer than a year, prevalence will be
higher; if on average they are shorter, incidence will be higher. Fairly obviously, the
two types of measures should not be mixed in comparative analysis.

Refinement of Demographic Rates and the Effect of Age

Structure

We encountered above several so-called ‘crude’ rates: the crude death rate, the crude
birth rate and the crude marriage rate. We noted that the latter two measures are not
really ‘rates’ at all, because some persons in their denominators are not at risk of
32 1 Basic Sources, Concepts, Definitions and Types of Measures

giving birth and marrying respectively. As a rule, crude rates do not provide much
information about the levels of the demographic phenomena to which they relate;
indeed they can be positively misleading. They involve a trade-off between precision
on the one hand, and their reliance on relatively unsophisticated data on the other.
Their main merit is that they can often be computed in circumstances where more
refined measures cannot be; they are better than nothing, although as we shall see
they can be so misleading that even that is sometimes debatable.
‘Crude’ rates deserve their name. Demographers tackle the problem their crude-
ness poses by seeking to calculate more refined rates. In doing this they aim to close
in on the population at risk by doing one or both of two things: (i) eliminating
from the denominator persons who could not possibly experience the event in
the numerator; and (ii) recognizing that certain personal characteristics make
someone more, or less, likely to experience that event.
In the case of the CBR, a refinement of the former type yields the general fertility
rate (GFR). This measure retains live births during year y as its numerator, but
its denominator excludes all men, and women who are outside the conventionally
defined reproductive ages. Hence we have a ‘true’ rate:
 
GFR D B=P.f;15–49/ :1;000 (1.10)

Where B D live births during year y; P(f,15–49) D mean (or mid-year) female popula-
tion aged 15–49 in year y.
This measure might also be subjected to a refinement of the second type. In most
societies marriage is culturally the most approved setting for bearing children, and
women are considerably more likely to give birth if they are married than if they
are unmarried. Hence there is a clear case for calculating specific rates; in this case
rates specific for marital status. Specific rates are calculated for subgroups of the
total population at risk by dividing the number of relevant events occurring to that
subgroup in the reference year by the mean, or mid-year, population in the subgroup.
In this case it makes sense to calculate both a marital fertility rate (MFR) and a non-
marital fertility rate (NMFR). These are given by:
 
MFR D Bm =P.f;m;15–49/ :1;000 (1.11)

Where Bm D marital live births during year y (live births to married women whose
husbands acknowledged paternity); P(f,m,15–49) D mean (or mid-year) married
female population aged 15–49 in year y.
And:
 
NMFR D Bn =P.f;u;15–49/ :1;000 (1.12)

Where Bn D non-marital live births during year y (live births to women who were
not married to the fathers of their children); P(f,u,15–49) D mean (or mid-year)
unmarried female population aged 15–49 in year y.
Refinement of Demographic Rates and the Effect of Age Structure 33

These two specific rates are less straightforward than most on two counts.
First, note that the numerators are not defined simply as ‘live births to married
women’ and ‘live births to unmarried women’ during year y respectively. This
is because it is possible for a married woman to have a non-marital birth (if
the father is someone other than her husband), and for an unmarried woman
to have a marital birth (if an event causing her to become unmarried occurred
between conception and confinement). Thus, strictly speaking the numerators
and denominators do not pertain to the same groups of individuals, although in
practice they are usually very close to doing so. Second, who should be considered
‘unmarried’ is not clearcut. Obviously never married, widowed and divorced women
qualify, but separated women could be considered either married (which legally
they are) or unmarried. Which option one takes depends on whether one judges that
births to separated women are more likely to be fathered by their legal husbands
(whence they should be recorded as marital births) or by some other partner
(whence they would be non-marital births). In Australia the latter may be more
likely.
The personal characteristic most commonly affecting a person’s risk of expe-
riencing a particular type of demographic event is age, and the most commonly
encountered specific rates are therefore age-specific rates. As a ‘true’ rate, the crude
death rate is not a candidate for the type of refinement undertaken in moving from
the crude birth rate to the general fertility rate, but it most certainly is a candidate
for the type of refinement that is accomplished through the calculation of specific
rates. Death is much more likely to occur at some ages than at others, and mortality
risks for men and women can also be quite different, both generally and at certain
ages (sex is a second personal characteristic by which demographic analyses
are routinely refined). So variable is the risk of dying by age, in particular, that
crude death rates, which take no account of the age structure of a population, can
give totally false impressions of the comparative underlying levels of mortality in
different populations.
Consider the populations of Malaysia and Australia. In 1988 Malaysia had a
CDR of 4.93 deaths per 1,000 mid-year population, while Australia had a CDR
of 7.25 deaths per 1,000 mid-year population. These figures suggest that mortality
was higher in Australia by a factor of 47 %. Intuitively, to anyone who knows
Malaysia and Australia, this sounds peculiar. While Malaysia is hardly among the
world’s least developed nations, of the two countries Australia in 1988 was the more
developed, and we would expect health services and survival chances to have been
superior.
In Table 1.2 the comparison of mortality levels in Malaysia and Australia has
been refined through the calculation of age-sex-specific death rates for the two
populations. The relevant equation is:
 
ASSDR.s;x/ D D.s;x/ =P.s;x/ :1;000 (1.13)
34 1 Basic Sources, Concepts, Definitions and Types of Measures

Table 1.2 Age-sex-specific Males Females

death rates for Malaysia and
Age group Malaysia Australia Malaysia Australia
Australia, 1988
0 16.6 9.8 12.8 7.6
1–4 1.7 0.5 1.1 0.4
5–9 0.6 0.2 0.5 0.2
10–14 0.6 0.3 0.4 0.2
15–19 1.0 1.1 0.5 0.4
20–24 1.5 1.6 0.6 0.5
25–29 1.7 1.5 0.9 0.5
30–34 2.0 1.4 1.1 0.6
35–39 2.4 1.5 1.5 0.8
40–44 3.4 2.2 2.3 1.2
45–49 5.5 3.4 3.2 2.1
50–54 9.9 6.0 5.9 3.4
55–59 15.4 10.0 9.9 5.5
60–64 25.5 17.3 17.0 8.7
65–69 38.3 27.2 27.8 13.8
70–74 62.7 45.3 46.6 23.5
75–79 86.2 71.9 69.8 40.7
80–84 144.3 110.7 119.0 71.4
85C 173.4 186.6 142.8 147.7
Source: United Nations Demographic Yearbook, 1990 and
Australian Bureau of Statistics, Deaths Australia 1988

Where s denotes sex (male or female); x denotes an age group; D(s,x) D deaths of
people of sex s at age x during year y; P(s,x) D mean (or mid-year) population of
sex s and age x in year y.
A quick inspection of Table 1.2 reveals that for the majority of age-sex groups
Australia’s age-sex-specific death rate was lower than Malaysia’s, often by a
substantial margin. Only for males aged 15–19, 20–24 and 85C, and for females
aged 85C, did Australia record a higher death rate. Marginally higher death rates for
Australian males in their late teens and early twenties reflect greater susceptibility
to involvement in fatal road accidents, while higher Australian death rates for both
sexes at ages 85 and over are due to older age structures within that age group.
Of men and women aged 85 and over, proportionately more in Australia than in
Malaysia were aged 90 and over, and 95 and over, and therefore subject to the
very highest risks of dying. But despite these exceptions, the conclusion invited by
Table 1.2 is the reverse of that invited by our earlier comparison of CDRs. The table
suggests very clearly that, age for age, mortality was lower, not higher, in Australia.
What causes the CDR for Malaysia to be lower than that for Australia, when for
almost all age-sex groups the death rate for Malaysia is higher? The answer is the
two populations’ different age structures. Higher fertility in Malaysia had given its
population in 1988 a younger age structure, and except in the first year of life, age-
sex-specific death rates are much lower at younger than at older ages. Malaysia’s
Refinement of Demographic Rates and the Effect of Age Structure 35

population was relatively more concentrated at younger ages where the risk of
dying is low; Australia’s was relatively more concentrated at older ages where the
risk of dying is high. The extent of this difference in population composition was
more than enough to offset Australia’s generally lower age-sex-specific death rates
and produce CDRs which convey a totally misleading picture of the comparative
underlying levels of mortality in the two populations.
In case you are still mystified as to how lower age-sex-specific death rates for
Australia could produce sufficient deaths to give rise to a higher CDR, Table 1.3
demonstrates what has happened. The table shows how each 1,000,000 members
of the populations of Malaysia and Australia in 1988 were distributed by sex and
age (for each population, total males C total females D 1,000,000). It also shows the
age-sex distributions of deaths occurring in 1988 to the two groups of 1,000,000
people, deaths in each age-sex group having been obtained by applying the relevant
age-sex-specific death rate from Table 1.2 to the corresponding number of persons
at risk from Table 1.3. We thus have two populations of identical size, each of which
has had the appropriate schedule of age-sex-specific death rates applied to it. What

Table 1.3 Age-sex distributions of 1,000,000 people and deaths occurring to those people in
Malaysia and Australia, 1988
Males Females
Malaysia Australia Malaysia Australia
Age group Persons Deaths Persons Deaths Persons Deaths Persons Deaths
0 14,254 237 7,597 74 13,464 172 7,245 55
1–4 56,773 97 30,324 15 53,804 59 28,954 12
5–9 61,207 37 37,762 8 58,012 29 35,803 7
10–14 56,002 34 38,712 12 53,613 21 36,758 7
15–19 52,776 53 43,626 48 50,772 25 41,743 17
20–24 50,193 75 40,771 65 49,060 29 39,297 20
25–29 42,748 73 42,914 64 44,839 40 41,982 21
30–34 35,704 71 40,234 56 38,326 42 39,933 24
35–39 30,427 73 38,778 58 31,269 47 38,391 31
40–44 23,858 81 36,248 80 23,075 53 34,537 41
45–49 20,473 113 27,923 95 19,767 63 26,357 55
50–54 17,052 169 23,855 143 17,190 101 22,798 78
55–59 12,615 194 22,708 227 13,604 135 21,924 121
60–64 10,043 256 21,764 377 10,513 179 22,336 194
65–69 7,373 282 17,701 481 8,498 236 19,973 276
70–74 4,608 289 12,857 582 5,246 244 16,200 381
75–79 3,430 296 8,712 626 3,976 278 12,555 511
80–84 1,399 202 4,471 495 1,675 199 7,827 559
85C 963 167 2,319 433 1,399 200 6,111 903
Total 501,898 2,799 499,276 3,939 498,102 2,152 500,724 3,313
Source: Prepared from data presented in United Nations Demographic Yearbook, 1990 and
Australian Bureau of Statistics, Deaths Australia 1988 and Estimated Resident Population by Sex
and Age: States and Territories of Australia June 1988 and Preliminary June 1989
36 1 Basic Sources, Concepts, Definitions and Types of Measures

do we find? We find that the population with the generally lower age-sex-specific
death rates (Australia) produces the larger total number of deaths (7,252 compared
to 4,951). How has this come about?
If you compare the two distributions of ‘Persons’ by age and sex in Table 1.3
you will note that Malaysia has substantially larger numbers at younger ages (up
to age group 20–24). After that Australia has the larger numbers, its advantage
becoming very pronounced beyond about age 60. There are at least twice as many
of each 1,000,000 people in Australia in each of these older age-sex groups as
in Malaysia, and by ages 80–84 and 85C the ratio for females is well over four
times as many. The older age groups are also where the great majority of deaths are
concentrated, and are clearly the source of Australia’s higher total number of deaths.
Although generally having lower death rates at these ages, those rates are applied
to significantly larger numbers of people at risk. In effect, the degree to which
Australia’s death rates at these ages lie below those of Malaysia is, by a considerable
margin, more than offset by the degree to which its populations at risk exceed those
for Malaysia. The result is very much larger absolute numbers of deaths at older
ages in Australia, purely because the age distribution of the population is biased
towards those ‘high risk’ ages.
In order to sensibly compare mortality levels in different populations we need
to eliminate such bias. The CDR does not do this, and is in consequence not only
not a very helpful index, but a positively dangerous one. Death rates refined by age
(and sex) do eliminate age-bias, and while the added insight gained is generally not
as spectacular or necessarily as clearcut as in the example just presented, there is a
case for similarly refining almost any summary rate used in studying a demographic
phenomenon.

The Concept ‘Person-Years Exposed to Risk’

While we have characterized the denominator of a demographic rate as the mean,

or mid-period (generally mid-year), population exposed to risk, we also noted that,
strictly speaking, this denominator should be the number of person-years for which
the population at risk was exposed to that risk. Mean and mid-period populations
at risk are estimates of person-years of exposure to risk. But what do we mean by
‘person-years exposed to risk’? Essentially we mean the total amount of time spent
at risk by all the individuals who were at risk of experiencing the event of interest
at some stage during the reference period.
Suppose we wish to compute some rate for calendar year y, and that this
rate pertains to a non-renewable event E. A non-renewable event is one that an
individual can experience only once (such as death, first marriage or having a first
birth); a renewable event is one that can recur in an individual’s life (such as giving
birth, getting married or divorcing). The number of person-years of exposure to the
risk of experiencing event E during year y is the sum of two components:
The Concept ‘Person-Years Exposed to Risk’ 37

1. Each person who was at risk of experiencing event E at the beginning of year y,
but who survived right through that year without experiencing it and remained at
risk at the end of the year, contributes one person-year of exposure to risk.
2. Each person who actually experienced event E during year y, or who only began
to be at risk part way through year y, or who ceased to be at risk part way through
year y for a reason other than that (s)he experienced event E (or who fits into
more than one of these categories) contributes part of a person-year of exposure
to risk. The fraction of a person-year each contributes equals the fraction of year
y spent at risk.
The second of these components clearly has several sub-components. Both
they and the first component can be summarized diagrammatically, the solid line
indicating the reference period (year y; the period of potential exposure to risk), the
dotted line indicating the period of actual exposure to risk, and the letters E, J and
L standing for ‘event’, ‘join’ and ‘leave’ respectively (i.e., occurrence of the event
of interest; a person joining the population at risk part way through the reference
period; and a person leaving the population at risk part way through for a reason
other than that event E was experienced). Component 1, using these conventions, is
represented as follows:

………………………………….

Component 2, subcomponent 1 consists of persons at risk at the beginning of

year y, who experienced event E during that year. Each was at risk for the fraction
of year y extending from the start of the year to the date when event E occurred. The
occurrence of event E terminated their exposure to risk.

………………………E

Component 2, subcomponent 2 consists of persons who began to be at risk part

way through year y, and continued to be at risk for the remainder of the year, not
experiencing event E. Each was at risk for the fraction of year y extending from
the date of joining the population at risk until the end of the year. Reasons for
joining part way through the reference period vary, depending on the rate being
calculated. In-migration is almost always a possibility; an in-migrant may have
been at risk of event E in some other population earlier in year y, but only begins
to be at risk in the population under study after joining it. Other reasons might
include, in relation respectively to a crude death rate, an age-specific death rate,
and a marriage-duration-specific divorce rate, for example, being born, attaining a
certain age, or attaining a certain marriage duration during year y. One isn’t at risk of
dying until born, or of dying at a particular age or divorcing at a particular marriage
duration until that age or marriage duration has been reached. In general, then, the
38 1 Basic Sources, Concepts, Definitions and Types of Measures

occurrence of some event J, causing one to join the population at risk, initiates a
person’s exposure to risk.

J………………….

Component 2, subcomponent 3 consists of persons who were at risk at the

beginning of year y, but ceased to be at risk during the year for a reason other than
that they experienced event E. Each was at risk for the fraction of a year that had
elapsed when the period at risk ended. Out-migration terminates a person’s exposure
to risk as a member of a population. So does death (even when it is not event E),
and so, too, for example, does attainment of a particular age or duration if it takes
a person out of the age or duration category for which a rate is being calculated.
One isn’t at risk of anything if dead; neither is one at risk of dying at a given age
or divorcing at a given marriage duration once having passed beyond that age or
marriage duration. In general, then, the occurrence of some event L, causing one to
leave the population at risk, terminates a person’s exposure to risk.

…………L

Component 2, subcomponent 4 consists of persons who combine elements of

the first and second subcomponents. They joined the population at risk part way
through year y and then experienced event E before the end of the year. Each was
at risk for the fraction of the year that elapsed between these two occurrences.

J……………..E

Component 2, subcomponents 5 and 6 consist of persons who combine elements

of subcomponents 2 and 3. Either they joined the population at risk part way through
year y, then left it again before the year ended without experiencing event E, or,
having been at risk at the start of year y, they left the population at risk without
experiencing event E, but rejoined it later in the year. These two subcomponents can
be represented diagrammatically as follows:

J………………L

……..L J………..
The Concept ‘Person-Years Exposed to Risk’ 39

It is possible to conceive also of more complex subcomponents entailing multiple

entries to and exits from the population at risk, in combination perhaps with ultimate
experience of event E, but we will ignore those. All of this probably makes obtaining
the denominator for a demographic rate in terms of person-years exposed to risk
seem rather complicated. We will try to clarify matters with a simple example.
Suppose we have a population of 100 people at 1 January in year y; that this
population is closed (i.e., migration is not a complicating factor); and that 1 person
dies and 3 are born in the middle of each month throughout year y. We will also
assume that all deaths occurred to people alive at the start of the year (eliminating
the complication of some individuals experiencing more than one event). What is
the number of person-years of exposure to the risk of dying during year y for this
population?
We will note for a start that we began with 100 people at 1 January in year y, but
that with 12 deaths and 36 births through the year the population at 31 December in
year y was 100  12 C 36 D 124. Having noted this figure, we proceed as follows:
1. Of our original 100 people, 88 survived right through year y (since all 12 deaths
occurred to this initial 100). They contributed 88 person-years of exposure to the
risk of dying, and this is component 1 of the answer we seek.
2. The 12 who died, one in the middle of each month, survived for 0.5, 1.5, 2.5, : : : ,
11.5 months. If we add these periods of survival we find that, in total, those who
died survived for 72 person-months, which divided by 12 comes to 6 person-
years. This is component 2, subcomponent 1 of our answer.
3. What about the 36 who were born during year y, three in the middle of each
month? They joined the population at risk part way through the year, and as none
later died during year y the three born in January each spent 11.5 months at risk,
the three born in February 10.5 months, : : : , and the three born in December
0.5 months. The total period spent at risk of dying during year y by persons born
during the year was therefore 3(11.5 C 10.5 C : : : C 0.5) person-months, which
is 216 person-months or 18 person-years. This is component 2, subcomponent 2
of our answer.
4. Because of the assumptions we have made (no migration, therefore nobody
leaving the population at risk other than by dying, or joining it other than by
being born; no individual experiencing more than one event) subcomponents 3–
6 of component 2 all contribute zero person-years of exposure to risk.
5. Hence the total number of person-years of exposure to the risk of dying during
year y for our population was 88 C 6 C 18 D 112.
This just happens to be the average of the beginning-of-year and end-of-year
populations we noted earlier ((100 C 124) / 2 D 112). In other words, the number of
person-years of exposure to the risk of dying is exactly equal to the mean population
exposed to risk, which is the denominator we have been using for demographic rates
until now. Why is this so?
It is so because the events which altered the size of our risk population during
year y, the 12 births and 36 deaths, were evenly distributed through the year. This
is the situation we assume prevails when we use the mean, or mid-year, population
40 1 Basic Sources, Concepts, Definitions and Types of Measures

at risk as the denominator of a rate. If it is a reasonable assumption to make, it is

very convenient. It is much easier and requires much less detailed data to calculate
the mean, or mid-year, population at risk than to calculate the number of person-
years of exposure to risk. It is, of course, possible to conceive of more complicated
examples than the one just presented; examples in which the population at risk might
be being affected by more than two types of events, including in-migrant arrivals and
out-migrant departures. Such complexity is no cause for alarm. The general finding
still holds. Provided that occurrences of each type of event affecting the size of a
population at risk are evenly spread through the reference period for which a rate
is to be calculated, the number of person-years of exposure to risk will equal the
mean, and the mid-period, population exposed to risk.
Before proceeding, we will pause to consider the implications of removing one
simplifying assumption from the example just discussed. Suppose that instead of
all 12 deaths during year y occurring to people alive at the beginning of the year,
one of them involved one of the people born during the year. Indeed, we will be
specific and say that the death in mid-July involved one of the three people born in
mid-February.
1. Under the original assumption that nobody born in year y also died in that year,
this birth and death involved two people, one of whom was at risk from mid-
February to the end of year y (i.e., for 10.5 months) and the other of whom was
at risk from the beginning of year y to mid-July (i.e., for 6.5 months). Together
these two individuals were at risk for 10.5 C 6.5 D 17 person-months.
2. Under the new assumption that the birth and death involved the same person, that
person was at risk from mid-February to mid-July (i.e., for 5 months). However,
because one of the 12 deaths during year y involved someone born during that
year, only 11 of the 100 persons alive at the start of year y died during that year.
An additional person therefore survived right through year y (i.e., for 12 months).
If we add the period at risk for this individual to that for the person who was born
in February and died in July we get a total of 12 C 5 D 17 person-months, the
same answer as before.
3. This tells us that while the reality may be that one person can experience more
than one event affecting his/her risk status during a year, for the purpose of
calculating total person-years of exposure to risk for a population we can assume
that no person experiences more than one such event and still get the same
answer. Errors in the periods over which individuals were exposed to risk
flowing from this assumption cancel out.
This finding is very helpful. Recall that the number of person-years exposed to
risk for a population has two components, the second having six subcomponents
(and then some). Subcomponents 4–6 relate to individuals who experienced two
events affecting their risk status during year y. Rarely are demographic data
sufficiently detailed to permit individuals with this sort of complex experience to be
identified. The finding we have made tells us, however, that for practical purposes
we can ignore subcomponents 4–6 and any more complicated subcomponents we
The Concept ‘Person-Years Exposed to Risk’ 41

may care to imagine. We cover them by treating the individuals concerned as if they
belonged to two (or more) of subcomponents 1–3.
We will also pause here to reflect on the assumption in our discussion that event
E was non-renewable. What difference does it make if event E is renewable (i.e.,
able to be experienced by an individual more than once)? The point about renewable
events is that experiencing them either does not end a person’s exposure to risk (as
experiencing a non-renewable event does), or alternatively ends it only temporarily.
When calculating person-years exposed to risk for a renewable event we need to
consider which of these categories the event fits into. The obvious example of a
renewable event that does not end exposure to risk is giving birth. While biology
dictates that several months must elapse after a woman has given birth before she
can do so again, fertility rates which are not parity-specific (parity-specific births –
first births, second births, etc. – are non-renewable events) assume that a woman
who gives birth is immediately at risk again. In terms of the component model
presented above it is therefore not appropriate to terminate the period of exposure to
risk at the date event E occurs; component 2, subcomponent 1 of the model becomes
irrelevant, and women giving birth are treated as contributing to component 1 (i.e.,
at risk throughout year y). Examples of renewable events which do end a person’s
exposure to risk, but perhaps only temporarily, are out-migration, marriage, divorce,
and remarriage following divorce. While a person may out-migrate, marry, divorce,
or remarry following divorce more than once, having once done so (s)he must
respectively in-migrate, divorce/be widowed, remarry, or be divorced again before
being able to experience the event a second time. This inevitably takes time, and the
experience of event E is treated as ending exposure to risk. Any renewal of exposure
during the reference period is covered under component 2, subcomponent 2 of the
model (i.e., the person rejoins the population at risk).
Let us now return to the finding from our example that if occurrences of each
type of event which can remove someone from, or add someone to, a population at
risk are evenly spread through the reference period, the number of person-years of
exposure to risk equals the mean (and mid-period) population exposed to risk. In
most instances an assumption that events altering risk status are evenly spread
through time is sufficiently accurate for a demographic rate to be computed using
the mean (or mid-period) population at risk, without it being necessary to compute
person-years of exposure to risk. This is a very useful principle, because it greatly
reduces the time and effort, and the complexity of the data, required to obtain
denominators for demographic rates.
To illustrate, let’s consider the example of calculating a crude death rate for
Australia for 2010. To calculate the number of person-years of exposure to the risk
of dying in Australia in 2010 we need to know:
1. The size of the population at risk at the beginning of 2010.
2. The number and timing of deaths during 2010 (since on dying a person ceased to
be at risk of death).
3. The number and timing of births during 2010 (since newborn children were at
risk of dying from the dates of their births).
42 1 Basic Sources, Concepts, Definitions and Types of Measures

4. The number and timing of arrivals and departures by immigrants and emigrants
during 2010 (since these groups respectively became at risk on arrival and ceased
to be at risk on departure).
These data, with events distributed by quarter (1st quarter D January–March, 2nd
quarter D April–June, etc.), are shown below. The timing of events could be given
in greater detail, but distributions by quarter are sufficient to indicate quite some
deviation from the even patterns through time we assume when using the mean (or
mid-period) population exposed to risk as the denominator for our crude death rate.
Population at 1:1:10 D 21,865,623.

Quarter Deaths Births Arrivals Departures

1 32,689 73,420 119,239 59,380
2 35,694 73,479 88,499 56,491
3 38,439 73,774 109,580 67,804
4 35,808 69,997 107,801 69,406
Total 142,630 290,670 425,119 253,081

While almost certainly some individuals will have experienced more than one
of the four types of events affecting the population at risk of dying during 1980,
we noted above that we can calculate person-years of exposure to risk assuming
that nobody experiences more than one of these events, confident that errors at the
individual level will cancel out. This assumption means we must treat all deaths
and departures as having involved persons who were members of the Australian
population at the beginning of 1980. Hence we assume that 21,865,623 – 142,630 –
253,081 D 21,469,912 persons survived right through 1980, and component 1 of the
required exposure to risk figure is 21,469,912 person-years.
Component 2, subcomponent 1 of this figure consists of person-years spent at risk
by those who died during 1980 (all of whom, under our assumption, were members
of the Australian population at the start of 1980). Making the further assumption
that persons who died in any quarter on average died in the middle of that quarter,
we obtain subcomponent 1 as follows:

Quarter Deaths Average person-months lived Total person-months lived

1 32,689 1.5 49,033.5
2 35,694 4.5 160,623.0
3 38,439 7.5 288,292.5
4 35,808 10.5 375,984.0
Total 873,933.0

Dividing 873,933 person-months by 12 gives approximately 72,828 person-

years.
The Concept ‘Person-Years Exposed to Risk’ 43

Component 2, subcomponent 2 consists of person-years spent at risk by those

who were born into the Australian population during 1980 or who joined it as
immigrants in that year. We assume that births and arrivals of immigrants in
any quarter on average occurred in the middle of the quarter again, and obtain
subcomponent 2 as follows:

Quarter Sum of births and arrivals Average person-months lived Total person-months lived
1 192,659 10.5 2,022,919.5
2 161,978 7.5 1,214,835.0
3 183,354 4.5 825,093.0
4 177,798 1.5 266,697.0
Total 4,329,544.5

Dividing 4,329,544.5 person-months by 12 gives approximately 360,795 person-

years.
Component 2, subcomponent 3 consists of person-years spent at risk by those
who emigrated during 1980. We assume that departures of emigrants in any quarter
on average took place in the middle of the quarter, and obtain subcomponent 3 as
follows:

Quarter Departures Average person-months lived Total person-months lived

1 59,380 1.5 89,070.0
2 56,491 4.5 254,209.5
3 67,804 7.5 508,530.0
4 69,406 10.5 728,763.0
Total 1,580,572.5

Dividing 1,580,572.5 person-months by 12 gives approximately 131,714 person-

years.
Total person-years of exposure to the risk of dying in Australia during
2010 is the sum of the components and subcomponents now obtained; i.e.
21,469,912 C 72,828 C 360,795 C 131,714 D 22,035,249 person-years.
What if, instead of basing our calculations on distributions of deaths, births,
arrivals and departures by quarter, we had assumed that events of each type were
evenly distributed through 2010? This is, as we have noted, the assumption we
would make in using the mean (or mid-year) population at risk as the denominator
for our crude death rate. Given our other assumption that no individual experienced
more than one event altering his/her risk status, this new assumption amounts to
assuming that, on average, everyone who died, was born, arrived as an immigrant
or departed as an emigrant during 1980 was at risk of dying in Australia for exactly
6 months. Under this assumption the components and subcomponents of our person-
years of exposure to the risk of dying become:
44 1 Basic Sources, Concepts, Definitions and Types of Measures

Component 1 D 21,469,912 person-years (unchanged)

Component 2, subcomponent 1 D 142,630 / 2 D 71,315 person-years (total deaths
divided by 2, since every two deaths corresponds to one person-year of exposure
to risk).
Component 2, subcomponent 2 D (290,670 C 425,119) / 2 D 357,894.5 person-
years (total births plus total arrivals divided by 2, since every two births and
every two arrivals corresponds to one person-year of exposure to risk).
Component 2, subcomponent 3 D 253,081 / 2 D 126,540.5 person-years (total
departures divided by 2, since every two departures corresponds to one person-
year of exposure to risk).
Total person-years of exposure to the risk of dying in Australia in 2010 is the sum
of these components and subcomponents, or 21,469,912 C 71,315 C 357,894.5 C
126,540.5 D 22,025,662 person-years.
This ‘mean population’ exposure to risk estimate differs from the ‘seasonally
adjusted’ one calculated using events distributed by quarter, but what happens
when we use each estimate as the denominator for our 2010 crude death rate? Our
‘seasonally adjusted’ CDR is:

.142;630=22;035;249/ 1;000 D 6:47 deaths per 1;000

Our ‘mean population’ CDR is:

.142;630=22;025;662/ 1;000 D 6:48 deaths per 1;000

The increased accuracy we have gained through adjusting seasonally isn’t worth
the effort. Although we also obtained the denominator for our ‘mean population’
CDR using the size of the population at the beginning of 2010 and data on deaths,
births, arrivals and departures during that year, we needn’t have gone to this trouble.
We’d have got exactly the same answer by just averaging the populations at risk at
the beginning and end of 2010. This would have been a very simple calculation –
certainly much simpler than the procedure we went through to get a ‘seasonally
adjusted’ denominator. It usually happens that very little error is introduced
by using a ‘mean population exposed to risk’ denominator when calculating a
demographic rate. Hence demographers normally make do with these (or with mid-
year populations exposed to risk as estimates of mean populations exposed to risk),
and don’t bother computing person-years exposed to risk in intricate detail.
So why bother introducing the concept of person-years of exposure to risk at all?
There are at least three reasons why it is an important concept to grasp.
1. Theoretically the denominator of any demographic rate should be in this form.
It just happens that most of the time demographers make do with simple
approximations that (i) use less detailed, and importantly more readily available,
data, and (ii) save a lot of work for little added precision.
2. The concept of person-years lived is a key one for understanding life tables.
The Principle of Correspondence 45

3. There are some circumstances in which the assumption that events occur evenly
over time is grossly inaccurate. In these circumstances, thinking in terms of
person-years of exposure to risk assumes more practical significance. The
obvious example, which we will dwell on at greater length in Chap. 4, is infant
mortality, because deaths during the first year of life are heavily concentrated
early in the year of exposure to risk (i.e., within the first month of life, and
within that month within the first week of life).

The Principle of Correspondence

When constructing demographic rates and probabilities, a commonsense sort of

principle tells us: (i) that if a person is part of the exposed-to-risk population forming
the denominator of such a measure, any occurrence of the relevant event to that
person during the reference period should be included in the numerator; and (ii)
that if an occurrence of the relevant event to a person during the reference period
forms part of the numerator, that person should also be included in the exposed-to-
risk population forming the denominator. This principle is known as the principle
of correspondence. It is so commonsense as to seem obvious, but its application
in practice often requires data of a complexity that is seldom available. Hence, it is
frequently violated, and approximations are used.
An obvious example arose in our earlier discussion of the infant mortality rate.
We noted that the deaths at age 0 during year y forming the numerator did not all
occur to the persons born (attaining exact age 0) during year y who formed the
denominator. Some occurred to children born the previous year. Likewise, not all
deaths at age 0 of children born during year y (the denominator population) were
included in the numerator. Some occurred the following year. A second example
flows from our discussion of use of the mid-year population exposed to risk as an
estimate of the mean population exposed to risk, and hence an estimate of person-
years exposed to risk in the calculation of a crude death rate. If a person dies prior to
1 July in the reference year, her/his death is recorded in the numerator of the CDR,
but (s)he is not alive at mid-year to be part of the exposed-to-risk denominator.
The first of these examples concerns an estimate of a probability; the second
concerns a rate. Rates (sometimes called central rates), when measuring mortality,
invariably breach the principle of correspondence in the manner described. Probabil-
ities, when constructed without using the type of estimation procedure the IMR uses,
do not. They are inherently cohort measures, which focus on a group of individuals
and demographic events experienced by that group. The distinction being made here
should become clearer once you are familiar with Lexis diagrams.
46 1 Basic Sources, Concepts, Definitions and Types of Measures

Other Types of Measures

We have focused at some length on demographic rates and probabilities because

of their central importance in demographic analysis. However, discounting absolute
measures, also already discussed, there is a range of other types of measures that
interest demographers. Many are, like rates and probabilities, ratios. One that is
commonly encountered is the sex ratio, giving the number of males per 100 females
in a population or population subgroup:

S D .M=F/ : 100 (1.14)

Where M D number of males in the population or subgroup; F D number of females.

While rates and probabilities have numerators which are counts of events and
denominators which are counts of people at risk (or person-years of exposure to
risk), both numerator and denominator of the sex ratio are counts of people. Several
other measures fit this pattern; for example, the dependency ratio (the ratio of
persons in the dependent age groups per 100 in the working age groups), and the
child-woman ratio (a fertility measure sometimes used if counts of birth events are
not available). These are calculated as follows:

D D ..E C Y/ =W/ : 100 (1.15)

Where E D persons in the age group defined to constitute the dependent elderly
(often ages 65 and over); Y D persons in the age group defined to constitute the
dependent young (often ages 0–14); W D persons in the ‘working’ age group
(often ages 15–64).
The dependency ratio can also be split into additive components, the old-age
dependency ratio:

D.old-age/ D .E=W/ : 100 (1.16)

and the youth dependency ratio:

D.youth/ D .Y=W/ : 100 (1.17)

The child-woman ratio, the number of children aged 0–4 per 1,000 women of
childbearing age, is given by:
 
CWR D P.04/ =P.f;1549/ : 1;000 (1.18)

Where P(0–4) D population (of both sexes) aged 0–4; P(f,15–49) D female population
aged 15–49.
Other Types of Measures 47

Other measures are not single ratios of any description, but sums of series of
ratios. The best known of these is the total fertility rate (TFR). In recent years this
widely used index has been renamed in some quarters the total fertility ratio, in
recognition that it does not conform to the strict definition of a demographic rate.
But even this name misrepresents it; as just indicated it is not a single ratio, but
the sum of several ratios. The TFR is the sum, over all reproductive ages, of the
age-specific fertility rates prevailing in a particular calendar year; that is:

TFR D †.xD15;49/ ASFRx (1.19)

Where ASFRx D age-specific fertility rate for women aged x; x D a single-year age
group.
Alternatively the TFR may be calculated using age-specific fertility rates for
women in five-year age groups:

TFR D 5†.xD15–19;4549/ ASFRx (1.20)

The TFR is, in fact, a measure of the number of children the average woman in
a population would have throughout her life, IF she experienced the age-specific
fertility rates that prevailed in the year for which the TFR is being calculated. The
multiplier of 5 that appears in Eq. 1.20 recognizes that it takes a woman 5 years to
pass through a 5-year age group, so she is therefore exposed to the annual fertility
rate for that age group five times.
The TFR and similar measures (e.g., the total first marriage rate (TFMR) – the
sum over all marriageable ages for a given sex of ratios first marriages at age x
in a calendar year divided by the mid-year total population of that sex aged x) are
summary measures of the level of a set of specific rates or ratios. Other measures
summarize the distribution of sets of specific measures, an important group of such
measures being what statisticians call ‘measures of central tendency’ – mean (or on
occasion, median) ages or durations at which specified demographic events occur in
a population. Examples of such measures are the mean age of women at first birth,
the mean age of childbearing, the median age at first marriage, and the median
duration of first marriage at divorce. Shifts in these indices for a population over
time, or differences between indices for discrete populations at a point in time,
indicate changes or differences in the timing of the demographic process in question.
This brief discussion by no means exhaustively covers ‘other’ types of measures
used by demographers. It is, however, indicative of some major types, and along
with the earlier discussion of rates and probabilities, is sufficient to demonstrate
that demographic analysis utilizes a range of distinctive categories of measure.
48 1 Basic Sources, Concepts, Definitions and Types of Measures

References

Australian Bureau of Statistics. (2007). Deaths Australia 2006. Catalogue 3302.0. Canberra:
Australian Bureau of Statistics.
Bogue, D. J. (1969). Principles of demography. New York: Wiley.
Carmichael, G. A. (1986). Birth order in Australian vital statistics. Journal of the Australian
Population Association, 3(1), 27–39.
Carmichael, G. A. (1993). Beware the passenger card! Australian and New Zealand data on
population movement between the two countries. International Migration Review, 27(4), 819–
849.
Cleland, J. G., & Scott, C. (1987). World fertility survey: An assessment of its contribution. New
York: Oxford University Press.
Corr, P., & Kippen, R. (2006). The case for parity and birth order statistics. Australian and New
Zealand Journal of Statistics, 48(2), 171–200.
Festy, P., & Prioux, F. (2002). An evaluation of the fertility and family surveys project. New
York/Geneva: United Nations.
Hauser, P. M., & Duncan, O. D. (1959). The nature of demography. In P. M. Hauser & O. D.
Duncan (Eds.), The study of population: An inventory and appraisal (pp. 29–44). Chicago:
University of Chicago Press.
Hinde, A. (1998). Demographic methods. London: Arnold.
International Union for the Scientific Study of Population. (1958). Multilingual demographic
dictionary: English section. New York: United Nations Department of Economic and Social
Affairs.
Preston, S. H., Heuveline, P., & Guillot, M. (2001). Demography: Measuring and modelling
population processes. Oxford: Blackwell.
Shryock, H. S., Siegel, J. S., & Associates. (1973). The methods and materials of demography. 2nd
Printing (revised). Washington, DC: U.S. Bureau of the Census.
Siegel, J. S., & Swanson, D. A. (2004). The methods and materials of demography (2nd ed.). San
Diego/London: Elsevier Academic Press.
United Nations. (1955). Manual II. Methods of appraisal of quality of basic data for population
estimates (Population Studies No. 23). New York: United Nations. Available online at http://
www.un.org/esa/population/techcoop/DemEst/manual2/manual2.html. Accessed 26 July 2013.
United Nations. (1983). Manual X. Indirect techniques for demographic estimation.
ST/ESA/SER.A/81. New York: United Nations. Available online at http://www.un.org/esa/
population/techcoop/DemEst/manual10/manual10.html. Accessed 26 July 2013.
United Nations. (2001). Principles and recommendations for a vital statistics system: Revision 2.
New York: United Nations.
United Nations Statistics Division. (2013). 2010 World population and housing census pro-
gram. Available online at http://unstats.un.org/unsd/demographic/sources/census/censusdates.
htm. Accessed 12 July 2013.
Verhoef, R., & van de Kaa, D. (1987). Population registers and population statistics. Population
Index, 53(4), 633–642.
Weeks, J. R. (1994). Population: An introduction to concepts and issues (5th ed.). Belmont:
Wadsworth.
Weinstein, J., & Pillai, V.K. (2001). Demography: The science of population. Boston: Allyn and
Bacon.
Wunsch, G. J., & Termote, M. G. (1978). Introduction to demographic analysis principles and
methods. New York: Plenum Press.
Chapter 2
Comparison: Standardization
and Decomposition

Population Composition and Comparison of Summary

Measures

Demographers regularly wish to make comparisons, either between two or more

populations at the same point in time (or thereabouts) or between the same popula-
tion at two or more points in time. They frequently seek to do this by way of some
single summary measure calculated for each of the populations to be compared.
Comparisons of summary measures can, however, be quite misleading because
differences between them reflect differences in population composition, as well
as differences in the underlying level, incidence or prevalence of the demographic
phenomenon under scrutiny. ‘Population composition’ refers to the proportionate
distribution of a population across categories of variables, or combinations of
variables, which capture demographically pertinent characteristics of individuals.
Such variables include age, sex, marital status, ethnicity, education and occupation,
but not, for example, variables like eye colour, or whether left or right handed. The
most familiar representation of population composition is the age-sex pyramid, a
graphical device for summarizing the proportionate distribution of a population by
age and sex, several examples of which were presented in Chap. 1 to illustrate the
capacity of census data to capture demographic change over time. Composition is,
however, a multidimensional concept that embraces several other variables besides
these two.
The range of ‘summary measures’ to which the problem addressed in this chapter
might apply is extensive. It includes everything from basic measures like the crude
death and birth rates through more refined measures such as the general fertility
rate to even more refined measures such as, for example, the age-specific fertility
rate for women aged 15–19. Note that the word ‘rate’ recurs in these examples.
One of them, however, is not a ‘true’ rate (the CBR), and there are other measures
which are neither called, nor satisfy our definition of, ‘rates’ to which the principles

© Springer International Publishing Switzerland 2016 49

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_2
50 2 Comparison: Standardization and Decomposition

and techniques about to be discussed also apply. Examples are the proportion of
women of marriageable age who have never married, and the mean number of
children ever born to women of childbearing age or older. So what does distinguish
the ‘summary measures’ that are at risk of yielding distorted comparisons, and
hence are of interest? Essentially they are ratio-type measures that have rate-type
denominators. That is, their denominators are, or can be approximated as, the size
of a population, or a population subgroup, at a point in time. They thus include all
genuine ‘rates’ (whose denominators typically are approximated as the population
at risk at the middle of the reference period – a point in time), plus a range of
other measures with similar types of denominators (but numerators which are not
counts of events during a reference period). Excluded, however, are probabilistic
types of measures, whose denominators are not populations at a point in time, but
populations attaining an exact age or duration over a period of time. Discussion of
Lexis diagrams in Chap. 3 should make this distinction clearer.
The values summary measures of the type just described take on are the
product of two things: (i) the underlying level, or intensity, of the demographic
phenomenon being measured; and (ii) the composition of the (denominator)
population for which a calculation is made (i.e., the extent to which that population
is concentrated in compositional categories where the phenomenon being measured
is especially likely or unlikely to occur). In comparing populations we generally
are interested in how underlying levels or intensities compare, but we have to
accept that summary measures for those populations normally differ only partly
as a function of differences therein; they also differ as a function of differences in
population composition. Obvious questions arise: What are the relative strengths
of the two sources of difference? Are they operating in the same direction, or is
the observed difference the net effect of opposing forces, each favouring a different
population? If the latter, which of the two forces has proved stronger and therefore
dictated the direction of the observed difference?
The claim that summary measures of the type under discussion have an ‘under-
lying level, or intensity’ component and a ‘composition’ component can be substan-
tiated by a simple example. Consider the crude death rate, the equation for which
was given in Chap. 1 (Eq. 1.6) as CDR D (D / P). 1,000, where D D deaths during
year y and P D the mid-year total population in year y. Death is a phenomenon the
incidence of which is known to vary markedly by age and, to a lesser extent, sex
(two dimensions of population composition). We can rewrite the equation for the
CDR as follows (dropping the multiplier of 1,000):

CDR D †s †x .d .s; x/ =P/ (2.1)

Where s denotes sex (male or female); x denotes an age group; d(s,x) D deaths of
persons of sex s at age x during year y.
All we have done in Eq. 2.1 is say that total deaths in year y (i.e., D) equals the
sum of deaths in each sex-age group during year y (i.e., †s †x d(s,x)). Effectively
Population Composition and Comparison of Summary Measures 51

we divide the deaths in each sex-age group by P then add up the answers, instead of
adding the deaths first and then dividing the total by P; the answer is the same.
We now perform a little trick and multiply the righthand side of Eq. 2.1 by
a quantity (p(s,x) / p(s,x)) (where p(s,x) is the mid-year population of sex s aged
x in year y). We can do this because plainly this quantity has the value 1 (its
numerator and denominator are identical and therefore cancel), so that the value
of the righthand side is unaltered. We get:

CDR D †s †x f.d .s; x/ =P/ .p .s; x/ =p .s; x//g

Which can be rearranged to give:

CDR D †s †x f.d .s; x/ =p .s; x// .p .s; x/ =P/g (2.2)

The righthand side of Eq. 2.2 now incorporates the product of two ratios,
d(s,x) / p(s,x), which is the sex-age-specific death rate for sex s and age group
x, and p(s,x) / P, which gives the proportion of the total mid-year population that has
sex s and age group x. Thus, we have derived an equation for the CDR in terms of
specific death rates (death rates specific for categories defined by two compositional
variables, sex and age, across which we expect the incidence of death to vary) and
population composition (here, sex-age composition).
It is possible to do exactly the same for any other summary measure whose
denominator is the size of a population, or a population subgroup, at a point in time.
With the CDR we have decided that age and sex are key compositional dimensions
across which variation in the level of mortality is likely, and have consequently
expressed it as the sum of the products of (i) death rates specific for categories
defined by those two variables, and (ii) proportions of total population in the
same categories. Similarly, we can express any other summary measure with the
appropriate type of denominator as the sum of the products of two things:
1. Equivalent specific measures for categories which are defined by all possible
combinations of the categories of individual compositional variables held to be
important. The criterion for selecting a compositional variable as important is
an expectation that the intensity of the process or status in question (mortality,
fertility, education, etc.) varies markedly across its categories. This series of
measures collectively captures the ‘underlying level, or intensity’ component of
the summary measure referred to above.
2. Measures of population composition; i.e., proportions of the population forming
the denominator of the summary measure in each of the categories for which
specific measures are calculated. Obviously these proportions should sum to 1.0.
This specification gives rise to the following generalized equation:

M D †c m .c/ :p .c/ (2.3)

52 2 Comparison: Standardization and Decomposition

Where M denotes a summary measure with the appropriate type of denominator; c

denotes a series of compositional categories defined as specified under point 1
above; m(c) D the specific measure equivalent to M for compositional category
c; p(c) D the proportion of the population forming the denominator of M who are
members of compositional category c.
Before moving on, some discussion of compositional variables and categories is
in order. The compositional variable demographers most often suspect of affecting
the values of summary measures, and thereby contaminating comparisons of
those measures, is age. Most demographic processes are stronger, and demographic
statuses more common, at some ages than at others, and it is routine to control
comparisons for differences in age composition. Do not, however, fall into the trap
of thinking that age is the only compositional variable you ever need to worry
about. In dealing with the CDR above we also took into account the variable sex,
and there is ample evidence that mortality levels also vary across compositional
dimensions such as ethnicity, education, occupation and marital status. We might,
in certain circumstances, wish to recognize one or more of these additional variables
in specifying the CDR in accordance with Eq. 2.3 (so, for example, our categories,
(c), might be defined in terms of age, sex and ethnicity).
Or take a summary measure like the age-specific fertility rate for females aged
15–19. Here a ‘specific’ measure might also be treated as a ‘summary’ measure,
because there is reason to suspect significant variation in the level of fertility among
subgroups of the female population aged 15–19. The compositional variable we
might be most interested in is marital status, since we expect married women in
this age group to more often give birth than unmarried women. Thus, in specifying
the equation for this age-specific fertility rate in accordance with Eq. 2.3 our
compositional categories (c) would be marital status categories. Age, though, might
also be a consideration. Within the age range 15–19 years, women are more likely to
have children if aged 19 than if aged 15. If we suspected populations of 15–19 year-
old females we wished to compare of having different single-year age distributions,
we might want to recognize both marital status and age in specifying the age-
specific fertility rate in accordance with Eq. 2.3. The compositional categories
would then be age-marital status categories.
Implicit in the previous two paragraphs is the fact that, in specifying a summary
measure in the form generalized in Eq. 2.3 (as a precursor to applying techniques
which allow one to control for the distorting (or confounding) influence of
differences in population composition on comparisons of summary measures), the
decision as to what variables to take into account in defining the compositional
categories (c) is arbitrary. Often, in practice, only one variable is deemed to be
of concern, most commonly age. If larger numbers of variables are thought to
be potentially important, a useful strategy is to take them into account serially
(exploring the impact of controlling for one variable, reworking the analysis to
also take a second variable into account, etc.), and in different combinations.
By a process of trial and error one can isolate, and rank, the variables that are
really important confounders. But it rarely makes sense to take large numbers of
Population Composition and Comparison of Summary Measures 53

compositional variables into account simultaneously, especially as a first step. A

common practical difficulty is obtaining the detailed cross-tabulated data needed for
analysis using multi-dimensional compositional categories. More fundamentally,
though, to conclude that having controlled for a ‘grab-bag’ of compositional
variables alters the conclusion drawn before controlling is not all that helpful if
one doesn’t know which variable(s) is/are primarily responsible for the finding.
It should also be appreciated that controlling a comparison for one or more
compositional variables using methods with Eq. 2.3 at their core does not guarantee
that distortion due to differences in population composition has been eliminated.
First, other compositional variables not taken into account may remain sources of
distortion. Sometimes, lack of data precludes controlling for an important variable;
on other occasions an analyst may fail to recognize the need to control for a variable.
Second, taking a compositional variable into account using Eq. 2.3 isn’t just a matter
of introducing it to the equation. It is also a matter of deciding what categories to
specify for the variable. Sometimes these are self-evident. With the variable sex, for
example, there isn’t much alternative to ‘male’ and ‘female’. But for a variable like
age, options exist. If calculations are to be performed manually, there is an incentive
to limit the number of categories, and hence the quantity of mechanical work.
However, if too few, rather broad, categories are designated, within which there
is significant variation in both the level/intensity of the demographic phenomenon
being studied and the distribution of individuals from population to population, the
variable could, in effect, be only partially controlled for. Thus, for example, using
5-year age groups to capture age composition provides a more robust control for age
than using, say, age groups 0–49 and 50 and over.
One final point relevant to selecting appropriate compositional variables should
be mentioned. The distorting impact of a compositional variable on comparisons of
a summary measure is a function of two things: (i) variation in the level/intensity of
the demographic phenomenon under scrutiny across its categories; and (ii) variation
in the distribution of members of the populations being compared among those
categories. If, for a given variable, only one of these conditions exists, there will be
no distorting impact of that variable on a comparison. In particular, should the first
condition exist without the second, the variable can be ignored. Take the example of
the age-specific fertility rate for women aged 15–19 discussed above. We noted that
fertility was likely to vary significantly between 15 year-olds and 19 year-olds. But
for this variation to be a problem when comparing fertility rates for 15–19 year-olds
there needs to also be significant variation from population to population in the
proportionate distribution of 15–19 year-olds among the five constituent single-year
age groups. If investigation showed these age distributions to be similar, the case for
controlling for age would dissipate.
54 2 Comparison: Standardization and Decomposition

Making More Meaningful Comparisons of Summary

Measures

Having recognized the potential for comparisons of some demographic summary

measures to lead us astray, we turn now to techniques for dealing with the problem
of distortion arising from differences in population composition. Three approaches
are covered: the calculation of a series of specific measures; standardization; and
decomposition (sometimes also referred to as ‘component analysis’ or ‘components
of difference analysis’).
The first of these approaches was in essence covered in Chap. 1. In the section
on ‘Refinement of Demographic Rates and the Effect of Age Structure’ (pp. 31–36),
an example was introduced in which crude death rates for Malaysia and Australia
in 1988 were compared. The CDR for Australia was higher, which seemed strange
given Australia’s status as a ‘developed’ country, but when age-sex-specific death
rates were examined (Table 1.2), with few exceptions they were lower for Australia
than for Malaysia. It was pretty clear that, age for age and for both sexes, mortality
was lower, not higher, in Australia (i.e., the underlying level of mortality was
lower). We then investigated how lower age-sex-specific death rates in Australia
could produce a larger number of deaths among each 1,000,000 Australians than
among each 1,000,000 Malaysians, thus resulting in a higher CDR.
The Malaysian-Australian CDRs example is a spectacular one. Because of
this, the move from comparing CDRs to comparing age-sex-specific death rates
is enlightening. It tells us that the CDRs are totally misleading, to the point of
suggesting that the mortality differential between the two countries runs in the
opposite direction to the one it actually runs in. It doesn’t, however, give us any real
sense of how much lower the underlying level of mortality is in Australia (or how
much higher it is in Malaysia). We could easily generate a series of such measures
for each age-sex group (by taking ratios of age-sex-specific death rates for the two
countries), but we have no single summary indicator of the relative underlying levels
of mortality.
Viewed more generally, the practice of calculating series of specific measures to
assess the accuracy of the picture gained from a comparison of summary measures is
also beset by other difficulties. First, with the Malaysian-Australian CDRs example
we were fortunate that most of the age-sex-specific mortality differentials flowed
in one direction. If differentials in specific measures flow in opposite directions,
some favouring one population and others the other, unless there is a clear majority
flowing one way it can become difficult if not impossible to tell in which direction
the net overall differential flows. The fact that differences of direction exist, and
their pattern and magnitude, are useful pieces of supplementary information that
one might wish to consider in conjunction with a measure of net overall differential
(which would suppress such detail). But they do not, of themselves, yield such a
measure. Second, while with the Malaysian-Australian CDRs example we find a
complete reversal of the direction of the mortality differential once allowance is
made for differences in population composition, a more common scenario sees no
Standardization 55

change to the direction of a differential, but some modification of its magnitude.

Under this scenario, differences in population composition simply exaggerate or
moderate a differential. This more subtle type of distorting influence can be difficult,
and very often impossible, to detect merely through the calculation and comparison
of specific measures. Differentials between specific measures for compositional
categories will tend to run in the same direction as that between the summary
measures, and judging whether the former are collectively more marked than,
similar to, or more subdued than the latter may be well nigh impossible.
Thus, calculating specific measures is a hit-or-miss approach to assessing the
distorting impact of population composition on a comparison of summary measures.
Moreover, it is probably more often ‘miss’ than ‘hit’, and even when informative,
lacks the ‘summary’ character sought when using summary measures.

Standardization

Standardization is a technique which enables us to take into account and control

for the varying compositions of populations we wish to compare in the course
of calculating a single index for each population. It thus has the dual advantage
of being both a more robust approach to dealing with compositional distortion
in comparisons, and offering ‘summary character’. It has two variants. The more
straightforward, as its name implies, is direct standardization; the alternative
approach is indirect standardization.
Both approaches have at their core the generalized equation expressing a sum-
mary measure in terms of specific measures and population composition presented
above as Eq. 2.3, and that equation is therefore a very important one. With direct
standardization a standard (i.e., common) population composition is assumed for
all the populations to be compared (i.e., the set of p(c) values in the righthand side
of Eq. 2.3 remains the same from calculation to calculation), and the observed set
of specific measures for compositional categories for each population is applied to
this standard composition in turn. With indirect standardization a standard set of
specific measures for compositional categories is assumed for all the populations
to be compared (i.e., the set of m(c) values is the same in each calculation), and is
applied, in turn, to the observed population composition for each population.
To put this in more concrete terms, suppose we want to use standardization to
eliminate the effect of differences in age structure from a comparison of summary
measures. To standardize directly for age we apply different sets of age-specific
measures observed for the populations in question to a standard age structure,
whereas to standardize indirectly for age we apply a standard set of age-specific
measures to the observed different age structures for the populations.
56 2 Comparison: Standardization and Decomposition

Direct Standardization

As the more straightforward variant, direct standardization best enables one to

understand how standardization works. With direct standardization, in place of
our unstandardized summary measures we calculate standardized equivalents;
equivalents standardized for the compositional variable(s) we have chosen to
take into account. It is important to appreciate that these standardized summary
measures are hypothetical summary measures. Their values are values we would
expect the summary measure to take if the standard population composition, rather
than the actual composition, prevailed.
What does this mean? It means that no significance can be attached to the
absolute levels of standardized summary measures. These will vary depending on
the standard selected, and this selection is an arbitrary decision, although one often
governed by fairly commonsense options. When we interpret standardized summary
measures it is not their absolute levels but their relative levels that are important.
We are interested in how one standardized measure compares with another –
whether it is higher or lower, and by what order of magnitude. Standardization
is a procedure designed solely to facilitate COMPARISON. It seeks to overcome
the reality that differences in population composition can render comparisons
of summary measures misleading by ELIMINATING those differences from
the comparison. As implied earlier, this elimination may be less than total. Its
completeness depends on our (i) choosing the appropriate variable(s) to standardize
for and (ii) selecting suitably refined categories for those variables. But the major
point to grasp here is that STANDARDIZED SUMMARY MEASURES HAVE NO
MEANING OUTSIDE A COMPARATIVE SETTING. Furthermore, they can only
be compared if all measures were standardized using the same standard. It is not
appropriate to choose one standard composition for one population and a different
one for the population you wish to compare it to. To do so would run counter to the
aim of eliminating compositional difference from a comparison.
What data are required to be able to undertake a direct standardization? Given
the summary measure to be standardized, we require:
1. For each population we wish to compare, a set of equivalent measures (rates,
proportions, etc.) specific for compositional categories defined using the vari-
able(s) we wish to standardize for.
2. Preferably, although not essentially, the composition (i.e., the proportionate
distribution among compositional categories) of at least one of the populations
we wish to compare. This population becomes the standard population. As
intimated above, selecting the standard population is an arbitrary decision. The
composition of some other population may be used, and in some circumstances
is preferable (e.g., the national population when comparing several geographic or
socio-demographic subgroups of a national population), but as discussed below
there is advantage in choosing one of the populations to be compared.
Standardization 57

If this is done, a set of specific measures for the population chosen as standard is
not needed. There is no need to calculate a standardized summary measure for that
population – it is identical to its unstandardized summary measure, whose value
will already be known. For this standard population we would be standardizing to its
own composition, which intuition tells us is a ‘no change’ situation. The advantage
of selecting one of the populations to be compared as the standard population should
now be clear – one less standardization calculation is necessary.
To illustrate, suppose we wanted to directly standardize the general fertility rates
of two populations for age and marital status (recall that the GFR is live births
during year y divided by the mid-year female population aged 15–49 in year y – Eq.
1.10). Suppose further that we decided that our age categories would be the seven
5-year age groups 15–19 through 45–49, and our marital status categories would
be ‘never married’, ‘currently married’ and ‘formerly married’. To carry out the
necessary calculations we would need:
1. Fertility rates for each possible combination of our seven age and three marital
status categories (i.e., 7  3 D 21 age-marital status-specific rates) for both of the
populations to be compared (unless we had determined to make one of them the
standard population – we would then not need specific rates for that population).
2. Preferably, but not essentially, the proportionate distribution of one of the
populations among the same 21 categories (we could use as our standard
population composition any plausible proportionate distribution among those 21
categories).
Generalizing, we will denote any summary measure for population i that we wish
to standardize by M(i). The equation for calculating a directly standardized value of
M(i) is:

Ms .i/ D †c mi .c/ :ps .c/ (2.4)

Where Ms (i) denotes the standardized summary measure for population i (standard-
ized to the composition of population s); c denotes compositional categories
defined using the variable(s) for which we are standardizing (they might be
age categories, marital status categories, age-sex categories, age-marital status
categories, etc.); mi (c) D the specific measure equivalent to M(i) for category c
and population i; ps (c) D the proportion of the standard population s in category
c.
It is of interest to compare the righthand side of Eq. 2.4 and the expression for
M(i) (the unstandardized summary measure for population i) which is obtained
using the generalized Eq. 2.3. Using Eq. 2.3, the value of M(i) is given by:

M .i/ D †c mi .c/ :pi .c/

Where mi (c) and pi (c) are the values of m(c) and p(c) for population i.
58 2 Comparison: Standardization and Decomposition

The only difference between the righthand sides of this expression and Eq. 2.4 is
that in Eq. 2.4 pi (c) has been replaced by ps (c). The element pi (c) gives the actual
composition of population i. You can thus perhaps appreciate that the way direct
standardization works is to eliminate differences in composition from a compar-
ison by replacing the actual population composition element by the composition
element for the standard population. This process is repeated in calculating directly
standardized summary measures for each population being compared. The result is
a series of (standardized) measures with identical composition elements. It follows
that any differences between them are attributable to differences in the schedules
of mi (c) values; that is, solely to differences in composition-specific levels of the
phenomenon of interest (mortality, fertility, etc.), or solely to differences in the
underlying level of that phenomenon.
You can now, perhaps, also better appreciate why there is no need to calculate
a standardized summary measure for the standard population. For the standard
population s, Eq. 2.3 gives the unstandardized summary measure M(s) as:

M .s/ D †c ms .c/ :ps .c/

The process of direct standardization entails replacing the p(c) element in the
righthand side of this equation with the p(c) element for the standard population;
ps (c) is replaced by ps (c), or by itself. Thus we have, applying Eq. 2.4:

Ms .s/ D †c ms .c/ :ps .c/

Clearly no change occurs. From the preceding expressions Ms (s) D M(s) (the
standardized and unstandardized summary measures for the standard population are
equal), and as the value of M(s) is already known, no calculation using Eq. 2.4 to
find Ms (s) is necessary.
How do we use standardized summary measures to make comparisons which
are uncontaminated by differences in population composition (at least insofar
as we have standardized for the appropriate variables, and chosen appropriate
categories for them)? A useful device is standardized ratios. A standardized ratio
is the ratio of standardized summary measures (based on the same standard) for
two populations we wish to compare (i.e., the standardized summary measure for
the first population divided by that for the second population). It can be placed
alongside the equivalent unstandardized ratio (the ratio of the unstandardized
summary measures for the two populations), interpretation then proceeding on two
fronts:
1. The standardized ratio indicates the ‘true’ relative underlying level of the
demographic phenomenon under investigation in the first population compared
to the second.
2. Both ratios, looked at together, enable us to assess how standardization modifies
the conclusion as to the relative levels of the phenomenon under investigation
invited by the two unstandardized summary measures.
Standardization 59

Standardized and unstandardized ratios give a value for the population providing
the numerator assuming a value of 1.0 for the population providing the denomi-
nator. If a ratio has a value less than 1.0, the phenomenon under investigation has
a lower intensity in the numerator population; if it has a value greater than 1.0
the phenomenon has a higher intensity in the numerator population. The difference
between a ratio and 1.0 gives a proportionate measure of the degree to which the
phenomenon is less intense, or more intense, in the numerator population compared
to the denominator population.
When examining how standardization modifies the conclusion invited by a
comparison of two unstandardized summary measures, we are basically interested
in how the standardized ratio differs from the unstandardized ratio relative to a
pivotal value of 1.0. First, is it closer to or further away from 1.0? Second, is it
on the same side of 1.0 or on the opposite side? The ratio value 1.0 is ‘pivotal’
because that is the value that implies no difference between the populations being
compared. We are interested in whether the standardized ratio (i) is closer to 1.0 than
the unstandardized ratio (in which case standardization shows the two populations
to be more similar than the original summary measures indicated) or (ii) is further
away from 1.0 (in which case standardization shows the two populations to be
more different than we had thought), and (iii) in whether the two ratios both lie
above, both lie below, or lie on opposite sides of 1.0. Should the latter be the
case, standardization has altered the DIRECTION of the differential between the
populations, as well as its MAGNITUDE.
Naturally, when calculating pairs of a standardized and an equivalent unstan-
dardized ratio, it is essential that the same populations provide the numerator and
denominator in each calculation. A common approach when carrying out this type
of interpretative exercise after standardizing several summary measures using the
same standard population is to compare each population in turn with the standard
population. We thus end up with a series of standardized ratios that are directly
comparable, because they are based on a common comparator population, the
standard population. In our earlier notation, such standardized ratios take the form:

SR D Ms .i/ =M .s/ (2.5)

Despite contrary appearances, this is a ratio of two standardized summary measures

because, as we have seen, M(s), the unstandardized summary measure for the
standard population, is equal to Ms (s), the standardized summary measure for that
population. The equivalent unstandardized ratio is given by:

UR D M .i/ =M .s/ (2.6)

When the summary measure being standardized is the crude death rate, and
it is being standardized for age, a standardized ratio calculated using Eq. 2.5 is
commonly referred to as the comparative mortality factor (CMF) for population
i. It is a measure of how much higher, or lower, mortality is in population i than
in the standard population after standardizing for age. It can be shown to equal the
60 2 Comparison: Standardization and Decomposition

ratio of the number of deaths expected in the standard population, assuming it had
experienced the age-specific death rates of population i, to the actual number of
deaths in the standard population. The name ‘comparative mortality factor’ captures
the capacity of this measure, calculated for a series of populations (i, j, k, l, etc.),
to facilitate legitimate comparison among those populations. If we substitute the
letters D and d for the letters M and m in our generalized Eq. 2.4 (in recognition that
we are now dealing with crude, and specific, death rates), we have:

Ds .i/ D †c di .c/ :ps .c/

This can be rewritten:

Ds .i/ D †c di .c/ : .Ps .c/ =†c Ps .c// D †c di .c/ :Ps .c/ =†c Ps .c/

Where Ps (c) (as distinct from ps (c)) is the observed population in category c in
the standard population (as distinct from the proportion of total population in
category c).
If we multiply the righthand side by †c ds (c). Ps (c)/†c ds (c). Ps (c) (which clearly
equals 1) and then rearrange, we get:

Ds .i/ D .†c ds .c/ :Ps .c/ =†c Ps .c// : .†c di .c/ :Ps .c/ =†c ds .c/ :Ps .c//

And as †c ds .c/ :Ps .c/ =†c Ps .c/ D †c ds .c/ : .Ps .c/ =†c Ps .c// D †c ds .c/ :
ps .c/ D D .s/ (the crude death rate for the standard population, from Eq. 2.3):

Ds .i/ D D .s/ : .†c di .c/ :Ps .c/ =†c ds .c/ :Ps .c//

Whence:

CMF D Ds .i/ =D .s/ D †c di .c/ :Ps .c/ =†c ds .c/ :Ps .c/ (2.7)

This is an equation for obtaining the comparative mortality factor directly, and
verifies the claim made above that the CMF is the ratio of the number of deaths
expected in the standard population, were it to experience the age-specific death
rates of population i, to the observed (actual) number of deaths in the standard
population. The latter figure (which we can denote by Od (s)) may be known directly.
If it is there is no need to evaluate the denominator of Eq. 2.7. That is:

CMF D †c di .c/ :Ps .c/ =Od .s/ (2.8)

Let us now proceed with an example. We will again take the comparison of crude
death rates for Malaysia and Australia in 1988 that was introduced in Chap. 1. The
CDRs for the two populations were:
Standardization 61

Malaysia CDR D 4:93 deaths per 1; 000 mid-year population

Australia CDR D 7:25 deaths per 1; 000 mid-year population

We will call Malaysia population 1 and Australia population 2. This is an arbitrary

designation; we could just as easily allocate these numbers in reverse order. As
above, since we are now dealing with a particular summary measure, the CDR, we
will adopt the notation D for the CDR and d for a specific death rate, instead of
the generalized notations M and m (standing for summary, and specific, measure).
Using this notation we can first specify the equations for the two CDRs in the form
generalized in Eq. 2.3. We get:

D.1/ D †c d1 .c/ :p1 .c/

and:

D.2/ D †c d2 .c/ :p2 .c/

We can make these equations more specific by explicitly recognizing that the
compositional variables we suspect of distorting a CDR-based comparison of
mortality, and for which we want to standardize, are sex and age. Denoting these
by s and x respectively, we get:

D.1/ D †s †x d1 .s; x/ :p1 .s; x/ (2.9)

and:

D.2/ D †s †x d2 .s; x/ :p2 .s; x/ (2.10)

The d1 (s,x) and d2 (s,x) values needed to evaluate Eqs. 2.9 and 2.10 are the age-sex-
specific death rates for Malaysia and Australia given in Table 1.2 in Chap. 1. The
required p1 (s,x) and p2 (s,x) values (proportions of total population in each age-sex
category) are shown in Table 2.1. Using these data Eq. 2.9 yields a CDR of 4.84
deaths per 1,000 mid-year population for Malaysia and Eq. 2.10 yields a CDR of
7.16 deaths per 1,000 mid-year population for Australia. These figures are slightly
different from those given above because of the rounding of age-sex-specific death
rates in Table 1.2 to one decimal place and of proportions of total population in
age-sex groups to three decimal places.
Suppose we ask the following question: ‘What would the CDR for Malaysia be
if its population had the same age-sex composition as Australia’s population?’
In line with our earlier more general discussion, we obtain a directly standardized
CDR for Malaysia using Australia as the standard population by replacing the
compositional element in Eq. 2.9 with that for Australia from Eq. 2.10. The equation
we need to evaluate is:
62 2 Comparison: Standardization and Decomposition

Table 2.1 Proportions of Males Females

total population in age-sex
Age group Malaysia Australia Malaysia Australia
groups for Malaysia and
Australia, 1988 0 .014 .008 .013 .007
1–4 .057 .030 .054 .029
5–9 .061 .038 .058 .036
10–14 .056 .039 .054 .037
15–19 .053 .044 .051 .042
20–24 .050 .041 .049 .039
25–29 .043 .043 .045 .042
30–34 .036 .040 .038 .040
35–39 .030 .039 .031 .038
40–44 .024 .036 .023 .035
45–49 .020 .028 .020 .026
50–54 .017 .024 .017 .023
55–59 .013 .023 .014 .022
60–64 .010 .022 .011 .022
65–69 .007 .017 .008 .020
70–74 .005 .013 .005 .016
75–79 .003 .009 .004 .013
80–84 .001 .004 .002 .008
85C .001 .002 .001 .006
Source: United Nations Demographic Yearbook, 1990 and
Australian Bureau of Statistics, Estimated Resident Popula-
tion by Sex and Age: States and Territories of Australia June
1988 and Preliminary June 1989

D2 .1/ D †s †x d1 .s; x/ : p2 .s; x/ (2.11)

You can probably appreciate that this equation has the same form as Eq. 2.4,
the general equation for a directly standardized summary measure. The summary
measure is the CDR (denoted by D in place of M); the compositional categories
are defined by two variables, sex and age (denoted by s and x in place of c); and
the standard population is population 2 (Australia’s population). We can use the
data in Tables 1.2 and 2.1 to evaluate this equation. In doing this we construct the
calculating table shown as Table 2.2. The first column of figures is taken from Table
1.2, the second column from Table 2.1. To obtain the third column we multiply the
figures from the first two columns for each sex and age group. We then add the
answers in this column to obtain D2 (1), our standardized death rate (this is what the
righthand side of Eq. 2.11 instructs us to do).
The answer we obtain, D2 (1) D 10.42, is much higher than our unstandardized
CDR for Malaysia (D(1) D 4.84). It is a hypothetical CDR. It tells us that if
Malaysia (population 1) had had the age-sex-specific death rates actually recorded
for that country in 1988, but in combination with Australia’s (population 2’s) rather
than its own age-sex composition, its CDR would have been not 4.84, but 10.42.
Standardization 63

Table 2.2 Calculating table for obtaining, using Eq. 2.11, a 1988 CDR for Malaysia directly
standardized to the age and sex composition of the population of Australia
Sex and age group d1 (s,x) p2 (s,x) d1 (s,x).p2 (s,x)
Males 0 16.6 .008 .1328
1–4 1.7 .030 .0510
5–9 0.6 .038 .0228
10–14 0.6 .039 .0234
15–19 1.0 .044 .0440
20–24 1.5 .041 .0615
25–29 1.7 .043 .0731
30–34 2.0 .040 .0800
35–39 2.4 .039 .0936
40–44 3.4 .036 .1224
45–49 5.5 .028 .1540
50–54 9.9 .024 .2376
55–59 15.4 .023 .3542
60–64 25.5 .022 .5610
65–69 38.3 .017 .6511
70–74 62.7 .013 .8151
75–79 86.2 .009 .7758
80–84 144.3 .004 .5772
85C 173.4 .002 .3468
Females 0 12.8 .007 .0896
1–4 1.1 .029 .0319
5–9 0.5 .036 .0180
10–14 0.4 .037 .0148
15–19 0.5 .042 .0210
20–24 0.6 .039 .0234
25–29 0.9 .042 .0378
30–34 1.1 .040 .0440
35–39 1.5 .038 .0570
40–44 2.3 .035 .0805
45–49 3.2 .026 .0832
50–54 5.9 .023 .1357
55–59 9.9 .022 .2178
60–64 17.0 .022 .3740
65–69 27.8 .020 .5560
70–74 46.6 .016 .7456
75–79 69.8 .013 .9074
80–84 119.0 .008 .9520
85C 142.8 .006 .8568
†s †x d1 .s; x/  p2 .s; x/ D 10:4239
64 2 Comparison: Standardization and Decomposition

We can now compare this standardized CDR for Malaysia directly with the
CDR for Australia that we began with (7.16 deaths per 1000 mid-year population).
Having selected Australia as our standard population, the standardized CDR for
that population is identical to its unstandardized CDR. In the notation we have been
using:

D2 .2/ D D.2/

We can see that the standardized CDR for Malaysia (D2 (1) D 10.42) is higher than
that for Australia (D2 (2) D 7.16), a differential running in the opposite direction to
that between the respective unstandardized CDRs (D(1) D 4.84; D(2) D 7.16). We
need, however, to carry our interpretation of this result further. It is the exception
for standardization exercises to yield results as dramatic as this, and we need some
more precise indicator of the impact of standardization. We need to calculate an
unstandardized ratio and the equivalent standardized ratio. The former measures
the relative levels of mortality in the two populations as indicated by the two
unstandardized CDRs, D(1) and D(2). We have:

D.1/=D.2/ D 4:84=7:16 D 0:68

In keeping with the principles for interpreting standardized and unstandardized

ratios outlined earlier, this number suggests that the level of mortality in population
1 (Malaysia) is 0.68  1.00 D 0.32, or 32 % lower, than in population 2 (Australia)
(recall that 1.00 is the pivotal value when dealing with unstandardized and
standardized ratios – the ‘no difference’ value against which we assess ratios we
calculate). What if we now calculate the equivalent standardized ratio? This is given
by:

D2 .1/=D2 .2/ D 10:42=7:16 D 1:46

Comparing the lefthand side of this equation with Eq. 2.7, and remembering
that D2 (2) D D(2), you should recognize this as a comparative mortality factor
(CMF) (although with only two populations being compared, one of which is
the standard population, there will in this case be no other CMF with which to
compare it). Performing the same simple calculation as before, the indication after
standardization is that the level of mortality in Malaysia is 1.46  1.00 D 0.46, or
46 % higher, than in Australia. Thus, while the unstandardized CDRs suggest that
mortality was about one-third lower in Malaysia than in Australia in 1988, after
controlling for differences in age-sex composition we find that it was, in fact,
almost 50 % higher.
As already noted, this is a spectacular result. It is more common to find that the
direction of the differential between two populations in levels of a demographic
phenomenon is unaltered by standardization, but the magnitude of the differential
is modified. But another question may have occurred to you. What if we had made
Australia population 1 and Malaysia population 2, or made Malaysia, rather than
Standardization 65

Australia, the standard population? These are alternative ways of saying the same
thing. Both are arbitrary decisions. For the purpose of illustration we will retain our
original designation of Malaysia as population 1 and Australia as population 2, but
will see what happens when we select population 1 as the standard population in
place of population 2.
Our initial question is now: ‘What would the CDR for Australia be if its
population had the same age-sex composition as Malaysia’s population? We obtain
a directly standardized CDR for Australia by replacing the compositional element
in Eq. 2.10 with that for Malaysia from Eq. 2.9. The equation we need to evaluate
is:

D1 .2/ D †s †x d2 .s; x/ : p1 .s; x/ (2.12)

As before, this is an equation with the same form as Eq. 2.4, the general equation
for direct standardization; we are standardizing the CDR for population 2 using
population 1 as the standard population, and the compositional variables we are
endeavouring to control for are again sex (s) and age (x). Evaluating this equation
using relevant data from Tables 1.2 and 2.1, and with the aid of a calculating table
modelled on Table 2.2, we find that D1 (2) D 3.19.
In the same manner as previously, we can compare this standardized CDR for
Australia directly with the CDR for Malaysia we began with because, as the standard
population, Malaysia’s standardized CDR is identical to its unstandardized CDR.
That is:

D1 .1/ D D.1/

We can see that the standardized CDR for Australia (D1 (2) D 3.19) is lower than that
for Malaysia (D1 (1) D 4.84), a differential again running in the opposite direction
to that between the respective unstandardized CDRs (D(2) D 7.16; D(1) D 4.84).
As before, however, to summarize our finding more precisely we need to calculate
an unstandardized ratio and the equivalent standardized ratio. Normally, in doing
this, we would make the standard population the denominator population. However,
in this instance we will reverse that to facilitate comparison of the result obtained
with Malaysia as the standard population, with that obtained with Australia as the
standard population (as a result our standardized ratio will not fit the definition of a
comparative mortality factor). Our unstandardized ratio is the same as before:

D.1/=D.2/ D 4:84=7:16 D 0:68

Our standardized ratio is:

D1 .1/=D1 .2/ D 4:84=3:19 D 1:52

Having selected the other population as the standard population, a slightly different
standardized ratio is obtained for comparison with the unstandardized ratio (which
66 2 Comparison: Standardization and Decomposition

naturally remains unchanged). This often happens, but the broad conclusion that
whereas the unstandardized CDRs suggest that mortality was about one-third lower
in Malaysia than in Australia in 1988, once differences in age-sex composition are
controlled for it was, in fact, about 50 % higher, remains valid. Although we have
chosen a different standard, standardized the CDR for a different population, and
calculated a different number, our general conclusion is the same.
The obtaining of different standardized ratios does, though, warrant comment.
It implies that the result of a direct standardization – that is, the comparative
conclusion reached (as distinct from the numerical value of a standardized summary
measure) – is partly dependent on the choice of standard population. What one
ends up with is an indication of how two (or a series of) populations compare
when A PARTICULAR common population composition is assumed for them. The
stability of the conclusions one draws when different standard compositions are used
depends on the degree of variability from compositional category to compositional
category in the extent of inter-population differentials in specific measures. If, for
example, age-specific death rates in population A are at all ages higher, or lower,
than those in population B by about the same percentage factor, the choice of a
standard age composition will have little bearing on conclusions drawn from a direct
standardization of crude death rates for age. If, however, there is, say, a much wider
disparity in death rates at younger ages than at middle and older ages, choice of
a very young standard age structure will weight that disparity more heavily in the
calculation of age-standardized death rates than will choice of a much older one,
and will yield a wider standardized differential.
It follows that while the selection of a standard composition is an arbitrary
decision, some options are more sensible than others. The value of choosing
one of the populations being compared to provide the standard has already been
demonstrated. It obviates the need to perform standardization calculations for that
population. Another common approach, when comparing geographic or socio-
demographic subgroups of a national population, is to use the national population
as the standard population. This effectively uses the average composition of the
populations being compared as the standard, and that principle can also be extended
to comparisons of other populations which are not sugroups of a national population.
Another lesson from the finding that choosing a different standard population can
lead to different standardized ratios is to be careful in interpretation not to make too
much of minor differences between an unstandardized ratio and a corresponding
standardized ratio.
Reverting to our example, another question you might be asking concerns which
populations formed the numerators and denominators in the unstandardized and
standardized ratios we calculated. Population 1 consistently provided the numerator;
population 2 the denominator. This was because (i) population 2 was initially
the standard population and (ii) when we later made population 1 the standard
we needed to maintain the same relativity to be able to compare outcomes.
The initial designation of population 2 as the denominator population flowed
from the principle that where multiple pairwise comparisons based on the same
standard are to be made, using the standard population as denominator provides a
Standardization 67

common comparator, thus enabling a series of standardized ratios to be legitimately

compared. This is the principle that underpins the comparative mortality factor
(CMF), and the same idea can be extended to other demographic phenomena. But
what if we had started with population 1 as the standard population? We’d have got
the following unstandardized ratio:

D.2/=D.1/ D 7:16=4:84 D 1:48

and the following standardized ratios:

D1 .2/=D1 .1/ D 3:19=4:84 D 0:66

D2 .2/=D2 .1/ D 7:16=10:42 D 0:69

We would have concluded that the unstandardized ratio suggested that the level of
mortality was 1.48  1.00 D 0.48, or 48 % higher in population 2 (Australia) than
in population 1 (Malaysia), whereas the two standardized ratios indicated, on the
contrary, that it was 0.69  1.00 D 0.31 and 0.66  1.00 D 0.34, or 31 and 34 %
lower respectively. Thus, generalizing, we might say that while our unstandardized
CDRs pointed to mortality being about 50 % higher in Australia, our standardized
CDRs showed it to be around one-third lower. This is the same general conclusion
as we reached before. Then, we treated Australia as the ‘base’ population and noted
how much higher or lower mortality in Malaysia was, whereas this time the roles of
the countries have been reversed. But if you can convince yourself that if A is 50 %
higher than B, then B is one-third lower than A, you should be able to see that the
two broad conclusions amount to the same thing.
Before moving on, let us briefly consider the case of comparing, not two, but
several populations (1, 2, 3, : : : , n). An example might be a comparison of regional
CDRs within a country. In this case it would make sense to treat the national
population (i.e., the aggregate of all the regional populations) as the standard
population. If we denote this national standard population by N, we would calculate
directly standardized CDRs for the n regions as follows:

DN .1/ D †s †x d1 .s; x/ : pN .s; x/

DN .2/ D †s †x d2 .s; x/ : pN .s; x/
::::::::::::::::::::::::::::::::::::::::::::
DN .n/ D †s †x dn .s; x/ : pN .s; x/

When seeking to compare regions before and after standardization the sensible
approach would be to adopt the national population as a common comparator
population in pairwise comparisons with all regions. Its unstandardized CDR
(D(N)) would equal its standardized CDR (DN (N)), and interpretation would be
based on comparing a series of unstandardized ratios D(1) / D(N), D(2) / D(N),
: : : , D(n) / D(N) with a corresponding series of standardized ratios (comparative
mortality factors) DN (1) / DN (N), DN (2) / DN (N), : : : , DN (n) / DN (N).
68 2 Comparison: Standardization and Decomposition

Indirect Standardization

Indirect standardization is conceptually more complicated than direct standardiza-

tion, but in no way an inferior technique. Indeed it has an extremely appealing
attribute in that, unlike direct standardization, it does not require specific measures
to be known for each of the populations to be compared. This is obviously an
attraction when, as is not infrequently the case, such measures are simply not
available. But it is also an attraction when, because of small numbers of events, cases
or persons at risk in compositional categories, specific measures, although able to
be calculated, are highly unstable and subject to significant random error. This quite
often happens, for example, when dealing with mortality, especially when we wish
to compare small population subgroups, such as, for instance, Australia’s myriad
birthplace groups. By the time annual deaths by birthplace are broken down by 5-
year age groups and sex, numbers can get very small and subject to appreciable
annual fluctuation. Age-sex-specific death rates consequently can become quite
unstable, and this instability can translate into unstable direct standardization results.
The reliance of indirect standardization on a single, standard set of specific
measures, which are likely to pertain to a much larger population (e.g., a national
population), and therefore to be much more stable, is a very valuable attribute.
We will again begin by noting the data requirements in order to be able to apply
this method. Given the summary measure to be standardized, we require:
1. A single set of equivalent measures (rates, proportions, etc.) specific for compo-
sitional categories defined using the variable(s) we wish to standardize for. This
set of specific measures ideally will pertain to one of the populations we wish to
compare, or to some aggregate of those populations (e.g., the national population
when a series of geographic or socio-demographic subgroup populations are
being compared). These specific measures are our standard specific measures.
2. The compositions (i.e., proportionate distributions among compositional cate-
gories) of each of the populations we wish to compare. However, if one of these
populations is the source of our standard set of specific measures, we do not need
the composition of that population.
3. The unstandardized summary measures for each of the populations being
compared.
You will note that the above specification makes no reference to a ‘standard
population’; rather it refers to the population that provides a standard set of
specific measures. It is a common misconception that this population in indirect
standardization is the population to which the standardization is performed. With
indirect standardization there is strictly speaking NO single standard population.
We are still calculating values that eliminate differences in population composition
from a comparison, but not in a way that simply substitutes for the composition
of each comparator population that of a standard population. The procedure we
use is indirect (hence its name), and we do not calculate standardized summary
measures comparable to those calculated in direct standardization. Rather, we jump
Standardization 69

directly to the calculation of standardized ratios, in which, it can be shown, direct

standardization is implicit. But it is incorrect to refer to the population providing
the standard set of specific measures as the ‘standard population’.
The crux of indirect standardization is the generation, for a population, of
an expected value for the summary measure of interest, or an expected value
for its numerator, on the assumption that the standard set of specific measures
(rather than the actual set, which may be unknown) applied to the population’s
distribution among compositional categories. The ratio of the observed value of
the summary measure (or its numerator) to this expected value is then taken, and
is a standardized ratio indicating how much higher or lower the intensity of the
demographic phenomenon in question is in the population of interest compared
to the population providing the standard. The reason we can deal with ratios of
observed to expected values of either the summary measure, or just its numerator,
is that both observed and expected summary measures have the same denominator.
Hence, either way, the standardized ratio has the same value, because the common
denominators, if present, cancel one another out.
In general terms, an indirectly standardized ratio (ISR) for population i is given
by:

ISR D M .i/ =†c ms .c/ : pi .c/ (2.13)

That is, ISR is the ratio of the observed summary measure for population i (M(i))
to its expected value were the standard set of specific measures (ms (c)), rather than
the actual set (mi (c)), to apply to its observed population composition (pi (c)). This
equation can be rewritten as follows:

ISR D †c mi .c/ : pi .c/ =†c ms .c/ : pi .c/

In this form we can see clearly that we have the ratio of two directly standardized
summary measures – the specific measures elements (the m elements) in the
numerator and denominator differ, but the population composition elements (the
p elements) are the same. But, here the common composition element is not that
of the population providing the standard. It is that of population i. That is why
the method is called indirect standardization, and why it is incorrect to call the
population providing the standard set of specific measures ‘the standard population’.
With indirect standardization we are once again seeking to make comparisons
in which differences in population composition have been eliminated. But the
common population composition ends up being that of population i, not that of
population s. The ‘s’ now means something different – it means the population
providing a standard set of specific measures, not the population providing a
standard composition. We will return to this issue below.
If we now multiply numerator and denominator by †c Pi (c) (taking into account
that pi .c/ D Pi .c/ =†c Pi .c/, where pi (c) is the proportion of population i in
compositional category c and Pi (c) is the number of persons from population i in
that category), we get:
70 2 Comparison: Standardization and Decomposition

ISR D †c mi .c/ : Pi .c/ =†c ms .c/ : Pi .c/

Whence:

ISR D O .i/ =†c ms .c/ : Pi .c/ (2.14)

Where O(i) is the observed number of events, or persons with the demographic
status of interest, in population i.
Equation 2.14 is a calculating equation for the ISR that yields the ratio of
observed to expected numbers of events or persons of a particular demographic
status, whereas Eq. 2.13 yields the ratio of observed to expected values of the
summary measure. Either equation may be used, the choice perhaps depending on
the nature of available data.
The above is a generalized outline of indirect standardization. Most demographic
analysis texts deal with the technique exclusively in the context of the analysis of
mortality, and specifically the standardization of crude death rates for age. They
do not present equations for the ISR, but for the standardized mortality ratio (or
SMR). Epidemiologists also use the acronym SMR to mean standardized morbidity
ratio, when the measures they are standardizing pertain to levels of disease or illness,
rather than mortality. Equations 2.13 and 2.14 can readily be rewritten to yield the
standardized mortality ratio, which is merely a specific example (albeit a widely
encountered one) of an ISR. In line with earlier notation, this rewriting entails
substituting the letters D and d (standing for crude, and specific, death rates) for
the letters M and m (standing for summary, and specific, measures). Thus, we have:

SMR D D .i/ =†c ds .c/ : pi .c/ (2.15)

And:

SMR D Od .i/ =†c ds .c/ : Pi .c/ (2.16)

Where Od (i) is the observed number of deaths in population i.

It is usual to calculate SMRs separately for each sex, so that the CDRs
standardized are sex-specific CDRs and the compositional categories, c, are age
groups. It is, though, also possible to standardize conventional CDRs, in which case
the categories, c, may be either age-sex groups or simply age groups.
We need now to return to the point made earlier that the calculation of an ISR, and
by implication a SMR, is equivalent to standardizing directly, but with population
i, not population s, providing the standard (common) composition (perhaps age, or
age-sex, structure). This means that if we calculate ISRs or SMRs for a series of
populations (i, j, k, l, etc.), although we may use a common standard set of specific
measures, each calculation uses a different standard population composition – the
compositions of populations i, j, k, l, etc. respectively. We noted earlier, in discussing
Standardization 71

the results of our direct standardizations of the CDRs for Malaysia and Australia
using first one, then the other, population’s composition as the standard, that the
values of standardized ratios don’t necessarily remain constant when different
standards are adopted. It follows that while an ISR or a SMR provides a legitimate
PAIRWISE comparison between population i and population s (the population
providing the standard set of specific measures) it is NOT, strictly speaking, valid
to compare ISRs or SMRs for a SERIES of populations i, j, k, l, etc., even though
they are calculated using the same population s.
That said, the reality is that such comparisons commonly are made. Thus, the
practical recommendation is that you should make them mindful of the technical
invalidity of doing so. The advantages of indirect standardization – the fact that
it can be used when specific measures for the populations of interest are either
unknown or based on such small numbers of cases as to be likely to introduce
serious error into a direct standardization – have to be weighed against the downside
of the technical invalidity of comparing ISRs or SMRs. One piece of advice is to
not place too much emphasis in interpretation upon relatively small differences
between ISRs or SMRs. Another is to be particularly careful in situations where
the compositions of the populations for which ISRs or SMRs are being compared
are known, or are likely, to be very different. A situation in point might be a
comparison of SMRs for Australian birthplace groups. Because migration tends to
take place in early adulthood, and different birthplace groups have largely come to
Australia over different periods of time, the age structures of different groups tend
to be quite different. This is the sort of environment in which considerable bias
deriving from the effective use of different standard age structures could enter a
comparison of SMRs.
We will again finish with an example, once again comparing the 1988 CDRs for
Malaysia and Australia. We will calculate a SMR in which we will again seek to
control for differences in age-sex composition, and will continue to regard Malaysia
as population 1 and Australia as population 2. Suppose further that we decide to
treat the age-sex-specific death rates for population 2 (Australia) as the standard set
of specific measures, so that we will calculate a SMR for Malaysia, comparing its
mortality to that of Australia.
Using Eq. 2.15 we have:

SMR D D.1/=†c d2 .c/ : p1 .c/ D D.1/=†s †x d2 .s; x/ : p1 .s; x/

The denominator of this expression is the righthand side of Eq. 2.12, which was
earlier evaluated using data from Tables 1.2 and 2.1 as having a value of 3.19.
D(1) D 4.84, so:

SMR D 4:84=3:19 D 1:52

This is identical to the standardized ratio we obtained earlier using direct standard-
ization with Malaysia as the standard population. If, in calculating our SMR, we
reverse things and make the age-sex-specific death rates for population 1 (Malaysia)
the standard set of specific measures, we have:
72 2 Comparison: Standardization and Decomposition

SMR D D.2/=†c d1 .c/ : p2 .c/ D D.2/=†s †x d1 .s; x/ : p2 .s; x/

The denominator of this expression is the righthand side of Eq. 2.11, which was
earlier evaluated using data from Tables 1.2 and 2.1 as having a value of 10.42.
D(2) D 7.16, so:

SMR D 7:16=10:42 D 0:69

This is the inverse of the standardized ratio (1.46) we obtained using direct
standardization with Australia as the standard population. It is the inverse because
indirect standardization always makes the population providing the standard set of
specific measures the comparator (denominator) population. In essence, though,
we have obtained the same results using indirect standardization as we did using
direct standardization. Our first SMR tells us that mortality was about 50 % higher
in Malaysia; our second tells us that it was about one-third lower in Australia.
The two results mean the same, and as with direct standardization, each SMR
could be compared with the equivalent unstandardized mortality ratio to assess how
standardization changed the picture presented by the unstandardized CDRs.

Some Additional Points About Standardization

The following are some additional theoretical points about the standardization
techniques we have discussed:
1. The categories defined for any compositional variable by which one is stan-
dardizing need to cover the entire population forming the denominator of the
summary measure being standardized. Take, for example, the crude birth rate
(CBR), which has as its denominator the mean (or mid-year) total population. If
we were standardizing the CBR for age and sex, many age-sex-specific fertility
rates (mi (c) values) would be zero – those for males and for females younger
than 15 and 50 or older. We would not, however, be justified in using population
composition (pi (c)) values which expressed females in childbearing age groups
as proportions only of total females of childbearing age. Because the CBR uses
total population as its denominator, females in childbearing age groups must
be expressed as proportions of total population for standardization purposes.
It would, however, be a different matter were we standardizing the general
fertility rate (GFR). Then, because the denominator consists only of women
of childbearing age, it is appropriate that population composition (pi (c)) values
should express females in childbearing age groups as proportions of total females
of childbearing age.
2. Some important theoretical questions arise in standardization when a population
composition variable is itself related to the demographic process to which the
Decomposition 73

summary measure being standardized relates. Direct standardization of the CDR

is a good example. The age-sex composition of a population can be considered
to be dependent on its level of mortality (and also its level of fertility) in the
past. If change in the mortality level usually precedes change in the population
(age-sex) composition, it may not really be reasonable to ask what the CDR in a
country would be if its population’s age-sex composition changed to that of the
population of another country. Consider the questions asked earlier as the basis
for direct standardization of the CDRs for Malaysia and Australia: ‘What would
the CDR for Malaysia be if its population had the same age-sex composition as
Australia’s population?’ (Answer: D2 (1)); ‘What would the CDR for Australia be
if its population had the same age-sex composition as Malaysia’s population?’
(Answer: D1 (2)). In both cases we are effectively assuming that population
composition can be changed without prior change in age-sex-specific mortality
rates. Directly standardized summary measures are almost always interpreted as
answering questions of the type we posed, but we could also interpret D2 (1)
to be the CDR Australia would have if Malaysia’s age-sex-specific death rates
prevailed, and D1 (2) to be the CDR Malaysia would have if Australia’s age-sex-
specific death rates prevailed. These are answers to different questions. The issue
is: What changes first, the death rates or population composition? If they change
together, in conjunction, say, with change in fertility levels, there may not really
be a reasonable question to answer using a standardized rate.

Decomposition

As already noted, decomposition is also called in the literature component analysis

and components of difference analysis. It is a variation of standardization and is
concerned with the same general problem – making comparisons between summary
measures more meaningful. While with standardization we endeavour to eliminate
from a comparison the influence of one or more compositional variables, so that
the comparison becomes, as far as is practicable, uncontaminated by compositional
differences, with decomposition we aim to break down (decompose) a difference
between two summary measures into components attributable to: (i) differences in
population composition, and (ii) differences in measures specific for compositional
categories. Having done this, the strategy is to assess the relative importance of
the differences in composition and in category-specific measures in producing the
overall difference observed.
This general description may lead you to think that we only ever decompose
the difference between two summary measures into two components. That is not
the case; decompositions into more than two components are not uncommon, but
that is a complication we will put to one side for now. Possibly the most frequent
application of decomposition is in situations where the difference between summary
measures that is of interest entails change in a measure over time for a single
74 2 Comparison: Standardization and Decomposition

population. In this application, change is decomposed into change due to changes in

population composition and ‘real’ change (in specific measures).
It can be demonstrated quite simply that the concept of separation into compo-
nents is implicit in direct standardization. Suppose we have two summary measures
M(1) and M(2) which relate to populations P(1) and P(2). If we adopt P(1) as the
standard population and calculate a directly standardized value of M(2) (i.e. M1 (2)),
then it is clear that:

M.1/  M.2/ D .M.1/  M1 .2// C .M1 .2/  M.2// (2.17)

This equation splits the difference between the summary measures M(1) and M(2)
into two components:
1. M(1)  M1 (2), which can be rewritten as M1 (1)  M1 (2). This component mea-
sures that part of the overall difference M(1)  M(2) which is attributable to
differences in measures specific for compositional categories, since it measures
the difference between the summary measures with differences in composition
controlled (directly standardized) for.
2. M1 (2)  M(2). This component measures that part of the overall difference which
is left after controlling for differences in composition. In other words it is the
component which is attributable to differences in composition.
In similar fashion, if we adopt P(2) as the standard population and calculate a
directly standardized value of M(1) (i.e., M2 (1)), then it is clear that:

M.1/  M.2/ D .M.1/  M2 .1// C .M2 .1/  M.2// (2.18)

This equation likewise splits the difference between M(1) and M(2) into two
components:
1. M2 (1)  M(2), which can be rewritten as M2 (1)  M2 (2), also measures that part
of the overall difference M(1)  M(2) which is attributable to differences in
measures specific for compositional categories. Again it measures the difference
between the summary measures with differences in composition controlled
(directly standardized) for.
2. M(1)  M2 (1), which again measures that part of the overall difference which
is left after controlling for differences in composition. In other words it, too,
measures the component of the overall difference which is attributable to
differences in composition.
We have obviously derived two ways of measuring the component of the differ-
ence between M(1) and M(2) which is due to differences in measures specific for
compositional categories, and two ways of measuring the component attiributable
to differences in composition. Each pair of expressions is likely to give slightly
different results, in the same way as choosing a different standard population earlier
gave rise to slightly different standardized ratios. Hence the usual practice is to take
Decomposition 75

averages. The component of the difference M(1)  M(2) that is due to differences in
composition-specific measures is given by:

CSM D 0:5 .M.1/  M1 .2/ C M2 .1/  M.2// (2.19)

and the component due to differences in population composition is given by:

CC D 0:5 .M.1/  M2 .1/ C M1 .2/  M.2// (2.20)

Where M(1) and M(2) are the summary measures for populations 1 and 2; M1 (2)
is the summary measure for population 2 directly standardized to the composition
of population 1; and M2 (1) is the summary measure for population 1 directly
standardized to the composition of population 2.
Let us now resurrect yet again our example involving CDRs for Malaysia and
Australia. In the notation we were using with that example:

M.1/  M.2/ D D.1/  D.2/ D 4:84  7:16 D 2:32

This is the difference between the two CDRs. The values of elements required to
evaluate Eqs. 2.19 and 2.20 which we obtained earlier in this chapter were:

D.1/ D 4:84I D.2/ D 7:16I D1 .2/ D 3:19I D2 .1/ D 10:42

From Eq. 2.19 we have that:

CSM D 0:5 .D.1/  D1 .2/ C D2 .1/  D.2//

D 0:5 .4:84  3:19 C 10:42  7:16/
D 2:455

And from Eq. 2.20 we have that:

CC D 0:5 .D.1/  D2 .1/ C D1 .2/  D.2//

D 0:5 .4:84  10:42 C 3:19  7:16/
D 4:775

The first thing we will note is that:

CSM C CC D 2:455 C .4:775/ D 2:32 D D.1/  D.2/

This is what we expect; the sum of the two components CSM and CC should always
equal the difference between the two summary measures we are decomposing. If
this is not the case it is a clear sign that something is wrong and that calculations
should be checked.
76 2 Comparison: Standardization and Decomposition

But how do we interpret our results? We examine the components (2.455 and
4.775) in relation to the difference between the two summary measures (2.32)
paying particular attention to their signs. What is important is not the sign on a
component itself (i.e., whether it is positive or negative), but how that sign compares
with the sign on the overall difference between the two summary measures. If the
sign on a component is the same as that on the overall difference, that component
helped produce the overall difference. If the sign on a component is opposite to
that on the overall difference, that component has partially offset, or moderated the
overall difference (i.e., made it less substantial than it would otherwise have been).
In our example, the difference between the two summary measures is negative
because we happened to designate the population with the smaller CDR (Malaysia)
population 1 and that with the larger CDR (Australia) population 2. The negative
difference equates with a higher CDR for Australia. Any component with a negative
sign helped to produce this outcome. The specific measures component CSM has
the opposite sign to the difference between the summary measures; this means that
differences in composition-specific measures operated in the opposite direction –
that is, in the direction of producing a higher CDR not for Australia, but for
Malaysia. As this component on its own indicates the direction of the differential in
the underlying levels of mortality in the two countries, we confirm earlier findings
that the underlying level of mortality was in fact higher in Malaysia. The other
component, CC , capturing differences in population composition, has a negative
sign. What is more, its absolute value (i.e., its numeric value disregarding the
negative sign) is as large as those of the CSM component and the difference between
the summary measures combined. This means that differences in population
composition both cancelled out the effect of differences in composition-specific
measures and produced the overall differential in the CDR favouring Australia.
The two components operated in opposite directions, one favouring a higher CDR
for Malaysia and the other a higher CDR for Australia. The latter (composition)
component was the stronger, and the overall difference between the summary
measures reflects this.
What if we had designated Australia as population 1 and Malaysia as population
2? The difference D(1)  D(2) would have been 7.16  4.84 D 2.32, and the two
components would have had the values CSM D 2.455 and CC D 4.775. In other
words, all the quantities of interest would have had the same numeric values, but
their signs would have been the opposite of what they were above. Our conclusion
would, however, be unaffected. Remember, it is not the signs themselves that matter,
but how they compare. The signs on components can only be interpreted relative
to the sign on the overall difference between our summary measures (in this case
relative to the sign on D(1)  D(2)). The sign on CSM is still opposite to that on
D(1)  D(2) (they are now just respectively negative and positive, instead of positive
and negative), and the sign on CC is still the same as that on D(1)  D(2) (they are
just both now positive, instead of both being negative). We are still going to conclude
that the CSM component favoured a higher CDR for Malaysia, but was cancelled out
and more by the CC component working in the opposite direction, giving rise to
an overall differential in favour of Australia. Clearly it doesn’t matter which way
Decomposition 77

round we designate the populations as populations 1 and 2. We arrive at the same

conclusion.
We noted earlier that each of our components could be measured in two ways,
and that the answers each gave were likely to be a little different. As a consequence,
in Eqs. 2.19 and 2.20 we averaged the two expressions for each component. The
phenomenon which produces (generally slight) inequalities between the answers
obtained from the alternate ways of measuring each component is known as
interaction. It is given, in the general case, by the equation:

I D 0:5 .M.1/  M1 .2/  M2 .1/ C M.2// (2.21)

One way of viewing the interaction term I is as a measure of the degree of faith we
can have in CSM and CC . We are saying that the precise values of CSM and CC are
uncertain, but that they respectively lie somewhere in the ranges:

CSM ˙ I and CC ˙ I

From a practical point of view the important consideration is whether possible

variations in the ‘true’ values of CSM and CC within the ranges defined here raise
any doubts over the general conclusion we have drawn. We need to bear in mind that
not all combinations of ‘true’ values of CSM and CC in these ranges are possible;
once the ‘true’ value of one component has been determined the ‘true’ value of
the other is fixed by the requirement that CSM C CC equals the difference between
the two summary measures that is being decomposed. We could not, for example,
have CSM C I in combination with CC C I; whatever is added to or subtracted from
CSM to get its ‘true’ value must in turn be subtracted from or added to CC . Usually,
however, ‘true’ values do not alter our general conclusion, since I tends to be small
in comparison with CSM and CC . It is nevertheless a wise precaution to check your
interaction term.
In our Malaysia/Australia example the interaction term has the value:

I D 0:5 .D.1/  D1 .2/  D2 .1/ C D.2//

D 0:5 .4:84  3:19  10:42 C 7:16/
D 0:805

Hence the ‘true’ values of CSM and CC lie in the ranges:

CSM W 1:65 to 3:26

CC W 3:97 to  5:58

While these ranges are not especially narrow, having regard to the requirement noted
above that only combinations of values such that CSM C CC D 2.32 are valid, no
major alteration to our earlier interpretation of the CSM and CC values we calculated
is called for. In this example, in the limiting cases, CSM D 1.65 would pair with
78 2 Comparison: Standardization and Decomposition

CC D 3.97 and CSM D 3.26 would pair with CC D 5.58, with intermediate values
of CSM paring with intermediate values of CC .

Decomposition by Reverse Subtraction

The decomposition approach outlined above requires two direct standardizations to

be carried out before it can be implemented. Reverse subtraction is an alternative
approach to decomposition which does not have this requirement and which also is
very useful in carrying out more complex decompositions (i.e., decompositions into
three or more components). We will, however, stick with the simple two component
situation in outlining the approach. We will also couch the discussion in terms of
change in the value of a summary measure over time for a single population,
although the approach is equally applicable to decomposing the difference between
summary measures for two populations at a point in time.
Suppose a summary measure for a population has the values V(1) at some earlier
date and V(2) at some later date. Making use of Eq. 2.3, these measures can be
written in the form:

V.1/ D †c m1 .c/ : p1 .c/ (2.22)

V.2/ D †c m2 .c/ : p2 .c/ (2.23)

Where c denotes a population composition category; m1 (c) and m2 (c) are

specific measures for compositional categories at times 1 and 2; p1 (c) and p2 (c)
are proportions of the populations comprising the denominators of V(1) and V(2) in
compositional categories at times 1 and 2.
Change in the measure V between time 1 and time 2 is given by:

V.2/  V.1/ D †c m2 .c/ : p2 .c/  †c m1 .c/ : p1 .c/ (2.24)

Suppose we define quantities m(c) and p(c) ( D ‘delta’) representing the

changes in m(c) and p(c) values between time 1 and time 2. Thus:

m .c/ D m2 .c/  m1 .c/ (2.25)

p .c/ D p2 .c/  p1 .c/ (2.26)

We can rearrange Eqs. 2.25 and 2.26 to give:

m2 .c/ D m1 .c/ C m .c/ (2.27)

Decomposition by Reverse Subtraction 79

p2 .c/ D p1 .c/ C p .c/ (2.28)

In other words, our m(c) and p(c) values at time 2 equal the sums of (i) their values
at time 1 and (ii) the changes in their values between times 1 and 2.
We can now return to Eq. 2.24 and substitute for m2 (c) and p2 (c) from Eqs. 2.27
and 2.28. We get:

V.2/  V.1/ D †c .m1 .c/ C m .c// .p1 .c/ C p .c//  †c m1 .c/ :p1 .c/
D †c .m1 .c/ :p1 .c/ C m1 .c/ :p .c/ C m .c/ :p1 .c/
Cm .c/ :p .c//  †c m1 .c/ :p1 .c/
D †c m1 .c/ :p .c/ C †c p1 .c/ :m .c/ C †c m .c/ :p .c/
(2.29)

We have now decomposed the change in summary measure V into three compo-
nents:
1. A component †c m1 (c).p(c) due to changes in population composition. This
is what p(c) refers to – change in population composition.
2. A component †c p1 (c). m(c) due to changes in composition-specific mea-
sures. This is what m(c) refers to – change in specific measures.
3. An interaction component, identifiable by the fact that it incorporates more than
one delta () item. This is a general rule with component analysis by reverse
subtraction – the substantive components are expressions featuring single delta
items; expressions featuring multiple delta items are interaction components, or
terms.
To obtain Eq. 2.29 we substituted for m2 (c) and p2 (c) in Eq. 2.24. We could
equally have used Eqs. 2.25 and 2.26 to find expressions for m1 (c) and p1 (c) and
substituted these instead. You may like to go through this exercise. It should lead
you to the following alternative equation for V(2) - V(1):

V.2/  V.1/ D †c m2 .c/ :p .c/ C †c p2 .c/ :m .c/  †c m .c/ :p .c/
(2.30)

As in Eq. 2.29, we have decomposed the change in summary measure V into three
components associated with changes in population composition and in composition-
specific measures, and with interaction. We can, however, proceed to a much simpler
expression of the components of change; one which has the added benefit that the
interaction terms disappear. Because the righthand sides of Eqs. 2.29 and 2.30 both
equal V(2)  V(1), we can say that half their sum also equals V(2)  V(1). Thus:
80 2 Comparison: Standardization and Decomposition

h
V.2/  V.1/ D 0:5 †c m1 .c/ :p .c/ C †c p1 .c/ :m .c/ C †c m .c/ :p .c/
i
C †c m2 .c/ :p .c/ C †c p2 .c/ :m .c/  †c m .c/ :p .c/

D 0:5 Œ†c .m1 .c/ C m2 .c// :p .c/ C †c .p1 .c/ C p2 .c// :m .c/
D †c m .c/ :p .c/ C †c p .c/ :m .c/ (2.31)

Where m .c/ and p .c/ are the averages of the m(c) and p(c) values at times 1 and 2
for each compositional category. In other words:

m .c/ D 0:5 .m1 .c/ C m2 .c//

p .c/ D 0:5 .p1 .c/ C p2 .c//

We now have the change in a summary measure V decomposed into two com-
ponents, one associated with changes in population composition (the p(c) com-
ponent) and the other associated with changes in specific measures (the m(c)
component). Equation 2.31 is the one you would use in working through a real
example, and if doing the calculations by hand it would again make sense to draw
up a calculating table. This (or an excel spreadsheet) would have columns headed:

m1 .c/ I m2 .c/ I p1 .c/ I p2 .c/ I m .c/ I m .c/ I p .c/ I p .c/ I m .c/ :p .c/ I
p .c/ :m .c/

Values of the two components would be found by summing the final two columns.
These should then be added and checked against the value of V(2)  V(1), which
their sum is supposed to equal. This provides a quick check to see whether some
simple mechanical error has been made.
When we use this methodology to decompose change over time it is obvious that
V(2) should relate to the later point in time. When using it to decompose a difference
between two populations at a single point in time the decision as to which population
should be population 1 and which population 2 is again arbitrary. The main thing
is to be consistent. Once a designation has been made, stick with it. Make sure all
the V(1), m1 (c) and p1 (c) values relate to one population and all the V(2), m2 (c) and
p2 (c) values relate to the other one. Switching designations and mixing up values
will cause all sorts of strife.
It was mentioned above that the reverse subtraction approach to decomposition
can be extended to deal with more complex situations than the two-component one
discussed here. The following is a recipe for applying the approach in any situation.
Step 1: If the measure to be decomposed is denoted by V, obtain an expression
(equation) for V in terms of the factors you wish to recognize in the decompo-
sition. This expression will invariably involve summing terms over a series of
compositional categories of some sort.
Decomposition by Reverse Subtraction 81

Step 2: Write the difference, or change over time, in the measure V (i.e.,
V(2)  V(1)) in terms of the expression for V obtained at step 1. Elements in
the resulting equation should take the form x(1) where they relate to population,
or time, 1 and x(2) where they relate to population, or time, 2, where x is any
variable recognized on the righthand side of the equation for V.
Step 3: Rewrite each element in the expression for V(2) in the form x(2) D x(1) C x,
where x is the difference, or change, in x between populations, or times, 1 and
2 (i.e., x D x(2)  x(1)).
Step 4: Expand the resulting expression for V(2)  V(1) and collect together terms
which feature the same combinations of  (delta) items.
Step 5: Go back to the expression for V(2)  V(1) obtained at step 2 and rewrite
each element in the expression for V(1) in the form x(1) D x(2)  x (note the
negative sign).
Step 6: Again expand the resulting expression for V(2)  V(1) and collect together
terms which feature the same combinations of  items.
Step 7: Add together the expanded expressions for V(2)  V(1) obtained at steps 4
and 6, then divide by 2 and simplify where possible.
In the expression obtained for V(2)  V(1) at step 7, those components which
are functions of single  items are the components of the overall difference or
change due to that particular factor (i.e., the x component is the component due to
factor x). Any component which is a function of two or more  items is a separate
interaction factor. Such components are commonly added together to form a single
(generally small) overall interaction component.
The data presented in Table 2.3 are results from a reverse subtraction decompo-
sition in which the mechanics of a continuous 25-year increase in the ‘illegitimacy
ratio’ for New Zealand’s non-Maori population were investigated (Carmichael,
1985). Although the example is dated its value in illustrating how to go about
interpreting a set of decomposition results is considerable. The nowadays rather
quaintly named ‘illegitimacy ratio’ measures the percentage of all live births in a
population that are non-marital; that is, it is the ratio of non-marital live births to
total (marital plus non-marital) live births. It is a measure whose value can change in
response to change in four factors: in age-specific rates at which unmarried women
bear children; in age-specific rates at which married women bear children; in the
propensity of women of reproductive age to be married at different ages; and in
the age structure of the female population of reproductive age. To perform the
decomposition it was necessary to obtain an equation in terms of these four factors,
and then to follow the steps outlined above. The equation used was:

R D .†xD15;49 fx :qx :ix / = Œ†xD15;49 fx :qx :ix C †xD15;49 fx : .1  qx / :jx  (2.32)

Where R D illegitimacy ratio; x D a single-year age group; fx D the proportion of

all females of childbearing age aged x; qx D the proportion of females aged x
not currently married; ix D the non-marital fertility rate per unmarried woman
for females aged x; jx D the marital fertility rate per currently married woman for
females aged x.
82 2 Comparison: Standardization and Decomposition

Table 2.3 Decomposition of change in the New Zealand Non-Maori illegitimacy ratio, 1951–
1976
1951–1962 1962–1966 1966–1971 1971–1976
Starting ratio (R(1)) 4:38 6:51 9:49 11:23
Finishing ratio (R(2)) 6:51 9:49 11:23 13:83
Total change (R(2)  R(1)) 2:13 2:98 1:74 2:60
Change due to:
Non-marital fertility rates 4:19 1:04 1:39 1:64
Marital fertility rates 0:15 1:64 0:87 3:42
Marriage pattern 1:45 0:28 0:28 1:31
Age structure 0:34 0:39 0:29 0:39
Interaction 0:81 0:19 0:05 0:10
Source: Carmichael (1985: 170)

The algebra involved in expanding Eq. 2.32 in accordance with steps 1–7 above
was complex and is not reproduced here. To interpret the results obtained (Table 2.3)
we need first to ascertain what the signs (C or ) on the various components
mean. For each of the four periods for which results are presented the illegitimacy
ratio increased; the ‘total change’ figures, found by subtracting the ‘starting ratio’
from the ‘finishing ratio’ (R(2)  R(1)), are all positive. In line with the principle
previously enunciated for interpreting components, positive components, having
the same sign as the ‘total change’ figures, contributed to those changes; i.e., they
contributed to the rise in the illegitimacy ratio over the relevant period. Negative
components partially offset, or moderated, this upward momentum.
Under what circumstances would we expect each component to be causing the
illegitimacy ratio to rise, and under what circumstances would we expect each to be
causing it to fall?
1. Other things remaining equal, increasing non-marital fertility rates will boost
the number of non-marital compared to marital births, and so tend to increase
the illegitimacy ratio. A positive ‘non-marital fertility rates’ component thus
attests to rising non-marital fertility rates tending to raise the illegitimacy ratio; a
negative component points to falling non-marital fertility rates tending to lower
it.
2. Other things remaining equal, increasing marital fertility rates will boost
the number of marital compared to non-marital births and, as marital births
contribute only to its denominator, tend to lower the illegitimacy ratio. Thus,
a positive ‘marital fertility rates’ component attests to falling marital fertility
rates tending to raise the illegitimacy ratio; a negative component points to rising
marital fertility rates tending to lower it.
3. A change in the marriage pattern that results in women of reproductive age
being less likely to be married increases exposure to the risk of giving birth while
unmarried and thus tends to raise the illegitimacy ratio. A positive ‘marriage
Decomposition by Reverse Subtraction 83

pattern’ component is an indication that such a change occurred (ages at marriage

rose, the propensity to marry fell, the rate of divorce increased, the rate of
remarriage following divorce fell, or a combination of these things happened) and
tended to raise the illegitimacy ratio; a negative component indicates an increased
tendency for women of reproductive age to be married, a consequent reduced
exposure to non-marital (and increased exposure to marital) childbearing, and
hence a tendency to lower the illegitimacy ratio.
4. A change in age structure among women of reproductive age is likely to result
in a rise in the number of non-marital compared to marital births if it renders
the younger reproductive ages (15–19 years and to some extent 20–24 years)
more important. This is so because, regardless of age-specific rates of non-
marital childbearing, these age groups provide the overwhelming bulk of women
who are unmarried (mostly never married) and thus exposed to the risk of
such childbearing. A positive ‘age structure’ component is likely to mean that
an increase in the importance of the younger reproductive ages has tended to
raise the illegitimacy ratio; a negative component probably indicates that the
importance of these age groups declined, and tended to lower the illegitimacy
ratio.
With this background we can proceed to interpret Table 2.3. Between 1951 and
1962, rising non-marital fertility rates were overwhelmingly the main reason for the
rise in the illegitimacy ratio. Indeed the ‘non-marital fertility rates’ component was
almost twice as large as the ‘total change’. It was offset principally by changes in the
marriage pattern unfavourable to non-marital childbearing. This period coincided
with the post-war marriage ‘boom’, with ages at marriage falling and the propensity
to ever marry increasing. These trends tended to reduce the period of theoretical
exposure to the risk of non-marital childbearing, but they could not completely
offset rising levels of sexual activity outside marriage and the increased non-marital
fertility rates that resulted.
Between 1962 and 1966 the pattern changed. Non-marital fertility rates con-
tinued to rise and therefore to boost the illegitimacy ratio, but they were not the
dominant force they had been earlier. Instead, falling marital fertility rates associated
with the advent of new contraceptive technology (principally the oral contraceptive
pill) in the early 1960s assumed this mantle. As during 1951–1962, an increasing
importance of the younger reproductive age groups was a minor factor in the
illegitimacy ratio rising.
A somewhat similar pattern prevailed between 1966 and 1971, although rising
non-marital fertility rates regained the status of dominant force from falling marital
fertility rates, and age structure changes became marginally unfavourable to the
upward trend. Then, between 1971 and 1976 a totally new causative pattern
emerged. Having for so long been an influence for increase in the illegitimacy ratio,
non-marital fertility rates fell and became a force for decline. This was a period
when induced abortion became more accessible, and with informal cohabitation
also proliferating it is possible that unmarried women became contraceptively more
adept as well. However, this reversal was more than compensated for by accelerated
84 2 Comparison: Standardization and Decomposition

marital fertility decline and a sharp turnaround in the role played by the ‘marriage
pattern’ component as the marriage ‘boom’ ended and ages at marriage began to
rise. Both trends were partly attributable to the aforementioned improved access to
abortion, which saw marriages precipitated by bridal pregnancy, and consequently
early marital fertility, both fall dramatically.
This example serves to graphically illustrate that considerable complexity may
underlie what on the surface appears to be a straightforward, ongoing secular
trend. Decomposition techniques can greatly enhance our understanding of such
demographic trends.

Reference

Carmichael, G. A. (1985). Non-marital pregnancies in New Zealand since the second world war.
Journal of Biosocial Science, 17, 167–183.
Chapter 3
The Cohort and Period Approaches
to Demographic Analysis

Conceptualizing Problems in Demographic

Analysis: Lexis Diagrams

As the title to this chapter suggests, demographers adopt either (or both) of
two perspectives, or approaches, in analysing demographic data. They are the
cohort approach to demographic analysis and the period approach to demographic
analysis. The difference between these two approaches is most readily understood
by making use of a graphical device known as the Lexis diagram. Named for
Wilhelm Lexis (1837–1914), a German statistician, actuary and economist who was
the first to use it, the Lexis diagram is also an invaluable tool for conceptualizing the
construction of demographic measures and the procedures followed in demographic
estimation and projection. The first part of this chapter is devoted to describing it
and the concepts on which its use rests.

Time and Its Straight Line Representation

For purposes of analysis, demographers must generally classify their data accord-
ing to certain criteria, and often those criteria involve some concept of time.
Demographic events (births, deaths, migratory moves, marriages, divorces, etc.)
are frequently classified by age or some sort of duration (of marriage, divorce,
residence, etc.), and are almost always classified by period (usually calendar year)
of occurrence. Populations are also frequently classified by age or some sort of
duration, and also by date of enumeration (the date at which the population count
was made). For example, deaths are frequently classified by age of the deceased,
births by age of mother and/or (in the case of marital births) duration of marriage,
divorces by duration of marriage; all three types of event are also classified by year

© Springer International Publishing Switzerland 2016 85

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_3
86 3 The Cohort and Period Approaches to Demographic Analysis

of occurrence. The female population of reproductive age for a country might be

classified by age, its currently married component might be classified by duration of
marriage, and both populations would further be classified by date of enumeration.
The Lexis diagram is a device for representing graphically demographic data
which are classified according to criteria which involve some concept of time.
In order to understand and use Lexis diagrams it is essential to appreciate the
distinction between time instants and time intervals. When our time scale is an
age or duration scale (and really an age scale is also a duration scale – age is simply
duration, or time elapsed, since birth), the distinction we make is between exact
ages or durations (which are time instants in the lives of individuals) and ages or
durations in completed years, or last birthday/anniversary (which are attributes that
apply to an individual for a period of a year).
When expressed as whole numbers of years, exact ages or durations are fleeting
moments – instants – in the lives of individuals which correspond to birthdays or
anniversaries. A person is ‘exact age 10’ at the time of his/her tenth birthday –
that is, exactly 10 years after (s)he was born; one moment (s)he hasn’t yet turned
ten, the next (s)he is exactly 10 years old, and the next (s)he is, by a fraction of
a second, more than 10 years old (‘exact age 10 years, 0.1 of a second’, say). An
important example of an exact age is exact age 0, which is attained at the instant of
birth. For a married couple, ‘exact marriage duration 4 years’ is their fourth wedding
anniversary. They might spend a whole day celebrating this anniversary, but only for
an instant can they say they have been married for exactly 4 years; either side of that
instant they have been married for less than 4 years and for more than 4 years.
On the other hand, when you are asked ‘How old are you?’, typically you will
give not your current exact age (‘I’m aged 25 years, 10 months, 6 days, 15 h, 23 min,
13.4 s’) but the age you attained at your last birthday (‘I’m 25 years old’). This age
is an attribute that remains with you for a full year (from the time you turn 25
until you turn 26 a year later). Similarly a married person, if asked ‘How long have
you been married?’ will typically respond along the lines ‘I’ve been married for 6
years’, meaning ‘My spouse and I have passed our sixth wedding anniversary, but
have yet to reach our seventh’. These 1-year periods during which a particular age
last birthday or duration last anniversary applies to a person are examples of age or
duration intervals.
Consider a straight line with a series of points located on it to represent years
elapsed since some important event occurred.

0 1 2 3

years since event

The points marked on this line represent time instants; they represent the instant the
event occurred and subsequent instants exactly 1, 2 and 3 years after it occurred. We
could also represent other time instants on the line. For example, the instant exactly
9 months after the event would be represented by a point three-quarters of the way
Conceptualizing Problems in Demographic Analysis: Lexis Diagrams 87

between the scaled points 0 and 1; the instant exactly 2 years 6 months after the
event would be represented by a point halfway between the scaled points 2 and 3.
While time instants are represented by points on our line, time intervals are
represented by segments of the line. Thus, for example, the segment stretching from
the point representing exactly 1 year since the event to that representing exactly
2 years since the event corresponds to a time interval of 1 year, throughout which
one complete year (but not yet two) had elapsed since the event occurred.
Another way of thinking of our line is as a scale of ages. Age for an individual is
simply years elapsed since the event of birth occurred.

0 1 2 3

age (years)

Points marked on this line represent exact ages or birthdays. We would locate an
individual at these points at birth (exact age 0) and when (s)he was exactly 1, 2 and
3 years old. We could equally locate an individual at any other exact age between
birth and exact age 3 years by identifying the appropriate point along the line; for
example, the point corresponding to exact age 2 years, 5 months, 3 weeks, 2 days,
15 h, 30 min. On the other hand, segments of the line stretching from one exact age
in whole years to the next represent 1-year age intervals or age groups.
Consider, for instance, the segment stretching from exact age 2 years to, but
excluding, exact age 3 years. How might we describe the group of individuals
whose exact ages at a point in time were located somewhere along this segment?
They are the group of individuals who at that point in time had passed their second
birthdays, but had yet to reach their third birthdays. In other words, they are the
group of individuals (the population) 2 years old, or aged two last birthday (aged
two completed years). Our segment therefore corresponds to (or represents) a 1-year
age group within the population.
We could equally think of our line as a scale of marriage durations; that is, a
scale of time elapsed since the event of marriage.

0 1 2 3

marriage duration (years)

Points on the line now represent exact marriage durations or wedding anniver-
saries. Segments of the line represent 1-year marriage duration groups (i.e., groups
of married people who were between one wedding anniversary and the next one).
We could similarly think of our line as a scale of durations of divorce (points would
represent exact durations of divorce or anniversaries of divorce; segments would
represent 1-year duration of divorce groups), as a scale of durations of residence
88 3 The Cohort and Period Approaches to Demographic Analysis

(points would correspond to exact durations of residence or anniversaries since

taking up residence; segments to 1-year duration of residence groups), etc.
Having discussed age and duration time scales, we will also look at another time
scale – a scale of calendar time.

2006 2007 2008

calendar year

Again we can think in terms of time instants and time intervals. On this occasion,
because it is more conventional and convenient, we have labelled intervals rather
than points on our time line. Those intervals, or segments, represent calendar years.
But what do the points represent? They represent time instants defined by dates and
hours of the day. Thus, the four points marked on our line represent midnight on
31 December in each of the years 2005, 2006, 2007 and 2008. We could equally
identify other time instants; for example, the date of a census held at midnight on
30 June, 2006 would be represented by a point halfway along the segment which
corresponds to calendar year 2006.

The Lexis Diagram

The ideas we have been discussing are the basis of the Lexis diagram. A Lexis
diagram consists of a grid of squares which has its horizontal axis scaled in
calendar time and its vertical axis scaled in age or duration time (Fig. 3.1).
On the horizontal axis time intervals are usually labelled, normally, although not
necessarily, in calendar year units. On the vertical axis you will sometimes find
age or duration instants labelled (i.e., exact ages or durations), and sometimes age
or duration intervals (i.e., age or duration groups). The vertical axis should have a
caption that accurately reflects the way in which it has been scaled (e.g., ‘exact age’,
or ‘age group’).
Three types of geometric image are used to represent demographic concepts or
quantities on a Lexis diagram: points, lines, and areas.

Points on Lexis Diagrams

A point on the Lexis grid may represent either of two things:

1. A demographic event classified according to date of occurrence and the exact
age or duration at which the individual concerned experienced it.
2. A member of a population classified according to date of enumeration and exact
age or duration at the time of enumeration.
Conceptualizing Problems in Demographic Analysis: Lexis Diagrams 89

Age

5 7/32
9/32
15/32
4 1/32

2 +

1956 1957 1958 1959 1960 1961

Midnight, 31 March '59

Fig. 3.1 The Lexis diagram

Thus, the point represented by the ‘C’ sign in Fig. 3.1 might represent a death
which occurred at midnight on 30 June, 1958 and involved a person who at the
time was exactly 2 years and 6 months old. Alternatively it could represent simply a
member of a population enumerated at midnight on 30 June, 1958 who was exactly
2 years and 6 months old at the time of enumeration.
In practical applications of the Lexis diagram we are rarely interested in
individual points. Our interest in points stems from the fact that lines and areas
on the Lexis grid can be considered to be collections of points. It is lines and areas
that normally interest us.

Lines on Lexis Diagrams

Three types of line on a Lexis diagram are of interest: horizontal lines, vertical lines
and diagonal lines running at 45ı from bottom left to top right. A horizontal line
represents the group of individuals who attained the exact age or duration defined
by the location of the line on the vertical scale (i.e., who reached that stage of the
life cycle) during the calendar period defined by the length of the line. Hence, in
Fig. 3.1, the line forming the base of the shaded parallelogram represents persons
who attained exact age 2 years (i.e., who reached their second birthdays) during
90 3 The Cohort and Period Approaches to Demographic Analysis

calendar year 1959; the line forming the base of the diagonal band through the
diagram represents persons who attained exact age 0 (i.e., who were born) during
calendar year 1957.
Reference to ‘reaching a stage of the life cycle’ should ring a bell. We learned in
Chap. 1 that the denominator of a demographic probability, in a common method of
estimating such measures, is the population at risk at the beginning of the relevant
life cycle phase of risk. Thus, while it is possible to conceive of exceptions, the
denominator of a demographic probability is normally a population quantity we
would represent on a Lexis diagram as a horizontal line.
A vertical line on a Lexis diagram represents the group of individuals who at
a time instant (date) defined by the location of the line on the horizontal scale
were members of the age or duration group defined by the length of the line. Thus,
in Fig. 3.1, the vertical line dividing the shaded parallelogram in half represents
persons who at midnight on 31 December, 1959 were aged 2 years (i.e., 2 years
last birthday); the portion of the vertical line drawn at midnight on 31 March, 1959
that lies within the diagonal band through the diagram represents persons who at
that point in time were aged between exactly 1 year 3 months, and exactly 2 years
3 months.
You may recall, again from Chap. 1, that it is common practice to use the mean,
or mid-year, population at risk as the denominator of a demographic rate. The mid-
year population will be a population at a particular time instant. The denominator of
a demographic rate is thus often approximated by a population quantity we would
represent on a Lexis diagram as a vertical line. Placing this statement alongside
that highlighted two paragraphs above, we now have a new way of differentiating
demographic rates from demographic probabilities – in terms of the ways their
denominators are normally, or often, represented on a Lexis diagram.
The third type of line conventionally found on Lexis diagrams, a diagonal line
set at 45ı to the horizontal axis, is often referred to as a life line. Life lines trace the
passage of individuals through life (Fig. 3.2). A person’s life line starts at the date on
the horizontal axis on which (s)he attained exact age 0 (i.e., was born). It then runs
diagonally upwards to the right at 45ı until either the current date or the date the
person died is reached. Figure 3.2 shows the life line of a person who was born, let
us say, on 31 March 1980, and who thereafter aged 1 year for each year of calendar
time that passed until dying around mid-February (let us say on 14 February) 1985,
about a month and a half short of his/her fifth birthday (i.e., aged four).
We can also think of life lines in the context of forms of life other than that which
begins at birth and ends at death. For example, we can think in terms of a person’s
(or a couple’s) life in a marriage. That ‘life’ begins at the date of marriage (at
exact marriage duration 0 years – the instant they are pronounced ‘man and wife’)
and ends when the marriage ends (due to either the death of one party or divorce).
Similarly we could think of life in the state of being a woman of parity two (i.e.,
having had two live births). That ‘life’ begins at the instant the second live birth
occurs (at exact duration in parity two 0 years) and ends when either the woman
dies or her third live birth occurs. You might like to reinterpret Fig. 3.2 in terms of
Conceptualizing Problems in Demographic Analysis: Lexis Diagrams 91

Exact Age
6

0
1979 1980 1981 1982 1983 1984 1985
Year

Fig. 3.2 Lexis diagram illustrating concept of a ‘Life Line’

these two forms of ‘life’, with the vertical axis taken first to be a scale of durations
of marriage, and then to be a scale of durations in parity two.
From a practical point of view we are rarely concerned with life lines for
individuals when using Lexis diagrams. However, if you look at Fig. 3.3 you can
probably appreciate that the collection of life lines for all people born (or married,
or attaining parity two, etc.) in a given calendar year (1980 in this instance) defines
a diagonal band through the Lexis grid. All life lines for people born in that year lie
within this band, and only life lines for people born in that year lie within it.
You can perhaps also appreciate that any demographic event an individual
experiences can be represented by a point lying on his/her life line at the relevant
date and exact age. It follows from this and the fact that life lines for individuals
born in a given year all lie within a discrete diagonal band that all events occurring
to individuals born in that year, when represented as points, lie within the same
diagonal band. In other words, the diagonal band summarizes, or embraces, the
entire demographic experience of persons born during the year in question.
Persons born during a particular year constitute, according to the definition given in
Chap. 1, a birth cohort (a group of individuals who all experienced the demographic
event ‘birth’ during a specified period of time). It follows that the demographic
experience of a birth cohort (and indeed of any other type of cohort) can be
represented as a diagonal band through a Lexis diagram. Normally we deal with
calendar-year cohorts; i.e., cohorts comprising individuals who experienced the
demographic event defining the cohort (birth for a birth cohort, marriage for a
marriage cohort, etc.) during a calendar year. The diagonal bands representing such
92 3 The Cohort and Period Approaches to Demographic Analysis

Exact Age
12

10

0
79 80 81 82 83 84 85 86 87 88 89 90 91 92
Year

Fig. 3.3 Demographic experience of a cohort represented by a diagonal band through a Lexis
diagram

cohorts on a Lexis diagram are of width 1 year, and the diagonal lines defining
those bands have common points of intersection with the horizontal and vertical grid
lines that are also 1 year apart (Fig. 3.3). However, we do sometimes also deal with
non-calendar year cohorts (e.g., cohorts who experienced the defining demographic
event between the beginning of July in 1 year and the end of June in the following
year), and with cohorts that are wider or narrower than 1 year (e.g., 5-year or 6-
month birth cohorts). You might like to try sketching cohorts of these types on a
Lexis diagram after the fashion illustrated in Fig. 3.3.

Areas on Lexis Diagrams

Like lines, areas on a Lexis diagram can be thought of as collections of points. Areas
on a Lexis diagram represent collections of events distributed by exact age/duration
at the time of occurrence and date of occurrence. Put slightly differently, they
represent collections of events that occurred either within a particular age or
duration interval over a specified period of time, or to members of a particular cohort
between specified exact ages/durations, specified dates, or a combination of both.
Consider Fig. 3.1 again, and imagine that we have drawn this diagram in
connection with an analysis of mortality patterns (i.e., the event to which the
diagram pertains is death). The square in which the ‘C’ sign is located represents the
collection of deaths at age 2 that occurred during 1958. The shaded parallelogram
Conceptualizing Problems in Demographic Analysis: Lexis Diagrams 93

represents deaths at age 2 of members of the 1957 birth cohort. The triangle
defined by the left side of this parallelogram and the ‘31 March, 1959’ vertical
line represents deaths at age 2 of members of the 1957 birth cohort which occurred
before midnight on 31 March, 1959. Finally, the unshaded parallelogram represents
deaths that occurred between 1 April 1959 and 31 March 1960 of members of the
birth cohort born between 1 April 1954 and 31 March 1955 (we determine these
latter dates either by extending the diagonal lines defining the cohort back to the
horizontal axis, or by recognizing that a cohort aged between exact age 4 and exact
age 5 at midnight on 31 March 1959 was exact age 0, and therefore born, between
5 and 4 years previously).
Understanding what areas on a Lexis diagram represent can often be aided by
defining the area of interest as the intersection of vertical and/or horizontal and/or
diagonal bands through the Lexis grid. You can then read off the calendar period
and/or age/duration group and/or cohort corresponding to these bands, and this will
give you the calendar time-age/duration-cohort dimensions defining the set of events
the area represents. For example, in Fig. 3.1 the square in which the ‘C’ sign is
located can be thought of as the intersection of the horizontal band corresponding to
age group 2 years and the vertical band corresponding to calendar year 1958; hence
it represents deaths at age 2 that occurred during 1958. The shaded parallelogram
can be thought of as the intersection of the horizontal band corresponding to age
group 2 years and the diagonal band representing the 1957 birth cohort; hence it
represents deaths at age 2 of members of the 1957 birth cohort. The triangle defined
by the left side of this parallelogram and the ‘31 March, 1959’ vertical line can be
thought of as the intersection of the same horizontal and diagonal bands and the
vertical band corresponding to the period 1 January, 1959 to 31 March, 1959; hence
it represents deaths at age 2 of members of the 1957 birth cohort which occurred
between 1 January, 1959 and 31 March, 1959 inclusive.
This approach can also be followed in reverse if the task is to represent a certain
set of events on a Lexis diagram. Identify the vertical and/or horizontal and/or
diagonal bands that correspond to the calendar time-age/duration-cohort dimensions
defining the set of events of interest. The relevant area on the Lexis diagram will then
be the intersection of these bands, or if you like, the area that is common to all of
them.
Note in passing that a similar approach can also be adopted when trying to
understand what lines on a Lexis diagram represent. Consider, for example, in
Fig. 3.1 the portion of the vertical line drawn through 31 March, 1959 that lies
within the diagonal band. This line is the intersection of the vertical line representing
the population distributed by age at midnight on 31 March 1959 and the diagonal
band representing the 1957 birth cohort. Hence it represents the 1957 birth cohort
as at midnight on 31 March, 1959 (at which date, by reading off against the vertical
axis, we can ascertain that its members were aged between exactly 1 year 3 months
and 2 years 3 months). What about the line that forms the base of the shaded
parallelogram in Fig. 3.1? It can be thought of as the intersection of the horizontal
line corresponding to exact age 2 years with either the vertical band representing
calendar year 1959 or the diagonal band representing the 1957 birth cohort. Thus it
94 3 The Cohort and Period Approaches to Demographic Analysis

represents both the population that attained exact age 2 during 1959 and the 1957
birth cohort at exact age 2, and because we are talking about one and the same line,
these two populations are identical.
Returning to areas, though, you might reasonably ask ‘Do we always have data
on numbers of events neatly classified so as to correspond to areas on the Lexis grid
in which we are interested? The answer is emphatically ‘No’. Invariably our event
data are classified so as to correspond to squares on the Lexis grid. That is, they
are classified according to calendar year of occurrence and age or duration (last
birthday or last anniversary) at the time of occurrence. How, then, do we estimate
numbers of events corresponding to areas on the Lexis grid other than the grid
squares?
A basic assumption underlying the use of Lexis diagrams in demographic
analysis is that events occurring, and individuals located, within any time, age
or duration interval are distributed EVENLY through that interval. When we
translate this assumption onto the two-dimensional Lexis surface we are assuming
that points representing demographic events which lie in any square of the Lexis
grid have an even spread (or density) over that square.
This enables us to estimate the number of events corresponding to any area on
the Lexis surface as follows:
1. Divide the area into segments lying in different squares of the Lexis grid. (If the
area lies entirely within one square this obviously becomes unnecessary.)
2. Determine what fraction, or proportion, of the total area of its grid square each
segment occupies.
3. Calculate that fraction, or proportion, of the total number of events represented
by the relevant grid square. This gives the estimated number of events represented
by the segment.
4. Add the answers obtained for the various segments identified at item 1.
The basis of this procedure is, of course, that if points representing demographic
events are evenly distributed over a Lexis grid square, the proportion of those events
a segment of the square represents will be identical to the proportion of the square’s
area occupied by that segment.
By way of illustration we can once again turn to Fig. 3.1. If we again think of this
diagram as having been drawn to aid an analysis of mortality, how many deaths are
represented by the unshaded parallelogram? Our data almost certainly will consist
of deaths classified by single years of age and calendar year of occurrence; that is,
they will relate to squares of the Lexis grid. Parts of our parallelogram lie in four
different grid squares; hence, in accordance with step 1 of the method just outlined,
we divide it into four segments.
1. The lower right segment is a right-angled triangle in which the sides forming
the right angle are each one-quarter of the length of the relevant side of the grid
square in which the segment lies. Hence, since the area of a triangle is given by
area D ½ (base  height), this triangle occupies:
Conceptualizing Problems in Demographic Analysis: Lexis Diagrams 95

1=  1=  1= D 1= of the area of the grid square

4 4 2 32
2. The upper left segment is also a right-angled triangle, but with the sides forming
the right angle each three-quarters of the length of the relevant side of the grid
square in which the segment lies. Hence the triangle occupies:

3=  3=  1= D 9= of the area of the grid square

4 4 2 32
3. The lower left segment can be thought of in either of two ways: as the sum of a
rectangle of dimension 1 /4 by 3 /4 of a grid square side and a right-angled triangle
the same size as the one in item 2 above; or as a rectangle of dimension 1 3 /4 of
a grid square side minus a right-angled triangle the same size as the one in item
2 above. Hence this segment occupies:
   
1=  3= C 3=  3=  1= D 3= C9=
4 4 4 4 2 16 32
.
D 15 32 of the area of the grid square

or:
   
1  3=4  3=4  3=4  1=2 D 3=4  9=32

D 15=32 of the area of the grid square

4. The upper right segment can likewise be thought of as the sum of a rectangle
and a right-angled triangle or the difference between a larger rectangle and a
right-angled triangle. Hence this segment occupies:
   
1=  3= C 1=  1=  1= D 3= C 1=
4 4 4 4 2 16 32

D 7=32 of the area of the grid square

or:
   
1  1=4  1=4  1=4  1=2 D 1=4  1=32

D 7=32 of the area of the grid square

Therefore, if deaths at age x during year y are denoted by Dx,y , the estimated
number of deaths represented by the unshaded parallelogram is given by:

Deaths D 15=32 D4;59 C 9=32 D5;59 C 1=32 D4;60 C 7=32 D5;60

96 3 The Cohort and Period Approaches to Demographic Analysis

Having demonstrated how Lexis diagrams can be used under the assumption
that events occurring within any time or age/duration interval are evenly distributed
through that interval, it is appropriate to warn that this is an assumption, and there
are occasions when it is patently invalid. The classic example concerns infant
deaths, which typically do not even remotely approach being evenly distributed
by exact age between exact ages 0 years and 1 year. Instead they are very heavily
concentrated towards the lower end of this age range. In such a circumstance, where
the key underlying assumption is clearly seriously violated, it is not acceptable to
go blindly on using a Lexis diagram as a tool of demographic analysis. That said,
most of the time the assumption of events being evenly distributed through time or
age/duration intervals is an acceptably reasonable one, and Lexis diagrams can be
very useful aids to conceptualizing demographic indices and demographic analysis
problems.

The Lexis Diagram in Operation

The primary use of Lexis diagrams is to help demographers visualize how to go

about solving problems in demographic analysis. A secondary application is as an
illustrative device. We will now examine some examples of these applications.

Example 1

Suppose we wanted to estimate the size of the 1980 birth cohort for a population at
the date of a census conducted at midnight on 31 March 1986. To do this we need
to know how old the 1980 birth cohort was at the date of the census. In tackling this
problem we initially apply an important principle in solving problems using Lexis
diagrams. We draw ourselves a Lexis diagram and represent on it the information
we have been given. We are interested in the 1980 birth cohort, so we draw on
our diagram the diagonal band representing the demographic experience of that
cohort. We are also interested in the population enumerated at a census held at
midnight on 31 March 1986, so we draw a vertical line at that date to represent
the census population. The result is Fig. 3.3, which we used earlier to illustrate the
representation of cohorts on Lexis diagrams.
We want to know how old the cohort was at the census. We are thus interested
in the intersection of the cohort band with the census line. Reading off the vertical
coordinates of the points of intersection of the life lines defining the extremities of
the cohort band with the census line we find that, at the date of the census, persons
born in 1980 were aged between exactly 5 years 3 months and 6 years 3 months.
How many individuals were in this age group? Our census data will not tell us
The Lexis Diagram in Operation 97

directly; they will tell us how many people were aged 5 last birthday (i.e., between
exact age 5 and just short of exact age 6) and how many were aged 6 last birthday
(between exact age 6 and just short of exact age 7). The group we are interested in
lies partly in each of these age groups. We therefore ask ourselves ‘What proportion
of 5 year-olds and what proportion of 6 year-olds at the 1986 census were members
of the 1980 birth cohort?’
Looking at that segment of the vertical census line that lies within the cohort
band we find that it consists of two parts: a lower part that extends three-quarters
of the way across the horizontal band representing age group 5 years last birthday,
and an upper part that extends one-quarter of the way across the horizontal band
representing age group 6 years last birthday. Thus, using the assumption that events
or individuals located within a time or age/duration interval are evenly distributed
through that interval (in this case by exact age), the census population that belonged
to the 1980 birth cohort consists of three-quarters of all 5 year-olds and one-quarter
of all 6 year-olds. Given the number of 5 year-olds and the number of 6 year-olds at
the census, we could then quite easily work out the size of the 1980 birth cohort at
the census.
We will pause here to briefly highlight an idea mentioned in passing above and
an important implication of that idea. Lexis diagrams are a way of visualizing
demographic problems, in the most literal sense – a way of ‘seeing’ them with one’s
eyes. It is therefore absolutely vital that they be drawn accurately. You are not going
to be able to ‘see’ that this population of interest is three-quarters of that population
for which you have data, or whatever the case may be, unless your Lexis diagram is
drawn carefully and accurately. We have referred so far to Lexis grid squares, and
to life lines running at 45ı to the horizontal axis. The ideal is probably to keep to
these prescriptions, but it is also possible (and not wrong) to use a Lexis grid that
consists of rectangles of equal size. This occurs when the intervals between vertical
and horizontal grid lines differ. The important thing in this instance is to ensure that
the interval between vertical grid lines is constant, and that the interval between
horizontal grid lines is also constant (albeit different). If a Lexis grid consists of
rectangles rather than squares, the angle of life lines to the horizontal axis will
also vary from 45ı . However, all life lines should be carefully drawn, straight and
parallel, and should increase exactly one unit of time on the vertical axis for every
one unit of time on the horizontal axis.

Example 2

Examine Fig. 3.4, which deals with infant deaths which occurred to the 2009 and
2010 birth cohorts of a hypothetical population. One thing should immediately strike
you about this diagram. For each cohort, numbers of infant deaths during the year
of birth and the year following the year of birth are shown, and these numbers are
98 3 The Cohort and Period Approaches to Demographic Analysis

Exact Age

G H I
2

D E F
1
33 30

77 70

0
A 2009 B 2010 C 2011 2012

Year

Fig. 3.4 Lexis diagram showing infant deaths to the 2009 and 2010 birth cohorts of a hypothetical
population

unequal (a person is at risk of infant death throughout the first year of his/her life,
which invariably lies partly in each of two consecutive calendar years). For each
cohort the first year of life is represented by a parallelogram (ABED and BCFE)
that is subdivided (by lines BD and CE respectively) into two equal-sized triangles.
Lexis diagram principles say that the two triangles BED and BCE, which are also
of equal size and lie in the same grid square, represent equal numbers of events, but
here they do not.
What we have here is a simple representation of the situation that always arises
with infant deaths, because they are not evenly distributed by exact age between
exact age 0 and exact age 1. Rather, deaths are heavily concentrated at the lower end
of this age range, and as the early part of the first year of life is lived during the year
of birth for most individuals, more infant deaths for any birth cohort occur during
the year of birth than occur during the following year.
This does not fatally undermine all we have said about the principles that apply
in using Lexis diagrams and the usefulness of such diagrams. We could not use
a Lexis diagram to split the 103 infant deaths that Fig. 3.4 indicates occurred in
2010 into those involving members of the 2009 birth cohort and those involving
members of the 2010 birth cohort. But we don’t have to; in this instance we have
been given those data, and the Lexis diagram is serving an illustrative rather than an
analytical purpose so far as the distribution of infant deaths among cohort-calendar
year segments is concerned.
The line BC in Fig. 3.4 represents births during 2010. Let us say there were 1,000
of them. What do the lines EC and EF represent? EC is the intersection of a vertical
line representing the population at midnight on 31 December 2010 and the diagonal
The Lexis Diagram in Operation 99

band representing the 2010 birth cohort; hence it represents the size of the 2010 birth
cohort at the end of 2010, or the number of children born alive during 2010 who
remained alive at the end of that year. How do we obtain this number? The size of
the cohort at EC is simply its size at BC (which we know equals 1,000) adjusted for
the events (deaths) represented by the area separating these two lines (the triangle
BCE, which Fig. 3.4 tells us represents 70 infant deaths); that is, the size of the
cohort at EC D 1,000  70 D 930 survivors at midnight on 31 December 2010. Note
the general principle involved here. If we want to know the size of a cohort at a
particular date or exact age/duration A, we take its KNOWN size at some other
date or exact age/duration B, and ADJUST for those demographic events which
altered the cohort’s size between A and B. These events are represented by THE
AREA SEPARATING the horizontal/vertical lines identifying points A and B in
the cohort’s life course.
We can apply this principle again to obtain the size of the 2010 birth cohort
at EF from its now known size at EC. EF is the intersection of a horizontal
line representing exact age 1 and the diagonal band representing the 2010 birth
cohort; hence it represents the size of the 2010 birth cohort at exact age 1. This
population is simply the population at EC adjusted for deaths represented by the
area separating EC from EF (i.e., by the triangle ECF); that is, the size of the cohort
at EF D 93030 D 900 survivors at exact age 1.
We will take the opportunity here to resurrect an earlier issue, a fuller understand-
ing of which can be obtained with the aid of Fig. 3.4. Recall that in Chap. 1 we noted
that the infant mortality rate (IMR) was not a rate but an estimate of the probability
that a person born during the year for which it was calculated would die in infancy
(i.e., before reaching his/her first birthday). We will now use Fig. 3.4 to demonstrate
why the IMR is a probability, and why it is only an estimate of that probability.
Suppose we denote the population at BC (the 2010 birth cohort at exact age 0) by
N0 (whence the population at EF would be denoted by N1 , that at HI by N2 , etc.).
We will also denote the deaths at age 0 of members of the 2010 birth cohort (i.e.,
the deaths represented by the parallelogram BCFE) by 1 D0 . This introduces a new
form of notation that will become very familiar as we proceed to examine life tables.
Any expression of the form n Vx means the value of the central variable V for the
age/duration group which starts at exact age/duration x and ends at (or, in strict
terms, marginally short of) exact age/duration x C n. Thus, 1 D0 denotes deaths (of
members of the 2010 birth cohort) between exact age 0 and exact age 0 C 1 (i.e.,
exact age 1). We will use it to define a quantity:

1 q0 D 1 D0 =N0

This quantity is the ratio (for the 2010 birth cohort) of deaths between exact ages 0
and 1 (i.e., deaths at age 0) to the population which attained exact age 0. What sort
of measure is it? It is a probability – the ratio of events (deaths) occurring during
100 3 The Cohort and Period Approaches to Demographic Analysis

a defined phase of the life cycle (the first year following birth) to the population at
risk at the beginning of that life cycle phase (the population attaining exact age 0, or
the number of live births). What probability do we have? We have the probability
of dying at age 0, which should immediately remind you of the IMR. We said in
Chap. 1 that the IMR gave an estimate of the probability of dying in infancy (i.e., at
age 0). In other words, it is an estimate of 1 q0 . So how does the IMR compare with
1 q0 in Lexis diagram terms?
The ‘true’ probability of dying at age 0 for the 2010 birth cohort, 1 q0 , is the
ratio of deaths in the parallelogram BCFE to the population at BC. The IMR for
2010 is the ratio of deaths at age 0 in 2010 (i.e., deaths in the square BCED)
to the population at BC. Both measures have the same denominator, a measure
of the population at risk at the beginning of the life cycle phase of risk (and a
population represented on the Lexis diagram by a horizontal line). Both are therefore
probabilities. The IMR is an estimate of the ‘true’ probability of dying in infancy
(for the 2010 birth cohort) because it assumes that the infant deaths represented by
the triangle BED on the Lexis diagram (which actually involved children born the
previous year) are a reasonable estimate of the infant deaths represented by the
triangle CFE (which although occurring in 2011, involved members of the 2010
birth cohort).
Practically speaking, most of the time we have to make do with the IMR as an
estimate of 1 q0 . The reason is that we rarely have sufficiently detailed data on infant
deaths to enable us to do what is done in Fig. 3.4 – divide those during year y into
deaths that involved children born in year y and deaths that involved children born
in year y1. We cannot use a Lexis diagram to help us make this split, because the
assumption we would invoke, that deaths at age 0 were evenly distributed between
exact ages 0 and 1, is clearly invalid and would produce a completely inaccurate
split.

Example 3

We will now examine a more complicated example of the use of a Lexis diagram
to aid the calculation of a demographic measure. If you can follow this example,
understanding the steps as you go, you probably have a good grasp of the principles
underlying the use of Lexis diagrams. Suppose we wished to calculate the following
two measures for a population:
1. The probability that a never married woman turning 20 in 1998 married at age
20.
2. The first marriage rate at age 20 for the cohort of never married women turning
20 in 1998.
The Lexis Diagram in Operation 101

Suppose further that we have the following data which will prove relevant to our
exercise:

Number of never married women aged 19 at 30 June 1998 104,923

Number of never married women aged 20 at 30 June 1998 86,860
Marriages of never married women aged 19 in 1998 10,924
Marriages of never married women aged 20 in 1998 11,848
Marriages of never married women aged 20 in 1999 11,874
Arrivals of never married women aged 19 in 1998 14,701
Arrivals of never married women aged 20 in 1998 15,249
Arrivals of never married women aged 20 in 1999 16,976
Departures of never married women aged 19 in 1998 13,413
Departures of never married women aged 20 in 1998 15,090
Departures of never married women aged 20 in 1999 16,423
Deaths of never married women aged 19 in 1998 74
Deaths of never married women aged 20 in 1998 52
Deaths of never married women aged 20 in 1999 54

To aid us we draw the Lexis diagram shown as Fig. 3.5. There are two general
features of this Lexis diagram worthy of note at the outset. First, we have drawn only
that portion of the Lexis grid that is relevant to our problem. We have not scaled
ages right back to exact age 0 because we are dealing with a cohort’s experience
at age 20, and most younger ages are irrelevant. There is no point drawing Lexis
diagrams that feature masses of grid squares that are never used. We have, in this
instance, scaled age from exact age 19 rather than exact age 20; the necessity to do
this might not be immediately apparent, but is the sort of thing you should develop
a ‘feel’ for as you become experienced in using Lexis diagrams. Second, we have
represented on the diagram the information we have been given. This is invariably
a sensible first step in tackling Lexis diagram problems. We want to calculate two
measures for the cohort of single women who turned 20 (reached exact age 20) in
1998, so we have identified that cohort on our Lexis diagram. It is the population
represented by the horizontal line EB, and we have then drawn the cohort diagonals
through E and B. We are also given some data on numbers of never married women
by age at 30 June 1998, so we have drawn the vertical dashed line to represent the
population of never married women at that date. Finally, we have data on various
types of events occurring to never married women aged 19 in 1998, aged 20 in 1998,
and aged 20 in 1999; these events are represented by the grid squares ALBE, EBDK
and BMCD, which are defined by the intersections of the relevant age groups and
calendar time bands through the diagram.
As just noted, the cohort of never married women turning 20 in 1998 is
represented on the Lexis diagram by the line EB, and its demographic experience
is contained within the diagonal band formed by drawing life lines through this
line’s end points. According to the definition of a probability previously given, our
102 3 The Cohort and Period Approaches to Demographic Analysis

Exact Age
K F D
21
C

3/8
E
1/8 H
20
B M
1/8

A J L
19

1998 1999
Year

Fig. 3.5 Lexis diagram to aid solution of ‘Example 3’

required probability is the ratio of marriages of never married women at age 20 for
this cohort to the size of the cohort at the beginning of the life cycle phase of risk:

marriages of never married women in EBCD

probability D
never married female population at EB

Our required rate is the ratio of first marriages at age 20 for the same cohort to the
mean size of the cohort during the life cycle phase of risk, our best estimate of which
is the never married female population at BD, or at midnight, December 31st, 1998.
Hence:
marriages of never married women in EBCD
rate D
never married female population at BD

We thus have three things to estimate: the number of marriages of never married
females in EBCD (which provides a common numerator for both measures); the
never married female population at EB; and the never married female population at
BD. In case you are a little mystified over our specifying the never married female
population at BD as the mean population exposed to risk, this mean will be the
average of the never married female populations at EB and DC (the beginning and
the end of the life cycle phase of risk). Under an assumption that points representing
events of a given type which modify the size of the cohort as it ages are evenly
The Lexis Diagram in Operation 103

distributed over the parallelogram EBCD, this mean will exactly equal the size
of the cohort at BD, because numbers of modifying events represented by the
triangles EBD and BCD will be exactly equal. Specifying the population at BD
as the denominator of our rate also makes intuitive sense in that (i) it is located
in the middle of the 2-year period during which some members of the cohort of
interest are in the relevant age group (recall that the denominator of a rate is often
approximated by the mid-period population at risk), and (ii) it is located at the only
point in time at which the population aged 20 is comprised entirely of members
of the cohort of interest. Any earlier and some cohort members were still only 19
while some 20 year-olds had turned 20 in 1997 rather than 1998; any later and some
cohort members had by now turned 21 while some 20 year-olds had turned 20 in
1999 rather than 1998.
To estimate the number of marriages of never married women in EBCD: We
must make use of the data we have on marriages of never married women, which
correspond to squares on the Lexis grid. To use those data we must express
parallelogram EBCD in terms of segments, each of which lies in a discrete square.
Clearly EBCD is the sum of the two triangles EBD and BCD, each of which
comprises half the area of the grid square in which it lies. The two squares, EBDK
and BMCD, correspond to events (marriages of never married women) at age
20 in 1998 (M20,98 ) and at age 20 in 1999 (M20,99 ). Thus our required common
numerator is:
1= .M 1
2 20;98 C M20;99 / D =2 .11;848 C 11;874/
D 11;861 marriages of never married women

To estimate the never married female population at EB: We must invoke an

important principle in the use of Lexis diagrams. We must work from the infor-
mation we have to the information we want. The only data we have on numbers
of never married women are the numbers aged 19 and 20 at 30 June 1998. This
information is represented on the Lexis diagram by the lines JH and HF respectively.
Somehow we must work from this information to obtain an estimate of the number
of never married women at EB.
The data we have on numbers of never married women pertain to the date 30 June
1998. Our strategy is to estimate the size of the cohort of never married women of
interest at that date (represented by the line GI), then adjust it for events which
modify the size of the never married female population over time to obtain the size
of the cohort at EB. The important point to appreciate here is that GI and EB both
represent the size of the never married female cohort we are interested in; GI just
represents it at a particular date, while EB represents it at a particular exact age, or
birthday.
We noted earlier that if we want to find the size of a cohort at a particular date or
exact age/duration A, we take its known size at some other date or exact age/duration
B, and adjust for demographic events which altered the cohort’s size between A and
104 3 The Cohort and Period Approaches to Demographic Analysis

B, these events being represented by the area separating the horizontal/vertical

lines identifying points A and B in the cohort’s life course. In this example there is
a complication, in that the lines identifying the points in the cohort’s life course for
which we have information (GI) and want information (EB) intersect one another.
The easiest way to get around this complication is to divide our cohort into two
sub-cohorts by drawing a third diagonal line through the point of intersection of GI
and EB (i.e., through H). This effectively divides the cohort of never married women
who turned 20 during 1998 into the cohort who turned 20 between January and June
1998 (represented in Fig. 3.5 by the line EH) and the cohort who turned 20 between
July and December 1998 (represented by the line HB).
You can perhaps now appreciate that the populations for which we have
information (GI) and want information (EB) can each be broken into two segments
(GI D GH C HI; EB D EH C HB). Moreover, we can estimate each of the segments
of EB from one of the segments of GI. The population at EH is the population at
GH adjusted backwards for relevant demographic events represented by the triangle
EHG (the area on the Lexis diagram that separates EH from GH). Similarly the
population at HB is the population at HI adjusted forwards for relevant events
represented by the triangle HIB (the area that separates HB from HI). In case
these references to adjusting ‘backwards’ and ‘forwards’ are a source of confusion,
remember that, with the passage of time, individual members of a cohort progress
up the cohort diagonal band from bottom left to top right. Thus GH represents the
first of our two sub-cohorts at a later point in the life cycle than does EH, and HI
represents the second sub-cohort at an earlier point in the life cycle than does HB.
It follows that we adjust backwards from GH and forwards from HI to obtain EH
and HB respectively.
1. The population of never married women represented by EH is the population
of never married women represented by GH adjusted for marriages, deaths,
in-migrations and out-migrations of never married women represented by the
triangle EHG. Since the line GH is half of the line FH, it represents half of the
never married female population aged 20 at 30 June 1998:

population at GH D 1=2 P20;98 D 1=2 .86;860/ D 43;430

As a cohort of never married women ages, it is reduced in size by marriages,

deaths and out-migrations of never married women, and increased in size by
in-migrations of never married women. On this occasion, however, we need to
adjust backwards from the population at GH to obtain the population at EH, so
we need to add back in the marriages, deaths and out-migrations of never married
women, and subtract the in-migrations of never married women represented by
the triangle EHG. The triangle EHG lies entirely within the grid square EBDK
(representing events that occurred at age 20 during 1998) and occupies one-
eighth of the area of that square. Hence the adjustment factor will be:
The Lexis Diagram in Operation 105

1= .M 1
8 20;98 CD20;98 CO20;98  I20;98 / D =8 .11;848C52C15;09015;249/
D 1;467:6 never married women

and:

population at EH D population at GH C EHG adjustment factor

D 43;430 C 1;467:6 D 44;897:6 never married women

2. The population of never married women represented by HB is the population

of never married women represented by HI adjusted for marriages, deaths,
in-migrations and out-migrations of never married women represented by the
triangle HIB. Since the line HI is half of the line HJ, it represents half of the
never married female population aged 19 at 30 June 1998:

population at HI D 1=2 P19;98 D 1=2 .104;923/ D 52;461:5

Again, as a cohort of never married women ages, it is reduced in size by

marriages, deaths and out-migrations of never married women, and increased in
size by in-migrations of never married women. We need to adjust forwards from
the population at HI to obtain the population at HB, so we need to subtract the
marriages, deaths and out-migrations of never married women, and add in the in-
migrations of never married women which are represented by the triangle HIB.
The triangle HIB lies entirely within the grid square ALBE (representing events
that occurred at age 19 during 1998) and occupies one-eighth of the area of that
square. Hence the adjustment factor will be:

1= .M 1
8 19;98 D19;98 O19;98 CI19;98 / D =8 .10;9247413;413C14;701/

D 1;213:7 never married women

and:

population at HB D population at HI C HIB adjustment factor

D 52;461:5  1;213:7 D 51;247:8 never married women

It follows that the required never married female population at EB is:

population at EH C population at HB D 44;897:6 C 51;247:8

D 96;145:4 never married women

To estimate the never married female population at BD: We again must work
from known to desired information. We have numbers of never married females
corresponding to the lines JH and HF. We can use these to estimate the size of our
106 3 The Cohort and Period Approaches to Demographic Analysis

cohort of never married females at 30 June 1998 (i.e., the population at GI), then
adjust for marriage, death, in-migration and out-migration events involving never
married women represented by the parallelogram IBDG to obtain the never married
female population at BD. This parallelogram is relevant as the area separating the
lines GI and BD.
We have that:

population at GI D population at HI C population at GH

D half the population at JH C half the population at HF

D 1=2 .P19;98 C P20;98 /

D 1=2 .104;923 C 86;860/

D 95;891:5 never married females

The parallelogram IBDG consists of two segments, the area HBDG which occupies
three-eighths of the square EBDK, and the triangle HIB which occupies one-eighth
of the square ALBE. In adjusting from the population at GI to that at BD we are
adjusting forwards (i.e., from an earlier to a later point in the cohort’s life cycle);
hence we need to subtract marriages, deaths and out-migrations of never married
women represented by the parallelogram IBDG and to add in the in-migrations of
never married women it represents. Our adjustment factor is:

3= .M
8 20;98  D20;98  O20;98 C I20;98 /

C 1=8 .M19;98  D19;98  O19;98 C I19;98 /

D 3=8 .11;848  52  15;090 C 15;249/

C 1=8 .10;924  74  13;413 C 14;701/

D 4;403  1;214
D 5;617 never married women

The required never married female population at BD is therefore:

population at GI C IBDG adjustment factor D 95;891:5  5;617

D 90;274:5 never married women

Having now estimated the three quantities we set out to obtain, we can calculate
the required probability that a never married woman turning 20 in 1998 married at
age 20:
The Lexis Diagram in Operation 107

marriages of never married women in EBCD

probability D
never married female population at EB
D 11;861=96;145:4
D 0:12337

and the required first marriage rate at age 20 for never married women who turned
20 in 1998:
marriages of never married women in EBCD
rate D  1; 000
never married female population at BD
D .11;861=90;274:5/  1;000
D 131:4 marriages per 1;000 never married women

Postscript to Example 3

In Chap. 1 we discussed the concept of ‘person-years of exposure to risk’. We noted

that, strictly speaking, the denominator of a demographic rate should be the number
of person-years of exposure to risk for the reference population (the population for
which the rate is being calculated) over the reference period (the calendar period,
or in the case of a cohort rate, the period of the life cycle for which it is being
calculated). However, we discovered that, on the assumption that events subjecting
individuals to, and removing them from risk were evenly distributed through the
reference period, we could make do with simpler and much more easily obtained
approximations (the mean, or mid-period, population at risk).
Similar types of issues arise in relation to the denominators of demographic
probabilities. In Example 3 we obtained as the denominator of the probability that
a never married woman turning 20 in 1978 married at age 20 the population of
never married women at EB (Fig. 3.5); that is, the size of the relevant cohort at
the beginning of the life cycle phase of risk. The definition of the denominator
of a probability with which this particular denominator conforms can be stated as:
The number of person-years that would have been lived between the reference
ages/durations by people subject to the risk of the reference event at the first
reference age/duration if none of them had experienced the reference event.
Strictly speaking, this definition is appropriate only when (i) the probability relates
to the reference event death, and (ii) one is dealing with a closed population. This
is so because:
1. If the reference event is something other than death, someone who enters the
relevant life cycle phase (the reference period) at risk and dies within that phase
contributes fully to the denominator of the probability, but cannot contribute to
108 3 The Cohort and Period Approaches to Demographic Analysis

its numerator during the portion of the reference period that remains after (s)he
dies.
2. Whatever the reference event, an out-migrant during the reference period there-
after can no longer contribute to the numerator of the probability, but contributes
fully to the denominator.
3. Similarly an in-migrant during the reference period can contribute to the
numerator after arrival, but not having been present at the beginning of the
reference period, does not contribute to the denominator.
In practice the mortality factor (item 1 above) is normally so small that it can be
disregarded. With the migration factor (items 2 and 3) there is obviously some
cancelling, in-migrants compensating for out-migrants, so that again it is often
possible to disregard the factor for all practical purposes. Only where there is
significant net migration might we have a problem.
If we do decide to make an adjustment for either factor we make it on
the assumption that each deceased person or migrant was at risk for half the
reference period. Suppose we were to decide to adjust the denominator of the
probability calculated in Example 3 for both mortality and migration. Our use of the
population of never married women at EB (Fig. 3.5) as this denominator implied
two assumptions: (i) that all members of that population thereafter spent a full
year during which they were potentially able to contribute to marriages of never
married women aged 20; and (ii) that they were the only women able to contribute
such marriages as members of the cohort in which we were interested. In reality
death and out-migration meant that some never married women at EB were able
to contribute marriages at age 20 for less than a year, while in-migration meant
that newcomers joined our cohort as never married 20 year-olds and were able to
contribute marriages for whatever time remained until they reached exact age 21.
In Fig. 3.5 the mortality and migration events we need to consider in making
an adjustment are those represented by the parallelogram EBCD. These are the
events that occurred to members of our cohort at age 20 (the parallelogram is the
intersection of the cohort band and the horizontal band corresponding to age group
20). On the assumption that each in-migration event adds half a person-year of
exposure to risk and each death and out-migration event subtracts half a person-
year of exposure, our adjustment factor is:
h i
1= 1= .I C I /  1= .O C O /  1= .D C D /
2 2 20;98 20;99 2 20;98 20;99 2 20;98 20;99
h i
D 1=2 1=2 .15; 249 C 16; 976/  1=2 .15; 090 C 16; 432/  1=2 .52 C 54/

D 1=2 .16; 112:5  15; 761  53/

D 149 person-years of exposure to risk

The adjusted denominator is therefore the sum of the denominator we estimated

above and this figure (96,145.4 C 149 D 96,294.4 never married women, or person-
Issues Pertaining to Cohort Demographic Processes 109

years of exposure to risk), and we can use it to calculate an adjusted probability that
single women turning 20 in 1978 married at age 20 (11,861 / 96,294.4 D 0.12317).
This figure does not greatly alter the result (0.12337) that we obtained without
adjusting, and as this is frequently the case adjustment is often done away with.
However, if the added precision is deemed desirable for a particular piece of
analysis, adjustment should be undertaken.

Issues Pertaining to Cohort Demographic Processes

From the foregoing discussion of Lexis diagrams and the examples presented
illustrating their use you should have gained a sound understanding of the nature of
cohort demographic processes. With the cohort approach to demographic analysis
we trace the demographic experience of groups of individuals as they age together
up diagonal bands through a Lexis diagram. We turn now to some general issues
raised by, or relevant to, this approach.

The Tempo and Intensity of Cohort Processes

The tempo of a demographic process refers to its distribution in age-time or

duration-time. You can think of ‘age-time’ or ‘duration-time’ as being a cohort’s
diagonal band through the Lexis grid. Tempo refers to the distribution of a process
up that band. How rapidly or slowly, or how early or late, does it occur through the
life course? Variations in tempo (between two or more populations at a point in time
or within a population over time) are variations in the age or duration distribution of
a process.
The intensity of a cohort process refers to its level. The concept can be applied
overall (i.e., throughout the life cycle), or with reference to defined phases of the
life cycle. The overall intensity of the process of mortality is identical for every
population, since everyone eventually dies and does so only once. Other processes,
however, involve experiences that are not inevitable in the lives of individuals
(e.g., first marriage) or that may be either avoided or repeated in the lives of
individuals (e.g., giving birth, or fertility), so that their overall intensity varies
between populations and over time for individual populations.
With the process of mortality, identical overall levels of intensity do not preclude
variable levels of intensity over defined life cycle phases, although because the
overall intensity of the process is known and has a value of unity, or 100 % (death is
a once-only, universal experience), such measures are effectively also measures of
tempo (of the amounts of a known total occurring within defined life cycle phases).
This, however, is a peculiarity of the process of mortality alone.
There are various ways of measuring the intensity of a demographic process.
Consider Fig. 3.6, in which a cohort has been defined and e0 , e1 , e2 , etc. refer to
110 3 The Cohort and Period Approaches to Demographic Analysis

Exact Age
5

e4
4

e3
3

e2
2

e1
1

e0
0
Year 1 Year 2 Year 3 Year 4 Year 5 Year 6

Fig. 3.6 Lexis diagram illustrating the concepts of the ‘Intensity’ and ‘Tempo’ of a cohort
demographic process

numbers of events of a particular type taking place within successive age-time (or
duration-time) parallelograms. We could measure the intensity of this process by:

Intensity D †iD0;¨ .ei =N0 / (3.1)

Where ei D the number of events occurring at age (or duration) i; N0 D the popu-
lation attaining exact age (or duration) 0 (i.e., the original size of the cohort);
¨ D some age (or duration) beyond which nobody lives.
In the case of mortality, assuming a closed population, this measure is exactly
equal to 1 for every human cohort. However, if we let ei be the number of person-
years lived by the cohort at each age, the measure gives the cohort’s expectation
of life at birth – a measure of the intensity of surviving rather than of dying. For
the process of women giving birth to female children the equivalent measure is
the cohort net reproduction rate; a measure of the extent to which a population
of women replaces itself in the next generation. Another frequently encountered
measure of intensity is:

Intensity D †iD0;¨ .ei =1 Pi / (3.2)

Where ei D the number of events occurring at age (or duration) i; 1 Pi D the number
of person-years lived by the cohort between exact ages (or durations) i and i C 1
(approximately the mean population at age (or duration) i); ¨ D some age (or
duration) beyond which nobody lives.
Issues Pertaining to Cohort Demographic Processes 111

For the process of women giving birth to female children this measure gives the
cohort gross reproduction rate. For the process of women giving birth to children
(of either sex) it gives the cohort total fertility rate. For the process of (men or
women) marrying for the first time it gives the (male or female) cohort total first
marriage rate.
A frequently used measure of the tempo of a demographic process takes the
general form:
   
Tempo D †iD0;¨ i C 1=2 : ei = .†iD0;¨ ei / (3.3)

Where ei D the number of events occurring at age (or duration) i; ¨ D some age (or
duration) beyond which nobody lives.
Equation 3.3 gives the mean age at which the events of interest in a process
occurred (e.g., the cohort mean age at childbearing for the process of women giving
birth to children; the male or female cohort mean age at first marriage for the process
of men or women marrying for the first time). Medians, together with other similar
measures (e.g., quartiles) and, less commonly, modes are also used as measures of
tempo. Medians are often recommended over means where the distribution of an
event in age-time or duration-time is heavily skewed (i.e., instead of approximating
the bell-shaped normal curve familiar from statistics, the peak of the distribution is
displaced decisively to the right or left, destroying symmetry about that peak). A
case in point is the process of first marriage, for which the distribution of events
(first marriages) by age often has a pronounced peak quite early in the range
of marriageable ages and a long ‘tail’ extending over older marriageable ages,
although this pattern has become less pronounced than formerly in many developed
populations as consensual partnering and later marriage have become more common
during the ‘second demographic transition’ (van de Kaa 1987).

Renewable and Non-renewable Events

The distinction between renewable and non-renewable events was drawn in passing
in Chap. 1. To recapitulate, a non-renewable event is one that an individual can
experience only once (such as death, first marriage or having a first birth); a
renewable event is one that can recur in an individual’s life (such as giving birth,
getting married or divorcing).
Renewable events introduce a complexity into the construction and interpretation
of demographic measures. They are not well suited to the calculation of probabilis-
tic types of measures. It does not, for instance, make much sense to talk about the
probability of giving birth between two exact ages x and x C n, because individual
women can contribute more than one event (birth) over this n-year interval. With a
probability we aim to measure the likelihood of an event occurring once. Certain
specialized types of probabilities can be calculated for renewable event processes;
112 3 The Cohort and Period Approaches to Demographic Analysis

for example, we could calculate the probability of a woman giving birth at least
once between exact ages x and x C n. This, however, would require detailed data
on births by age of mother, birth order and birth interval, and because it eliminated
second and subsequent births to the same woman in the age interval would not be
a particularly good measure of fertility anyway. Demographic rates do not pose
the same difficulties for measuring renewable event processes, and are therefore the
more commonly used type of basic summary measure with such processes.
When it is desired to undertake detailed analysis of renewable event processes,
a common ploy is to split them into series of non-renewable processes. Thus, the
fertility process can be split into separate processes of first birth, second birth, third
birth, and so on. The marriage, or nuptiality, process can be split into processes
of first marriage, second marriage, etc., and if we want to recognize the different
modes of marriage dissolution that intervene between these processes and their
possible implications for remarriage prospects we might think in terms of processes
of first marriage, dissolution of first marriage by widowhood and by divorce, second
marriage preceded by widowhood and by divorce, dissolution of second marriage
by widowhood and by divorce, etc. Splits of this kind pave the way for construction
of meaningful probability measures.
For example, if in Eq. 3.1 ei is the number of first births occurring at age i, the
equation gives, for a closed population, the probability of a woman giving birth at
least once, or in other words the probability of becoming a mother. This measure is
also known as the cohort lifetime parity progression ratio from parity 0 to parity
1. The concept of parity, or the number of live births a woman has had, has been
touched on previously, and parity progression will be dealt with more fully in Chap.
6 when the analysis of fertility is discussed. But in similar fashion probabilities
of progressing between any other pair of successive parities can be computed. In
general, any cohort lifetime parity progression ratio beyond parity 1 is given by:

PPR D †iDy;¨ bi;kC1 =†iDy;¨ bi;k (3.4)

Where k and k C 1 are the two parities between which progression is being
measured; k ¤ 0; b D live births; i D a single-year age group; y, ¨ D respectively
the youngest and oldest ages at which any woman gives birth.
Equation 3.4 is the ratio of the total number of parity k C 1 births to the total
number of parity k births for a cohort throughout its life. It can also be applied to
progression from parity 0 to parity 1 by making the denominator N0 (the original
size of the female cohort), since all women attain parity 0.

Attrition

Attrition is the process by which a risk population is depleted (i.e., reduced in

size) over time as it experiences a non-renewable event. In a non-renewable
demographic process, once a person has experienced the reference event (s)he
Issues Pertaining to Cohort Demographic Processes 113

ceases to be at risk of experiencing it and is no longer part of the population under

analysis. Through age-time or duration-time the population at risk therefore gets
progressively smaller. It is to this process of getting smaller that the word ‘attrition’
refers.
Demographers often study attrition processes, or non-renewable event processes,
by means of attrition tables, or life tables. Strictly speaking a life table is an attrition
table which summarizes the non-renewable process of death, or mortality. However,
the term ‘life table’ has acquired a more general usage, such that these days any
attrition table is called a life table, and indeed the latter term is far more frequently
used than the former.
Attrition tables are constructed from probabilities of experiencing the reference
event between successive exact ages or durations (normally at single-year intervals,
although for a process like the termination of breastfeeding following birth, intervals
of a month would be more appropriate while mortality attrition is often summarized
using 5-year intervals). These tables typically feature the following columns:
1. The number of reference events occurring in each age or duration group, or
interval. When constructing an attrition table for a cohort, this column, obtained
empirically, along with knowledge of the initial size of the cohort (N0 in the
notation introduced with Eq. 3.1) is often the basis from which the rest of the
table is constructed. In the conventional, or standard, life table (summarizing the
non-renewable event process of death), values in this column are the 1 dx values –
1 d0 , 1 d1 , 1 d2 , etc. (see Chap. 4).
2. The population at the start of each age or duration interval; that is, the initial
population at risk at exact age/duration 0, and thereafter the survivors from
that population yet to experience the reference event. The term ‘survivors’ is
obviously especially apt when the reference event is death, and remains quite
appropriate when it is something like divorce, but it borders on comical when the
reference event is, say, first marriage (‘survivors in the never married state’). In
the notation introduced with Eq. 3.1 these values are denoted by N0 , N1 , N2 , etc.,
while in the notation of a conventional life table they are the lx values (l0 , l1 , l2 ,
etc.). If N0 and 1 dx values have been obtained empirically, remaining Nx values
can be found using the relationship Nx D Nx1 1 dx1 .
3. The age-specific or duration-specific probabilities of experiencing the reference
event within each age or duration interval. Denoted in the conventional life table
by 1 qx (1 q0 , 1 q1 , 1 q2 , etc.), this column can be obtained using the relationship
1 qx D1 dx / Nx (or 1 dx / lx ). However, as we shall see when the conventional life
table is discussed more fully in Chap. 4, it may also be derived independently
by other means and become the key to constructing the remainder of the attrition
table.
4. The number of person-time units (person-years, person-months, etc., depending
on the age/duration interval employed in the table) spent at risk (or ‘lived’) by
survivors from the initial population at risk within each age/duration interval. In
conventional life table notation this is the 1 Lx column, and provided a uniform
distribution of reference events by exact age/duration within an age/duration
114 3 The Cohort and Period Approaches to Demographic Analysis

interval can be assumed, these values can be obtained from the relationship
1 Lx D (Nx C NxC1 ) / 2 (or (lx C lxC1 ) / 2).
5. Cumulated person-time units spent at risk above each exact age/duration. These
are the Tx values in conventional life table notation, and are obtained by summing
all values of 1 Li where i  x.
6. The mean number of person-time units (person-years, person-months, etc.)
remaining before the event occurs per ‘survivor’ at each exact age/duration. This
is the eo x , or ‘expectation of life’, column of the conventional life table and is
obtained using the equation eo x D Tx =Nx (or Tx / lx ).
7. It is also not uncommon to find a column of age-specific (or duration-specific)
rates of experiencing the reference event. Denoted in the conventional life table
by 1 mx (1 m0 , 1 m1 , 1 m2 , etc.), this column is generated using the relationship
1 mx D1 dx /1 Lx .

The conventional life table is the most common type of attrition table in
demography, but the principles it employs can be used in analysing any other non-
renewable demographic process. Table 3.1 is an example of an attrition table for
the non-renewable process of termination of breastfeeding after birth of an infant.
Breastfeeding patterns are of interest to demographers in their search for under-
standing of fertility levels and trends in different populations, since breastfeeding
acts as a form of contraceptive, delaying the resumption of ovulation after birth of a
child. Aside from any nutritional arguments, the encouragement of breastfeeding is
often advocated to lengthen birth intervals, and thus lower fertility, in high fertility
populations.
The input data in Table 3.1 are the 1 dx values together with N0 , which were
obtained from a survey. The table relates to second-last births of women (to avoid
the complication of women still breastfeeding their youngest child at interview), and
excludes a few cases where the child was not breastfed or died before breastfeeding
ended. There are clear questions over the quality of the data, with obvious tendencies
for respondents to have opted for breastfeeding durations of 12 and 18 months
over durations either side of those ones (see 1 dx column). The degree to which
this reflects them having set those durations as targets in advance or having
approximated their durations of breastfeeding in retrospect is unknown. Note also
that the table has been truncated at x D 24 months, so that (i) the last row deals with
terminations of breastfeeding at durations of 24 months or longer (‘24C’ months),
and (ii) since all women who reached exact duration 24 months still breastfeeding
had to terminate at some duration thereafter, the probability of termination in this
‘open’ duration interval (q24C – the first subscript is omitted because the interval is
not of width 1 month) is 1.0000. Columns of the table apart from 1 dx were generated
using relationships specified above. The final value in the 1 Lx column assumes that
all women still breastfeeding at exact duration 24 months terminated breastfeeding
during the twenty-fifth month after on average breastfeeding for half that month.
The ‘attrition’ in Table 3.1 is evident in the Nx column, where the initial 2,164
breastfeeding women are depleted month by month until after 24 months only 32 are
still breastfeeding. The eı x column gives average periods of breastfeeding remaining
Period Analysis and Synthetic Cohorts 115

Table 3.1 Breastfeeding Month (x) eı x

1 dx NX 1 qx 1 Lx Tx
attrition table for Sudan, 1979
0 4 2;164 0.0018 2;162 15;700 7.3
1 27 2;160 0.0125 2;147 13;538 6.3
2 79 2;133 0.0370 2;094 11;392 5.3
3 150 2;054 0.0730 1;979 9;298 4.5
4 337 1;904 0.1770 1;736 7;319 3.8
5 274 1;567 0.1749 1;430 5;584 3.6
6 323 1;293 0.2498 1;132 4;154 3.2
7 389 970 0.4010 776 3;022 3.1
8 142 581 0.2444 510 2;247 3.9
9 111 439 0.2528 384 1;737 4.0
10 72 328 0.2195 292 1;353 4.1
11 10 256 0.0391 251 1;061 4.1
12 154 246 0.6260 169 810 3.3
13 6 92 0.0652 89 641 7.0
14 6 86 0.0698 83 552 6.4
15 6 80 0.0750 77 469 5.9
16 6 74 0.0811 71 392 5.3
17 1 68 0.0147 68 321 4.7
18 25 67 0.3731 55 254 3.8
19 2 42 0.0476 41 199 4.7
20 2 40 0.0500 39 158 4.0
21 3 38 0.0789 37 119 3.1
22 2 35 0.0571 34 83 2.4
23 1 33 0.0303 33 49 1.5
24C 32 32 1.0000 16 16 0.5
Source: Sudan Fertility Survey, 1979

for women still breastfeeding at each exact duration of breastfeeding since birth.
Thus the first value in it tells us that the 2,164 women on average breastfed for
7.3 months. This figure then generally declines, but rises sharply again at exact
duration 13 months – i.e., immediately after the attrition of women who claimed to
have breastfed for 12 months. This indicates that women continuing to breastfeed
beyond that point were very committed to doing so. Those still breastfeeding after
13 months on average continued to do so for a further 7.0 months, so on average
breastfed for 20.0 months.

Period Analysis and Synthetic Cohorts

We began this chapter by noting that demographic analysis adopted either of

two approaches: the cohort approach or the period approach. Thus far we have
concentrated exclusively on the cohort approach, which in many ways is the
116 3 The Cohort and Period Approaches to Demographic Analysis

preferred analytic perspective in demography. It has a realism that derives from

the fact that it focuses on an actual group of individuals and traces their experience
through the life course. Moreover, most demographic measures are based in some
way or other on measures of the intensity or tempo of demographic processes for
cohorts.
The cohort approach does, however, have drawbacks. Demographic data are
usually not collected on a cohort basis. Registration data tend to be tabulated
annually, the events covered corresponding to vertical columns in the Lexis grid.
Similarly census and survey data usually refer to a time instant (the latter, perhaps,
in an approximate sense since interviewing may take several days or weeks), and
therefore correspond to a vertical line through the Lexis grid. Surveys which collect
retrospective life histories take a cohort approach to the gathering of data. But cohort
analysis (running diagonally through the Lexis grid) frequently draws on many
separate annual series of data, adding to its complexity. More crucially, cohorts
take time to live out their lives, and we cannot afford to wait around until they
have finished their lives before starting to study demographic processes affecting
them. We can, and do, study the partial, or incomplete, experience of cohorts, up to
whatever stages of their lives they have reached at the time of analysis. This does
not necessarily obscure demographic change as it occurs, but it can be an inefficient
way of demonstrating it.
Given these drawbacks, demographers often engage in period analysis. This is a
cross-sectional approach to analysis in which the focus is either on events occurring
in a particular year (or consecutive group of years) or on events occurring to a
population at a particular point in time in the age/duration groups its members
occupied at that time. Thinking in terms of the Lexis diagram, we saw that cohort
analysis is concerned with events located along diagonal bands drawn through
the Lexis grid. Period analysis is concerned with events arranged VERTICALLY
through the Lexis grid (Fig. 3.7). Either they are located in columns of the Lexis
grid (analysis is focused on events that occurred in a particular year or series of
consecutive years – the upper diagram in Fig. 3.7), or they are located in a series
of parallelograms stacked vertically through the grid about a line representing a
focal population (analysis is concerned with events that occurred to members of this
population in the age/duration groups they occupied at the focal time, each group’s
experience being represented by its relevant parallelogram of cohort experience –
the lower diagram in Fig. 3.7).
It is common with period analysis to treat events occurring in either of the two
Lexis configurations just described AS IF they were experienced by a cohort. This
practice gives rise to a distinction between real cohorts and synthetic cohorts. Real
cohorts are the ‘genuine’ cohorts we have been dealing with thus far. They are the
ones represented by diagonal bands through a Lexis diagram. Synthetic cohorts,
by comparison, are hypothetical cohorts, represented on a Lexis diagram by either
of the configurations illustrated in Fig. 3.7 and, as those two diagrams illustrate,
composed of segments of the experience of a series of real cohorts. You can see
in Fig. 3.7 that diagonal bands representing several successive real cohorts have
been drawn in, and that the shaded synthetic cohorts corresponding to demographic
Period Analysis and Synthetic Cohorts 117

Exact Age
4

0
2006 2007 2008 2009 2010 2011

Year
Exact Age
4

0
2006 2007 2008 2009 2010 2011

Year

Fig. 3.7 Alternate Lexis diagram representations of a synthetic cohort

experience during calendar year 2010 (upper diagram) and of the population at the
end of 2010 (lower diagram) comprise little bits (segments) of the experience of
each of these real cohorts. Thus the calendar year 2010 synthetic cohort consists of:
Half the experience at age (or duration) 0 of the 2010 real cohort
Half the experience at each of ages (or durations) 0 and 1 of the 2009 real cohort
Half the experience at each of ages (or durations) 1 and 2 of the 2008 real cohort
Half the experience at each of ages (or durations) 2 and 3 of the 2007 real cohort
Etc.
Similarly the 31 December 2010 synthetic cohort consists of:
The experience at age (or duration) 0 of the 2010 real cohort
The experience at age (or duration) 1 of the 2009 real cohort
118 3 The Cohort and Period Approaches to Demographic Analysis

The experience at age (or duration) 2 of the 2008 real cohort

The experience at age (or duration) 3 of the 2007 real cohort
Etc.
Note that the little bits of real cohort experience that we cobble together to
form synthetic cohorts collectively cover all possible life course phases, without
duplicating any phase. Note also that when period analysis is concerned with
measures constructed for synthetic cohorts it essentially asks the following question:
What would be the life cycle experience of this demographic event of a real
cohort which happened to encounter the age-specific (or duration-specific) rates
or probabilities of experiencing the event which prevailed (i) in this particular
year, or (ii) cross-sectionally among the population at this particular date?
This is a ‘what if?’ question. What would a real cohort’s life cycle experience be
if it happened to experience this cross-sectional regime of age-specific (or duration-
specific) rates or probabilities. You can perhaps now appreciate why synthetic
cohorts are sometimes called hypothetical cohorts.

The Tendency of Synthetic Cohort Measures to Exaggerate

Change

Period measures that purport to measure the lifetime experience of cohorts (i.e.,
that measure the lifetime experience of synthetic cohorts) are apt to take on values
that are abnormally high or low. If a trend line plotting such a measure over time
is compared with one plotting the equivalent measure for successive real cohorts,
typically the line for the synthetic cohort measure has more pronounced peaks and
troughs than that for the real cohort measure.
Figure 3.8 illustrates this phenomenon. It shows for New Zealand over a lengthy
period the total fertility rate (TFR) and the cohort completed fertility rate (CCFR).
The TFR is a synthetic cohort measure which, we noted in Chap. 1, is found
by summing age-specific fertility rates in a year over all reproductive ages, and
which indicates the number of children women would on average have during their
reproductive lives assuming they experienced that year’s schedule of age-specific
fertility rates (in other words it gives the average completed family size for the
synthetic cohort). The CCFR gives the actual average completed family size for
(real) birth cohorts of women. It is plotted in Fig. 3.8 with birth cohorts offset
27 years against calendar year synthetic cohorts, age 27 being an approximate mean
age of childbearing for the period covered. Thus the CCFR plot at calendar year
1921 on the horizontal axis is the plot for the 1894 birth cohort (the cohort that
turned 27 in 1921). For the birth cohorts of 1962–1971, CCFRs (plotted against
calendar years 1989–1998) are estimates based on recorded fertility to at least age 40
and then, increasingly as cohorts become more recent, projections of future fertility
at ages 40–49. Because fertility is relatively low at ages 40–49 these projections
Period Analysis and Synthetic Cohorts 119

Fig. 3.8 Trends in New Zealand’s total fertility rate and cohort completed fertility rate (Source:
Statistics New Zealand. Note: Cohort completed fertility rates cover 1894–1971 birth cohorts)

will almost certainly ultimately prove to be quite accurate, and allow the CCFR
trend line to be extended a further 10 years. It is not, however, extended beyond the
1971 birth cohort (plotted against calendar year 1998) because the birth cohorts in
question were in 2011 (the final year for which the TFR is plotted) still aged less
than 40. This, of course, serves to illustrate one of the drawbacks of cohort analysis
noted above – the availability of only partial data for some cohorts.
You can see in Fig. 3.8, however, that the TFR ‘troughs’ in the mid-1930s and
briefly in the early 1940s, then again in the early 1980s at levels lower than the
CCFR for birth cohorts then of peak childbearing age (i.e., cohorts of the early
1900s and late 1950s respectively). Similarly it ‘peaks’ in 1960 at over 4.2 children
per woman, well above the CCFR peak of 3.6 children per woman recorded for the
1931 birth cohort. A difference of 0.6 of a child per woman in estimates of the peak
average family size associated with the post-war baby boom may not seem large,
but it is a substantial one.
So what causes this tendency for synthetic cohort measures like the TFR to
exaggerate the extent of demographic change? We noted above that synthetic
cohorts add together bits, or segments, of the experience of series of real cohorts,
and one answer to the question is that, at times, they add together particularly
favourable or unfavourable segments. This happens when there is cross-sectional
heaping of demographic events, or when cross-sectional deficits of events occur.
In the former circumstance cross-sectional age-specific or duration-specific rates
or probabilities of event occurrence are generally higher than any real cohort
contributing part of the synthetic cohort will experience across its life; in the latter
120 3 The Cohort and Period Approaches to Demographic Analysis

Fig. 3.9 Cumulative divorce trajectories to 1994 for selected Australian marriage cohorts (Source:
Adapted from Carmichael et al. (1997: Figure 4))

circumstance they are generally lower. Two mechanisms, either separately or in

combination, can bring these circumstances about.
1. Demographic behaviour often responds to historical events or periods which
are either especially conducive or especially unfavourable to that behaviour.
For example, marriage activity and childbearing may be reduced in times of war
or severe economic depression, and conversely may be strongly stimulated by the
cessation of war or economic recovery. The Great Depression and the outbreak
of the Second World War clearly had the former effect in New Zealand (Fig. 3.8).
Liberalizing divorce laws similarly invariably gives rise to an immediate, across-
the-board increase in synthetic cohort measures of the incidence of divorce, these
implying a level of instability of marriage higher than is characteristic of any
real marriage cohort. These sorts of ‘period effects’, as they are known, tend
to affect entire populations when they occur; different real cohorts are affected
simultaneously at whatever stages of the life cycle they happen to have reached
at the time, leading to a sudden surge or decline in the cross-sectional incidence
of the demographic behaviour in question.
Figure 3.9 illustrates this. It shows for selected financial year (July–June)
Australian marriage cohorts, followed until 1994, cumulative divorce rates
to successive exact marriage durations. Each cohort trajectory at one point
steepens sharply, this steepening occurring at successively earlier marriage
durations for successively more recent marriage cohorts. What is in evidence
Period Analysis and Synthetic Cohorts 121

Fig. 3.10 Total first marriage rate for Australian females, 1921–2002 (Source: Adapted from
Carmichael (2002: Figure 1))

is the introduction in 1976 of Australia’s Family Law Act, which substantially

liberalized the grounds for divorce. There was a rapid, in part temporary, surge
in the number of divorces granted (i) as couples who would have had to wait to
divorce under the old law became immediately eligible to do so under the new
one, and (ii) as people took the new law to signal greater public acceptance of
dissolving unhappy marriages. Marriage duration-specific divorce rates recorded
in 1976 implied that, in a synthetic marriage cohort constructed for that year,
62 % and almost 71 % of marriages would end in divorce within 30 years and
50 years respectively. But these were one-off figures, unlikely to be remotely
approached by any real marriage cohort in the foreseeable future. The equivalent
1977 synthetic cohort figures were much lower at 44 and 50 %, and by 1979 they
were down to 36 and 40 % as backlogs of disrupted marriages at all marriage
durations were cleared.
When a synthetic cohort measure is calculated for a period or a time instant
that is especially auspicious or inauspicious for the demographic behaviour
being studied, the inference is that a real cohort could sustain the abnormally
high or low cross-sectional incidence of the behaviour prevailing at that time
throughout an actual life cycle. In reality this is rarely the case. Real cohorts
tend to compensate for abnormally high or low levels of experience of a
demographic event at one stage of the life cycle by lowering or raising their
levels of experience at later stages. Figures 3.10 and 3.11 are again illustrative.
The former shows the total first marriage rate (TFMR) for Australian females
since 1921, a synthetic cohort measure analogous to the TFR which indicates the
proportion of a birth cohort who would ever marry if it experienced through life
the age-specific ratios of first marriages to total population recorded in the year
for which the calculation is performed. The feature of this graph to note at this
point is the huge trough centred on 1931. This reflects the economic disincentive
to marriage generated by the Great Depression.
122 3 The Cohort and Period Approaches to Demographic Analysis

Fig. 3.11 Cumulative first marriage rates to selected exact ages for Australian female birth cohorts
(Source: Adapted from Carmichael (1988: Figure 4))

If we then focus on Fig. 3.11 we find proportions of successive female birth

cohorts in Australia ever married by exact ages 18, 20, 22, 25, 30 and 35 plotted.
The effect of the Depression is evident in ‘dips’ in trend lines between time
lines drawn at 1:1:30 and 1:1:40. But if we note the birth cohorts for which the
dips are evident at exact ages 20 and 22 (those of 1911–1919 and 1909–1917
respectively), then examine the trend for the same cohorts at exact age 35, the
dip has disappeared. By this life cycle stage these cohorts had not only made up
for reduced marriage activity in their teens and early twenties; they’d become
integrated into a trend toward more universal marriage.
2. The second mechanism that causes synthetic cohort measures to exaggerate the
extent of change over time is change in the tempo, or timing, of a demographic
process; that is, a tendency, over time, for the process to occur earlier or later in
the life cycle. Consider Fig. 3.12, on which is represented (i) the first marriage
experience of a 1990 synthetic cohort of women attaining marriageable age in
Australia (exact age 16) (the shaded vertical column), and (ii) the first marriage
experience of a series of real cohorts of women attaining marriageable age
between 1976 and 1990, each of which contributed part of their experience to
Period Analysis and Synthetic Cohorts 123

Fig. 3.12 Lexis diagram illustrating how a change in the tempo of the process of first marriage
leads to abnormally high or low numbers of first marriages occurring cross-sectionally in synthetic
cohorts

the synthetic cohort (the dashed diagonal bands – note that before 1986 only
selected bands are shown, but you can imagine where others would run). The
overlap between the diagonal bands and the shaded vertical column defines the
parts of the experience of the various real cohorts that are added together to create
the synthetic cohort.
Suppose that a trend to earlier marriage was in progress in this population
around 1990. This means that, in that year, the number of first marriages at
younger ages will be relatively high (compared to recent years), because real
cohorts that reached marriageable age more recently (those of 1986–1990,
say) will be leading the trend to earlier marriage (and such a trend implies
proportionately more people marrying at younger ages). However, assuming
that the propensity to ever marry remains fairly constant, the number of first
marriages at older marrying ages in 1990 will also be relatively high (compared
to what women in real cohorts supplying marriages at younger ages to the 1990
synthetic cohort will experience 10 years or so into the future when they reach
the older marrying ages). Why? We have to think of what the real cohorts
supplying first marriages to our synthetic cohort at older ages (late 20s and
early 30s) experienced when they were younger. Figure 3.12 indicates that the
124 3 The Cohort and Period Approaches to Demographic Analysis

cohorts concerned reached marriageable age 10–15 years previously, in the mid
to late 1970s. Since we have a trend to earlier marriage in progress we can
assume that these cohorts married less freely when they were young than the
real cohorts of 1986–1990. It follows that when these earlier cohorts reached
the older ages at which they were contributing to the 1990 synthetic cohort,
relatively large numbers of their members were still unmarried and at risk of
marrying for the first time at those older ages (compared to the situation that
would pertain in the future when the 1986–1990 real cohorts reached their late
20s and early 30s). Thus, the 1990 synthetic cohort combines the experience
of younger real cohorts whose first marriage experience is relatively heavily
weighted toward younger ages with that of older real cohorts whose experience is
relatively heavily weighted toward older ages. The result is an abnormally high
number of female first marriages occurring cross-sectionally in 1990. We call
this phenomenon cross-sectional heaping.
Alternatively, suppose that a trend to later marriage is in progress around
1990. This means that, in that year, the number of first marriages at younger
ages will be relatively low (compared to recent years), because real cohorts
that reached marriageable age more recently will be leading the trend to later
marriage (and such a trend implies proportionately fewer marriages at younger
ages). However, the number of first marriages at older marrying ages in 1990
will also be relatively low (compared to what women in real cohorts supplying
first marriages at younger ages to the 1990 synthetic cohort will experience
10 years or so into the future when they reach the older marrying ages). Why?
Again we have to think of what the real cohorts supplying first marriages to
our synthetic cohort at ages in the late 20s and early 30s experienced when
younger. Since a trend to later marriage is in progress we know that these cohorts
married more freely when they were young than the real cohorts of 1986–1990
(Fig. 3.12). It follows that when they reached the older ages at which they were
contributing marriages to the 1990 synthetic cohort, relatively small numbers
of their members were still unmarried and at risk of marrying for the first
time at those older ages (again, compared to the situation that would pertain
in the future when the 1986–1990 real cohorts reached their late 20s and early
30s). Thus, the 1990 synthetic cohort combines the experience of younger real
cohorts whose first marriage experience is relatively lightly weighted toward
younger ages with that of older real cohorts whose experience is relatively lightly
weighted toward older ages. The result is an abnormally low number of female
first marriages occurring cross-sectionally in 1990, a phenomenon we refer to as
a cross-sectional deficit.
The translations between real and synthetic cohort experience which accom-
pany changes in tempo and have been outlined here for the process of first
marriage apply equally in respect of other demographic processes. They can
be complicated if a change in intensity accompanies the change in tempo.
But ignoring this possibility and maintaining the focus on the process of first
marriage, the capacity of changes in tempo to cause synthetic cohort measures to
assume exaggeratedly high or low values is amply demonstrated by Fig. 3.10.
Period Analysis and Synthetic Cohorts 125

This graph features a horizontal line at TFMR D 1,000. This is the value of
the TFMR that implies universal marriage, and clearly over long periods the
TFMR for Australian females exceeded that level. That this is possible is at first
perplexing, but all a TFMR above 1,000 means is that first marriages have heaped
cross-sectionally in the year in question at a level that is unsustainable in the
longer term. Somewhere in the recent past, or down the track, there has been, or
will be, a compensating deficit of first marriages.
The first period during which the TFMR exceeds 1,000 coincides with the
Second World War, at which time extreme cross-sectional heaping of first mar-
riages was due to the circumstances of war (i.e., the ‘historical events/periods’
mechanism was operating). There was a ‘now or never’ rush to the altar (or
registry office) by couples seeking to marry before grooms departed for war
service overseas, and this was aided by the fact that, with separation imminent,
the usual economic prerequisites to marriage could be largely disregarded. Then,
later during the war, first marriage activity received a considerable boost from the
presence of American servicemen in Australia on leave. It is estimated that there
were 12,000–15,000 marriages of American servicemen to Australian women
during the War, much having been written about the attraction their uniforms,
their wallets, their readiness for a good time and their capacity to provide gifts
of merchandise that was in short supply (e.g., nylon stockings) represented
(Aitchison 1972; Moore 1981; Sturma 1989).
The second period during which the female TFMR exceeded 1,000 lasted
from the end of the war (when the return of servicemen temporarily created
another special stimulus to marriage) to the early 1960s. This sustained period
of high TFMRs, which really lasted through to the early 1970s, was in large part
the product of declining ages at first marriage, which were an integral part of
Australia’s post-war marriage boom and therefore its baby boom. In other words
it was largely a product of the ‘changing tempo’ mechanism outlined above.
Later on, through the 1970s, the female TFMR fell steeply, until by the end of
the decade it was at a level that suggested that only about two-thirds of a cohort
attaining marriageable age would ever marry. The evidence for real cohorts is that
this figure is exaggeratedly low; that perhaps 20–25 %, rather than a third, will
fail to marry. Again the changing tempo mechanism was primarily responsible
for this exaggeration, because commencing in the early 1970s the previous trend
to earlier marriage reversed emphatically and was replaced, as part of the second
demographic transition, by a strong trend to later marriage.

Other Issues in Period Analysis

We often make assumptions and use approximations in constructing period mea-

sures. This is especially so when period analysis follows the model illustrated by
the upper diagram in Fig. 3.7. Figure 3.13 illustrates the translation of real cohort
concepts in the construction of demographic measures using this model. In a real
126 3 The Cohort and Period Approaches to Demographic Analysis

Fig. 3.13 Lexis diagram illustrating the translation between real cohort and synthetic cohort
demographic probabilities and rates

cohort the events considered in constructing a rate or probability for any age or
duration group lie in a parallelogram; for example, in the parallelogram A0 B0 C0 D0 .
In a period synthetic cohort of the type depicted in the upper diagram of Fig. 3.7 they
lie in a square (e.g., ABCD). In the real cohort case, the events in A0 B0 C0 D0 all occur
to the population at A0 B0 ; that is, to the population attaining the exact age/duration
marking the start of the relevant age or duration interval. However, in the period
synthetic cohort case events in the triangle ACD do not occur to members of the
population at AB. The period equivalent of the cohort probability events in A0 B0 C0 D0
divided by the population at A0 B0 , that is, events in ABCD divided by the population
at AB, assumes that events in ACD give a reasonable estimate of events in the
equivalent triangle whose left side is the line BC. Similarly the period equivalent of
the cohort rate events in A0 B0 C0 D0 divided by the population at B0 D0 , that is, events
in ABCD divided by the population at a line joining the midpoints of AB and DC
(the dashed vertical line), is perhaps best thought of as approximating the rate for the
cohort represented by this line. The relevant events for the ‘true’ rate for this cohort
are contained not within ABCD, but within the dashed parallelogram centred on
the dashed vertical line. This parallelogram overlaps substantially with the square
ABCD, but its left and right extremities extend into the grid squares immediately
adjacent to ABCD on its left and right. Our period rate essentially assumes that
the number of events lying inside ABCD but outside the dashed parallelogram
is a reasonable approximation of the number lying inside the parallelogram but
outside ABCD.
All measures used in cohort analysis have equivalents in period analysis, but
many summary measures which have simple constructions as cohort measures must
References 127

be calculated in different ways as period measures so as to make allowance for

the differences between period and cohort populations. Measures such as the net
reproduction rate and expectation of life, which are simple measures of intensity for
a cohort, become more complicated when calculated from period data.
Period measures should not, however, be seen as second-best to cohort measures,
despite their derivative status. Many influences on demographic processes, as we
have already seen, are time-specific rather than specific to the life experience of
particular cohorts. The impact of a famine, a war, an influenza pandemic, the
introduction of a new treatment for a particular disease, a major change in divorce
legislation, etc. will be felt by all age/duration cohorts in a population at the time it
occurs, not by a particular cohort or group of cohorts. Similarly, where demographic
processes are affected by economic conditions or changes in social attitudes these
effects also tend to be time-specific rather than cohort-specific (although there may
be variability across cohorts in a matter such as revision of social attitudes – younger
cohorts are often more receptive to social change than are older ones). There is a
good case to be made that most influences on demographic processes are period,
or cross-sectional, or time-specific influences, and this means that a need for period
measures would exist even if cohort measures were available in a timely fashion
(which they are not). As already noted, cohort analysis smoothes out sudden peaks
and troughs in period measures that are caused by transient circumstances. It tends
to focus on permanent, as opposed to transient, change. But demographers are
interested in both types of change, and so there is a place for both the period and the
cohort approaches to demographic analysis.
There are cases where the calculation process that gives meaningful results for a
cohort gives very misleading results when applied to period analysis. For example,
if we have cohort data on numbers of children ever born classified by mothers’
ages, it is possible to calculate age-specific fertility rates for the cohort of women
from the differences between mean numbers of children ever born to the cohort at
different ages. However, because fertility is a very volatile demographic variable
the same technique applied cross-sectionally to a population, seeking to combine
the experience of different cohorts of women with possibly very different fertility
histories, is usually spectacularly unsuccessful.
With any type of period analysis, the interpretation of measures is helped by
considering the principles underlying the construction of equivalent measures in the
cohort case. Nearly all measures have very simple interpretations for cohorts, which
may be difficult to discern for equivalent period measures.

References

Aitchison, R. (1972). Thanks to the Yanks? The Americans and Australia. Melbourne: Sun Books.
Carmichael, G.A. (1988). With this ring: First marriage patterns, trends and prospects in Australia.
Canberra: Department of Demography, Australian National University and Australian Institute
of Family Studies.
128 3 The Cohort and Period Approaches to Demographic Analysis

Carmichael, G. A. (2002). Taking stock at the millennium: Family formation in Australia. In G.

A. Carmichael & A. Dharmalingam (Eds.), Populations of New Zealand and Australia at the
millennium (pp. 91–104). Canberra/Wellington: Australian Population Association/Population
Association of New Zealand.
Carmichael, G. A., Webster, A., & McDonald, P. (1997). Divorce Australian style: A demographic
analysis. In C. A. Everett (Ed.), Divorce and remarriage: International studies (pp. 3–37). New
York: The Haworth Press.
Moore, J. H. (1981). Over-sexed, over-paid and over here: Americans in Australia 1941–1945. St
Lucia: University of Queensland Press.
Sturma, M. (1989). Loving the alien: The underside of relations between American servicemen and
Australian women in Queensland, 1942–1945. Journal of Australian Studies, 13(24), 3–17.
Van de Kaa, D. (1987). Europe’s second demographic transition. Population Bulletin, 42(1), 1–57.
Chapter 4
Analysis of Mortality: The Life Table
and Survival

Introductory Matters

In earlier chapters we have played around a bit with the crude death rate (CDR). We
have seen that, while it is the most readily available measure of mortality, true to its
name it is crude, to the point of being at times extremely misleading. Indeed, other
than as a component of population growth, it is a measure to be avoided.
In our efforts to sort out a perplexing differential between CDRs for Australia
and Malaysia we discovered that any meaningful analysis of mortality levels and
patterns must pay heed to the extreme variability in the incidence of death by
age in particular, and also by sex. We looked initially at calculating death rates
specific for age and sex, and followed up by making use of such rates in applying
techniques of standardization and decomposition to expose the capacity for serious
distortion in a simple comparison of CDRs. Age and sex are not the only dimensions
of population composition along which mortality varies; as was noted in Chap.
2, mortality levels in some populations vary across such additional dimensions as
ethnicity, education level, occupation and marital status. In Australia, for instance,
there are major differences between the mortality patterns of the Indigenous and
non-Indigenous populations; mortality can be higher among the least well educated
sector of a population because of poorer living conditions, more limited knowledge
of behaviours and environmental conditions that endanger health, and more limited
awareness of and access to modern medical care; some occupations (e.g., mining,
building and construction) are inherently more hazardous than others are; and being
unmarried sometimes carries a higher mortality risk, in part because poor health
selects certain people out of marriage. But variation in mortality levels by age and
sex is universal.
One consequence of this is that the analysis of mortality typically deals with
patterns for males and females separately. In this chapter we will examine the
life table, the major tool of the demographer in the analysis of general mortality

© Springer International Publishing Switzerland 2016 129

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_4
130 4 Analysis of Mortality: The Life Table and Survival

patterns and trends. Life tables are usually constructed separately for males and
females. Concerning differentials in the level of mortality by age, probably the
most important is the invariably higher level of mortality in infancy (the first year
following live birth) than at other childhood ages. Infant mortality is an especially
important category of mortality. It is both a major component of total mortality in
many (particularly less developed) populations, and by far the category of mortality
where the largest number of avoidable deaths occurs. Thus, the greatest potential
for reducing overall mortality often rests with reducing infant mortality. Moreover,
in some populations a reduction in infant mortality is doubly desirable in that
added certainty over the survival of children is likely to be a force for fertility
decline. Infant mortality is also important because, since they occur at the very
beginning of life, infant deaths averted have the greatest potential for improving
overall survivorship in a population as measured by the well known summary
measure average expectation of life at birth. In many ways the infant mortality
rate is itself a summary measure of the overall level of mortality in a population;
high mortality and low mortality populations typically have high and low infant
mortality rates respectively. Infant mortality is also central to the study of child
survival, a prominent specialist field of policy-relevant research within demography.
First, though, it is necessary to deal with general measures and methods in mortality
analysis.
We need to be clear at the outset just what it is we are studying when we study
mortality. What is ‘death’? The World Health Organization (1977: 15) defines it as
‘the permanent disappearance of all evidence of life at any time after live birth has
taken place.’ The main point to note about this definition is its specific exclusion
of foetal deaths (deaths ‘prior to the complete expulsion or extraction from its
mother of a product of conception : : : [such that] after : : : separation the foetus
does not breathe or show any other evidence of life’ (World Health Organization
1977: 13) – in other words, miscarriages, abortions and stillbirths). The deaths in
which demographers are interested for the purpose of studying general mortality
patterns and trends presuppose ‘live birth’ (the definition of which will be left until
the analysis of fertility is discussed).

The Life Table: A General Perspective

As noted above, the life table is the major tool available to demographers for the
general analysis of mortality. A life table is constructed from age-sex-specific death
rates. While life tables typically are regarded as being concerned with mortality,
they should more properly be thought of as being concerned with survival; that is,
with not dying rather than with dying. Viewed this way a life table becomes a tool
for focusing on the health of a population, as indexed by the quantity of life its
members enjoy. There are also, of course, issues of quality of life. In its basic form
the life table does not directly address these issues, although quantity and quality
The Life Table: A General Perspective 131

are to a considerable degree correlated, and the life table approach can be extended
to encompass issues of quality of survival.
The life table is also appropriately thought of as a device concerned with survival
in the context of population projections and associated planning applications.
Demographers with planning responsibilities often need to estimate numbers of
people who will occupy the age groups for which particular types of facilities and
services will be needed at nominated dates in the future, and estimating survivors
from a census population and/or projected future birth cohorts who will form the
relevant age groups at the relevant dates is, together with a consideration of the
likely impact of migration, fundamental to such an exercise. The life table provides
us with a theoretical framework from which to calculate survivors.
A life table is an attrition table that traces the process of mortality (or survival).
In its usual period form it is a good example of the application of cohort concepts to
period data. It is obviously impractical to wait around for over 100 years until a birth
cohort dies out before constructing a complete ‘cohort’ record of mortality. Cohort
life tables can be constructed for partial cohorts, but they are partial life tables in
two senses. Not only do they pertain to partial cohorts; they do not include some
of the more important life table functions (in particular the ‘average expectation
of life remaining’ function), which cannot be calculated until a cohort’s complete
mortality experience is known. Given these difficulties, conventional practice is to
use cross-sectional data to construct life tables which trace the mortality experience
of synthetic birth cohorts. These product-limit life tables, as they are known because
of their distinctive method of construction, address the following question:
What would be the lifelong process of mortality attrition of a birth cohort of
men or women which experienced the age-specific death rates that prevailed
for the relevant sex in year y?
An obvious further question arises. Is it reasonable to assume that persons born in
year y will, 60 and more years into the future, say, experience the year y death rates
at ages 60 and over? People forming these age groups in year y might have been
exposed to risk factors injurious to their health earlier in their lives that younger
people may be less affected by; for example, younger people might be more aware
of the health risks associated with smoking than were older people when they
were younger, and may therefore smoke less. Medical advances might in the future
assist inherently more frail individuals to survive to older ages, reducing death
rates at younger ages but increasing them at older ages. These are very plausible
propositions, but without a crystal ball, can we do better than construct our life
tables using the latest information available? This seems to satisfy life insurance
companies, whose premiums are based on detailed analyses of life tables. They
probably console themselves by assuming that most future changes in mortality
regimes are likely to be improvements, and therefore to their advantage, although
the emergence during the 1980s of HIV-AIDS warned against complacency on that
front. Then again, it is not as if, for populations covered by the insurance industry,
new life tables are not produced at regular intervals, so that mortality conditions at
132 4 Analysis of Mortality: The Life Table and Survival

different ages are updated (and premiums adjusted) as any real cohort approaches
those ages.
There are two types of period life table: single-year-of-age life tables (also
known as standard or complete life tables) which, as their name indicates, treat
each year of life separately; and abridged life tables, which are based on death rates
for fewer, broader age intervals. While other schemes of abridgement are possible,
abridged life tables typically use death rates for 5-year age groups, except that (i) the
first year of life is treated separately, so that age group 0–4 is split into age groups
0 and 1–4, and (ii) the oldest ages (frequently ages 85 and over) are grouped into a
single category.

The Single-Year-of-Age Life Table

Although excessively detailed for many demographic purposes, the single-year-of-

age life table is the more straightforward variant when it comes to understanding the
mechanics of life table construction. It begins with a group of people (of a given sex)
at birth and follows them through life, subtracting each year the number who die.
We have already seen, in the example using Sudanese breastfeeding data presented
in the previous chapter, that the principles of attrition can be readily extended to the
study of non-renewable demographic processes other than mortality, but restricting
ourselves to mortality, the sorts of questions a single-year-of-age life table permits
us to answer include:
1. What is the probability of surviving (or dying) between two nominated birthdays,
from one single-year age group to some older single-year age group, or from a
nominated birthday to an older single-year age group?
2. What is the average life expectancy of men, or women, at birth, or at any
subsequent birthday? Life expectancy at birth is a fairly familiar concept given
the frequency with which it is used to index general population health and
improvements or deterioration therein, but one can also, using a life table, calcu-
late the average remaining expectation of life given survival to any nominated
birthday.
3. What would be the age structure of a stationary population experiencing con-
stantly the mortality regime represented by the life table? A stationary population
is one that experiences a constant annual number of births, a constant mortality
regime (i.e., a constant set of age-specific death rates) that annually produces
exactly the same number of deaths as births, and is closed to migration. Such a
population has a constant size, a zero growth rate and a constant age structure. We
will return to stationary populations when addressing stable population theory. In
the meantime it is sufficient to note that what we call a life table population is a
stationary population. We will return to the idea of a life table population shortly.
The Single-Year-of-Age Life Table 133

Obtaining Age-Sex-Specific Death Rates to Construct a Life

Table

The age-sex-specific death rates from which we construct a single-year-of-age life

table are given by:

1 Mx D .1 Dx =1 Px / 1;000 (4.1)

Where 1 Mx denotes the death rate for males or females aged x last birthday (i.e.,
between exact ages x and x C 1); 1 Dx D male or female deaths at age x during
the year for which the calculation is being performed (i.e., deaths between exact
ages x and x C 1); 1 Px D the mid-year male or female population aged x last
birthday (i.e., the population aged between exact age x and exact age x C 1).
We noted in Chap. 1 that more exact denominator terms for demographic rates,
expressed in terms of person-years of exposure to risk, are possible. However, unless
extreme precision is required, and unless one’s data are of sufficient quality for the
necessary calculations to be performed, a denominator more sophisticated than the
one used in Eq. 4.1 is rarely justified.
What may be justified is an adjustment to the numerator term of Eq. 4.1, in
recognition of the fact that relatively small numbers of deaths of persons of a given
sex occur annually in some single-year age groups. This can introduce an element of
random fluctuation into single-year-of-age life tables, and as a means of smoothing
out such fluctuations a common practice is to use, not deaths at age x in the year
for which a life table is being constructed, but the annual average of deaths at age
x over a period of years centred on the year of interest. Thus in Australia and New
Zealand, for example, new life tables are prepared after each census using not just
deaths in the census year, but averages of deaths over a 3-year period straddling the
census year (i.e., for the census year and one year either side of it). In this scheme
the equation equivalent to Eq. 4.1 becomes:
nh  i o
D 1= .y1/
C 1 Dx.y/ C 1 D.yC1/ =1 P.y/ 1;000
1 Mx 3 1 Dx x x (4.2)

Where y denotes the year on which the life table is to be centred; other items have
the same meanings as in Eq. 4.1.
For actuarial purposes (i.e., when life tables are constructed for use by, for
example, life insurance companies) much more smoothing of data than Eq. 4.2
accomplishes typically is engaged in, with denominators as well as numerators
of age-sex-specific death rates likely to receive attention. The aim of actuaries
is to generate a smooth underlying set of age-specific mortality probabilities that
eliminates irregularities reflecting temporary or chance factors. Demographers,
however, are much less worried about these irregularities. Indeed, they are even
somewhat suspicious of the lengths to which actuaries go to eliminate them, because
sometimes their sources lie in demographically interesting phenomena.
134 4 Analysis of Mortality: The Life Table and Survival

Components of a Life Table

A life table begins with an arbitrary number of individuals at birth (i.e., at exact
age 0). This arbitrary number is known as the radix of the life table. In theory it
could have any value one might care to give it, but for the sake of convenience
a number which is some power of 10 is almost always used, and by far the most
common radix to encounter is 100,000.
Having selected a radix (and if you don’t choose 100,000 you are probably being
awkward), a single-year-of-age life table is constructed by applying probabilities of
dying between successive birthdays initially to the radix population, and thereafter
to survivors from that population until the point is reached at which there are no
longer any survivors. Thus we start by applying the probability of dying between
birth (exact age 0) and exact age 1 to the radix population and calculate survivors at
exact age 1. We then apply the probability of dying between exact age 1 and exact
age 2 to survivors at exact age 1 and obtain survivors at exact age 2. And so on
through each successive single year of age.
Life tables consist of a series of columns of numbers, each column giving
values of a particular life table function. These columns are related to one another
by a series of equations and can be derived from one another. Indeed, much of
the information in a life table can be said to be redundant; different columns
are equivalent, and merely say the same thing in different ways. While you will
encounter different numbers of columns in life tables, there is a basic core of six
life table functions that are found in most single-year-of-age life tables. They have
a standard, internationally recognized notation, and you should familiarize yourself
with that notation and with what the various life table functions measure. Some of
these notations were introduced in passing when we addressed the general concept
of ‘attrition’ in Chap. 3.
The six core life table functions are:
1 qx – the probability of dying between exact age x and exact age x C 1 (or if you
like, the proportion of people who reach exact age x alive who die before reaching
exact age x C 1).
lx – the number of members of the initial radix population surviving at exact age x
(the radix itself is denoted by l0 ). (Note that in the breastfeeding attrition table for
Sudan presented as Table 3.1 in Chap. 3 the equivalent function was denoted by
Nx , but in a mortality attrition table (i.e., a ‘true’ life table) lx is the conventional
notation.)
1 dx – the number of members of the initial radix population who die between exact
ages x and x C 1.
1 Lx – the number of person-years lived between exact ages x and x C 1 by members
of the initial radix population (or, if you like, by those of them who survive to
exact age x).
Tx – the number of person-years lived at all ages above exact age x by members of
the initial radix population (or, if you like, by those of them who survive to exact
age x).
The Single-Year-of-Age Life Table 135

eo x – the average number of years of life remaining beyond exact age x for each
member of the initial radix population who survives to exact age x.
Other life table functions that you may encounter include:
1 mx – the life table death rate between exact ages x and x C 1. (Note that this life
table death rate should be distinguished from 1 Mx , which is the observed age-
specific death rate between exact ages x and x C 1. The two are not necessarily
equal in some methods of life table construction, although methods described in
this chapter assume equality. Be aware, however, of the conceptual difference
between them. One (1 mx ) is a life table function; the other (1 Mx ) is obtained not
by manipulating other life table functions, but from empirical data – see Eqs. 4.1
and 4.2 above.)
x – the instantaneous force of mortality at exact age x.
1 px – the probability of surviving between exact age x and exact age x C 1 (or if you
like, the proportion of people who reach exact age x alive who survive to reach
exact age x C 1).
1 Sx – the survival ratio; the proportion of people who survive to the age group
bounded by exact ages x and x C 1 from the immediately younger single-year
age group, or from birth to the very youngest age group (that bounded by exact
ages 0 and 1).
Note in the above list of functions that some are prefixed by a subscript 1, and
others are not. Functions that do have a prefixed subscript 1 are quantities that
pertain to a 1-year interval of age – the interval stretching from exact age x to
exact age x C 1; functions that do not have a prefixed subscript 1 are quantities
that pertain either to exact age x (lx , x and eo x are in this category) or to all ages
beyond exact age x (as with Tx ). Having said this, however, it is not uncommon
for single-year-of-age life table functions which strictly speaking should be written
with a prefixed subscript 1 to in fact be written without it. Thus you may encounter
qx instead of 1 qx , dx instead of 1 dx , etc. Don’t be confused by this. If the prefixed
subscript is missing, you can assume it to be 1, and that you are dealing with a
life table function for a single-year-of-age life table. The prefixed subscript is often
dropped as an unnecessary piece of detail when its value is invariably 1 (i.e., when
it is clear that the life table in question is a single-year-of-age life table).

Generating a Single-Year-of-Age Life Table: The 1 qx Column

The normal starting point when constructing a single-year-of-age, or product-limit,

life table is the 1 qx column. This is the column we derive from raw age-specific data
on deaths and population size. There are a number of approaches to the task. Which
one we use depends to a degree on the nature of the input data available and the ages
for which calculations are to be performed. We will begin with a simple estimator
of 1 qx in a closed population (one whose size is not affected by migration) which is
136 4 Analysis of Mortality: The Life Table and Survival

Exact Age
I D G C
x+1

E H
1Px(y )

A F B J
x
Nx
Year y-1 Year y Year y+1

Fig. 4.1 Lexis diagram illustrating derivation of standard equation for obtaining values of 1 qx

applicable provided x ¤ 0 (the calculation of 1 q0 is invariably treated as a special

case for reasons that will be discussed shortly).
Consider a closed population at the middle of some reference year y. The
probability of dying at age x for the population aged x in the middle of year y
(which we’ll denote by 1 Px (y) ) is given by:

1 qx D 1 Dx =Nx (4.3)

Where 1 Dx D deaths at age x to the population aged x in the middle of the reference
year y (i.e., to the population 1 Px (y) – note that the ‘D’ is italicized to distinguish
it from the one that appears in Eqs. 4.1 and 4.2); Nx D the size of this cohort at
exact age x (i.e., at the beginning of the life cycle phase during which its members
were at risk of dying at age x).
The Lexis diagram presented as Fig. 4.1 will assist us to evaluate this equation,
because almost certainly we will not have directly available to us data to substitute
for 1 Dx and Nx in Eq. 4.3. The line BD in Fig. 4.1 represents the population aged x in
the middle of year y (i.e., 1 Px (y) ); it is the intersection of the vertical line representing
the population in the middle of year y and the horizontal band corresponding to age
group x (the age group bounded by exact ages x and x C 1). It is this population, or
cohort, for which we wish to obtain a value for 1 qx .
We next draw in the diagonal lines passing through AD and BC which define, or
enclose, the demographic experience of our cohort. The deaths at age x experienced
by this cohort (1 Dx ) are those that lie within the parallelogram ABCD; this
parallelogram is the intersection of the diagonal band representing the cohort, and
the horizontal band corresponding to age group x. Data on deaths available to us
will be tabulated by single years of age and year of occurrence; i.e., they will
correspond to squares of the Lexis grid. Clearly the parallelogram ABCD lies partly
within three different Lexis grid squares (it is the sum of the triangle AFE, the area
EFBHGD and the triangle GHC), so that applying principles learned in Chap. 3:
The Single-Year-of-Age Life Table 137

1 Dx D 1=8Dx.y1/ C 3=4Dx.y/ C 1=8D.yC1/

x (4.4)

(y) (y)
Where Dx D deaths at age x during year y (the equivalent of 1 Dx in Eq. 4.2 above).
In conformity with our definition of a probability from Chap. 1, the denominator
Nx of the righthand side of Eq. 4.3 is the size of the cohort of interest at exact age
x, which marks the beginning of the life cycle phase when a person is at risk of
dying aged x. Thus Nx is represented in Fig. 4.1 by the line AB; the intersection of
the diagonal band representing the cohort and the horizontal line representing exact
age x. How do we estimate Nx ? Well, again applying principles learned in Chap. 3,
the size of our cohort at AB is its size at BD (for which we are likely to have data)
plus the deaths represented by the shaded triangle ABD. This triangle captures the
demographic experience (in this case the mortality) of the cohort between exact age
x and the middle of year y, during which time mortality was reducing its size. But
we are working backwards from the middle of year y to exact age x (an earlier point
in the cohort’s life course), and so we have to add back in the deaths represented
by the shaded triangle. This triangle lies partly in each of two squares of the Lexis
grid (the triangle AFE plus the area EFBD), and so applying standard Lexis diagram
principles we have:

Nx D 1 P.y/ 1 .y1/ C 3= D.y/

x C =8Dx 8 x (4.5)

We have now obtained expressions for both elements in the righthand side of Eq. 4.3
in terms of the size of the population aged x in the middle of year y (1 Px (y) ) and
(y)
numbers of deaths at age x in specified calendar years (Dx ). These are quantities
for which data are typically available, and so we can proceed to calculate a value for
1 qx .
The equation expressing 1 qx as the ratio of the righthand sides of Eqs. 4.4 and
4.5 is, however, cumbersome, and a simplified form is often used. This rests on
the assumptions, first, that deaths in the triangle AFE in Fig. 4.1 are reasonably
approximated by deaths in the triangle EDI, and second, that deaths in the triangle
GHC are reasonably approximated by deaths in the triangle BJH. Under these
assumptions Eqs. 4.4 and 4.5 reduce to:

1 Dx D D.y/
x

and:

Nx D 1 P.y/ 1 .y/
x C =2Dx

The equation for 1 qx thus becomes:

 
1 qx D D.y/
x =
.y/
1 Px C 1=2Dx.y/
138 4 Analysis of Mortality: The Life Table and Survival

or, simplifying by dropping superfluous superscripts and subscripts:

 
qx D Dx = Px C 1=2Dx (4.6)

Where Dx D deaths at age x during the year for which qx is required; Px is the mid-
year population aged x in that year.
It is also possible to express Eq. 4.6 in terms of the observed age-sex-specific
death rate at age x, which we denoted in Eqs. 4.1 and 4.2 by 1 Mx . If we define this
death rate by 1 Mx D Mx D Dx / Px (i.e., don’t multiply by 1,000 as in Eq. 4.1), we
can rearrange to obtain Dx D Mx .Px . Substituting for Dx in Eq. 4.6 yields:
 
qx D Mx : Px = Px C 1=2Mx : Px

whence, cancelling:
 
qx D Mx = 1 C 1=2Mx (4.7)

or:

qx D 2Mx = .2 C Mx / (4.8)

Equation 4.7 or 4.8 (they are equivalent) is commonly used to obtain values of qx
from age-sex-specific death rates when x ¤ 0. Remember, though, that if you have
death rates per 1,000 persons at risk, you will need to divide them by 1,000 (i.e.,
shift the decimal point three places to the left) before substituting in Eq. 4.7 or 4.8.
Note also that Eq. 4.7 or 4.8 yields standard approximations of qx values given
values of Mx . They are approximations on two counts. First, we derived the equation
assuming that we were dealing with a closed population. In fact populations rarely
are closed, but departures from such a situation normally are not substantial enough
to be of consequence. Second, we introduced assumptions to allow us to base our
calculation on deaths in a single calendar year when, strictly speaking, we should
have based it on deaths over a 3-year period. Be aware that other, more complex
methods exist for obtaining qx values. They need not concern you now, save to note
(i) that they may be called for if precision is required in circumstances of appreciable
net migration and/or rapid annual change in mortality levels, and (ii) that their
use may be indicated if you are unable to reproduce qx values in a published life
table from raw deaths and population data. Another possible explanation, should
you strike the latter situation, is that Mx values were subjected to graduation, or
smoothing, before they were used in evaluating Eq. 4.7 or 4.8.
The Single-Year-of-Age Life Table 139

Obtaining 1 q0 : Separation Factors

But, what about 1 q0 , or q0 , which we noted needed to be treated as a special case?

This is so because the Lexis diagram principles that underpin the derivation of
Eqs. 4.6 and 4.7 (or 4.8) are invalid at age 0. They assume even distributions of
events (deaths in this case) within age-time squares of the Lexis grid. As we have
already noted, for infant mortality this assumption is patently false. Infant deaths
are very heavily concentrated, by exact age, at the lower end of the age interval
bounded by exact ages 0 and 1; that is, the great majority of infant deaths occur
very shortly after birth.
This means that, whereas for any cohort reaching exact age x (x ¤ 0) during a
reference year approximately half the deaths at age x usually occur in that year and
half in the following year, for persons born in the reference year (i.e., reaching exact
age 0) the majority of deaths at age 0 occur during the year of birth. Because of
this a satisfactory estimate of q0 is normally provided by the infant mortality rate
(IMR) (remember, it was pointed out in Chap. 1 that the IMR was an estimate of the
probability that a person born during the year in question died in infancy; all we are
saying now is that, usually, it is a pretty good estimate). Thus:

q0 D IMR D D0 =B (4.9)

Where D0 D deaths at age 0 in the reference year y; B D live births during year y.
An estimate of q0 may also be obtained where we have available data on deaths
at age 0 and the mid-year population aged 0, but not data on births. From the
former data, using Eq. 4.1 (without the multiplier of 1,000), we can calculate
the death rate at age 0, M0 , which we can use in conjunction with a quantity
called a separation factor to compute q0 . The demographic literature provides
various ways of calculating this separation factor, but a simple approximation, which
conveniently also makes use of M0 , can be obtained using the following formula
derived empirically by Keyfitz (1970):

f D 0:07 C 1:7M0 (4.10)

q0 is then given by:

q0 D M0 = .1 C .1  f/ M0 / (4.11)

Note, however, that in practice this equation can pose problems because of
differential under-enumeration at age 0 in registration and census data. Other
equations for q0 (not least Eq. 4.9) use only registration data (on births and infant
deaths). Because it is based on M0 , Eq. 4.11 uses both registration and census
data (for the numerator and denominator, respectively, of M0 ), and in developed
countries under-enumeration at age 0 tends to be more of a problem with census
data. Respondents filling out census forms sometimes overlook children aged 0,
140 4 Analysis of Mortality: The Life Table and Survival

perhaps because of difficulty understanding the concept of someone being aged ‘0’.
Young babies are typically thought of as being x days, weeks or months old, where
x is a non-zero number, not 0 years old.
Setting this issue to one side, the obvious question is, what is a separation
factor? Note at the outset, lest any misconception arise due to the concept being
dealt with at this juncture, that separation factors are not only used in deriving the
qx column of a life table, and are not only used when dealing with infant mortality
(i.e., mortality at age 0). Both points should become clear as this chapter proceeds.
For a 1-year birth cohort attaining exact age x, deaths at age x occur partly
during the year in which cohort members attain exact age x and partly during the
following year. In this single-year-of-age context, a separation factor measures the
proportion of deaths at age x to a cohort attaining exact age x during year y which
occur during year y C 1. If you compare the righthand sides of Eqs. 4.11 and 4.7
above you will note a similarity. The latter is a generalized version of the former in
which the separation factor f takes the value ½, or 0.5 (since 1  f D ½, then f D ½).
A separation factor is therefore implicit in the standard equation for qx when x ¤ 0.
That its value is 0.5 reflects the equation’s underlying assumption that deaths to the
cohort at age x are evenly distributed through the relevant age-time parallelogram
on a Lexis diagram (see Fig. 4.1); in this circumstance half the deaths at age x occur
in the year during which the cohort attains exact age x (between the middle of year
y  1 and the middle of year y in Fig. 4.1), and half occur during the following year
(between the middle of year y and the middle of year y C 1).
In the case of infant mortality we have already noted that deaths concentrate
heavily at the beginning of the 1-year interval of exposure to risk (i.e., towards
exact age 0 rather than towards exact age 1). The effect of this is, if you like, to
distort the cohort diagonals on the Lexis diagram after the fashion illustrated in
Fig. 4.2. Deaths at age 0 of members of the year y birth cohort lie within what we
might conveniently conceptualize as a curved band rather than a diagonal band, the

Exact Age
1

1-f

Year y Year y+1

Fig. 4.2 Lexis diagram illustrating the separation factor f used in generating certain life table
functions at age 0
The Single-Year-of-Age Life Table 141

majority (a proportion measured by 1  f) occurring during the year of birth and a

minority (a proportion measured by f) occurring during the following year. Thus, the
separation factor f that we might use in calculating q0 (depending on which equation
for q0 we use), and will use in calculating other life table functions pertaining to the
first year of life, always has a value less than 0.5. This is true regardless of which
of the variety of available approaches to obtaining a value for f is used.
Note, though, that this is a characteristic of this particular separation factor
(i.e., that used in calculating life table functions pertaining to the first year of life).
Separation factors are also used at older ages, especially in constructing abridged
life tables where, because we are dealing with age intervals wider than 1 year
(typically 5 years), the assumption that deaths are distributed evenly through an
interval by exact age can also be untenable. Many of these separation factors take on
values greater than 0.5 (but less than 1.0). In general, if the tendency is for mortality
to concentrate at the YOUNGER end of an age interval a separation factor for the
age interval will have a value LESS THAN 0.5; if the tendency is for mortality to
concentrate at the OLDER end of an age interval a separation factor for the age
interval will lie BETWEEN 0.5 AND 1.0. This statement should suggest to you
another general interpretation for separation factors. A separation factor measures
the proportion of deaths in an age interval that occur in the second half of the
age interval. If deaths are evenly distributed through the age interval by exact age
the separation factor has a value of 0.5, or one-half. If they are concentrated early in
the interval or late in the interval the separation factor is respectively lower than, or
greater than, this value (with the proviso it can never lie outside the range 0.0–1.0).
Having moved from a discussion of methods of estimating q0 to one of separation
factors, we will now briefly backtrack and note that we can use the separation factor
f to modify our original equation for q0 , Eq. 4.9. This sets q0 equal to the infant
mortality rate, the ratio of deaths at age 0 in the reference year y to the number of
live births in year y. But as we noted previously, and as Fig. 4.3 illustrates, some of
the deaths in the numerator of this ratio (a proportion of them equal to the separation
factor f) involved children born not in year y, but in year y  1 (the previous year)
(this is a slightly different interpretation of f to that illustrated in Fig. 4.2). Because
our deaths derive partly from live births in year y  1 and partly from live births in
year y, there is a case for respecifying Eq. 4.9 so that its denominator, instead of
being live births in year y, becomes a weighted average of live births in years y  1
and y, with the separation factor f providing the weight. Hence we have:
 
q0 D D0 = fBy1 C .1  f/ By (4.12)

When annual numbers of births are reasonably constant from year to year there is
no need to bother going to the trouble of using Eq. 4.12. The simpler Eq. 4.9 is
adequate. However, where the sizes of successive birth cohorts differ substantially
it is preferable to use Eq. 4.12.
Note in passing that where Eq. 4.12 or Eq. 4.9 is used in preference to Eq. 4.11
to obtain q0 , care should be taken that the value of m0 , the life table death rate at age
0, is compatible with q0 . It was noted above that problems can arise with Eq. 4.11
142 4 Analysis of Mortality: The Life Table and Survival

Exact Age

1-f

0
Year y-1 Year y

Fig. 4.3 Lexis diagram illustrating how a proportion equal to the separation factor f of deaths at
age 0 in a year involve children born the previous year

because it uses both census and vital registration data, in respect of which there
may be differential under-enumeration at age 0. If using Eq. 4.9 or Eq. 4.12, both
of which require only registration data, to obtain q0 you should not assume that a
value for m0 obtained from Eq. 4.1 or Eq. 4.2 (under the assumption that m0 D M0 ;
i.e., that the life table death rate at age 0 equals the observed death rate at age 0)
is satisfactory. If you do this m0 may be affected by inconsistency between the two
sources of data while q0 is not, and the two may be incompatible. If either Eq. 4.9
or Eq. 4.12 is used to obtain q0 , ensure that your value of m0 satisfies Eq. 4.11,
under the assumption that m0 D M0 . In other words, after rearranging Eq. 4.11 to
make M0 its subject, obtain m0 from:

m0 D M0 D q0 = Œ1  .1  f/ q0  (4.13)

Returning to the separation factor f, Keyfitz’s formula for f in terms of the death
rate at age 0, M0 (Eq. 4.10), can be used in conjunction with our earlier equation
linking q0 to M0 (Eq. 4.11) to yield a table of values of f corresponding to values of
the infant mortality rate q0 . This is presented as Table 4.1. It can be used to estimate
a value for f given the infant mortality rate using linear interpolation (which we will
deal with shortly). The introduction of Table 4.1 at this point probably seems rather
strange, because we are discussing ways of estimating q0 , some of which use the
separation factor f, and yet we are introducing a table that requires us to already
know q0 in order to obtain f. We would only use Table 4.1 to obtain a value of f for
use in generating life table functions other than q0 . The reason for introducing it
here is that it makes an interesting theoretical point.
The relationship between the infant mortality rate q0 and the separation factor f
in Table 4.1 is positive. As the infant mortality rate increases, so does the separation
factor; if q0 is low f is low, and if q0 is high f is high. Since f measures the proportion
of infant deaths that occur in the year following the year of birth, this relationship
The Single-Year-of-Age Life Table 143

Table 4.1 Values of the IMR D q0 f IMR D q0 f

Keyfitz separation factor f
corresponding to selected .01 .085 .11 .273
values of the infant mortality .02 .105 .12 .293
rate .03 .122 .13 .313
.04 .140 .14 .333
.05 .159 .15 .352
.06 .178 .16 .372
.07 .196 .17 .392
.08 .215 .18 .412
.09 .235 .19 .432
.10 .254 .20 .452
Source: P.F. McDonald, unpublished
notes

means that the lower the infant mortality rate, the greater is the concentration of
infant deaths early in the first year of life, and therefore in the year of birth. Thus,
in a country like Australia in which the infant mortality rate is very low, the few
infant deaths that do occur (4.66 per 1,000 live births during 2005–2010) mostly
occur soon after birth, and therefore overwhelmingly (well over 90 % according
to Table 4.1, since q0 D 0.00466 < 0.01) in the year of birth. But in countries with
much higher infant mortality rates (e.g., Chad with 131.17 infant deaths per 1,000
live births during 2005–2010 and Afghanistan with 135.95) the concentration of
infant deaths in the year of birth is less marked (more like 65–70 %).
Why is this? A measure of understanding can be achieved by thinking in terms of
causes of death. While specifying cause of death can be complicated in individual
cases, with respect to infant deaths a basic distinction can be made between
endogenous and exogenous causes of death. Endogenous causes of death are those
that arise from the genetic makeup of the child and the circumstances of prenatal
life and the birth process. Exogenous causes of death reflect the physical, social and
medical environment to which a baby is exposed after birth.
Infant deaths due to endogenous causes tend to be heavily concentrated in
the first few hours and days after birth and to be not readily preventable. They
are in many ways extensions of foetal mortality (miscarriage and stillbirth), with
conditions like congenital deformities and extreme prematurity often implicated.
Those due to exogenous causes are much less concentrated early in the first year of
life and are more readily preventable. As infant mortality falls from a high level in
a population, it is primarily deaths from exogenous causes occurring later in the first
year of life that are eliminated (by basic, often relatively inexpensive, public health
initiatives – improved sanitation, educating mothers to avoid, detect and treat child
illnesses, providing rudimentary post-natal medical services, etc.). Reducing deaths
from endogenous causes, by contrast, tends to depend on medical advances and
the availability of sophisticated medical technology, both of which are expensive
and not a high priority until deaths that can be prevented more cheaply have been
prevented. Thus, as infant mortality declines endogenous causes of death become
144 4 Analysis of Mortality: The Life Table and Survival

Table 4.2 Values of the IMR D q0 Male f Female f

separation factor f
corresponding to selected .006 .061 .070
values of the infant mortality .008 .067 .076
rate in Coale-Demeny ‘West’ .01 .072 .081
and United Nations ‘General’ .02 .100 .110
model life tables .03 .128 .138
.04 .156 .169
.05 .185 .198
.06 .214 .229
.07 .243 .260
.08 .272 .290
.09 .303 .321
.10 .329 .350
.11 to .20 .330 .350
Source: Derived from Coale et al.
(1983) and United Nations (1982)

proportionately more prominent. Infant deaths accordingly increasingly occur very

early in the first year of life, and the separation factor f (the proportion of infant
deaths occurring in the year following the year of birth) declines.
An alternative table of separation factors associated with different values of q0
is presented in Table 4.2. These derive from model life tables prepared by Coale
and Demeny (1966) (see also Coale et al. 1983) and the United Nations (1982) –
specifically the Coale-Demeny ‘West’ and United Nations ‘General’ model life
tables. Although based on life tables for actual populations, these ‘model’ life
tables relate to hypothetical populations. They generalize mortality patterns found
in families of actual life tables, their purpose being to aid the analysis of mortality
patterns in populations lacking adequate data to construct their own reliable life
tables. The Coale-Demeny model life tables are based on 192 of 326 life tables
examined (the other 134 were discarded for various reasons) for predominantly
Western populations. The ‘West’ variant (there are also ‘North’, ‘South’ and ‘East’
variants) represents the most general mortality pattern, and reflects populations
of Western Europe, North America, Australasia, Japan, Taiwan and White South
Africa. The United Nations model life tables, to be distinguished from an earlier
set (United Nations 1955), have a developing country focus, their ‘General’ variant
averaging four more regionally specific variants. Despite their respective developed
and developing country foci, values of f in relation to values of q0 underpin-
ning the Coale-Demeny ‘West’ and UN ‘General’ model life tables are virtually
identical.
Comparing Table 4.2 with Table 4.1 we can note three things. First, again values
of f increase as values of q0 do, but only until q0 D 0.10 (i.e., 100 infant deaths
per 1,000 live births). Thereafter they remain constant at higher infant mortality
levels. Second, values of f increase more rapidly with increasing q0 in Table 4.2
The Single-Year-of-Age Life Table 145

to q0 D 0.10, suggesting a more rapid process of concentration of infant deaths in

the year of birth as infant mortality declines below 100 deaths per 1,000 live births.
Finally, for any value of q0 , f is a little higher for females than for males, meaning
that, at any level of q0 , male infant deaths are slightly more concentrated in the year
of birth than female infant deaths are. If seeking to obtain a value of f corresponding
to a given value of q0 either Table 4.1 or Table 4.2 could be used. The latter has
the attraction of providing sex-specific estimates; the former differentiates at values
of q0 > 0.10 which the latter does not. The important thing is to make clear which
option has been taken; is f based on the Keyfitz equation or the Coale-Demeny
‘West’ and UN ‘General’ model life tables.

Generating a Single-Year-of-Age Life Table: The lx , 1 dx and 1 px

Columns

Our digression to explain and elaborate on the concept of a separation factor

over, we will refocus on the generation of a single-year-of-age life table. Once a
set of 1 qx values has been obtained (as in Table 4.3), three other closely related
functions can be generated quite simply. Recalling that l0 , the life table radix, is
conventionally arbitrarily set at 100,000, we can obtain the lx , 1 dx and 1 px columns
by applying the following equations, starting with x D 0 and increasing its value by
1 for each subsequent application. Note that, for simplicity, leading subscripts have
been dropped in these equations.

dx D lx qx (4.14)

lxC1 D lx  dx (4.15)

px D 1  qx (4.16)

Other equations may also be useful in manipulating these same life table functions:

qx D dx =lx (4.17)

lxC1 D lx px D lx .1  qx / (4.18)

dx D lx  lxC1 (4.19)
146 4 Analysis of Mortality: The Life Table and Survival

Table 4.3 Single-year-of-age life table for Australian males, 2009–2011

Age (x) 1 Mx 1 qx 1 px lx 1 dx 1 Lx Tx eo x
0 0.00498 0.00496 0.99504 100000 496 99555 7974745 79.75
1 0.00041 0.00040 0.99960 99504 41 99481 7875190 79.14
2 0.00019 0.00019 0.99981 99463 19 99453 7775709 78.18
3 0.00014 0.00015 0.99985 99444 14 99437 7676256 77.19
4 0.00013 0.00013 0.99987 99430 13 99423 7576819 76.20
5 0.00012 0.00012 0.99988 99417 12 99411 7477396 75.21
6 0.00010 0.00011 0.99989 99405 10 99400 7377985 74.22
7 0.00010 0.00010 0.99990 99395 10 99390 7278585 73.23
8 0.00010 0.00009 0.99991 99385 10 99380 7179195 72.24
9 0.00009 0.00009 0.99991 99375 9 99371 7079815 71.24
10 0.00009 0.00009 0.99991 99366 9 99362 6980444 70.25
11 0.00009 0.00009 0.99991 99357 9 99353 6881082 69.26
12 0.00010 0.00010 0.99990 99348 10 99343 6781729 68.26
13 0.00012 0.00012 0.99988 99338 12 99333 6682386 67.27
14 0.00016 0.00017 0.99983 99326 16 99319 6583053 66.28
15 0.00025 0.00025 0.99975 99310 25 99298 6483734 65.29
16 0.00036 0.00037 0.99963 99285 36 99268 6384436 64.30
17 0.00047 0.00047 0.99953 99249 47 99226 6285168 63.33
18 0.00054 0.00055 0.99945 99202 54 99176 6185942 62.36
19 0.00061 0.00060 0.99940 99148 60 99118 6086766 61.39
20 0.00063 0.00062 0.99938 99088 62 99057 5987648 60.43
21 0.00063 0.00063 0.99937 99026 62 98995 5888591 59.47
22 0.00063 0.00063 0.99937 98964 62 98933 5789596 58.50
23 0.00064 0.00064 0.99936 98902 63 98871 5690663 57.54
24 0.00065 0.00065 0.99935 98839 64 98807 5591792 56.57
25 0.00067 0.00067 0.99933 98775 66 98742 5492985 55.61
26 0.00070 0.00070 0.99930 98709 69 98675 5394243 54.65
27 0.00074 0.00074 0.99926 98640 73 98604 5295568 53.69
28 0.00077 0.00077 0.99923 98567 76 98529 5196964 52.73
29 0.00081 0.00081 0.99919 98491 80 98451 5098435 51.77
30 0.00085 0.00085 0.99915 98411 84 98369 4999984 50.81
31 0.00090 0.00089 0.99911 98327 88 98283 4901615 49.85
32 0.00094 0.00094 0.99906 98239 92 98193 4803332 48.89
33 0.00099 0.00098 0.99902 98147 97 98099 4705139 47.94
34 0.00103 0.00103 0.99897 98050 101 98000 4607040 46.99
35 0.00108 0.00108 0.99892 97949 106 97896 4509040 46.03
36 0.00114 0.00114 0.99886 97843 111 97788 4411144 45.08
37 0.00120 0.00120 0.99880 97732 117 97674 4313356 44.13
38 0.00126 0.00126 0.99874 97615 123 97554 4215682 43.19
39 0.00132 0.00133 0.99867 97492 129 97428 4118128 42.24
40 0.00141 0.00140 0.99860 97363 137 97295 4020700 41.30
41 0.00149 0.00149 0.99851 97226 145 97154 3923405 40.35
42 0.00159 0.00159 0.99841 97081 154 97005 3826251 39.41
(continued)
The Single-Year-of-Age Life Table 147

Table 4.3 (continued)

Age (x) 1 Mx 1 qx 1 px lx 1 dx 1 Lx Tx eo x
43 0.00170 0.00170 0.99830 96927 165 96845 3729246 38:47
44 0.00183 0.00183 0.99817 96762 177 96675 3632401 37:54
45 0.00197 0.00197 0.99803 96585 190 96491 3535726 36:61
46 0.00214 0.00213 0.99787 96395 206 96293 3439235 35:68
47 0.00231 0.00231 0.99769 96189 222 96080 3342942 34:75
48 0.00252 0.00252 0.99748 95967 242 95848 3246862 33:83
49 0.00274 0.00274 0.99726 95725 262 95596 3151014 32:92
50 0.00299 0.00299 0.99701 95463 285 95322 3055418 32:01
51 0.00326 0.00325 0.99675 95178 310 95025 2960096 31:10
52 0.00355 0.00354 0.99646 94868 336 94702 2865071 30:20
53 0.00385 0.00384 0.99616 94532 363 94353 2770369 29:31
54 0.00417 0.00416 0.99584 94169 392 93976 2676016 28:42
55 0.00452 0.00451 0.99549 93777 423 93568 2582040 27:53
56 0.00490 0.00488 0.99512 93354 456 93129 2488472 26:66
57 0.00530 0.00529 0.99471 92898 491 92656 2395343 25:78
58 0.00574 0.00573 0.99427 92407 529 92146 2302687 24:92
59 0.00626 0.00623 0.99377 91878 573 91595 2210541 24:06
60 0.00682 0.00680 0.99320 91305 621 90999 2118946 23:21
61 0.00746 0.00744 0.99256 90684 674 90351 2027947 22:36
62 0.00819 0.00815 0.99185 90010 734 89648 1937596 21:53
63 0.00901 0.00897 0.99103 89276 801 88881 1847948 20:70
64 0.00993 0.00988 0.99012 88475 874 88045 1759067 19:88
65 0.01096 0.01090 0.98910 87601 955 87131 1671022 19:08
66 0.01211 0.01204 0.98796 86646 1043 86132 1583891 18:28
67 0.01341 0.01332 0.98668 85603 1140 85041 1497759 17:50
68 0.01484 0.01473 0.98527 84463 1244 83850 1412718 16:73
69 0.01643 0.01629 0.98371 83219 1356 82551 1328868 15:97
70 0.01815 0.01800 0.98200 81863 1473 81137 1246317 15:22
71 0.02009 0.01989 0.98011 80390 1599 79601 1165180 14:49
72 0.02220 0.02195 0.97805 78791 1730 77937 1085579 13:78
73 0.02456 0.02426 0.97574 77061 1870 76138 1007642 13:08
74 0.02725 0.02689 0.97311 75191 2022 74194 931504 12:39
75 0.03034 0.02990 0.97010 73169 2187 72090 857310 11:72
76 0.03392 0.03335 0.96665 70982 2368 69814 785220 11:06
77 0.03803 0.03733 0.96267 68614 2561 67350 715406 10:43
78 0.04276 0.04188 0.95812 66053 2766 64688 648056 9:81
79 0.04819 0.04707 0.95293 63287 2979 61816 583368 9:22
80 0.05437 0.05296 0.94704 60308 3193 58729 521552 8:65
81 0.06143 0.05960 0.94040 57115 3405 55430 462823 8:10
82 0.06935 0.06705 0.93295 53710 3601 51925 407393 7:59
83 0.07828 0.07536 0.92464 50109 3776 48235 355468 7:09
84 0.08828 0.08456 0.91544 46333 3918 44384 307233 6:63
85 0.09935 0.09467 0.90533 42415 4015 40414 262849 6:20
(continued)
148 4 Analysis of Mortality: The Life Table and Survival

Table 4.3 (continued)

Age (x) 1 Mx 1 qx 1 px lx 1 dx 1 Lx Tx eo x
86 0.11145 0.10557 0.89443 38400 4054 36374 222435 5.79
87 0.12444 0.11713 0.88287 34346 4023 32329 186061 5.42
88 0.13815 0.12920 0.87080 30323 3917 28353 153732 5.07
89 0.15258 0.14166 0.85834 26406 3741 24518 125379 4.75
90 0.16798 0.15479 0.84521 22665 3509 20890 100861 4.45
91 0.18491 0.16908 0.83092 19156 3238 17511 79971 4.17
92 0.19999 0.18148 0.81852 15918 2889 14446 62460 3.92
93 0.22132 0.19888 0.80112 13029 2591 11707 48014 3.69
94 0.24273 0.21582 0.78418 10438 2253 9282 36307 3.48
95 0.26433 0.23257 0.76743 8185 1904 7203 27025 3.30
96 0.28050 0.24484 0.75516 6281 1538 5483 19822 3.16
97 0.29555 0.25623 0.74377 4743 1215 4111 14339 3.02
98 0.31104 0.26755 0.73245 3528 944 3035 10228 2.90
99 0.32563 0.27815 0.72185 2584 719 2208 7193 2.78
100C 0.37412 1.00000 0.00000 1865 1865 4985 4985 2.67

Generating a Single-Year-of-Age Life Table: The 1 Lx , Tx and eo x

Columns

We noted earlier that the 1 Lx (often abbreviated to Lx ) column of the life table gives
the number of person-years lived between exact ages x and x C 1 by members of the
life table radix population who survive to exact age x. Another way of putting this
is that it gives the number of person-years of exposure to the risk of dying between
exact ages x and x C 1 for the life table population. This statement should sound
familiar; it should sound to you like the specification of the denominator of some
demographic rate, and it is. We can note in passing that the life table death rate 1 mx
(often abbreviated to mx ) is given by:

mx D dx =Lx (4.20)

Another interpretation of the Lx column is that it gives the age structure of the life
table stationary population. We mentioned this concept of the ‘life table population’
and noted that it was a stationary population in passing earlier in this chapter. The
Lx column gives the life table population. A population experiencing 100,000 births
annually (the life table radix), which was subject to the probabilities of dying at
each age specified in the qx column (which would result in exactly 100,000 deaths
annually, since these qx values cause the life table radix population to completely
die out), and which was closed to migration would have constant numbers in each
single-year age group equal to the schedule of Lx values.
The Single-Year-of-Age Life Table 149

The general equation for calculating Lx is:

Lx D fx lx C .1  fx / lxC1 (4.21)

Where fx D a separation factor appropriate to the age group bounded by exact ages
x and x C 1.
However, when x ¤ 0 (i.e., for all values of x except x D 0), the value of fx
normally is ½, because it is reasonable to assume an even distribution of deaths
by exact age, otherwise known as linear survivorship, between exact ages x and
x C 1. Thus, except where x D 0, Eq. 4.21 reduces to:

Lx D 1=2 .lx C lxC1 / (4.22)

It does sometimes happen with published life tables that you are unable to reproduce
given values of Lx from given values of lx using Eq. 4.22. You are most likely
to encounter this situation at early childhood ages in life tables for high mortality
populations, when an argument similar to that on which the use of a separation
factor of less than ½ at age 0 is based may be advanced for immediately older ages
as well. Non-use of Eq. 4.22 at these ages is an indication that, within the relevant
single-year age interval, deaths concentrate at its lower end, and a separation factor
with a value less than ½ has been used (e.g., the values 0.43, 0.45, 0.47 and 0.49
which Chiang (1984) uses for x D 1, 2, 3 and 4 respectively).
We should also pause at this point over the introduction of the concept of
‘linear survivorship’. We noted earlier that separation factors are used at ages other
than 0, especially in the construction of abridged life tables, and the notion of
linear survivorship provides a useful way of specifying why, in general, separation
is an issue in life table construction. Separation factors are used in life table
construction to adjust for non-linear survivorship over an age interval. With
single-year-of-age life tables this problem occurs mainly at age 0, although as just
indicated, it may also occur at slightly older childhood ages. As we shall see, though,
it is a more general issue in the construction of abridged life tables because of the
wider age intervals over which survivorship is measured.
Resuming our discussion of Lx values, for x D 0 we have:

L0 D fl0 C .1  f/ l1 (4.23)

Where f, as previously, is the separation factor appropriate to mortality during the

first year following birth.
If data are available that classify infant deaths by both year of occurrence and
year of birth it is possible to calculate f directly (and absolutely accurately) as the
proportion of infant deaths of children born in year y that occurred in year y C 1.
However, it is usual to have data classified only by year of occurrence, whence
we must estimate f. In using Eq. 4.23 we might wish to obtain a value for f from
150 4 Analysis of Mortality: The Life Table and Survival

either of Tables 4.1 and 4.2. These tables specify f values for only a restricted
number of values of the infant mortality rate (or q0 ), but as indicated earlier, linear
interpolation can be used to estimate f values corresponding to other q0 values.
Linear interpolation assumes that relationships between f and q0 between successive
pairs of q0 and f values in Tables 4.1 and 4.2 are linear (i.e., follow a straight line).
To apply this principle, identify in either table the values of q0 nearest to, but lower
than, and nearest to, but higher than, your observed q0 . These define the IMR interval
within which interpolation will occur. Then evaluate the equation:

f D fl C fŒ.IMRo  IMRl / = .IMRu  IMRl / : .fu  fl /g (4.24)

Where IMRo D the observed value of q0 ; IMRl and IMRu D the values of q0 defining
the lower and upper limits, respectively, of the IMR interval in Table 4.1 or 4.2
within which IMRo lies; fl and fu are the values of f associated with IMRl and
IMRu , respectively, in Table 4.1 or 4.2.
An alternative approach to calculating Lx values uses the equation:

Lx D dx =Mx (4.25)

This formula can be used for all ages, including age 0 provided that q0 has been
obtained from M0 . Its use assumes equality between observed death rates Mx and
life table death rates mx . A rearrangement of Eq. 4.25 to make Mx its subject and a
comparison of the result with Eq. 4.20 should convince you of this.
We turn next to the function Tx , which gives total person-years of exposure of the
life table population to the risk of dying at all ages beyond exact age x. It is of little
use in itself, but vital to the calculation of average expectations of life remaining (eo x
values), measures which are very useful in comparing mortality levels in different
populations. It is also valuable in simplifying problems which require calculation
of the number of person-years of exposure to the risk of dying (or the number
of persons in the life table stationary population) between exact ages x and x C n
(where n > 1). Rather than having to sum Lx values over the range x to x C n  1 (a
tedious process if n is not small) we need only subtract TxCn from Tx . That is:

†iDx;xCn1 Li D Tx  TxCn

Where n gives the width in years of the age interval over which person-years of
exposure to risk is required.
The basic equation for Tx is:

Tx D †iDx;¨ Li

Where ¨ D the oldest age reached by anyone in the life table population.
Normal practice in generating Tx values, however, is to begin at the bottom of the
life table (i.e., at the oldest age reached by the life table population, where Tx D Lx )
The Single-Year-of-Age Life Table 151

and work from there to the top of the table (age 0) shuffling back and forth between
the Tx and Lx columns using the equation:

Tx D TxC1 C Lx (4.26)

And so to the eo x column. The concept of average expectation of life at birth (eo 0 )
probably has a reasonable level of public recognition in developed populations, but
the single-year-of-age life table also provides measures of average expectation of
life remaining at each subsequent birthday as well. These address the question,
‘Given that a man or woman has survived to exact age (birthday) x, how many more
years, on average, can he or she expect to live?’ The calculation is very simple;
total person-years lived by the life table population beyond exact age x (Tx ) is
apportioned among survivors at exact age x (lx ). Hence:

eo x D Tx =lx (4.27)

Always remember, though, that eo x is an average; some persons attaining age x live
longer, others not as long, and very few live for exactly eo x further years. Values
from this column are very useful in comparing the general health of populations (as
this is manifested by their survival capacities) and in monitoring changes in general
health levels over time. They are measures of longevity (i.e., length of life).

Generating a Single-Year-of-Age Life Table: Other Issues

In this subsection we deal with a number of miscellaneous issues that arise in

relation to the generation of single-year-of-age life tables. We focus on the bottom
of the life table (the oldest ages); taking migration into account in the calculation of
qx values; and the calculation of the separation factor f from detailed data on infant
deaths by age.

The Bottom of the Life Table

Particular difficulties can arise in generating single-year-of-age life tables at the

oldest ages. Because so few people live to very advanced ages (ages greater than
100 years, say), we can strike phenomena such as: (i) single-year age groups in
which there was a death (or more) during the year in question, but nobody at risk of
dying at mid-year, so that a value for Mx , and hence one for qx , cannot be calculated;
(ii) single-year age groups which yield qx values equal to or greater than 1 (implying
that all survivors to exact age x die at age x) when there are also deaths and persons
at risk at older ages; and (iii) single-year age groups in which there are neither deaths
nor persons at risk, when there are deaths and persons at risk at older ages.
Sometimes this sort of problem is dealt with by truncating the life table at some
single-year age group beyond which difficulties of the type described occur. In
152 4 Analysis of Mortality: The Life Table and Survival

other words, entries for the final few single-year age groups are simply omitted.
A second strategy is to group the final few ages, creating an age group xC, or ‘x and
over’. Under this strategy, for the last line of the life table qx D 1.0, since everyone
attaining exact age x dies at some subsequent age. Also:

LxC D Tx D lx =mxC D lx =MxC (4.28)

Where MxC is the observed death rate for the age group stretching from x to
whatever the final age featuring in our data happens to be, and is assumed to
equal mxC .
Other life table functions can then be obtained by conventional means.
A common approach of this type to dealing with the bottom of a life table
is to group all ages above exact age 85. This approach is frequently used in the
construction of abridged life tables in particular, but may also be used to complete
single-year-of-age tables, especially for high mortality populations in which rela-
tively few people survive into their late 80s and beyond. Clearly q85C D 1.0, and
provided l0 D 100,000 we have as an approximation:

L85C D T85 Š l85 :log10 l85 (4.29)

The approximation this equation provides for T85 (note that log10 l85 means ‘loga-
rithm of l85 to base 10’) is not, however, always a good one. If it is used to estimate
T85 in Table 4.3, for example, it yields an answer of 186,168 compared to the
value shown in the table of 262,849. Table 4.3 is a life table for a low mortality
population (expectation of life at birth 79.8 years), not the type of population
for which grouping above exact age 85 would ordinarily be undertaken. Better
approximations seem to be provided by Eq. 4.29 for higher mortality populations
of developing countries. Considering 1970, 1980 and 1990 female life tables for the
Philippines with life expectancies at birth of respectively 61.5, 65.1 and 67.4 years
constructed by Flieger and Cabigon (1994), for example, values of T85 (hence L85C )
of 61,711, 70,439 and 95,979 compare with estimates from Eq. 4.29 of 62,286,
70,473 and 81,480 respectively. Equation 4.29 relies on an observed empirical
‘regularity’ rather than a theoretical relationship. Other equations can be used,
including Eq. 4.28 with x set equal to 85. If all of this sounds disquietingly
imprecise, there is some consolation in the fact that the degree of inaccuracy that
might exist in life table functions at these advanced ages has little impact on
functions at younger ages, even though there is a feedback to those ages through
the Tx and eo x columns.

Taking Migration into Account

Equation 4.6 above indicated that values of qx could be obtained from:

The Single-Year-of-Age Life Table 153

 
qx D Dx = Px C 1=2Dx

Where Dx D deaths at age x during the year for which qx is required; Px is the mid-
year population aged x in that year.
This equation, however, assumes a closed population. It is often argued that this is
usually not an unreasonable assumption, and that adjustment for migration is rarely
called for. But where migration, and more especially net migration, is significant,
an adjustment for it probably should be made in calculating qx values. Denoting
out-migrants aged x during the year for which a life table is being constructed by
1 ox D ox , and in-migrants by 1 ix D ix , the necessary adjustment is achieved by using
in place of Eq. 4.6 the following alternative:
h i
qx D Dx = Px C 1=2 .Dx C ox  ix / (4.30)

Calculating Separation Factors from Detailed Data on Infant Deaths

by Age

Earlier discussion of the separation factor f claimed that it summarized the distribu-
tion of infant deaths between the year of birth and the following year, and relied on
approximations for f obtained from the Keyfitz equation or model life tables which
linked f to q0 (Tables 4.1 and 4.2). As illustrated in Fig. 4.3 an alternative, essentially
equivalent but strictly more correct conceptualization says that f apportions infant
deaths in a year between those involving children born in that year (1  f) and
in the PREVIOUS year (f). Moreover, where detailed data on infant deaths by age
such as are presented in Table 4.4 are available, it is possible to use Lexis diagram
principles to calculate the separation factor f empirically, thus obtaining a value
likely to be more accurate than other estimation techniques will yield.
‘Hang on’, you say. ‘Didn’t we decide that Lexis diagram principles were invalid
at age 0?’ Indeed we did, but that was in the context of treating the first year of life
as a single age interval. Deaths were so heavily concentrated early in the interval
that to assume they were evenly distributed through it by exact age was unrealistic.
Having data such as those in Table 4.4, however, allows us to slice the first year
of life into a series of much narrower age intervals – seven that are each only
1 day wide (covering the first week after birth), three more that are each 1 week
wide (covering the rest of the first month after birth), and eleven that are each
1 month wide (covering the remainder of the first year after birth). While Lexis
diagram assumptions are unreasonable when applied over the whole of the first year
following birth, they are much more reasonable when applied separately within
each of the ‘slices’ just defined. That is, it is reasonable to assume that deaths aged
less than 1 day were evenly spread through the first 24 h after birth, those aged 1 day
were evenly spread through the second 24 h, : : : , those aged 1 week were evenly
154 4 Analysis of Mortality: The Life Table and Survival

Table 4.4 Calculation of the separation factor f for Australian infant deaths during 2010 using
detailed data on deaths by age
Persons dying
Age at death Mean age at death (years) Deaths in 2010 Born 2009 Born 2010
<1 day 1
/730 D 0.001370 382 1 381
1 day 3
/730 D 0.004110 68 0 68
2 days 5
/730 D 0.006849 47 0 47
3 days 7
/730 D 0.009589 33 0 33
4 days 9
/730 D 0.012329 25 0 25
5 days 11
/730 D 0.015068 18 0 18
6 days 13
/730 D 0.017808 8 0 8
7–13 days 21
/730 D 0.028767 61 2 59
14–20 days 35
/730 D 0.047945 35 2 33
21–27 days 49
/730 D 0.067123 30 2 28
1 month 3
/24 D 0.125000 68 9 59
2 months 5
/24 D 0.208333 59 12 47
3 months 7
/24 D 0.291667 41 12 29
4 months 9
/24 D 0.375000 28 11 17
5 months 11
/24 D 0.458333 24 11 13
6 months 13
/24 D 0.541667 17 9 8
7 months 15
/24 D 0.625000 14 9 5
8 months 17
/24 D 0.708333 11 9 2
9 months 19
/24 D 0.791667 10 8 2
10 months 21
/24 D 0.875000 9 8 1
11 months 23
/24 D 0.958333 10 10 0
Total 998 115 883
Separation factor D 115/998 D 0.115

spread through days 8–14 following birth, : : : , those aged 1 month were evenly
spread through the second month following birth, etc.
Using these assumptions it is possible, with reasonable precision, to allocate
infant deaths in each age ‘slice’ between those of children born during the year in
which death occurred and those of children born the previous year. This is the type
of exercise performed in Table 4.4, and its basis is illustrated in Fig. 4.4. This Lexis
diagram shows a portion of one square of the conventional Lexis grid; a portion
covering (i) the months of January and February in the year (2010) for which we
wish to calculate the separation factor f, and (ii) demographic experience between
exact ages 0 and 2 months. Try to imagine the remainder of this grid square –
another ten monthly divisions along each axis (March, April, : : : , December along
the horizontal axis and exact ages 3 months, 4 months, : : : , 12 months up the
vertical axis).
The horizontal lines through Fig. 4.4 divide mortality at age 0 during 2010 into
the ‘slices’ just described. The bottom ‘slice’ corresponds to the top row of Table 4.4
(deaths aged <1 day), the next ‘slice’ corresponds to the next row (deaths aged
The Single-Year-of-Age Life Table 155

Exact Age
2 months

4 weeks

3 weeks

2 weeks

1 week

0
January 2010 February 2010

Fig. 4.4 Lexis diagram illustrating calculation of the separation factor f for Australian infant
deaths during 2010 using detailed data on deaths by age

1 day), and so on. The diagonal line through the diagram is the life line that separates
the demographic (in this case mortality) experience of the 2010 birth cohort (to the
right of the line) from that of the 2009 birth cohort (to the left of it).
Consider the bottom ‘slice’ in Fig. 4.4; that representing deaths aged <1 day.
Most of it lies to the right of the diagonal dividing the 2009 birth cohort from the
2010 birth cohort; i.e., most deaths aged <1 day during 2010 involved babies also
born in 2010. All that lies to the left of the diagonal is a small triangle. This triangle
represents the deaths aged <1 day during 2010 that involved babies born during
2009. What fraction of the area of the entire ‘slice’ is the area of this triangle?
Well, the slice is 1 day wide on the vertical scale, and the diagonal forming the
hypotenuse of the triangle runs at 45ı , implying that the triangle is half of a square
that also extends 1 day along the horizontal axis. The ‘slice’ comprises 365 such
squares – one for each day of 2010. Half of 1 /365 of the ‘slice’ is 1 /730 of the ‘slice’,
and this fraction is the first entry in the ‘Mean age at death’ column of Table 4.4.
We estimate that 1 /730 of the deaths aged <1 day during 2010 involved children born
during 2009. As there were 382 such deaths, calculating to the nearest whole death
(since dealing with fractions of deaths makes no intuitive sense) we conclude that 1
156 4 Analysis of Mortality: The Life Table and Survival

of them occurred to a member of the 2009 birth cohort and the other 381 to members
of the 2010 birth cohort (top row of Table 4.4).
Moving to the second ‘slice’ – that corresponding to deaths aged 1 day – the
area to the left of the diagonal consists of a complete ‘1 day by 1 day’ square plus a
triangle half the area of another such square. That is, 1 /365 C1 /730 D3 /730 of the deaths
aged 1 day during 2010 were of children born during 2009. With only 68 deaths in
total to be apportioned between the two birth cohorts, to the nearest whole death
NONE of them involved a child born during 2009; all involved children born during
2010 (second row of Table 4.4).
And so on. When we get to the wider ‘slice’ corresponding to deaths aged 1 week
(but less than 2 weeks) the entire slice consists of 365 rectangles 1 day wide by
1 week high placed side by side, and the area to the left of the diagonal consists of
the first 7 of these plus half of the next 7; i.e., 7 /365 C1 /2 (7 /365 ) D14 /730 C7 /730 D21 /730
of the deaths aged 1 week, or 7–13 days, were deaths of children born during 2009
(row 8 of Table 4.3). When we have completed apportioning deaths in each age
‘slice’ during 2010 between the 2009 and 2010 birth cohorts in this fashion, we
add the deaths allocated to the earlier (2009) cohort and express the answer as a
proportion of total infant deaths during 2010.
This is our separation factor. Its value (Table 4.4) is somewhat different from that
we would have calculated using the Keyfitz equation (Eq. 4.10) (0.115 compared to
0.076). This emphasizes that estimates provided by approaches such as Keyfitz’s are
just that – estimates, and not always very good ones. Fortunately, some imprecision
in the value of the separation factor f usually is of limited consequence for the value
of q0 (Shryock, Siegel, and Associates 1973), and hence for the values of other
life table functions. This is just as well, because it is relatively rare to have data as
detailed as those in Table 4.4, whence we are forced to use approximations for f such
as those presented in Tables 4.1 and 4.2 (the literature contains a range of approaches
to calculating separation factors, and to fully understand how any particular life table
was constructed it is necessary to check how separation was done). Imprecision in f
does, for example, tend to be of less consequence for q0 than using Eq. 4.11 rather
than Eq. 4.12 as the basic calculating equation; i.e., it tends to be less of a problem
than the former equation’s reliance on a mixture of registration and census data,
between which there may be different levels of under-enumeration at age 0.
One final point about Table 4.4 that may require some clarification is the heading
‘Mean age at death’ used for the column in which the fractions of deaths in each age
‘slice’ to be allocated to births during the previous year are given. It happens that
these fractions equal the exact ages, in fractions of a year, that form the mid-points of
the age intervals defining the various age ‘slices’. And since we are assuming even
distributions of deaths by exact age within each ‘slice’, these mid-points correspond
to mean ages at death for those dying in each age interval. For example, children
dying aged <1 day (the bottom ‘slice’ in Fig. 4.4) are assumed to live on average
for half a day, which is 1 /730 of a year (top row of Table 4.4); those dying aged
1 day are assumed to live for one and a half days on average, or 3 /730 of a year
(second row of Table 4.4), etc. Thus we have an alternative interpretation of these
fractions to the one based on areas on the Lexis diagram. As a general rule, under
The Single-Year-of-Age Life Table 157

the assumptions of the Lexis diagram, the mean age in years of infants dying in
any age interval gives the proportion of deaths in that interval during a year which
involved children born during the previous year. Of course, the qualifier ‘under the
assumptions of the Lexis diagram’ in this statement is all-important. The value of the
principle just enunciated depends on age intervals being used which are not so broad
as to render these assumptions unreasonable. For example, it is clear from Table 4.4
that there is nothing like an even distribution of infant deaths by exact age at death
through the first week following birth, so that to treat the first week as a single age
interval and apply the ‘general rule’ above would be to invoke an assumption that
plainly is invalid, and therefore to potentially introduce significant error into one’s
separation factor. Note also the word ‘infants’ in the italicized statement above. It
says that we are talking only about deaths between exact ages 0 and 1.

Recipe for Constructing a Single-Year-of-Age Life Table

Having now completed our discussion of the single-year-of-age life table we can
set out the steps to be followed in constructing such a life table in recipe form.
Assuming, for a given sex in a given year, that we have single-year-of-age data on
deaths (1 Dx ) and population at mid-year (1 Px ), and perhaps data on live births (B)
during the year in question and the previous year, we can proceed as follows:
1. Calculate values of Mx using either Eq. 4.1 or Eq. 4.2.
2. Calculate the separation factor for age 0, f, either directly from a detailed
distribution of infant deaths by age in the manner illustrated by Table 4.4 or,
if such data are not available, by using an empirical formula such as Keyfitz’s
equation (Eq. 4.10).
3. Calculate q0 either using Eq. 4.9 or Eq. 4.12, in which case Eq. 4.13 should be
used to obtain m0 , or using Eq. 4.11, in which case m0 D M0 .
4. Calculate values of qx for x > 0 using Eq. 4.7, and set mx D Mx for each value
of x.
5. Set l0 D 100,000 and calculate values of lx for x > 0 using Eq. 4.18.
6. Use Eq. 4.19 to calculate values of dx .
7. Calculate values of Lx using Eq. 4.23 for x D 0 and Eq. 4.22 for x > 0.
Alternatively Eq. 4.20 may be rearranged to give Lx D dx / mx for all values
of x.
8. Use Eq. 4.28 to calculate LxC and Tx for the last row of the life table.
9. Derive the remainder of the Tx -column using Eq. 4.26, working from the last
row of the life table upwards.
10. Calculate values of eo x using Eq. 4.27.
158 4 Analysis of Mortality: The Life Table and Survival

Abridged Life Tables

It is not always possible, or desirable, to construct single-year-of-age life tables.

Available data may not be of sufficient quality to warrant, or indeed even to permit,
their construction. Single-year-of-age data may simply not be available. Or they may
be so contaminated by age misstatement and/or other defects as to be unusable at a
single-year-of-age level of detail. These practical problems create one incentive for
constructing abridged life tables. A second important incentive is that single-year-
of-age tables are far too detailed for most demographic purposes. Other than at age
0 demographers rarely make comparisons of mortality, whether across populations
or through time, by single years of age. Similarly, they frequently use abridged life
tables in population projection exercises.
As was indicated earlier, abridged life tables summarize a population’s prospects
of survival over fewer, longer age intervals – typically 5-year intervals, with the
exception that age 0 is split from ages 1–4 and the oldest ages are grouped into
a broad open interval, frequently 85 years and over (85C). Other schemes of
abridgement are possible (for example, treating ages 0, 1 and 2–4 separately, 10-
year intervals between exact ages 5 and 85, or specifying the terminal open interval
differently), but the one just described is the most common. An abridged life table of
this type is almost as accurate as a complete life table, provides all the information
needed for most demographic analysis purposes, and requires less data and, if the
job is being done manually, less ‘donkey’ work to construct.
In introducing the functions (columns) of a single-year-of-age life table a
distinction was drawn between those that pertained to exact age x (lx , eo x and x )
or all ages beyond exact age x (Tx ), and therefore didn’t take a prefixed subscript
of 1, and those that pertained to the 1-year age interval x (the interval from exact
age x to exact age x C 1), which strictly speaking did take a prefixed subscript of 1
(1 qx , 1 dx , 1 Lx , 1 mx and 1 px ). The former functions are sometimes referred to as the
stock functions of the life table, and the latter as the flow functions. In theory stock
functions in an abridged life table remain the same as in the equivalent single-year-
of-age life table, and can be directly transferred from the latter to the former. All that
happens is that only values for selected x’s are transferred. However, if you compare
an abridged life table with an equivalent single-year-of-age life table you will often
find departures from this theoretical equality. If you do, the major reason is likely
to be the manner in which separation has been undertaken. As has already been
intimated, separation becomes an issue at all ages in the construction of abridged
life tables, and the ‘departures’ referred to typically arise from errors introduced by
the use of a separation factor of one-half beyond exact age 5 in generating abridged
life tables. This is a frequently adopted approach which assumes linear survivorship
over 5-year age intervals, but its straightforwardness does sacrifice some precision.
Flow functions in an abridged life table (n qx , n dx , n Lx , n mx and n px ) have no
counterpart in a corresponding complete life table, except where n D 1 (generally
the case only in the first row of an abridged life table). Other than when n D 1 they
cannot be transferred directly from a complete to an abridged life table. The notation
Abridged Life Tables 159

Table 4.5 Abridged life table for Australian males, 2009–2011

Age (x) n n Mx n qx lx n dx n Lx Tx eo x
0 1 0.00498 0.00496 100000 496 99561 8014468 80:14
1 4 0.00021 0.00082 99504 82 397722 7914907 79:54
5 5 0.00012 0.00058 99422 58 496965 7517185 75:61
10 5 0.00011 0.00056 99364 56 496680 7020220 70:65
15 5 0.00046 0.00230 99308 229 495968 6523540 65:69
20 5 0.00062 0.00309 99079 306 494630 6027572 60:84
25 5 0.00074 0.00367 98773 362 492960 5532942 56:02
30 5 0.00094 0.00469 98411 461 490903 5039982 51:21
35 5 0.00118 0.00587 97950 575 488313 4549079 46:44
40 5 0.00158 0.00787 97375 767 484958 4060766 41:70
45 5 0.00229 0.01140 96608 1101 480288 3575808 37:01
50 5 0.00344 0.01709 95507 1632 473455 3095520 32:41
55 5 0.00520 0.02568 93875 2411 463348 2622065 27:93
60 5 0.00800 0.03921 91464 3586 448355 2158717 23:60
65 5 0.01303 0.06313 87878 5548 425520 1710362 19:46
70 5 0.02179 0.10331 82330 8506 390385 1284842 15:61
75 5 0.03776 0.17250 73824 12735 337283 894457 12:12
80 5 0.06758 0.28907 61089 17659 261298 557174 9:12
85 ¨ 0.14678 1.00000 43430 43430 295876 295876 6:81

used for these flow functions has been explained before, but to reiterate, ‘x’ gives
the exact age marking the lower limit of the age interval to which the value of a
function pertains, and ‘n’ gives the width of the interval in years (so that its upper
limit is exact age x C n). Table 4.5 gives values of x and n pertaining to each row of
the life table in its first two columns (often only x values are given and the user is
left to deduce the n values). The final n value shown is ‘¨‘, which means an interval
width such that, added to x, it defines the youngest exact age in completed years
beyond which no-one lives.
One interesting difference concerning flow functions in an abridged life table
is that values in the n mx (or n Mx ) and n qx columns are substantially different
(Table 4.5), whereas in the single-year-of-age life table values of 1 mx and 1 qx (mx
and qx ) were similar and often identical (Table 4.3). Why the change? Values of
1 mx and n mx are annual life table death rates, respectively for age groups bounded
by exact ages x and x C 1 and by exact ages x and x C n. Values of 1 qx and n qx are
probabilities of dying between exact ages x and x C 1, and exact ages x and x C n,
respectively. In the case of 1 mx and 1 qx we are comparing an annual death rate
for a single-year age group with a probability of dying over the equivalent 1-year
phase of the life cycle, so near-equality is not surprising. But in the case of n mx and
n qx we are comparing an annual death rate (albeit for an n-year age group) with
a probability of dying over an n-year period. Any individual takes n years to live
through the relevant age group, and is exposed afresh to the death rate n mx in each
one of those years, but n qx is a single measure of a person’s prospects of dying over
160 4 Analysis of Mortality: The Life Table and Survival

the entire n-year phase of the life cycle. Thus, n qx is expected to be of the order of
n-times larger than n mx ; in effect it is equivalent to experiencing the death rate n mx
n times – once for each year it takes to age from exact age x to exact age x C n.
Essentially, then, values of n qx in an abridged life table are larger than values
of 1 qx in a single-year-of-age life table because they give probabilities of dying
over longer periods of the life cycle. Similarly values of n dx and n Lx are larger than
values of 1 dx and 1 Lx (giving life table deaths and person-years of exposure to the
risk of dying, respectively, over longer periods of the life cycle). Values of n px are,
however, smaller than values of 1 px because the probability of surviving decreases
as the length of the age interval for which it is calculated increases.
It was noted above that abridged life tables usually terminate with a broad open
age interval. This terminal open interval may begin at exact age 75, 80, 85, 90 or 95,
and on occasion even at a younger age than 75. Methods for dealing with the bottom
of a life table have already been discussed, and grouping the final ages was noted
as being one option. When choosing this option the main challenge was to obtain a
value for LxC . We can do this using Eq. 4.28 or, if grouping beyond exact age 85,
Eq. 4.29. The latter can only be used when x D 85 (i.e., you cannot, for example,
substitute ‘80’ for ‘85’ and ‘80C’ for ‘85C’ in Eq. 4.29), and Eq. 4.28 is in any case
preferable to Eq. 4.29. Note that, given that for a terminal open age group lx D dxC ,
Eq. 4.28 can be extended as follows:

LxC D lx =MxC D dxC =MxC

If you compare Table 4.5, the abridged life table for Australian males for 2009–
2011, with Table 4.3, the equivalent single-year-of-age life table, you will note
that the average expectations of life at birth (eo 0 ) are slightly different (80.14 years
compared to 79.75 years). This is a difference of 0.39 years. Now compare values
of eo 85 (the expectation of life remaining for survivors at exact age 85). They are
respectively 6.81 and 6.20 years, a difference in the same direction of 0.61 years.
This suggests that the higher expectation of life at birth in the abridged life table
is largely a product of inaccuracy in the estimate of L85C D T85 compared to what
single-year-of-age data yield. This reverberates up the life table as n Lx values are
added for successively younger ages to generate the Tx column used in calculating
eo x values, with other bits and pieces of error arising from the assumption of linear
survivorship over 5-year age intervals partly offsetting this inaccuracy along the
way. The terminal age group in Table 4.5 has been set at 85C because that is the
most common terminal age group to encounter in abridged life tables. However,
because Australian males in 2009–2011 had low mortality, and therefore a high
proportion (over 43 %) surviving to exact age 85, it would have made sense to
have used, say, 95C as the terminal age group. Ideally the terminal age group in
an abridged life table should commence at an exact age to which only a small
proportion of the population survives, and Table 4.3 shows only a little over 8 %
of Australian males in 2009–2011 surviving to exact age 95.
Abridged Life Tables 161

Calculating Abridged Life Tables from Single-Year-of-Age Life

Tables

Should the need and the opportunity arise, we can readily collapse a complete
life table into an abridged life table. We may wish to do this either to present the
essential features of a population’s mortality regime more succinctly or to facilitate
comparison with some second population for which only an abridged life table is
available.
Functions lx , Tx and eo x do not change. It is simply a matter of extracting relevant
values from the single-year-of-age life table once the age intervals to be used in the
abridged life table have been selected. Other functions can then be obtained from
the following equations:

n qx D 1  .lxCn =lx / (4.31)

n px D 1  n qx D lxCn =lx (4.32)

n dx D dx C dxC1 C    C dxCn1 D lx  lxCn D lx :n qx (4.33)

n Lx D Lx C LxC1 C    C LxCn1 (4.34)

Generating an Abridged Life Table: Estimating Values of n qx

More commonly demographers are interested not in collapsing complete life tables
but in generating abridged life tables from scratch. As it was with the generation
of single-year-of-age life tables, the key to this exercise is the calculation of
probabilities of dying (n qx values) from observed age-sex-specific death rates (n Mx
values). By direct analogy with our earlier Eq. 4.1 we have:

n Mx D .n Dx =n Px / 1;000 (4.35)

Where n Mx denotes the death rate for males or females aged between exact ages x
and x C n (i.e., aged x to x C n  1 last birthday); n Dx D male or female deaths
between exact ages x and x C n during the year for which the calculation is being
performed (i.e., deaths at ages x to x C n  1); n Px D the mid-year male or female
population aged between exact ages x and x C n (i.e., the mid-year population
aged x to x C n  1 last birthday).
We could also, if we wished, modify Eq. 4.35 in the same manner as we earlier
modified Eq. 4.1 to obtain Eq. 4.2; i.e., we could use as our numerator the average
annual number of deaths between exact ages x and x C n over the 3-year period
162 4 Analysis of Mortality: The Life Table and Survival

straddling the reference year for which we require n Mx values. There is less reason to
do this when constructing abridged life tables because random volatility in mortality
data is reduced by the use of wider age intervals. That said, this is the approach taken
in generating Table 4.5, simply because it was the approach taken in generating the
equivalent Table 4.3. The indication that both life tables pertain to 2009–2011 says
that the age-sex-specific death rates on which they are based relate average deaths
over the 3-year period 2009–2011 to mid-year risk populations for the central year,
2010.
Equation 4.35 gives age-sex-specific death rates per 1,000 mid-year population
at risk, and as was the case when using results from Eqs. 4.1 and 4.2 to obtain
qx values for the single-year-of-age life table, we ignore the multiplier of 1,000 in
specifying n Mx values to be used in obtaining n qx values for an abridged life table.
Thus, n Mx Dn Dx /n Px .
Recall Eqs. 4.7 and 4.8 earlier in the chapter, which gave values of qx for
the single-year-of-age life table from values of Mx . Analogous equations for an
abridged life table are:
 
n qx D .n:n Mx / = 1 C n=2:n Mx (4.36)

and:

n qx D .2n:n Mx / = .2 C n:n Mx / (4.37)

As with Eqs. 4.7 and 4.8, these are alternative forms of the same equation. They
assume linear survivorship between exact ages x and x C n, and hence assume a
separation factor of one-half. As previously noted, separation is an issue over all
age intervals in an abridged life table. However, a frequently adopted approach (and
the one taken in constructing Table 4.5) is to assume linear survivorship over 5-year
age intervals between exact age 5 and the exact age marking the lower bound of
the terminal open age interval (exact age 85 in Table 4.5). Error introduced by this
approach is tolerable for most demographic purposes, and the appropriate version
of Eq. 4.36 becomes:
 
5 qx D .5:5 Mx / = 1 C 5=2:5 Mx (4.38)

Use of this equation to obtain n qx values beyond exact age 5 (remember that
q85C D 1.0) still leaves us needing to find 1 q0 and 4 q1 . The former can be calculated
exactly as it was for the single-year-of-age life table (see the alternative Eqs. 4.9,
4.11 and 4.12 and associated discussion). In the case of 4 q1 an assumption of linear
survivorship typically is unsound, and a separation factor of less than one-half is in
order. The situation over this age interval is directly analogous to that during the
first year of life. Just as infant deaths are concentrated at the younger end of the age
interval bounded by exact ages 0 and 1, so, too, deaths between exact ages 1 and 5
tend to be concentrated at the younger end of that 4-year age interval.
Abridged Life Tables 163

Fig. 4.5 Schematic

illustration of how the
survivorship function lx
approximates linearity over
1-year age intervals between
exact ages 1 and 5, but does
not do so over the combined
4-year interval or between
exact ages 0 and 1

lx

0 1 2 3 4 5
Exact Age

Figure 4.5 helps to explain the situation. It shows a schematic plot of the survivor-
ship function lx against age (the dashed curve), and straight line approximations
of that function over successive one-year age intervals (the solid lines) and over
the interval from exact age 1 to exact age 5 (the dot-dashed line). These straight
lines represent linear survivorship over the age intervals to which they relate. The
vertical axis is broken to signify that only the upper end of the full range of lx
values is portrayed. The point at which the survivorship curve meets the vertical
axis corresponds with lx D l0 D 100,000.
You can see that between exact ages 0 and 1 the solid straight line is a poor
approximation of the dashed curve. The survivorship function falls much more
steeply early in this 1-year age interval than it does later in the interval. This is the
situation invariably encountered with infant mortality, and the reason we introduced
the concept of separation. An assumption of linear survivorship over this age interval
amounts to an assumption that, on average, those who died in the interval did so at
its mid-point or, to put it another way, survived for half a year. Clearly their average
survival period was appreciably less than half a year, and we needed to adjust for
this reality by using the separation factor f in calculating certain life table functions.
If we next focus on the age interval bounded by exact ages 1 and 5 we see
that, when broken into 1-year segments, the series of four solid straight lines that
traverse the interval in Fig. 4.5 collectively are quite a good approximation of the
dashed survivorship curve. In other words, treating single years of age within the
interval one by one (as the single-year-of-age life table does), an assumption of
linear survivorship is not unreasonable. However, when we combine the four one-
year age intervals into one four-year interval (as the abridged life table does) the
straight line joining the points of intersection of the survivorship curve with exact
164 4 Analysis of Mortality: The Life Table and Survival

ages 1 and 5 (the dot-dashed line) is again a poor approximation of the survivorship
curve. We have the same sort of situation we have discussed at length with respect
to infant mortality – deaths are concentrated at the younger end of the age interval,
with the result that the survivorship function falls more steeply at that end of the
interval. To assume linear survivorship through the interval (i.e., that those who die
in the interval on average survive for half its width – 2 years) is not reasonable, and
a separation factor with a value less than one-half is called for (because there are
fewer deaths in the second half of the interval than in the first half).
It is helpful to pause at this juncture and re-conceptualize life table separation
factors. When dealing with the separation factor used at age 0 we were able to
conceive it as either (i) the proportion of deaths at age 0 in a birth cohort that occur
in the year following the year of birth, or (ii) the proportion of deaths at age 0 in a
calendar year that occur to children born the previous year. Because in an abridged
life table we are no longer dealing with single-year age groups beyond exact age
1 an alternative conceptualization of the separation factor (equally applicable in a
single-year age group context) is more readily understood. The separation factor
may be thought of as the average proportion of the relevant age interval survived
by persons who die in that age interval. Chiang (1984: 141) refers to it as the
‘fraction of the last age interval of life’.
Bearing this conceptualization in mind, the equation for 4 q1 may be specified as:
   
4 q1 D .4:4 M1 / = 1 C 4: 1  f :4 M1 (4.39)

Where f* is a separation factor appropriate to mortality between exact ages 1 and 5.

This separation factor can be computed for any population for which a complete
life table is available using a method outlined by Chiang (1984); the method makes
use of values of qx , px and separation factors for the four constituent single-year age
groups that make up age group 1–4. The equation is:

f D .q1 f1 C p1 q2 .1Cf2 / C p1 p2 q3 .2Cf3 / C p1 p2 p3 q4 .3Cf4 // =4 .1  p1 p2 p3 p4 /

(4.40)

Where qx and px are standard single-year-of-age life table functions for x D 1, 2, 3

and 4; fx values are separation factors for the same four single-year age groups.
To date we have assumed them to all have a value of 0.50 (i.e., we have assumed
linear survivorship across those age intervals), but Chiang (1984) suggests values
of f1 D 0.43, f2 D 0.45, f3 D 0.47 and f4 D 0.49 are more accurate. These values
reflect slight concentrations of mortality at the lower ends of the four single-year
age groups.
Frequently, however, the desire is to use Eq. 4.39 in the construction of an
abridged life table for a population for which no complete life table is available,
so that Eq. 4.40 cannot be evaluated. In that circumstance it is common to use
an estimate of f*. Fortunately values of f* generally lie in a fairly narrow range.
As a guide to selecting an appropriate value Table 4.6 shows values of f* from
Abridged Life Tables 165

Table 4.6 Values of f* from Coale-Demeny and United Nations model life tables by sex and
expectation of life at birth
Expectation of life at birth
Model 50 55 60 65 70 75 80 85
Male
C-D East 0.33 0.33 0.32 0.34 0.36 0.38 0.38 0.38
C-D North 0.39 0.39 0.41 0.43 0.44 0.45 0.46 0.46
C-D South 0.31 0.31 0.33 0.35 0.36 0.38 0.39 0.40
C-D West 0.34 0.34 0.36 0.38 0.39 0.40 0.41 0.41
UN Chilean 0.34 0.34 0.34 0.36 0.38 0.39 0.40 0.41
UN Far Eastern 0.34 0.36 0.37 0.39 0.40 0.41 0.41 0.41
UN General 0.34 0.34 0.35 0.37 0.39 0.40 0.40 0.41
UN Latin American 0.34 0.34 0.35 0.37 0.38 0.39 0.40 0.41
UN South Asian 0.34 0.34 0.34 0.36 0.38 0.39 0.40 0.41
Female
C-D East 0.33 0.33 0.31 0.32 0.33 0.34 0.35 0.35
C-D North 0.39 0.39 0.40 0.41 0.42 0.43 0.43 0.43
C-D South 0.31 0.31 0.33 0.34 0.35 0.35 0.36 0.37
C-D West 0.34 0.34 0.35 0.36 0.37 0.37 0.38 0.38
UN Chilean 0.34 0.34 0.34 0.35 0.36 0.36 0.37 0.38
UN Far Eastern 0.34 0.35 0.36 0.36 0.37 0.37 0.38 0.38
UN General 0.34 0.34 0.35 0.36 0.36 0.37 0.37 0.38
UN Latin American 0.34 0.34 0.35 0.35 0.36 0.37 0.37 0.38
UN South Asian 0.34 0.34 0.34 0.35 0.36 0.37 0.37 0.38

Coale-Demeny and United Nations model life tables by sex for expectations of life
at birth ranging from 50 to 85 years. The Coale-Demeny model life tables (Coale
and Demeny 1966; Coale et al. 1983) comprise an ‘East’ variant reflecting Eastern
European populations, a ‘North’ variant based on Nordic populations, a ‘South’
variant based on Southern European populations and the ‘West’ variant previously
noted to be based on populations from Western Europe, North America, Australasia,
Japan, Taiwan and White South Africa. The United Nations model life tables for
developing countries (United Nations 1982) offer ‘Chilean’, ‘Far Eastern’, ‘Latin
American’ and ‘South Asian’ variants along with the ‘General’ variant that averages
the other four. They are based on 36 pairs of male and female life tables for 10 Latin
American, 11 Asian and one African population (Tunisia).
You can see from Table 4.6 (i) that f* values generally lie in the range 0.31 to
0.41 (higher for the Coale-Demeny North model, but this is an unusual model based
on historical Scandinavian life tables), (ii) that they generally rise as expectation of
life at birth increases, and (iii) that they tend to be higher for males, especially at
higher life expectancies. A plausible estimate of f* for an abridged life table can be
made having regard to the sex for which it is being constructed, an educated guess
at the life expectancy at birth of the population in question, and an assessment of
the most appropriate model to use. Thus, in constructing Table 4.5 above a value of
166 4 Analysis of Mortality: The Life Table and Survival

0.41 was used – it is a male life table for which eo 0 has a value around 80, and the
population is one for which the Coale-Demeny ‘West’ model is likely to be the most
appropriate.
Note that f* is distinct from the separation factor f that we used in dealing with
infant mortality. We cannot use the Keyfitz equation (Eq. 4.10) or the tables of
separation factors presented as Tables 4.1 and 4.2 to obtain a separation factor for
use in Eq. 4.39.
We can note at this point that in Eqs. 4.11, 4.39 and 4.38 we now have a family
of equations for the various values of n qx required to generate an abridged life table
which satisfy the same general equation:

n qx D .n:n Mx / = .1 C n: .1  F/ :n Mx / (4.41)

Where F is a separation factor relevant to mortality between exact ages x and x C n.

It remains the case, however, that problems may attach to the use of this equation
to calculate 1 q0 , and that equations requiring only registration data are preferable
(see earlier discussion of Eq. 4.11). The correspondence between Eqs. 4.41 and 4.39
is clear, while that between Eqs. 4.41 and 4.38 is also clear when it is remembered
that the latter assumes a separation factor of one-half.
You will sometimes (e.g., Preston et al. 2001) see Eq. 4.41 written as follows:

n qx D .n:n Mx / = .1 C .n  n ax / :n Mx / (4.42)

This introduces the function n ax , which gives average person-years lived between
exact ages x and x C n by persons dying in that age interval. It is the product of
the width of the age interval and the separation factor F relevant to mortality in that
interval, or the interval width multiplied by the proportion of deaths in the interval
that occur in the second half of the interval (i.e., n ax D n.F, so that F Dn ax /n, which
if substituted for F in Eq. 4.41 yields Eq. 4.42).
The foregoing discussion has provided a basic approach to estimating n qx values
for an abridged life table. Its chief drawback is assuming linear survivorship, and
hence a separation factor of one-half, over 5-year age intervals beyond exact age 5.
Other methods have been developed which take the issue of separation over these
age intervals more seriously. Five will be discussed here – the Reed-Merrell method,
the Greville method, the Fergany method, the Keyfitz-Frauenthal method and the
method of reference to a standard life table.

The Reed-Merrell Approach to Estimating n qx

This method, once widely used but nowadays largely superseded, dates from an
analysis of early U.S. life tables (Reed and Merrell 1939) that suggested there were
equations that linked values of n qx and n Mx across wide ranges of observations.
While this ‘traditional’ method of life table construction (Preston et al. 2001: 45)
Abridged Life Tables 167

has been criticized for its ‘questionable generality’ (Fergany 1971: 333) and being
based on life tables whose widespread applicability there is ‘no special reason to
believe in’ (Preston et al. 2001: 46), its equations are still sometimes used, especially
to obtain values of 5 qx . It is thus desirable to be aware of them:

1 q0 D 1  exp Œ1 M0 .0:9539  0:55091 M0 / (4.43)

4 q1 D 1  exp Œ44 M1 .0:9806  2:0794 M1 / (4.44)

5 qx D 1  exp Œ55 Mx .1 C 0:25 Mx / (4.45)

To the non-mathematical these equations will look horrendous. Equation 4.43 is

almost never used and can be ignored, but the other two, especially Eq. 4.45, are
encountered. Evaluating them manually should not prove difficult for competent
users of a hand calculator offering the function ‘ex ’. Simply evaluate the expression
in square brackets following ‘exp’, hit ‘ex ’, and subtract the answer from 1. An
alternative is to use Reed-Merrell conversion tables, which allow values of n qx
corresponding to values of n Mx to be read off. There are separate tables for 1 q0 ,
4 q1 and 5 qx , and also for 1 q1 and 3 q2 if an abridged table splits the interval between
exact ages 1 and 5 into two – exact ages 1–2 and 2–5. There are separate equations
analogous to Eqs. 4.43, 4.44 and 4.45 for 1 q1 and 3 q2 as well. However, the
Reed-Merrell tables were prepared prior to the advent of electronic calculators and
computers, and are nowadays superfluous. They can be found in Reed and Merrell
(1939) or in appendices to Shryock, Siegel and Associates (1973) and Siegel and
Swanson (2004). Note that Eqs. 4.43, 4.44 and 4.45 and the Reed-Merrell tables
assume values of n Mx calculated to a base of 1.0. Thus, if an age-specific death rate
is specified as, say, 7.48 deaths per 1,000, the equivalent n Mx is found by dividing
by 1,000, or shifting the decimal point three places left (i.e., n Mx D 0.00748).

The Greville Approach to Estimating n qx

This approach to dealing with non-linear survivorship between exact ages specified
in an abridged life table (Greville 1943) focuses on the general equation for
n qx presented above as Eq. 4.41, and provides a further equation for obtaining
the appropriate separation factor F for substitution in this equation. The Greville
equation for the separation factor is:

F D 1=2  Œn .n Mx  0:095/ =12 (4.46)

To obtain a value of n qx corresponding to an observed value of n Mx using Greville’s

approach, evaluate Eq. 4.46 for the observed n Mx , then substitute back in Eq. 4.41
for both F and n Mx to obtain n qx . Although this approach is sometimes claimed to
168 4 Analysis of Mortality: The Life Table and Survival

Table 4.7 Abridged life table for Australian males, 2009–2011 constructed using Greville
separation factors beyond exact age 5
Age (x) n n Mx n Fx n qx lx n dx n Lx Tx eo x
0 1 0.00498 0.11500 0.00496 100000 496 99561 8019757 80.20
1 4 0.00021 0.41000 0.00082 99504 82 397722 7920196 79.60
5 5 0.00012 0.53953 0.00058 99422 58 497348 7522474 75.66
10 5 0.00011 0.53953 0.00056 99364 56 496688 7025126 70.70
15 5 0.00046 0.53939 0.00230 99308 229 496013 6528438 65.74
20 5 0.00062 0.53933 0.00309 99079 306 494690 6032425 60.89
25 5 0.00074 0.53928 0.00367 98773 362 493030 5537735 56.07
30 5 0.00094 0.53919 0.00469 98411 462 490991 5044705 51.26
35 5 0.00118 0.53909 0.00587 97949 575 488420 4553714 46.49
40 5 0.00158 0.53892 0.00787 97374 767 485103 4065294 41.75
45 5 0.00229 0.53863 0.01140 96607 1102 480494 3580191 37.06
50 5 0.00344 0.53815 0.01710 95505 1633 473754 3099697 32.46
55 5 0.00520 0.53742 0.02570 93872 2413 463779 2625943 27.97
60 5 0.00800 0.53625 0.03926 91459 3591 448968 2162164 23.64
65 5 0.01303 0.53415 0.06327 87868 5559 426392 1713196 19.50
70 5 0.02179 0.53050 0.10364 82309 8531 391520 1286804 15.63
75 5 0.03776 0.52385 0.17322 73778 12780 338465 895284 12.13
80 5 0.06758 0.51142 0.29003 60998 17691 261772 556819 9.13
85 ¨ 0.14678 1.00000 43307 43307 295047 295047 6.81

be appropriate for all age intervals in an abridged life table it is not recommended
for the calculation of 1 q0 or 4 q1 , because it is based on a model of mortality that
is unsound at those ages. It is best to use standard procedures at these younger
ages, and to treat Greville’s equation as a useful method of allowing for non-linear
survivorship in generating n qx values beyond exact age 5.
Table 4.7 shows the abridged life table presented above as Table 4.5 reconstructed
using Greville separation factors beyond exact age 5. In other words Eqs. 4.46
and 4.41 have been used to calculate 5 qx values, where in Table 4.5 Eq. 4.38
was used. Separation factors have been indicated in a new column (n Fx ). The
first separation factor (1 F0 ) is the one calculated in Table 4.4, while the second
(4 F1 ) is the separation factor for that age interval from the Coale-Demeny male
‘West’ model life table for eo 0 D 80 extracted from Table 4.6. There are differences
between the final columns of Tables 4.5 and 4.7, but they are minor. Allowing
for non-linear survivorship at the oldest ages (the reality that those dying in these
age groups on average survive for a bit longer than 2.5 years) adds marginally to
life expectancies at those older ages, and these increments feed through to slightly
higher life expectancies at younger ages and at birth in Table 4.7 as well.
Abridged Life Tables 169

The Fergany Method of Estimating n qx

A deceptively simple approach to converting n Mx values to life table n qx values was

suggested by Fergany (1971). He was reacting to perceived deficiencies in the Reed-
Merrell and Greville approaches, both of which he claimed (1971: 333) ‘involve
assumptions that are either unjustifiable or of questionable generality and rough
approximations’. He was also unimpressed with a method suggested by Chiang
(1968) because of its ‘exacting data requirements’. The Fergany conversion equation
is:

n qx D 1  exp .n:n Mx / (4.47)

Reed and Merrell had actually derived this equation and discarded it as unsatis-
factory for transforming age-specific death rates to probabilities of death, Fergany
(1971: 333) describing it as ‘very strange indeed that their justification for this
went unquestioned by demographers for more than thirty years.’ Essentially that
justification was a perceived poor fit to 1910 US life tables, but these had been
derived using ‘a host of : : : intricate actuarial techniques’ that made them dubious
models against which to assess the merit of a ‘universal’ technique for transforming
age-specific death rates into life table probabilities of death. Fergany saw the
simplicity of his approach as a major attribute – it required no special conversion
tables, did not involve a complicated formula, and had no parameters that needed
to be estimated empirically or borrowed from a population other than the one the
life table was being constructed for. All it needed was age-specific death rates and,
at the time, widely available exponential function tables. The latter have since been
superseded by electronic calculators and computers.

The Keyfitz-Frauenthal Method of Estimating n qx

This approach to converting n Mx values into life table n qx values has been shown by
Keyfitz and Frauenthal (1975) to embrace the Reed-Merrell and Greville methods
as special cases. It provides the following equation for converting age-specific death
rates to n qx values for an abridged life table:

n qx D 1  exp Œn: fn Mx C .n Pxn  n PxCn / .n MxCn  n Mxn / =48n Px g (4.48)

Where n Px is the observed population aged between exact ages x and x C n (i.e., it
is the denominator of n Mx – see Eq. 4.35).
You will note that Eq. 4.48 makes use of values of P and M for n-year age
intervals immediately below and above the n-year interval for which a value of
n qx is being calculated. This implies that it cannot be used to obtain 1 q0 and 4 q1 ,
as there are no 1-year and 4-year age intervals, respectively, below the age intervals
to which these probabilities of dying apply. Questions also arise with respect to the
170 4 Analysis of Mortality: The Life Table and Survival

equation’s application to obtain 5 q5 and 5 q80 . In the former instance we need to

combine the 1-year and 4-year intervals that typically begin an abridged life table
to yield a 5-year interval below the one of interest. In the latter we would need data
for the interval between exact ages 85 and 90, which is part of the terminal open
interval in the standard abridged life table format, in order to have a 5-year interval
above the one of interest.

Using Reference to a Standard Life Table to Estimate n qx

A further approach to the conversion of observed n Mx values into life table n qx

values bases it on the relationship that exists between these values in an existing,
standard life table. The standard life table may be either a single-year-of-age life
table or an abridged life table, with the proviso that it cannot be more abridged
than the life table one wishes to construct. It should be selected so that the levels
of mortality by age (the patterns of mortality) in the standard population and the
population of interest (as measured by observed age-specific death rates n Mx ) are
roughly comparable. It would not, for example, be appropriate to use a life table for
Australian males in 2009–2011 as the standard to construct a life table for males in
a developing African country around the same date; the patterns of mortality would
be very different. It is also usual to employ a standard life table for the same sex as
the population for which a life table is being constructed, since mortality patterns
often differ significantly by sex.
Frequently, judicious use is made of model life tables (United Nations 1955;
Coale and Demeny 1966; Brass 1971; United Nations 1982; Coale et al. 1983; Coale
and Guo 1989; Murray et al. 2003; Zhao 2007) to provide appropriate standards.
As previously explained, model life tables generalize mortality patterns found in
families of actual life tables, the skill in using them lying in selecting the ‘family’
to which mortality conditions in a population of interest are most likely to belong.
A major initial reason for developing model life tables was to assist in analysing
the mortality of developing populations, for which reliable mortality data were
often lacking. In particular they were often used to gain insight into likely adult
mortality conditions in such populations, on the basis of more reliable empirical
evidence about infant and childhood mortality. More recent developments have
sought to address such things as the very low levels of mortality now prevalent in
many developed countries, which did not exist when the United Nations Population
Division, Coale and Demeny, and Brass were pioneering the field, and the huge
impact of HIV/AIDS on mortality patterns in African and some Asian nations in
particular.
To use the method of reference to a standard life table we compute a quantity n gx
using the equation:
   
n gx D n=n qsx  1=n Msx (4.49)
Abridged Life Tables 171

Where n gx actually represents the average number of person-years of life lost

between exact ages x and x C n by persons dying in that age interval in the
standard population; n qsx and n Msx are the n qx and n Mx values for the age interval
bounded by exact ages x and x C n for the standard population (assuming that
n Mx D n mx ).
s s

Having evaluated Eq. 4.49, n gx is substituted back in the following equation for
n x:
q

n qx D n:n Mx = .1 C n gx :n Mx / (4.50)

As an aside, comparison of Eq. 4.50 with the generalized equation for n qx , Eq. 4.41,
indicates that n gx D n (1  F) (where F is the separation factor appropriate to the
age interval in question). It is from this relationship that the interpretation of n gx
given above derives. Where deaths are concentrated at the lower end of an age
interval, those dying in the interval obviously survive, on average, for less than
half the interval; or putting it another way, they fail to survive for more than half
the interval. Since we noted earlier that when deaths are concentrated at the lower
end of an age interval F <1 /2 , it follows that n gx D n (1  F) >n /2 , which is consistent
with n gx measuring average failure to survive, or life lost. Similarly, if deaths are
concentrated at the upper end of an age interval those who die in the interval on
average survive for more than half the interval and fail to survive for less than half of
it. Since in this circumstance F >1 /2 , n gx D n (1  F) <n /2 , which again is consistent
with n gx measuring average failure to survive, or life lost.
Equation 4.50 can also be compared with Eq. 4.42 to reveal that n gx D n n ax .
So the average number of person-years of life lost between exact ages x and x C n
by persons dying in that age interval (n gx ) is equal to the width of the interval (n)
less the average number of person-years of life lived between exact ages x and x C n
by persons dying in that age interval (n ax ). We can use this relationship to develop
an alternative approach to obtaining n qx values by reference to a standard life table.
Substituting (n n ax ) for n gx in Eq. 4.49 yields:
   
n ax D n  n=n qsx C 1=n Msx (4.51)

After evaluating this equation we can substitute for n ax in Eq. 4.42 to obtain n qx .
Other schemes for converting observed n Mx values into life table n qx values also
exist in the literature, and there have in addition been techniques developed for
graduating, or smoothing, series of n Mx values prior to translation to n qx values.
This sort of technical detail, however, need not concern us in this introductory text,
the aim of which is to provide a basic coverage of life table construction. All manner
of methodological refinements are discussed in the literature, but many are of more
concern to actuaries than to demographers.
You might, however, reasonably ask when, assuming that equations which
assume linear survivorship have been ruled out, each of the approaches to obtaining
172 4 Analysis of Mortality: The Life Table and Survival

n qxvalues just outlined should be used. There are no hard and fast rules. Differences
between results the methods produce ordinarily are not large (e.g., both Fergany
(1971) and Keyfitz and Frauenthal (1975) show their methods to yield similar
results to the methods of Reed and Merrell, and Greville), and when dealing with
populations of developing countries, data deficiencies leading to unreliable n Mx
values will probably have far greater bearing on the accuracy of a life table than
will the choice of a method of life table construction. The method of reference to
a standard life table can be especially convenient if, for example, there is a desire
to construct a life table at some other date for a population for which a life table
already exists to act as a standard.

Generating an Abridged Life Table: Estimating Other Life Table

Functions

Once a set of n qx values have been calculated for an abridged life table, the
remainder of the table can be constructed using the following relationships and the
fact that l0 D 100,000.

lxCn D lx  lx :n qx D lx .1  n qx / (4.52)

n dx D lx :n qx D lx  lxCn (4.53)

n px D 1  n qx (4.54)

and, after calculating 1 L0 using Eq. 4.23:

 
4 L1 D f :l1 C 4  f :l5 (4.55)

and, for x  5, assuming linear survivorship:

n Lx D n=2 .lx C lxCn / (4.56)

or, for x  5, assuming n mx Dn Mx :

n Lx D n d x = n Mx (4.57)

and, for the last row of the life table:

LxC D Tx D dxC =MxC D lx =MxC (4.58)

Abridged Life Tables 173

then, for each row before the last, working up the life table:

Tx D TxCn C n Lx (4.59)

and finally:

eo x D Tx =lx (4.60)

The Reed-Merrell and Keyfitz-Frauenthal approaches to the construction of

abridged life tables and the method of reference to a standard life table do,
in addition to providing procedures for obtaining values of n qx , provide special
procedures for calculating the n Lx and Tx columns. However, they are not widely
used and will not be elaborated on here. The interested reader is referred to Siegel
and Swanson (2004: 312–315). The equations above are satisfactory for present
purposes, but note that Eq. 4.56 should be used in conjunction with Eq. 4.38, while
Eq. 4.57 is appropriate when values of n qx have been obtained by a method which
does not assume linear survivorship between exact ages x and x C n.

The Shapes of Life Table Functions

Having now completed discussion of the mechanics of constructing single-year-of-

age and abridged life tables it is appropriate to pause and examine the shapes of life
table functions and how these vary between high mortality (low life expectancy) and
low mortality (high life expectancy) populations. Figure 4.6 shows plots of several
life table functions from Coale and Demeny ‘West’ model life tables for female
populations with life expectancies at birth of 40, 60 and 80 years. With respect
to the mortality transition (a component of the familiar ‘demographic’ transition)
from high to low mortality these life tables can be regarded as respectively pre-
transitional, mid-transitional and post-transitional, or tables pertaining to high,
intermediate and low mortality populations. Function plots derive from abridged
life tables which recognise age groups 0, 1–4, then 5-year age groups 5–9 to 95–99.
The open interval 100C is excluded.
Life table probabilities of dying (n qx values) derived from age-specific death rates
(n Mx values) are much higher at ages 0 and 1–4 for the high mortality population
than for the low mortality population, with the mid-transitional population falling
in between. It is typical of higher mortality populations that they have much higher
rates of infant and early childhood mortality (i.e., under-5 mortality) than do low
mortality populations, and this reality is a major factor in the differential life
expectancies at birth they exhibit, because in higher mortality populations so many
members forfeit (through premature death) nearly all of their potential contribution
to the Tx column of the life table. At older ages n qx values rise earlier and to
ultimately higher levels for the low and intermediate life expectancies than for
life expectancy at birth D 80 years, with values for the latter, post-transitional,
population only beginning to move noticeably above zero in middle age (late 40s
and 50s).
174 4 Analysis of Mortality: The Life Table and Survival

Fig. 4.6 Graphs of life table functions for west model abridged life tables: Female life expectan-
cies at birth of 40, 60 and 80 years
Abridged Life Tables 175

Survivorship (lx ) functions for the three populations are very different. Under
conditions of high mortality there is an initial steep fall over the infant and early
childhood ages (a loss of almost 18 and almost 28 % of the radix population by
exact ages 1 and 5 respectively), a noticeable though less steep decline through the
late childhood and early and middle adult ages, then a steepening again beyond
age group 50–54 until sheer lack of survivors intervenes at the very oldest ages. By
contrast under conditions of low mortality the survivorship curve is flat to about ages
50–54, and only then begins to fall increasingly rapidly. It is sobering in comparing
these lx plots to realize that the declines in lx over the first year of life for eo 0 D 40 in
Fig. 4.6 are not equalled for eo 0 D 80 until beyond exact age 70. This should strongly
reinforce in your mind the huge potential that initiatives to reduce infant mortality
have for improving longevity in high mortality populations, and indeed their key
importance historically in bringing about mortality transitions across the globe.
Once again the mid-transitional population follows a path intermediate between
those of the high and low mortality populations.
The importance of infant and early childhood mortality in high mortality
populations is again clear in the comparative plots of n dx in Fig. 4.6. Very large
numbers of deaths occur at the youngest ages in a high mortality population. By
contrast very few occur before about age 40 in low mortality populations, where
deaths peak at ages in the late 70s and 80s. For the high mortality population there
is a secondary peak at ages 70–74, but it is nowhere near as pronounced as the
infant mortality peak, and we are comparing here a 5-year age group with a 1-year
age group. The n dx plot for the mid-transition population is once again intermediate.
There are still appreciable numbers of deaths in infancy and early childhood for this
population, but the peak at older ages is more pronounced and located at ages 75–79,
between the peaks for the high and low mortality populations.
Age structures for the three life table stationary populations are depicted by the
n Lx plots. The comparatively larger size of the life table stationary population as life
expectancy at birth increases is clear, and while the high mortality population loses a
lot of members very young its stationary population also has a comparatively young
age structure due to the cumulative effect with increasing age of having higher
probabilities of dying at every age. Its value of 5 L80 , for example, the number aged
80–84 in the life table stationary population, is barely one-tenth of that for the low
mortality post-transitional population.
Unsurprisingly the graphs of person-years lived by the life table stationary
population beyond exact age x in Fig. 4.6, represented by the Tx columns of the
respective life tables, show much larger numbers at each age in the low mortality
population than in the high mortality population, although at the very oldest ages
numbers in all three populations naturally converge toward zero. The nature of the
Tx function of course dictates that all trend lines fall progressively from the youngest
to the oldest ages.
Aside from life expectancies being higher, age for age, in the low mortality
population than in the high mortality population, with those for the mid-transitional
population in between, the most interesting difference between graphs of eo x
in Fig. 4.6 is that for the high mortality and mid-transitional populations life
176 4 Analysis of Mortality: The Life Table and Survival

expectancy increases from its level at birth for survivors at exact age 1 and (for
the high mortality population) exact age 5. This is a common feature of mortality
conditions in higher mortality populations. Those who manage to survive through
the high mortality infant and early childhood age groups actually improve their
subsequent survival chances by doing so.

Using Life Tables: Survival

How do demographers use life tables? In a variety of ways, but a number of

very important applications hinge on measures known as life table survival ratios.
They are used in such exercises as making population projections, estimating net
intercensal migration, calculating the net reproduction rate (NRR), checking the
accuracy of age statement and completeness of enumeration at a census, checking
the completeness of birth registration, and estimating fertility from a population’s
age distribution. A survival ratio expresses the proportion of persons who reach
a given age alive who survive to attain a nominated subsequent age. These ages
may be exact ages or ages in completed years (i.e., age groups).

Population Survival Ratios

Survival ratios need not necessarily derive from life tables. Indeed many useful
applications of life tables entail equating a life table survival ratio with an equivalent
survival ratio derived from census and/or vital registration data in which one element
(numerator or denominator) is unknown, then solving for that unknown. Population
survival ratios, usually based on census data, give the ratio of survivors in an age
group at some later time to the population in the equivalent younger age group at
some earlier time. Thus they measure the proportion surviving, or the probability
of survival, from the younger to the older age group.
For a single-year age group x, the population survival ratio over an n-year period
from age x to age x C n is given by:

PSRx;xCn D P.y C n/xCn =P.y/x (4.61)

Where P(y)x is the population aged x in year y; n is the interval in years over which
the population is being survived.
For an age group x to x C t  1 (where t is the width of the age group in years)
the population survival ratio over an n-year period is given by:

PSRx to xCt1;xCn to xCt1Cn D P.y C n/xCn to xCt1Cn =P.y/x to xCt1 (4.62)

Using Life Tables: Survival 177

Where P(y)x to xCt1 is the population aged x to x C t  1 in year y; n is once again

the interval in years over which the population is being survived.
In case specifying an age group as ‘x to x C t  1’ is a source of confusion, think
of the manner in which we conventionally specify 5-year age groups – 0–4 years, 5–
9 years, 10–14 years, etc. Although these are 5-year age groups, because we specify
the upper and lower limits in terms of ages in completed years rather than in terms
of exact ages, the difference between the upper and lower limits is four, or one less
than the true age group width. The ‘1’ in ‘x C t  1’ simply takes account of this
in, following convention, specifying the upper limit of an age group as an age in
completed years rather than an exact age.
Note that in Eqs. 4.61 and 4.62 the numerators refer to survivors at the
later point in time (year y C n) from the populations forming the denominators.
This implies that ideally one should make allowance for migration over the n-
year survival period; numerators should be adjusted upward or downward from
straight population figures to take account of net out-migration or net in-migration
respectively. Having made this point we will proceed for the moment as if dealing
with a closed population. The migration factor is, however, important, to the point
where a major application of population survival ratios is the estimation of net
(generally intercensal) migration.
Note also that when we refer in Eqs. 4.61 and 4.62 to a population in year y and
an equivalent population in year y C n it is assumed that those populations refer to
the same date in each year (or very close to it). If there is some significant departure
from this assumption (as, for example, when we find that successive censuses were
held on different dates during the year) the value of n, the survival interval, becomes
something other than a whole number of years and calculations are not quite so
straightforward. One approach to this problem is to adjust one of the sets of census
data to the same date during the year as applies to the other set, then calculate
survival ratios.
Note further that population survival ratios are multiplicable. The survival ratio
over any age range is equal to the product of survival ratios over any number of
discrete component age ranges (i.e., age ranges which do not overlap, but which
together exactly make up the larger age range). For example:

PSR0;5 D PSR0;1 : PSR1;5

since:

P.y C 5/5 =P.y/0 D .P.y C 1/1 =P.y/0 / : .P.y C 5/5 =P.y C 1/1 /

Similarly:

PSR04;1014 D PSR0–4;5–9 : PSR59;1014

178 4 Analysis of Mortality: The Life Table and Survival

since:

P.y C 10/1014 =P.y/04 D .P.y C 5/59 =P.y/04 / : .P.y C 10/1014 =P.y C 5/59 /

In each case elements on the righthand sides of these expressions cancel. These
are straightforward examples, with the survival age ranges split into just two
components, but the multiplicative principle extends to splits into larger numbers
of components.

Survival Ratios from Birth

Survival ratios from birth are a variation on population survival ratios. They measure
the ratio of survivors in an age group at some later time to the size of the cohort
forming that age group at birth. In other words they measure the proportion of the
birth cohort who survive to form the age group, or the probability of survival from
birth to the age group.
The distinction between survival ratios from birth and population survival ratios
can best be understood using the Lexis diagram. Population survival ratios pertain
to situations of the type illustrated by Fig. 4.7a (note that both Lexis diagrams in
Fig. 4.7 feature grid squares with dimensions 5 years of age by 5 years of calendar
time, rather than the more familiar one year by one year squares). This diagram
depicts survival from age group 0–4 years at a census held on 30:6:96 to age group
10–14 at a census held 10 years later on 30:6:06. This is an example of survival
between age groups (populations we represent on a Lexis diagram by vertical lines),
and that is what population survival ratios measure.
Figure 4.7b, on the other hand, illustrates survival from birth to the age group 0–
4 years at a census held on 30:6:06. Age group 0–4 years is a 5-year age group, and
in a closed population clearly comprises survivors from births over the 5-year period
immediately preceding the census – that is, births between 1:7:01 and 30:6:06.
These births constitute a population attaining exact age 0, which we represent on a
Lexis diagram by a horizontal line. The population (age group) to which this cohort
is being survived, however, is one we represent on a Lexis diagram by a vertical line.
Thus in Lexis diagram terms survival from birth is concerned with survival from a
population we represent by a horizontal line at exact age 0 to one we represent by a
vertical line at some later point up the cohort diagonal band which the extremities
(the end points) of the former line define.
The survival ratio from birth to a single-year age group x is given by:

SRB;x D P.y/x =B .y  x  1/ (4.63)

Where P(y)x is the population aged x at a specified date in year y; B(y  x  1) is

births during the year commenced the date of the year y population count in year
y  x  1.
Using Life Tables: Survival 179

a
Exact Age
15

10

0
1996-2000 2001-05 2006-10

b
Exact Age

0
2001-05 2006-10

30:06:01 30:06:06

Fig. 4.7 Lexis diagram illustrations of survival between age groups and survival from birth (a)
Survival from age group 0–4 in mid-1996 to age group 10–14 in mid-2006 (b) Survival from birth
to age group 0–4 in mid-2006
180 4 Analysis of Mortality: The Life Table and Survival

The survival ratio from birth to an age group x to x C t  1 (where t is the width
of the age group in years) is given by:

SRB;x to xCt1 D P.y/x to xCt1 =Bt;yxt (4.64)

Where y D year of the population count which includes survivors aged x to

x C t  1; x D lower limit of age group survived to; P(y)x to xCt1 is the population
aged x to x C t  1 in year y; Bt,yxt is births during a t-year period commenced
the date of the year y population count in year y  x  t.
Taking the example illustrated in Fig. 4.7b, yD 2006, x D 0 and t D 5. Therefore:

SRB;04 D P.2006/04 =B5;2001

In other words the survival ratio from birth to age group 0–4 is the ratio of the
population aged 0–4 at the date of the 2006 census to the births that occurred over
a 5-year period commenced midnight on 30th June (the date of the 2006 census) in
2001.

Life Table Survival Ratios

The true value of population survival ratios and survival ratios from birth becomes
apparent when they are linked with life table survival ratios. There are several
important applications of this sort of strategy, which basically involve:
1. Equating a population survival ratio or a survival ratio from birth with an
equivalent life table survival ratio (derived from a life table that summarizes the
mortality experience of the population in question during the period over which
it is being studied).
2. Using this equality to obtain one of the quantities (numerator or denominator)
from the population survival ratio or the survival ratio from birth which is
unknown.
In other words we don’t actually calculate the population survival ratio or the
survival ratio from birth, because we can’t – we don’t know one of the quantities
needed. Instead we obtain an expected value of the survival ratio from a life table
and use this to estimate the unknown quantity (either a population in some age group
or a number of births over a defined period).
Recall that the life table population is a stationary population, the age structure
of which is given by the Lx column. Use of this column enables us to calculate life
table survival ratios which express the probability of surviving from any age to any
later age under the mortality conditions summarized by the life table. In general, for
a single-year-of-age life table:
Using Life Tables: Survival 181

LTSRx;xCn D LxCn =Lx (4.65)

Where LTSRx,xCn means the life table survival ratio from age x to age x C n.
Like population survival ratios, life table survival ratios are multiplicable; the life
table survival ratio over any age range is equal to the product of life table survival
ratios over any number of discrete component age ranges. Thus, for example:

LTSR0;5 D LTSR0;1 : LTSR1;5

The life table survival ratio from age 0 to age 5 is the product of the life table survival
ratios from age 0 to age 1 and from age 1 to age 5.
Life table survival ratios may also be computed from abridged life tables. In
general:

LTSRx to xCt1;xCn to xCt1Cn D t LxCn =t Lx (4.66)

Where t D width of age group for which the life table survival ratio is being
calculated; n D number of years over which the life table population aged x to
x C t  1 is survived.
So, for example:

LTSR04;1014 D 5 L10 =5 L0 D 5 L10 = .1 L0 C 4 L1 /

Since 5 L0 D1 L0 C4 L1 , because an abridged life table begins with a 1-year, then a

4-year age interval.
The life table survival ratios discussed to this point have been equivalents of
population survival ratios; they have measured survival from one age group in the
life table stationary population to another. We can also calculate life table survival
ratios which measure survival from a given exact age to an older age group. In
particular, we can calculate the life table equivalent of a survival ratio from birth;
a ratio measuring the probability of surviving from exact age 0 to some subsequent
age group.
For survival from birth to a single-year age group x we have:

LTSRB;x D Lx =l0 (4.67)

For survival from birth to an age group of width t years we have:

LTSRB;x to xCt1 D t Lx =t:l0 (4.68)

Where t D width in years of the age group survived to; x D lower limit of age group
survived to.
182 4 Analysis of Mortality: The Life Table and Survival

So, for example:

LTSRB;1519 D 5 L15 =5 : l0

Why do we need to multiply l0 in the denominator by t? Because the population

in any t-year age group are survivors from births over a t-year period. And, of
course, the life table stationary population experiences l0 births each year. The
population aged 0–4, for example, comprises survivors from five single-year (though
not necessarily calendar year) birth cohorts.
Note that when calculating a life table survival ratio a numerator or denomi-
nator is obtained from the n Lx column of the life table if it refers to an age group,
and from the lx column if it refers to an exact age. If confused, try to think of your
problem in Lexis diagram terms; if your survival ratio numerator or denominator
is a population you would represent on a Lexis diagram by a vertical line, use the
Lx column; if it is a population you would represent by a horizontal line, use the lx
column. There is one modification to this principle which we make for practical
reasons. If an age group of interest is of the type ‘ages x and older’, while we could
use the n Lx column in accordance with our principle (summing all values in the
column from exact age x to the end of the life table), it is simpler to consult the Tx
column, which gives precisely this summation directly. Thus the life table survival
ratio from ages x and older to ages x C n and older is simply TxCn / Tx .

Relationship Between Population and Life Table Survival Ratios

If migration is zero during a period, then a population survival ratio equals the
equivalent life table survival ratio; that is, the equivalent survival ratio derived from
a life table representative of the mortality of the population in question over the
period in question. Thus, from our earlier equations for a PSR and a LTSR from age
x to age x C n (Eqs. 4.61 and 4.65) we have:

P.y C n/xCn =P.y/x D LxCn =Lx

It follows that:

P.y C n/xCn D P.y/x : .LxCn =Lx / (4.69)

Here we are obtaining the population in a given single-year age group at some later
point in time as the product of the population in an equivalent younger age group at
an earlier point in time and an appropriate life table survival ratio. This procedure
is known as FORWARD SURVIVAL. Equation 4.69 is one for use when forward
surviving single-year age groups.
Using Life Tables: Survival 183

Similarly, from our earlier equations for a PSR and a LTSR from age group x
to x C t  1 to age group x C n to x C t  1 C n (i.e., for an age group of width t
survived over a period of n years – see Eqs. 4.62 and 4.66) we have:

P.y C n/xCn to xCt1Cn =P.y/x to xCt1 D t LxCn =t Lx

It follows that:

P.y C n/xCn to xCt1Cn D P.y/x to xCt1 : .t LxCn =t Lx / (4.70)

Equation 4.70 is the one to use when forward surviving age groups wider than a
single year. An important application of Eqs. 4.69 and 4.70 for forward survival of
age groups in year y to corresponding age groups in year y C n is the projection
of populations into the future. The crux of projecting a population is to take that
population at a given date (normally a census date) and survive it to some future date
n years beyond the first date by applying the appropriate forward survival equation
to each of its constituent age groups. Having survived our census population in year
y to year y C n, there are two other general elements to be attended to in order to
complete our projection.
1. We must consider what is likely to happen to births over the projection period.
Clearly the youngest members of our census population in year y (those aged
0 years) will be aged n years at the end of the n-year projection period. Younger
members of our projected population will be survivors of birth cohorts born
between the census and the projection date. We must therefore estimate the
sizes of these birth cohorts and then survive those to the projection date to get
the projected population aged 0 to n  1 years. Of course most of the births that
will take place over the projection period have yet to take place at the time the
projection is made. It follows that an integral part of the projection exercise is
making assumptions about the sizes of birth cohorts (i.e., about fertility trends)
over the projection period. Commonly a series of projections is made, each
incorporating a different assumption about likely fertility trends, so that a range
of options as to what the future population size and distribution by age might be
is established.
The type of exercise just described entails forward survival from birth. If
this is done for 1-year birth cohorts (which will form 1-year age groups in the
projected population) we equate the righthand sides of Eqs. 4.63 and 4.67:

P.y/x =B .y  x  1/ D Lx =l0

and rearrange to get:

P.y/x D B .y  x  1/ : .Lx =l0 / (4.71)

Where y is the projection year; x is a single-year age group in the projected pop-
ulation such that x < n (where n is the projection period in years); B(y  x  1)
is estimated births in the year commenced the date in year y to which the
184 4 Analysis of Mortality: The Life Table and Survival

population projection pertains (e.g., mid-year) in year y  x  1; Lx comes from

an appropriate single-year-of-age life table.
If forward surviving t-year birth cohorts (which will form t-year age groups
in the projected population) we equate the righthand sides of Eqs. 4.64 and 4.68:

P.y/x to xCt1 =Bt;yxt D t Lx =t : l0

and rearrange to get:

P.y/x to xCt1 D Bt;yxt : .t Lx =t : l0 / (4.72)

Where y is the projection year; x is the lower limit and t is the width of an
age group in the projected population to be estimated using forward survival
from birth (it is necessary that x C t  1 < n, where n is the projection period in
years); Bt,yxt is estimated births during a t-year period commenced the date
in year y to which the population projection pertains (e.g., mid-year) in year
y  x  t; t Lx comes from an appropriate abridged life table.
Equations 4.71 and 4.72 can be used to forward survive birth cohorts born
during the n-year projection period which will form ages 0 to n  1 in the
projected population. Figure 4.8 illustrates how population projection involves
both forward survival of a census population and forward survival of projection
period birth cohorts.
2. We must consider whether, and if so how, migration is likely to modify our
population over the projection period. In other words, in addition to making
assumptions about future fertility we must make assumptions about future
migration. These assumptions can be important when projecting populations at
the national level, but are likely to be even more critical when making projections
at the sub-national level (i.e., for regions of a country, cities, local government
areas, etc.).
As with making assumptions about future fertility, it is common in population
projection exercises to prepare series of projections based on a range of migra-
tion assumptions. The mathematics of incorporating migration into population
projections are not dealt with here. It is sufficient at this juncture to appreciate
the need in projecting populations to weld onto the forward survival of an initial
census population and births during the projection period a consideration of
likely migration trends and patterns during that projection period. Population
projection is dealt with in greater detail in Chap. 9.

Estimation of Births

Just as we can use life table survival ratios to project populations into the future
using forward survival, so we can use them to work backwards from a population
Using Life Tables: Survival 185

Exact Age

Age groups n and

over projected by
forward survival
of year y census
population

Age groups 0 to
n-1 projected by
forward survival
of births
between census
and projection
dates in years y
and y+n

0
n-year projection period

Year y census count Year y+n

projection date

Fig. 4.8 Schematic Lexis representation of the role of forward survival in population projection
186 4 Analysis of Mortality: The Life Table and Survival

distribution by age to obtain information about the past. This procedure is known as
REVERSE SURVIVAL.
We could, if we wished, use this method to derive a population distribution by age
at some earlier point in time, assuming no migration and with the proviso that only
age groups for which there were survivors in the later population would be covered.
We might, for example, be interested in estimating the size and age distribution of
a population at some date between two censuses. We could use a combination of
forward survival from the earlier census and reverse survival from the later one,
averaging the results of the two exercises. Or if the later census was felt to be more
reliable we might rely solely on reverse survival. We could even use reverse survival
from that census to assess the extent of unreliability in the earlier, less reliable,
census.
A more useful application of reverse survival, however, is in estimating births
for the period preceding the date of a population count (normally a census). This
technique uses the idea of survival from birth in reverse, and is especially useful
when dealing with populations for which there are no birth registration data or for
which birth registration data are incomplete. It enables us to estimate births from
census data. However:
1. The technique is only as good as the quality of the census and the life table one
uses.
2. If birth registration is incomplete or non-existent, so may death registration be.
This may impede the generation of a life table for the population of interest,
although this problem can be overcome by resorting to an appropriate model life
table. Obviously, when this approach is taken, the accuracy of birth estimates
will depend on how well the model table summarizes mortality conditions in the
population of interest.
Suppose a census is conducted in year y. Then under an assumption of zero
migration we have, equating the righthand sides of Eqs. 4.63 and 4.67:

P.y/x =B .y  x  1/ D Lx =l0

This is the same equation as the one from which we developed Eq. 4.71 for
forward survival of a birth cohort to age x. We again rearrange it, but this time
treating B(y  x  1) as the unknown and making it the subject of the new equation:

B .y  x  1/ D P.y/x : .l0 =Lx / (4.73)

Where y is the census year; x is a single-year age group at the census;

B(y  x  1) D births over the 1-year period commenced the date of the census
count in year y  x  1; Lx comes from an appropriate single-year-of-age life
table.
Using Life Tables: Survival 187

So, for example, from the population aged 5 at a census held on 1 April 2006
we can estimate the number of births during the year commenced 1 April 2000
(y D 2006, x D 5 so that y  x  1 D 2000).

B.2000/ D P.2006/5 : .l0 =L5 /

Note that, except where the population (census) count was taken at midnight on
31 December, estimates of births obtained from this procedure do not relate to
calendar years. They relate to years commenced the date (day and month) when
the population count was taken. However, having estimated births for a series of
successive non-calendar years it is possible to convert these to estimates for calendar
years assuming even distributions of births through non-calendar years.
In Eq. 4.73 we have a method for estimating births over a 1-year period through
reverse survival of a 1-year age group. It is equally possible to estimate births over
a t-year period through reverse survival of a t-year age group. Suppose again that
a census is conducted in year y. Under an assumption of zero migration we have,
equating the righthand sides of Eqs. 4.64 and 4.68:

P.y/x to xCt1 =Bt;yxt D t Lx =t : l0

This is the same equation as the one from which we developed Eq. 4.72 for
forward survival of a t-year birth cohort to age group x to x C t  1. We again
rearrange it, but this time treating Bt,yxt as the unknown and making it the subject
of the new equation:

Bt;yxt D P.y/x to xCt1 : .t : l0 =t Lx / (4.74)

Where y is the census year; x is the lower limit and t is the width of an age group
at the census; Bt,yxt D births over the t-year period commenced the date of the
census count in year y  x  t; t Lx comes from an appropriate abridged life table.
So, for example, from the population aged 0–4 at a census held on 1 April 2006
we can estimate the number of births during the 5-year period commenced 1 April
2001 (y D 2006, x D 0, t D 5 so that y  x  t D 2001).

B5;2001 D P.2006/04 : .5 : l0 =5 L0 / D P.2006/04 : Œ5 : l0 = .1 L0 C 4 L1 /

Since 5 L0 D1 L0 C4 L1 , because an abridged life table begins with a 1-year, then a

4-year age interval.
Equations 4.73 and 4.74 for estimating numbers of births by reverse survival
from a census have assumed zero net migration. In using reverse survival for this
purpose it is important to assess how reasonable the zero migration assumption is.
Normally there isn’t much of a problem when estimating births over the few years
immediately preceding a census. This is done by reverse survival of the youngest
age groups at the census, whose youth means that there has been limited opportunity
for migration, although obviously parental migration can draw in children as well.
188 4 Analysis of Mortality: The Life Table and Survival

However, more importantly, as one attempts to estimate births further back in time
one reverse survives older and older age groups at the census, which have had
more and more time in which to be influenced by migration. The zero migration
assumption becomes increasingly problematic, and one’s estimates of births become
increasingly prone to error. The technique is thus best reserved for estimating births
in the few years immediately preceding a census, unless a credible basis exists for
incorporating migration into the calculation.

Estimation of Intercensal Migration

The estimation of intercensal migration is another important application of life table

survival ratios. The general strategy is to forward survive an age group from an
earlier census to find the expected size of the corresponding age group at a later
census under an assumption of zero migration. This is then compared with that
age group’s actual size at the later census, the difference between the two being net
migration.
Using forward survival we have, for single-year age groups:

P .y/x =P.y  n/xn D Lx =Lxn

Where P(y  n)xn is the population aged x  n at a census in year y  n which

marks the beginning of an n-year intercensal period; P*(y)x is the expected
number of survivors from P(y  n)xn at the end of the intercensal period
assuming no migration; Lx and Lxn come from a single-year-of-age life table
which summarizes intercensal mortality conditions. It might also help to realize
that y D the later census year, x D the relevant single-year age group at the later
census, and n D the width of the intercensal period in years.
It follows that:

P .y/x D P.y  n/xn : .Lx =Lxn / (4.75)

Whence net migration over the intercensal period for the cohort aged x at the later
census is given by:

NM D P.y/x  P .y/x (4.76)

This is the actual, or observed, population aged x at the later census less the
population aged x expected had there been no intercensal migration.
If our interest is in estimating intercensal migration for t-year age groups (rather
than single-year age groups) we have, using forward survival:

P .y/x to xCt1 =P.y  n/xn to xCt1n D t Lx =t Lxn

Using Life Tables: Survival 189

Where P(y  n)xn to xCt1n is the population aged x  n to x C t  1  n at a

census in year y  n which marks the beginning of an n-year intercensal period;
P*(y)x to xCt1 is the expected number of survivors from P(y  n)xn to xCt1n at
the end of the intercensal period assuming no migration; t Lx and t Lxn come from
an abridged life table which summarizes intercensal mortality conditions over t-
year age intervals (they can also be obtained by summing values of Lx from
an appropriate single-year-of-age life table over relevant t-year age intervals). It
might also help to realize that y D the later census year, x D the lower limit of the
relevant t-year age group at the later census, and n D the width of the intercensal
period in years.
It follows that:

P .y/x to xCt1 D P.y  n/xn to xCt1n : .t Lx =t Lxn / (4.77)

Whence net migration over the intercensal period for the cohort aged x to x C t  1
at the later census is given by:

NM D P.y/x to xCt1  P .y/x to xCt1 (4.78)

This is the actual, or observed, population aged x to x C t  1 at the later census less
the population aged x to x C t  1 expected had there been no intercensal migration.
As an example, let us assume we have censuses conducted on 1 April 1996 and
1 April 2006, and that we want to specify equations following Eqs. 4.77 and 4.78
which will permit net intercensal migration for the cohort aged 20–24 at the 2006
census to be estimated. We have that y D 2006 (the later census year), x D 20 (the
lower limit of the relevant age group at the later census), t D 5 (the width of the
relevant age group) and n D 10 (the length of the intercensal period). Therefore,
following Eq. 4.77:

P .2006/2024 D P.1996/1014 : .5 L20 =5 L10 /

And, from Eq. 4.78:

NM D P.2006/2024  P .2006/2024

Note that a rather intimidating equation like Eq. 4.77 looks a good deal less
intimidating once we get down to a real example. The expected population aged
20–24 in 2006 equals the observed population aged 10–14 (10 years younger) in
1996 (10 years earlier) multiplied by a life table survival ratio which is the ratio of
the sizes of the same two age groups in the life table stationary population (the n Lx
values for those age groups).
Equations 4.75, 4.76, 4.77 and 4.78 facilitate the calculation of intercensal net
migration as a cohort ages from a younger age group at the earlier census to a
corresponding older age group at the later census. It is in theory also possible
190 4 Analysis of Mortality: The Life Table and Survival

to estimate net migration as a cohort ages between exact ages, between an exact
age and an age group, or between an age group and an exact age. In practical
demographic terms the problem of this nature most likely to be of interest is one
requiring net migration between birth (exact age 0) and the relevant age group at
a later census to be estimated for a cohort. The principle is the same as before.
We obtain the expected size of the age group at the census, but this time using the
principles of forward survival from birth, and compare this with the observed size
of the age group.
For net migration from birth to a single-year age group x in year y we have,
equating the righthand sides of Eqs. 4.63 and 4.67 with P*(y)x written for P(y)x :

P .y/x =B .y  x  1/ D Lx =l0

Where B(y  x  1) is births during the year commenced the date of the year y
census count in year y  x  1; P*(y)x is the expected number of survivors from
birth forming age group x at the census in year y assuming no migration.
Whence:

P .y/x D B .y  x  1/ : .Lx =l0 / (4.79)

After which net migration would be given by Eq. 4.76.

And for net migration from birth to a t-year age group x to x C t  1 in year y we
have, equating the righthand sides of Eqs. 4.64 and 4.68 with P*(y)x to xCt1 written
for P(y)x to xCt1 :

P .y/x to xCt1 =Bt;yxt D t Lx =t : l0

Where Bt,yxt is births over the t-year period commenced the date of the year y
census count in year y  x  t; P*(y)x to xCt1 is the expected number of survivors
from birth forming age group x to x C t  1 at the census in year y assuming no
migration; x is the lower limit and t is the width of this age group at the census.
Whence:

P .y/x to xCt1 D Bt;yxt : .t Lx =t : l0 / (4.80)

After which net migration would be given by Eq. 4.78.

Unless the year y census to which net migration from birth was being estimated
happened to have taken place at midnight on 31 December, the birth cohorts
corresponding to age cohorts at the census for which migration estimates were
sought would be non-calendar year birth cohorts. Unless detailed data on births
by month/week/day were available, the sizes of these cohorts would need to be
estimated from data on births during calendar years by assuming even distributions
of births by date of occurrence through those years.
Using Life Tables: Survival 191

It is perhaps worth concluding this section by recalling that life tables usually are
constructed separately for males and females. Thus in using life tables to estimate
intercensal net migration, to estimate births or to project a population we normally
have to deal separately with the male and female components of the population.

Survival Between Birthdays (Exact Ages): Joint Survival

We have looked at survival from one age group to another, and at survival from
birth (exact age 0) to an age group. In the former instance a life table survival ratio
was the ratio of two Lx values; in the latter it was the ratio of an Lx value to l0 . It is
also possible to calculate survival ratios which measure the probability of surviving
between two birthdays, or exact ages, and consistent with the principle enunciated
earlier, such survival ratios are the ratio of two life table lx values. The probability
of surviving from one’s xth birthday to one’s x C nth birthday (from exact age x to
exact age x C n) is given by:

n SRx D lxCn =lx (4.81)

This application of the life table tends to be of actuarial significance more than of
demographic significance, but it does raise the issue of joint survival. Joint survival
refers to the probability that two (or more) persons will both (all) survive over a
given period, or indeed that they will survive and fail to survive in some nominated
combination. It is measured by the product of the probabilities of survival or
failure to survive for each individual person.
As an example, suppose a man is now aged 30 years 3 months and his wife is
aged 27 years 9 months. What is the probability that the husband will survive to
retirement age (his 65th birthday) with his wife still alive? The probability that the
husband survives a further 65–30.25 D 34.75 years is given by:

34:75 SR30:25 D l65 =l30:25

Where l65 and l30.25 are taken from an appropriate life table for males.
The probability that the wife survives over the same period of 34.75 years, at the
end of which she will still be 2 years 6 months younger than her husband is given
by:

34:75 SR27:75 D l62:5 =l27:75

Where l62.5 and l27.75 are taken from an appropriate life table for females.
To find the probability of joint survival to the husband’s retirement we would
evaluate these two survival ratios and multiply them together. The spectre of lx
values where x is not a whole number will be new. They can be evaluated by linear
192 4 Analysis of Mortality: The Life Table and Survival

interpolation between the lx values of a single-year-of-age life table, since other than
at very young ages it is reasonable to assume linear survivorship between exact age
x and exact age x C 1. So, for example, l30.25 is one-quarter of the way between l30
and l31 .
If you are having difficulty imagining ‘demographic’ applications of this type
of calculation, think in terms of studies of phenomena such as widowhood,
orphanhood, the likelihood as a parent of experiencing the death of a child, and the
effects of mortality on family composition (e.g., in producing sole parent families).
These are all topics that can be greatly illuminated by illustrative joint survival
calculations, especially comparatively over time or between different populations
at a point in time.

Further Issues in the Analysis of Mortality

In the final section of this chapter we look briefly at several further issues that
become relevant as mortality analysis is expanded beyond the basics covered thus
far. The aim is simply to alert you to these issues, not to develop them in detail. The
issues in question are: demographers’ interest in causes of death when undertaking
detailed studies of mortality trends and patterns; the nature of differential mortality
analysis; the concept of perinatal mortality; more sophisticated types of life tables;
the concept of years of potential life lost; and approaches to studying not just
the quantity of life lived (survivorship), but the quality of life lived (i.e., taking
morbidity as well as mortality into account in assessing the health of populations).

Causes of Death

It is a feature of detailed studies of mortality trends and patterns that they often focus
on different causes of death. In theory, at least, such a focus has the potential to help
explain changes in mortality levels apparent from more general studies of life tables.
One can sometimes identify increases or decreases in mortality from particular
causes as the major contributors to general mortality trends, thereby enhancing
understanding of those trends. It may even happen that a cause-of-death study may
show rather more change in the pattern of mortality than an examination of all-
cause life tables would lead one to suspect, with offsetting changes having occurred
in different cause categories. Cause-specific studies of mortality, in which the focus
is exclusively on deaths from a particular cause or group of causes, are also possible.
The obvious example in recent decades has been studies of HIV-AIDS, which has
wreaked havoc with the mortality levels and patterns, and hence with the life tables,
of many developing countries since the 1980s.
Classifications of death by cause are normally based on an international classifi-
cation standard established by the World Health Organization (WHO). The WHO’s
Further Issues in the Analysis of Mortality 193

International Classification of Diseases, Injuries and Causes of Death (ICD) is

revised every 10 years, a reality that can create complications for trend analysis
of cause of death data. Deaths that over one period were allocated to one category
suddenly, because the opinion of medical personnel with input into the classification
system has changed, are allocated to another, and it is not always possible to
reclassify data on a consistent basis for analysis purposes. Thus, always approach
time series data on causes of death, especially detailed time series data, with caution,
and check the consistency of the basis of classification over time. Be especially
suspicious of sudden change in the level of mortality from a particular cause; the
first question to ask yourself in such a circumstance is, ‘Is this change likely to be
an artefact of a change in classification procedures?’ Classifying deaths by cause
can be problematic at the best of times, but when deaths that once went into one
category suddenly are switched to another life becomes doubly difficult.
Many countries, including most African countries and some Asian countries, do
not maintain systems of death certification requiring a medical practitioner to fill
out a ‘Certificate of Cause of Death’. For such countries comprehensive analyses of
mortality by cause of death are obviously impossible. Some data may be available
from records of individual medical establishments, but the potential for bias in these
due to factors such as their pertaining to geographically local areas and perhaps only
to persons who die in a hospital is acute.
A second problem with cause of death data is that the quality of classification
by cause of death can be extremely variable, both from country to country and over
time. Data on causes of death published in the 2011 edition of the United Nations
Demographic Yearbook are illustrative. The fairly broad ICD categories used
include one labelled ‘Symptoms, signs and abnormal clinical laboratory findings
not elsewhere classified’. Variable proportions of deaths allocated to this essentially
‘dump’ category are reflective of the variable quality of cause of death certification
in different countries. In Australia in 2006, for example, just 1.2 % of male deaths
and 1.0 % of female deaths were allocated to this category. In South Africa in 2009
the equivalent figures were respectively 12.6 % and 14.8 %, in El Salvador in 2008
they were 14.2 % and 19.5 %, in Sri Lanka in 2006 they were 13.4 % and 21.0 %,
and in Thailand in 2006 they were fully 33.5 % and 44.7 %. These differences
are clear evidence of considerable disparity in the capacities and/or willingness
of the medical systems in different countries to accurately certify causes of death.
Problems can especially arise with deaths in old age, which are apt to be attributed
to medically imprecise ‘causes’ like ‘senility’. Indeed this may explain why, in the
pairs of figures just quoted (except that for Australia) the female figure is always
higher. Female mortality, because women are less prone to premature death through
accidents, violence and conditions that reflect unhealthy lifestyles (smoking, alcohol
consumption, etc.), tends to be more concentrated at advanced ages where deaths
are expected, and being precise as to their causes seems consequently in some
jurisdictions to not be a high priority.
Do not imagine, however, that even data for a country like Australia are problem-
free. A person may suffer from several medically adverse conditions in the period
immediately prior to death. In such circumstances the medical practitioner certifying
194 4 Analysis of Mortality: The Life Table and Survival

the death, while perhaps listing a range of causes, is normally required to nominate
from among them the underlying cause of death, which is the one taken notice of
in classifying deaths by cause. It should be recognised that this choice may be a
matter of medical opinion or even guesswork. Then again, a doctor may describe in
intricate medical detail the injuries sustained by an accident victim or a victim of
violence, without saying anything about the cause of the accident or the nature of the
violence which occasioned death. A statistical agency may have other information,
such as a coroner’s report, to assist in detecting and correctly classifying such
cases, but otherwise misclassification can easily occur. Suicide can be an especially
difficult cause of death to detect – if someone deliberately has a fatal motor vehicle
accident, for example, and fails to leave any indication of premeditation, their death
will probably be deemed to be due to a motor vehicle accident rather than to
suicide. We can also conceive of the notion of the ‘underlying cause’ of a death
operating at different levels. At one level the cause may be long-term exposure
to an environmental hazard – radioactivity or the dust encountered in working as
a miner, for example – or the long-term practice of a dangerous personal habit –
smoking or excessive drinking, perhaps. At another level the cause may be poverty,
which precluded access to an unpolluted water supply, to adequate knowledge of
preventive health behaviour, and to basic medical facilities. At yet another level the
cause may be cancer, cholera or coronary disease, or at still another it may be that
the heart stopped beating and the brain ceased its activity. Thus, even in the best of
statistical environments, cause-of-death statistics have a certain fuzziness about
them, and it is important to deal with them warily because of that.
In making comparisons of mortality levels by cause of death it is always impor-
tant to be aware of ‘differences’ that are wholly, or substantially, due to differences
in age structure. Cause-specific death rates should routinely be standardized for
age, for just as general mortality levels are highly dependent on age structure,
so, too, are rates of death from particular causes. In a comparison of cause of
death data for Sri Lanka in 1986 and Australia in 1994, for example, the female
death rate from ‘Other diseases of the digestive system’ was 11.0 per 100,000
population in Australia and 5.3 per 100,000 in Sri Lanka. Age-specific rates were,
however, except at ages 75 and over, all higher for Sri Lanka. The exception at
the oldest ages was probably due partly to a tendency to attribute deaths at those
ages to ‘senility’ in Sri Lanka, and partly to an older age structure within the
‘75 and over’ age category for Australia. But even setting these considerations
aside, standardization of the Australian death rate to the age structure of the Sri
Lankan population reduces it to 3.5 per 100,000, well below, rather than more than
double, the Sri Lankan rate. You will often, in cause of death and other mortality
studies, encounter the standardized mortality ratio, a measure introduced in Chap.
2 as a particular example of standardized ratios, measures that facilitate legitimate
comparison of summary measures after controlling for differences in population
composition. The standardized mortality ratio for mortality from ‘Other diseases
of the digestive system’ among Australian and Sri Lankan females in the present
example was 3.5/5.3 D 0.66; i.e., the level of mortality from this cause in Australia
in 1994 was two-thirds of what it was in Sri Lanka in 1986. The unstandardized
Further Issues in the Analysis of Mortality 195

mortality ratio was 11.0/5.3 D 2.08, suggesting falsely that female mortality from
this cause in Australia was more than double what it was in Sri Lanka.

Analysis of Differential Mortality

Studies of differential mortality establish and attempt to explain different levels of

mortality in discrete subgroups of a population. This notion really gets back to
compositional variables other than age and sex by which it was noted earlier that
mortality levels often vary. Thus, studies have established differential risks of dying
by ethnicity, geographic location (reflecting environmental factors), occupation
(reflecting risk factors associated with particular types of employment), marital
status (marriage has been claimed to select the healthy), education (the better
educated supposedly are better informed in matters of preventive health care
and more knowledgeable of how to seek medical help when needed), parental
educational attainment (which can affect infant and child mortality in similar ways),
and income (which affects ability to access medical services). In high income
countries mortality differentials may also be associated with what have been termed
‘social roles’, or particular lifestyles. These may reduce or increase risks of death
from conditions related to, for example, excessive alcohol consumption, smoking,
the taking of other drugs and diet, and from involvement in accidents and violence
of various descriptions.
Studies of differential mortality are undertaken largely using techniques of
standardization and life table measures, so you are already quite well equipped
to tackle them. Some do, however, make use of more sophisticated life table
techniques – for example, multiple increment-decrement life tables which, in
relation to a variable like marital status, permit changes of marital status over the
life cycle to be dealt with in conjunction with cross-marital status differences in
mortality risks. These are dealt with briefly below.

Perinatal Mortality

At the beginning of this chapter we defined death as the cessation of life in a person
born alive. Demographers also have an interest in foetal deaths. More commonly
referred to as stillbirths, these are defined as involving children born weighing at
least 500 grams (or if body weight is unavailable, at least 22 weeks gestation) who
did not, at any time after being born, breathe or show any other sign of life. The
arbitrary limits of 500 g or 22 weeks gestation are supposed to signify theoretical
viability – a theoretical ability to survive outside the uterus.
One reason demographers take an interest in stillbirths is that whether or not
a child breathes after birth is often a fine line; there frequently isn’t much to
distinguish a stillbirth from an early infant death, and it can be debatable which
196 4 Analysis of Mortality: The Life Table and Survival

category a particular child should be classified to. Treating the two groups together
conveniently avoids this classification problem. Secondly, stillbirths and very early
infant deaths often have similar causes, and thus are likely to respond similarly to
antenatal medical initiatives, be they ones that enhance survival chances or ones
(such as the prescribing of thalidomide to pregnant women to alleviate morning
sickness in the 1950s and early 1960s) that have the opposite effect (thalidomide
was later shown to cause severe birth defects). On both of these counts, too, their
simultaneous treatment is recommended.
We have noted several times the tendency for infant deaths to concentrate very
early in the first year of life. Those that occur within 28 days following birth are
given a special name. They are termed neonatal deaths, and within this category
those that occur within 7 days of birth are termed early neonatal deaths. In the
broadest definition, it is to the amalgam of foetal deaths (stillbirths) and neonatal
deaths that the term perinatal mortality is given. By this definition a perinatal death
is a foetal death (stillbirth) or a neonatal death. One does, however, encounter
variations in the definition of a perinatal death. Sometimes, for example, only early
neonatal deaths are included, while in the other direction only late foetal deaths
(those occurring after the 28th week of gestation) may be defined as perinatal. Two
things follow. First, in general terms perinatal mortality is concerned with deaths
of children occurring just before, during, and just after childbirth (or parturition).
Second, one must always be careful to check the specific definition underlying
perinatal mortality data, especially if comparative analysis between populations or
over time is contemplated.
Perinatal death rates are ratios of perinatal deaths to total births (i.e., live births
plus stillbirths). They often are divided into stillbirth (or late foetal death) rates and
neonatal (or early neonatal) death rates, but the latter can be a source of confusion.
Just as, in the broadest definition, perinatal deaths can be divided into foetal deaths
and neonatal deaths, so, too, infant deaths can be divided into neonatal deaths and
post-neonatal deaths. When neonatal (or early neonatal) death rates are calculated
in the contexts of perinatal and infant mortality their denominators are different,
respectively including and excluding stillbirths (or late foetal deaths). Thus, you
should always check which denominator was used to calculate such death rates. It
is more usual for them to be calculated as a component of infant mortality, and
hence relative to live births only, but this is not invariably the case. Be careful, in
particular, not to compare neonatal or early neonatal death rates when one figure has
a live births denominator and the other has a total births denominator.

More Sophisticated Life Tables

We have dealt in this chapter with standard, or single-year-of-age, life tables and
abridged life tables as vehicles for the study of mortality, or survival. These are
examples of what are known as single decrement life tables, which deal with an
Further Issues in the Analysis of Mortality 197

attrition process (recall Chap. 3) in which only one transition is of interest – from
the transient state ‘alive’ to the absorbing state ‘dead’. The state ‘alive’ is said to be
‘transient’ because nobody remains in it indefinitely. The state ‘dead’ is by contrast
‘absorbing’ because, once entered, there can be no transition back in the reverse
direction (at least not for mere mortals).
Without going into detail it is appropriate at this juncture to alert you to two
things, one of which was touched on in passing in Chap. 3. First, life tables can
be used to study processes other than mortality. Thus when we discussed attrition
in Chap. 3 the example we used focused on breastfeeding mothers transitioning
from being breastfeeders to having terminated breastfeeding. We could similarly
use life table techniques to study the process of transition with age from being never
married to being married, or the process of transition with marriage duration from
being married to being divorced. Second, life tables can be used to study processes
where transitions to more than one absorbing state are possible, and also to study
processes allowing transitions to and fro between multiple transient states as well
as to one or more absorbing state. Life tables that permit the former are known as
multiple decrement life tables. Those that permit the latter are multistate life tables.

Multiple Decrement Life Tables

The notion of multiple decrement life tables really comes to mind once one
appreciates that life tables can be used to study processes other than mortality.
Clearly, for example, if we are thinking about attrition from a radix of never married
males or females through marriage, there is the possibility, indeed the near certainty,
that some will be lost along the way to death prior to marriage. Hence we are
thinking in terms of a double decrement, and in the case of marriage about what
is known as a net nuptiality table – one that captures the attrition of a never married
radix population through either marriage or death as a single person. Sooner or
later everyone ceases to be a member of the never married population, and they do
this either by marrying or by dying. Marriage is an absorbing state in this context
because, once married, one can never revert to being ‘never married’.
A common application of multiple decrement life tables (three or more modes
of decrement) is to the study of mortality patterns by cause of death. Instead of
simply studying the transition from ‘alive’ to ‘dead’ we can study transitions from
‘alive’ to ‘dead from cause A’, ‘dead from cause B’, ‘dead from cause C’, etc. These
alternative absorbing states are known as competing risks.
Construction of a multiple decrement life table that recognizes different causes of
death, or groups of causes of death, begins with an all-causes standard or abridged
life table for the relevant population. The values of n dx in this life table are then
partitioned into values n dx,1 , n dx,2 , : : : n dx,c (where c is the number of discrete cause
of death categories to be recognized in the analysis) on the basis of the proportions of
total observed deaths at each age attributed to each cause in the relevant population
in the year (or group of consecutive years) to which the all-causes life table pertains.
198 4 Analysis of Mortality: The Life Table and Survival

Thus:

n dx;c D n dx .n Dx;c =n Dx / (4.82)

Where D stands for observed deaths in the relevant population.

These columns of life table deaths partitioned by cause are then used to construct
life tables conditional on dying from each cause. The radices (plural of ‘radix’) of
these conditional life tables are given by the sums of the relevant columns of life
table deaths, which means that they sum to 100,000, the radix of the initial all-cause
life table (i.e., l0 is partitioned into additive components l0,1 , l0,2 , : : : l0,c ). Other
values in these columns are then found from:

lxCn;c D lx;c  n dx;c (4.83)

with columns so generated describing the survival experience of persons who

eventually die from cause of death c.
Cause-specific probabilities of dying are given by:

n qx;c D n dx;c =lx (4.84)

Note that the denominator on the righthand side of this equation is lx (from the
initial all-causes life table), not lx,c (from the cause-specific life table for cause c). It
follows, since the sum of values of n dx,c across all cause of death categories c is n dx
from the initial all-causes life table, that the sum of values of n qx,c across all cause
of death categories c is n qx from the initial all-causes life table. That is, for a given
age interval the probabilities of dying by cause sum to the all-causes probability of
dying.
Other columns of the cause-specific life tables (n Lx,c , Tx,c and eo x,c ) are generated
using the same equations as are used to generate equivalent all-cause life table
functions, but with all terms being specific for cause c. Columns of eo x,c values
give expectations of life remaining for persons who will ultimately die of cause c.

Multistate Life Tables

Multistate life tables are mentioned only in passing. Also known as increment-
decrement life tables, they are a type of life table you may encounter reference to in
the literature, so it is as well to have a basic understanding of what they are. Their
distinguishing feature is that they have two or more transient states, for each of
which transitions both to and from other transient states (although not necessarily
all of them) are possible. They may also feature one or more absorbing states, but
the label ‘increment-decrement life tables’ captures their distinctive characteristic –
their allowing for flows of population both into as well as out of some states, instead
of in only one direction.
Further Issues in the Analysis of Mortality 199

Multistate life tables are not used for studies of mortality per se, because death is
the ultimate absorbing state. They can be used to study health status, with population
moving between states defined by being diagnosed with, or declared recovered from,
various health conditions, as well as, potentially, to the absorbing state deceased.
More commonly, however, they are used in analyses of (i) marital statuses and
relationships (where individuals, over time, can move into and out of states such
as ‘married’, ‘divorced’, ‘widowed’, ‘in a relationship’ and ‘not in a relationship’
multiple times), (ii) internal migration (where multiple movements into and out of
geographic regions are possible), and (iii) labour force participation (where people
can move into and out of the labour force, and back and forth between full-time
and part-time employment, multiple times over their working lives). Multistate life
tables designed to study labour force activity are often known as working-life tables.

Competing Risks and Cause-Eliminated Life Tables

The decrement probabilities in multiple decrement life tables are described vari-
ously as crude, or dependent probabilities. Their values for a particular mode of
decrement, or risk, depend partly on those for the competing risks. If probabilities
of succumbing to one risk change over time or differ between two populations,
that has implications for probabilities of succumbing to other risks. Where the
competing risks are alternative causes of death, for example, because everyone must
die from one cause or another, rising or higher dependent probabilities of dying from
one cause or group of causes must be compensated by falling or lower dependent
probabilities of dying from one or more other causes or groups of causes, and vice
versa.
Often the real interest is in the independent or underlying probabilities asso-
ciated with a particular risk; that is, in the probabilities that would apply in
the absence of competing risks. In the context of the analysis of mortality by
cause of death, such probabilities, and cause-eliminated life tables constructed
using them, provide a sounder basis for comparison of mortality from particular
causes in different populations or over time. They also provide a sounder approach
to the assessment of past mortality patterns as a basis for projecting cause-
specific mortality, and thereby future life expectancy. Independent probabilities of
dying from a particular cause are invariably higher than corresponding dependent
probabilities, since people die from that cause who, in the presence of competing
risks, would die earlier from some other cause.
Cause-eliminated life tables of the type just described eliminate all causes
of death other than the one that is the focus of attention. Other variants of
this procedure are also followed. In particular it is common to construct life
tables in which individual causes are eliminated, so that improvements in life
expectancy that would result from elimination of that cause, or alternatively the
reduction in longevity to which the cause gives rise, can be assessed. The effects
of cause-elimination on mortality from other causes can also be studied, as can the
implications of partial elimination of individual causes (addressing questions of the
200 4 Analysis of Mortality: The Life Table and Survival

type ‘What would be the gain to life expectancy if mortality from cause x could
be reduced by y percent at all ages?’). For many if not most causes, reduction in
mortality will be a more realistic scenario to examine than complete elimination.
It is also, of course, possible to conceive of cause-specific mortality rising, and
of wanting to assess the implications of this type of scenario, too. The obvious
contemporary example has been projecting the impact on longevity of the spread of
HIV/AIDS, although many countries in recent years have been bringing epidemics
more under control.
The mathematics associated with the procedures just summarized are not
straightforward, and will not be canvassed here. Anyone keen to pursue them is
referred to Chiang (1968: Chapter 11), Smith (1992: 159–169) and Siegel and
Swanson (2004: 329–331).

Years of Potential Life Lost (YPLL)

YPLL is a measure that is sometimes calculated for particular causes of death or

groups of causes to capture the aggregate loss of life due to a cause or group of
causes. It originated with the argument of Dempsey (1947) that measuring time
lost due to mortality provided more insight into the burden it imposed on society
than simply relying on crude or age-standardized death rates, which captured only
the amount of death in a population. YPLL recognized that, beyond the mere
occurrence of a death, the age at which it occurred had an important bearing on its
societal implications. The method weighted deaths at younger ages more heavily
than those at older ages and thus became a favoured metric for, in particular,
advocates for a greater share of health resources to address causes of death that
took lives at younger ages. It made those causes look more important in the overall
scheme of competing healthcare demands than they otherwise would.
The literature contains a number of methodological variants of YPLL measures
(Gardner and Sanborn 1990). In its most straightforward form, the number of deaths
from a cause (or group of causes) at each single year of age is multiplied by the
average expectation of life remaining at the mid-point of that age interval obtained
from an appropriate life table. These values are then summed over all single-year
age intervals for which deaths are recorded to yield total YPLL. Mid-interval life
expectancies typically are not directly available, but can be estimated as averages
of those associated with the upper and lower bounds of each interval. If only an
abridged life table is available the same methodology can be applied using the age
intervals represented in that life table rather than single-year intervals.
This YPLL approach effectively sets what has variously been described as a
‘reference’ age or an ‘upper cut-off’ age for each age interval (the age defining how
long a person would be expected to live in the absence of premature death) as the
interval midpoint plus average expectation of life remaining at that midpoint. Other
approaches have set reference or upper cut-off ages that were constant for deaths in
all age intervals, variously at (i) the expectation of life at birth for the population in
Further Issues in the Analysis of Mortality 201

question, (ii) arbitrary values such as 65, 70, 75 or 85 years, and (iii) ages by which
a nominated percentage of all deaths (e.g., 90 %) have occurred. In addition there
has been debate over whether deaths at age 0 should be excluded (the consensus
seems to be that they shouldn’t) and some studies have focused on loss of working
life, excluding life lost before some minimum working age (e.g., 15 years) as well
as that lost beyond some upper age limit for working life (e.g., 65 or 70).
These methodological variants need to be appreciated, as do some issues to
which they give rise. First, different methodologies are apt to yield different results,
and it is important to be mindful of the particular methodology used. Second, the
younger the reference or upper cut-off age used the greater will be the weighting in
results toward health conditions (e.g., accidents and suicide/homicide) that are more
prominent causes of death at younger ages.
More refined calculations of YPLL make use of life expectancies from cause-
eliminated life tables rather than from all-causes life tables. Strictly speaking, for
a given cause or group of causes, c, one should use in YPLL calculations life
expectancies from a life table in which c is eliminated as a cause. The exercise
is endeavouring to determine how many more years would be lived by a population
if cause c was not a risk, and elimination of that risk will raise life expectancies for
all other risks.
Discounts are sometimes introduced to YPLL calculations, the argument for
them being that a year of life gained (or lost) immediately is of greater value (or
cost) than one gained (or lost) in, say, 20 years’ time. The effects of discounting can
be substantial. Discounting at 3 % per year, for example, reduces the YPLL due to
a male infant death in Australia from 81 years to 30 years. It is therefore important
to be aware of whether, and what, discounts have been applied.
A variant on (total) YPLL is to calculate the cost of different causes of death
in loss of productive life. Some might wish to debate the definition of ‘productive’,
but conventionally productive life is defined as working life, or life before retirement
age (exact age 65). The calculation proceeds exactly as before, except that partial
life expectancies (i.e., expectations of life remaining before exact age 65) are used
instead of total life expectancies. These partial life expectancies at any exact age
are readily calculated. To obtain the average expectation of life remaining before
retirement at exact age x, simply sum, in the relevant life table, the Lx columns
between exact ages x and 65, and divide by lx .

Focusing on the Quality of Life: Health Expectancies

As life expectancies in developed countries have increased to historically unprece-

dented levels in recent decades, demographers and public health practitioners have
become increasingly interested not just in the quantity of life lived (survivorship),
but also in its quality. There is interest in the extent to which total life expectancies
represent, on the one hand, healthy life, and on the other, health-impaired life.
Rising life expectancies give rise to questions as to the extent to which these
202 4 Analysis of Mortality: The Life Table and Survival

increases in longevity represent quality life, or merely extensions of health-

impaired life. And more recently burden of disease concepts have sought to extend
assessments of the health of populations beyond life table-based analyses of cause-
specific mortality to also take into account non-fatal health conditions.
Attempts to introduce notions of quality of life into assessments of population
health date from a paper by Sanders (1964), in which he wrote, after expressing
dissatisfaction with existing approaches to measuring community health levels:
It is proposed, therefore, that we develop a modified life table method of analysis in
measuring comparative adequacy of health services for different population groups. In such
an analysis we would not only determine for each age the probability of survival, but also
the subsidiary probabilities of those surviving on the basis of their functional effectiveness.
This would range from individuals who are completely dependent on others, even for
carrying on their daily living activities, to those fully equipped to carry on with no apparent
handicaps all the functions characteristic of their age and sex. (Sanders 1964: 1067)

The first method for calculating health expectancies as foreseen by Sanders

was proposed by Sullivan (1971). Subsequent methodological developments can be
traced in publications such as those by Mathers et al. (1994), Robine et al. (1999),
Murray et al. (2002), Robine et al. (2003), Albert et al. (2008), Guillot and Yu (2009)
and Bruggink (2011). At the core of the concept of health expectancy, or disability-
free expectation of life as it is sometimes called, is the idea of taking the life table
notion of total expectation of life, whether at birth or at some other exact age (e.g.,
65), and splitting it into expected years of disabled, or unhealthy life and expected
years of disability-free, or healthy life – healthy life years (HLY). The former
health status category is also often further split by, for example, distinguishing (in
ascending order of severity) ‘impairment’ from ‘disability’ and ‘handicap’ (various
other labels are also given by different studies to subcategories of unhealthy life –
e.g., ‘moderate activity limitation’ and ‘serious activity limitation’). The splitting
process uses age-specific rates of disability (and impairment, handicap etc.) from
cross-sectional surveys to partition the Lx , and hence the Tx , values in a population’s
life table. This in turn allows total life expectancy to be partitioned by health status
according to:

eo x;h D Tx;h =lx (4.85)

Where h is a health status category.

A major field of debate in this type of analysis is, of course, how one defines
and measures health status concepts such as ‘disability-free’ or ‘healthy’ on the
one hand, and ‘impairment’, ‘disability’, ‘handicap’ and other subcategories falling
under the rubric ‘unhealthy’ on the other. Does one, for example, allow individuals
to self-select into these categories, or does one attempt to impose more objective
classificatory criteria? If the latter, what questions should be asked as the basis
for classification? It naturally follows from these issues that comparability between
studies of health expectancy is a major consideration. Full comparability requires
common health status categories defined in a standard way based on a standard
survey questionnaire.
Further Issues in the Analysis of Mortality 203

As an example of the sorts of findings health expectancy analysis can yield, data
for the 27 countries of the European Union in 2010 showed that while male life
expectancy at birth ranged from 67.77 years in Lithuania to 79.79 years in Italy
(a difference of 12.02 years) and female life expectancy ranged from 77.37 years
in Bulgaria to 85.32 years in Spain (a difference of 7.96 years), healthy life
expectancies at birth ranged from 52.33 years in Slovakia to 71.68 years in Sweden
(males – a difference of 19.35 years) and from 52.04 years in Slovakia to 71.64 years
in Malta (females – a difference of 19.60 years). Thus the cross-national disparity in
healthy life expectancy was considerably greater for both sexes than that in overall
life expectancy. The analysis further split expectation of life with a disability into
two subcategories – expectation with ‘moderate activity limitation’ and expectation
with ‘severe activity limitation’. Life expectancy at birth with moderate activity
limitation ranged from 15.31 years for Slovenia to 4.71 years for Sweden (males –
a threefold disparity) and from 18.80 years for Finland to 7.17 years for Sweden
(females – well over a twofold disparity). Similarly life expectancy at birth with
severe activity limitation ranged from 7.98 years for Slovenia to 1.88 years for
Bulgaria (males) and from 10.18 years for Slovenia to 2.45 years for Bulgaria
(females). The figures for Bulgaria are suspiciously low given its comparatively
low overall life expectancies, unless of course people with severe activity limitation
simply don’t survive long in that country.

Disability-Adjusted Life Years (DALYs) and the Burden

of Disease

Developed by Christopher Murray and Alan Lopez (Murray 1994; Murray and
Lopez 1994, 1996; Lopez 2005), although with roots in earlier literature, the
disability-adjusted life year, or DALY, has been adopted by WHO as a key public
health metric combining morbidity and mortality in a single index. It has been
an evolving metric, its elaborate methodology subject to constant review and fine
tuning in response to the intense debate generated since its initial development
in the early 1990s (see Murray (1996) and Murray et al. (2012a), including their
substantial appendix). Thus not all studies based on DALYs define them in exactly
the same way. The measure is designed to give effect to the concept of the burden
of disease (BOD – sometimes extended to the burden of disease and injury) in a
population, either overall or from particular health conditions or health risk factors,
or groups of conditions or risk factors. There have thus far been three Global Burden
of Disease (GBD) studies, GBD 1990, GBD 2000 and GBD 2010. A summary of
methodological developments across these studies can be found in the appendix
to Murray et al. (2012a), but note that GBD 2010 made a point of generating
comparable results for 1990 as well as 2010 (Lim et al. 2012; Murray et al. 2012b;
Salomon et al. 2012b; Vos et al. 2012).
204 4 Analysis of Mortality: The Life Table and Survival

Burden of disease can be thought of as a measure of the gap between current

health status and an ideal health status under which individuals universally live to
old age free from disease and disability. It is promoted as an aid in setting health
service and health research priorities and in identifying disadvantaged groups and
targeting health interventions appropriately, and as providing a measure of output
for intervention, program and health sector evaluation that facilitates legitimate
comparison (before/after, through time and across populations). The quest that
yielded DALYs as a measure had three main objectives. The first was to facilitate
the proper incorporation of non-fatal health outcomes into health policy debate.
Some of these outcomes, such as depression and musculoskeletal conditions, ranked
far lower as health priorities than they should have when those priorities were
set on the basis of mortality data alone. The second objective was to separate
epidemiological assessment from advocacy by providing an objective measure of
the mortality/disability arising from a health condition or risk factor. This was
desirable to place in perspective the arguments of lobby groups with vested interests
seeking funding to address particular health conditions or risk factors, and thereby
avoid their advocacy skills dominating and distorting debate. The third objective was
to produce a measure that could be used for both quantifying burden and analysing
the cost-effectiveness of intervention options.
Broadly speaking, the number of DALYs for a disease, health condition or health
risk factor is the sum of (i) years of life lost due to premature mortality caused by
that disease, condition or risk factor (YLLs) and (ii) years of healthy life lost due
to disability attributable to the disease, condition or risk factor (YLDs). The DALY
thus extends the concept of potential years of life lost due to premature death, which
it refines to also include years of ‘healthy’ life lost through being in poor health or
disabled. One DALY is one lost year of ‘healthy’ life. In case you are wondering
what exactly is meant by ‘health risk factors’, the ten globally most significant
among 67 that were identified in GBD 2010 were, in rank order by number of
lost DALYs, high blood pressure, tobacco smoking (including passive inhalation of
second-hand smoke), household air pollution from solid fuels, diet low in fruits,
alcohol use, high body-mass index, high fasting plasma glucose (indicative of
diabetes), childhood underweight, ambient particulate matter pollution, and physical
inactivity and low physical activity. This rank ordering varied (i) from those for
males and females separately and (ii) from that for 1990 (Lim et al. 2012: Figures 1
and 2).
The methodology for calculating DALYs seeks to embrace four general concepts.
First, as far as possible, any health outcome that represents a loss of welfare
should be included as an indicator of health status. The GBD ‘cause list’ identifies
discrete conditions in a four-level hierarchy that begins at level 1 with three broad
categories. These then subdivide into level 2, 3 and 4 subcategories. The base-
level categories are (i) communicable, maternal, neonatal and nutritional disorders,
(ii) non-communicable diseases and (iii) injuries. Level 2–4 subcategories have
multiplied over time from 131 in GBD 1990 to 159 in GBD 2000 and 288 in
GBD 2010. Beyond these subcategories ‘sequelae’ are also identified (i.e., direct
consequences of disease or injury that are not otherwise captured in the cause list) –
Further Issues in the Analysis of Mortality 205

483 of them in GBD 1990, 474 in GBD 2000 and 1,160 in GBD 2010 (Murray et
al. 2012a: Appendix). Second, the characteristics of an individual affected by a
health outcome that should be considered in calculating the associated burden
of disease should be restricted to age and sex alone. Third, like health outcomes
should be treated alike – i.e., the premature death of a person of a given age and
sex should contribute equally to the global burden of disease regardless of where,
geographically, that person lives (Dhaka slum or salubrious Sydney suburb). This
concept, advanced for GBD 1990, was modified for GBD 2010 by adjusting for
‘comorbidity’ – i.e., recognizing that some individuals’ health loss has multiple
contributing conditions, and that their health loss should be apportioned among
those conditions rather than each condition being assessed separately, leading to
exaggerated estimates of health loss at the individual level. Finally, the unit of
measurement for the burden of disease should be time. The second and third
of these concepts in particular rule out discrimination by socio-economic status,
whether in favour of the better off, on the ground that they contribute more to
societal wellbeing, or in favour of the disadvantaged, on the ground that their very
disadvantage justifies such favouritism. Murray (1996: 7) observes that they ‘give
DALYs a strongly egalitarian flavor.’ Another debate concerning GBD methodology
has been over whether incidence or prevalence measures (see Chap. 1) should be
used in assessing disability from particular causes. The initial position of Murray
(1996) was that death rates were incidence rates and measures of prevalence made
no sense in a mortality, or YLL, context. Therefore consistency recommended also
using an incidence approach in performing YLD calculations. However, by GBD
2010 the debate had swung strongly in favour of taking a prevalence approach to
measuring YLDs (Murray et al. 2012a: Appendix).
Five important ‘social choices’ are built into the construction of DALYs.
1. Estimating the duration of life lost due to a death at each age. Since Dempsey
(1947) first raised the issue of time-based measures of premature mortality, an
array of possibilities have been suggested. Two of these were the alternative
measures of YPLL discussed above – the one where life lost is determined
with reference to a fixed upper cut-off age for deaths at all ages and the one
where it is set for deaths at each age as the midpoint of the age group plus
average remaining expectation of life at that age. Murray (1996) refers to these,
respectively, as ‘potential years of life lost’ (PYLL) and ‘period expected years
of life lost’ (PEYLL). A third option he considers is ‘cohort expected years of
life lost’ (CEYLL), which differs from PEYLL only in that the life expectancies
employed are from a relevant cohort life table rather than a standard, or period
(cross-sectional) life table. The argument for this approach is that twentieth
century experience is that cohort mortality beyond a given age rarely matches
what a standard life table constructed at the time a cohort reaches that age
predicts, because mortality in most populations has been declining (and hence
survival prospects have been improving) over time. However, Murray rejects
both PEYLL and CEYLL as violating the concept that like health outcomes
should be treated alike, since life expectancies employed in respect of deaths
206 4 Analysis of Mortality: The Life Table and Survival

at a given age for a given sex will vary from population to population. Instead
he opts for ‘standard expected years of life lost’, or SEYLL, in estimating the
YLL component of DALYs. This approach uses an ideal standard life table
to represent the level of survivorship corresponding to optimal health and to
provide, for all populations of a given sex, the remaining life expectancy lost
through a premature death at each age. The approach allows deaths at the same
age in all populations to contribute equally to the burden of disease in that
population, so that like outcomes are treated as like. For GBD 1990 the female
and male standard schedules of remaining life expectancies were derived from
Coale and Demeny ‘West’ model life tables for females with life expectancies at
birth of 82.5 years and 80.0 years respectively. These were chosen on the basis
that (i) the highest achieved life expectancy at birth at the time, for Japanese
females, was over 82 years, and (ii) there was solid evidence that a comparable
male life expectancy should be a little lower. The use of a female model life table
to set the male standard schedule reflected the fact that no male model table with
a life expectancy as high as 80 existed in the Coale and Demeny system at that
time (although it was in the process of being added (Coale and Guo 1989)). By
the time of GBD 2010, however, a single standard, or reference, life table with a
life expectancy at birth of 86.0 years had been developed and was used for YLL
calculations for both sexes (Murray et al. 2012a).
2. Comparing time lost due to premature death and time lived with a non-
fatal health outcome. The task of quantifying time lived with non-fatal health
outcomes in a manner comparable to time lost due to premature mortality and
standardized across populations and over time requires considerable simplifi-
cation. While death is death and its measurement is relatively straightforward,
non-fatal health outcomes are each unique in many ways and ‘a certain level of
reductionism must be accepted’ (Murray 1996: 23). One genre of approaches
to conceptualizing non-fatal health outcomes is a huge array of ‘health-related
quality of life’ (HRQL) approaches. This genre was rejected as a basis for
measuring the YLD component of DALYs on several grounds, but not least
its substantial reliance on self-reports of health status. These have commonly
yielded counter-intuitive patterns that suggest the measures in question do
not meet the criterion of treating like outcomes alike in all populations. The
alternative conceptualization that informs measuring YLDs is embedded in
the International Classification of Impairments, Disabilities and Handicaps
(ICIDH), a WHO classification that envisages a linear progression from ‘disease’
to ‘pathology’ to ‘manifestation’ to ‘impairment’ to ‘disability’ and then to
‘handicap’. The key distinctions are between impairment (defined at the level
of the organ system), disability (defined in terms of impact on the performance
of the individual) and handicap (defined in the context of overall consequences,
which depend on a person’s social environment). Measurement of the YLD
component of DALYs is focused on the measurement of disability, not handicap
(whose social environment element is incompatible with treating like outcomes
alike), and also focused wherever possible on observations about non-fatal health
Further Issues in the Analysis of Mortality 207

outcomes rather than self-reports (although the latter are the only option for
outcomes categorized under ‘pain and suffering’).
The process for calculating YLDs is complex and has undergone significant
revision with each new GBD study in light of debate generated by previous
studies. Central to it is the determination of disability weights for each non-
fatal health state that effectively rank those health states from least to most
disabling on a scale ranging from 0 (no disability – perfect health) to 1 (complete
disability – death). Murray et al. (2012a: Appendix, p. 11) describe disability
weights as ‘the key mechanism in the GBD approach through which disease and
injury sequelae are made comparable with each other and with time lost due to
premature mortality.’ For GBD 1990, after an initial methodology had attracted
criticism (Murray 1996: 34), weights were established using a deliberative
process that first asked a group of health professionals to participate in an
exercise to achieve a consensus ranking of 22 ‘indicator conditions’ chosen
to reflect different dimensions of non-fatal health outcomes by their disabling
severity. ‘Dimensions’ traversed conditions with largely physical manifestations
such as blindness and below the knee amputation, neuro-psychiatric condi-
tions, conditions representing varying degrees of pain, conditions that affect
sexual/reproductive function, and one condition, vitiligo on the face, that has
exclusively social or group interaction consequences (vitiligo causes depigmen-
tation of parts of the skin, thereby affecting physical attractiveness). The ranking
obtained was then used to establish seven ‘disability classes’, each associated
with a discrete range of severity, or disability, weights (Murray 1996: Table
1.4) ranging from class 1, weight range 0.00–0.02 (indicator conditions vitiligo
on the face, weight-for-height lower than two standard deviations below the
mean) to class 7, weight range 0.70–1.00 (indicator conditions active psychosis,
dementia, severe migraine, quadriplegia). Disability weights for hundreds of
other conditions were then agreed on by expert panels with reference to these
disability classes and the weight ranges associated with them (what class did a
condition belong in, then where within the associated range of disability weights
did it fit?).
For GBD 2010 disability weights were comprehensively revised in the wake
of considerable debate about the GBD 1990 weights. The new method (Salomon
et al. 2012a) clarified its objective as being to measure health loss rather
than welfare loss and moved to using the general public (rather than health
professionals or the third option, individuals actually in a particular health state)
to make the comparative assessments required to establish disability weights.
The 1,160 disease and injury sequelae referred to above were mapped into 220
distinct ‘health states’ capturing the most salient differences in symptoms and
functioning, and household surveys in five deliberately contrasting countries –
Bangladesh, Indonesia, Peru, Tanzania and the USA – were used, along with
a web-based survey posted online for almost 9 months, to elicit responses on
paired comparison questions. These presented respondents with descriptions
of two hypothetical people, each with a particular health state, and asked them
which person they deemed healthier. In the household surveys each respondent
208 4 Analysis of Mortality: The Life Table and Survival

was asked to respond in respect of 15 paired comparisons randomly selected

from a pool of 108 health states. These were chosen from the broader pool of
220 on the basis that (i) their salient features could be easily communicated to lay
respondents in 35 or fewer words, and (ii) they could be presented realistically
as chronic states that would persist throughout a lifetime. Respondents to the
web-based survey were randomly assigned to one of four survey versions,
one of which mimicked the household survey while the others focused on
different subsets of the 220 health states. One of these was a subset of 30
health states in respect of which respondents were asked to compare the health
benefits of different life-saving or disease-prevention programs as an aid to
anchoring disability weights to the necessary zero to one scale. There were
13,902 participants in the five household surveys and 16,328 in the web-based
survey. The complex analysis of these data to generate disability weights with
95 % ‘uncertainty’ intervals about them (another new innovation compared to
GBD 1990) is described in Salomon et al. (2012a), where the weights for the
220 health states are also presented. Health states with the highest disability
weights were acute schizophrenia (0.756), severe multiple sclerosis (0.707),
untreated spinal cord lesion at the neck (0.673), severe epilepsy (0.657), severe
major depressive disorder (0.655), heroin and other opioid dependence (0.641),
severe traumatic brain injury with long-term consequences (0.625) and severe
musculoskeletal problems (0.606). On the other hand 26 % of all health states
had disability weights below 0.05, and 11 % had weights below 0.01. These
were obviously minimally disabling states.
3. Discounting future health. Discounting is an economic concept that reflects the
fact that most people prefer benefits now rather than in the future. There has
been considerable debate among health economists over whether future health
should be discounted in the calculation of YLLs and YLDs, and hence of DALYs.
For summaries of this debate see Murray (1996) and Murray et al. (2012a:
Appendix). In Murray et al.’s view the more persuasive arguments in favour of
discounting future healthy life years focus upon decision-making paradoxes that
without discounting invite unpalatable choices. In particular, the ‘research and
eradication’ paradoxes that, in the absence of discounting, suggest that spending
money on research and eradication could yield extremely large streams of healthy
life years in the future, therefore justifying devoting nearly all resources today to
those areas, at the expense of properly treating those currently suffering from
a health condition. Thus, the arguments for discounting are not arguments that
health in the future should intrinsically be valued less than health in the present.
They are arguments that one should discount to avoid unreasonable sacrifice by
those currently living with a condition.
For GBD 1990 a 3 % discount rate was adopted, but vigorous debate ensued
subsequently that by the time of GBD 2010 had generated a consensus view
that a year of healthy life should be deemed equally important to population
health, regardless of the year in which it was lived. Hence GBD 2010 used a
zero discount rate. The important thing to bear in mind is that whether or not
Further Issues in the Analysis of Mortality 209

one discounts makes a non-trivial difference to YLL, YLD and DALY results.
A 3 % compounding discount rate makes a big difference by the time one gets to
older ages for someone who dies in infancy, and significantly alters the relative
importance attached to deaths in childhood compared to those in adulthood in
BOD studies.
4. Age-weighting. This became an issue for GBD studies on the basis that empirical
evidence from a range of sources showed that people tend to assign greater
importance to saving the lives of adolescents and young adults than to saving
those of very young children and older adults (Murray 1996). In other words,
there is a tendency to value life-years lived at different ages differently. This may
have a variety of rationales for different individuals, but it is likely to reflect such
things as the greater value to society of years lived through the working ages,
when a return is obtained on the investment in a person’s education and people
are productive economically and actively or potentially involved in producing
and raising the next generation, than of those lived through childhood and old age
in varying degrees of dependence. As a result of these considerations GBD 1990
built into the computation of DALYs age weights that assigned greater value to
a year of young or middle-aged adult life than to a year lived as a young child or
one of the elderly. These weights were determined via a continuous mathematical
function that increased from a value of 0 at birth to a peak of around 1.5 at ages
in the mid-20s, then tailed off thereafter to around 0.5 at age 80, and less at still
older, more dependent, ages.
Yet again, however, methodological debate led to simplification of the cal-
culation of YLLs, YLDs and DALYs for GBD 2010. Viewing these as strict
summary measures of population health it was eventually agreed that arguments
for weighting years of healthy life lived at different ages differently were not
as compelling as previously thought. In consequence it was concluded that ‘we
should treat a year of healthy life as equal irrespective of the age at which it
is lived’ (Murray et al. 2012a: 2064). In other words, no age weighting was
employed in GBD 2010.
5. Considerations of equity. In developing the methodology for GBD 1990 there
was consideration given to a couple of issues of equity. Empirical evidence
has suggested that people generally have a preference for a smaller number of
people to receive a substantial health benefit than for a larger number of people
to receive a smaller one. The possibility was explored of incorporating this sort
of distributional preference into DALY methodology, but it was deemed not to be
justified. Another argument has been that DALYs should somehow be weighted
socio-economically; that life-years for the disadvantaged should be more heavily
weighted than life-years for the better off. This proposition, too, was rejected
after consideration, it being concluded that the needs of disadvantaged groups
could be met simply by studying and reporting on differential burden patterns
by socio-economic status. Neither of these issues appears to have been seriously
revisited for GBD 2010.
210 4 Analysis of Mortality: The Life Table and Survival

References

Albert, F. S., Bragg, J. M., & Brooks, J. C. (2008). Health expectancy. Paper presented at the Living
to 100 and Beyond Symposium, Orlando, Florida.
Brass, W. (1971). On the scale of mortality. In W. Brass (Ed.), Biological aspects of demography
(pp. 69–110). London: Taylor and Francis.
Bruggink, J.-W. (2011). Towards a better health expectancy. The Hague: Statistics Netherlands.
Chiang, C. L. (1968). Introduction to stochastic processes in biostatistics. New York: Wiley.
Chiang, C. L. (1984). The life table and its applications. Malabar: Robert E. Krieger Publishing
Company.
Coale, A. J., & Demeny, P. G. (1966). Regional model life tables and stable populations. Princeton:
Princeton University Press.
Coale, A., & Guo, G. (1989). Revised regional model life tables at very low levels of mortality.
Population Index, 55(4), 613–643.
Coale, A.J., Demeny, P.G., & Vaughan, B. (1983). Regional model life tables and stable
populations (2nd ed.). New York/London: Academic.
Dempsey, M. (1947). Decline in tuberculosis. The death rate fails to tell the entire story. American
Review of Tuberculosis, 56, 157–164.
Fergany, N. (1971). On the human survivorship function and life table construction. Demography,
8(3), 331–334.
Flieger, W., & Cabigon, J. (1994). Life table estimates for the Philippines, its regions and provinces,
by sex: 1970, 1980 & 1990. Manila: Department of Health and USAID.
Gardner, J. W., & Sanborn, J. S. (1990). Years of potential life lost (YPLL) – What does it measure?
Epidemiology, 1(4), 322–329.
Greville, T. N. E. (1943). Short methods of constructing abridged life tables. Record of the
American Institute of Actuaries, 32(65), 29–42.
Guillot, M., & Yu, Y. (2009). Estimating health expectancies from two cross-sectional surveys:
The intercensal method. Demographic Research, 21, 503–534.
Keyfitz, N. (1970). Finding probabilities from observed rates or how to make a life table. The
American Statistician, 24(1), 28–33.
Keyfitz, N., & Frauenthal, J. (1975). An improved life table method. Biometrics, 31, 889–899.
Lim, S., Vos, T., Flaxman, A. D., & 206 others. (2012). A comparative risk assessment of burden
of disease and injury attributable to 67 risk factors and risk factor clusters in 21 regions, 1990–
2010: A systematic analysis for the Global Burden of Disease Study 2010. The Lancet, 380,
2224–2260.
Lopez, A. D. (2005). The evolution of the global burden of disease framework for disease, injury
and risk factor quantification: Developing the evidence base for national, regional and global
public health action. Globalization and Health. doi:10.1186/1744-8603-1-5.
Mathers, C., McCallum, J., & Robine, J.-M. (1994). Advances in health expectancies. Canberra:
Australian Institute of Health and Welfare.
Murray, C. J. L. (1994). Quantifying the burden of disease: The technical basis for disability-
adjusted life years. Bulletin of the World Health Organization, 72(3), 429–445.
Murray, C. J. L. (1996). Rethinking DALYs. In C. J. L. Murray & A. D. Lopez (Eds.), The global
burden of disease: A comprehensive assessment of mortality and disability from diseases,
injuries, and risk factors in 1990 and projected to 2020 (pp. 1–98). Boston: Harvard University
Press for the Harvard School of Public Health, WHO and the World Bank.
Murray, C. J. L., & Lopez, A. D. (1994). Quantifying disability: Data, methods and results. Bulletin
of the World Health Organization, 72(3), 481–494.
Murray, C. J. L., & Lopez, A. D. (1996). The global burden of disease: A comprehensive
assessment of mortality and disability from diseases, injuries, and risk factors in 1990 and
projected to 2020. Boston: Harvard University Press for the Harvard School of Public Health,
WHO and the World Bank.
References 211

Murray, C. J. L., Salomon, J. A., Mathers, C. D., & Lopez, A. D. (2002). Summary measures of
population health: Concepts, ethics, measurement and applications. Geneva: WHO Press.
Murray, C. J. L., Ferguson, B. D., Lopez, A. D., Guillot, M., Ahmad, O., & Salomon, J. A. (2003).
Modified logit life table system: Principles, empirical validation and application. Population
Studies, 57(2), 165–182.
Murray, C. J. L., Ezzati, M., Flaxman, A. D., Lim, S., Lozano, R., Michaud, C., Naghavi, M.,
Salomon, J. A., Shibuya, K., Vos, T., Wikler, D., & Lopez, A. D. (2012a). GBD 2010: Design,
definitions and metrics. The Lancet, 380, 2063–2066. Includes a substantial appendix by the
same authors separately titled ‘Comprehensive systematic analysis of global epidemiology:
Definitions, methods, simplification of DALYs, and comparative results from the Global
Burden of Disease Study 2010’.
Murray, C. J. L., Vos, T., Lozano, R., & 358 others. (2012b). Disability-adjusted life years (DALYs)
for 291 diseases and injuries in 21 regions, 1990–2010: A systematic analysis for the Global
Burden of Disease Study 2010. The Lancet, 380, 2197–2223.
Preston, S. H., Heuveline, P., & Guillot, M. (2001). Demography: Measuring and modelling
population processes. Oxford: Blackwell.
Reed, L. J., & Merrell, M. (1939). A short method for constructing an abridged life table. American
Journal of Hygiene, 30(2), 33–62.
Robine, J.-M., Romieu, I., & Cambois, E. (1999). Health expectancy indicators. Bulletin of the
World Health Organization, 77(2), 181–185.
Robine, J.-M., Jagger, C., Mathers, C. D., Crimmins, E. M., & Suzman, R. M. (2003). Determining
health expectancies. Chichester/Hoboken: Wiley.
Salomon, J. A., Vos, T., Hogan, D. R., & 125 others. (2012a). Common values in assessing health
outcomes from disease and injury: Disability weights measurement study for the Global Burden
of Disease Study 2010. The Lancet, 380, 2129–2143.
Salomon, J. A., Wang, H., Freeman, M. K., Vos, T., Flaxman, A. D., Lopez, A. D., & Murray, C.
J. L. (2012b). Healthy life expectancy for 187 countries, 1990–2010: A systematic analysis for
the Global Burden of Disease Study 2010. The Lancet, 380, 2144–2162.
Sanders, B. S. (1964). Measuring community health levels. American Journal of Public Health,
54(7), 1063–1070.
Shryock, H. S., Siegel, J. S., and Associates. (1973). The methods and materials of demography.
2nd Printing (revised). Washington, DC: U.S. Bureau of the Census.
Siegel, J. S., & Swanson, D. A. (2004). The methods and materials of demography (2nd ed.). San
Diego/London: Elsevier Academic Press.
Smith, D. P. (1992). Formal demography. New York/London: Plenum Press.
Sullivan, D. F. (1971). A single index of mortality and morbidity. HSMHA Health Reports, 86(4),
347–354.
United Nations. (1955). Age and sex patterns of mortality: Model life tables for under-developed
countries. New York: United Nations.
United Nations. (1982). Model life tables for developing countries. New York: United Nations.
Vos, T., Flaxman, A. D., Naghavi, M., & 354 others. (2012). Years lived with disability (YLDs) for
1160 sequelae of 289 diseases and injuries 1990–2010: A systematic analysis for the Global
Burden of Disease Study 2010. The Lancet, 380, 2163–2196.
World Health Organization. (1977). Manual of mortality analysis: A manual on methods of analysis
of national mortality statistics for public health purposes. Geneva: WHO.
Zhao, Z. (2007). Interpretation and use of the United Nations 1982 model life tables: With
particular reference to developing countries. Population-E, 62(1), 89–116.
Chapter 5
Marriage, Marital Status and Relationships

Why Study Marriage?

Unlike births and deaths, the two types of events of primary interest to demog-
raphers, marriage is not a biological event but a social event. Historically it has
been of interest to demographers principally because of its close connection with
the process of fertility. In most societies marriage has traditionally signalled the
commencement of regular sexual activity and hence of serious exposure to the risk
of childbearing. Equally, provided the biological procceses associated with getting
older have not already done so, marriage dissolution, whether through divorce or
widowhood, tends (sometimes only temporarily) to terminate exposure to the risk
of childbearing.
This is not, of course, to suggest that unmarried people are not sexually
active and do not have children. Indeed, in many of the populations of Europe,
North America and Australasia non-marital sexual activity and childbearing have
increased drammatically over recent decades as the second demographic transition
(Van de Kaa 1987) has seen consensual partnering proliferate, greatly weakening the
link between formal marriage and fertility and raising questions as to the definition
of ‘marriage’. The link between marriage and childbearing is also less than complete
in many African societies. However, it is still true to say that the great majority of
fertility in most populations is marital fertility, as distinct from non-marital, or ex-
nuptial, or, as it was often pejoratively labelled historically, illegitimate fertility.
It is often the case when studying changes in fertility levels and patterns over time
that a major part of the explanation for those changes is to be found in concurrent
changes in marriage patterns. People have been getting married earlier or later in
life (i.e., at younger or older ages), and/or proportionately more of them or fewer of
them have been marrying at all. In Australia, for example, there was considerable
alarm over declining fertility in the late nineteenth and early twentieth centuries.
Much of that decline was due to later and less universal marriage, and to consequent

© Springer International Publishing Switzerland 2016 213

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_5
214 5 Marriage, Marital Status and Relationships

diminished exposure to risk. Likewise the well-known post-war baby boom in

Australia has been attributed primarily not to increased fertility within marriage, but
to women marrying younger and to proportionately fewer of them failing to marry
at all. In developing countries with high fertility levels, some of the most effective
strategies for reducing fertility have been strategies which sought to delay entry into
marriage.

Defining Marriage

A ‘marriage’ is defined by the Multilingual Demographic Dictionary (IUSSP 1958)

as a ‘union between persons of opposite sexes which involves rights and obligations
fixed by law or custom’. Another view is that marriages are legally recognized
unions between a man and a woman, recognition being accorded under one or more
of religious law, civil law or customary/tribal law. In most societies a couple marries
by passing through a ceremony or ritual, prior to which they are not married but
after which they are. There are, however, also societies in which becoming married
is more a process than an event, so that it is not necessarily easy to date precisely
when the married state was entered, or indeed to say whether or not it has been
reached at a given point in time. Indigenous Australians are a case in point, and the
marriage practices of some African societies also have been claimed to fit this type
of model.
In addition to marriages demographers also are interested, and increasingly
so these days in developed countries in particular, in consensual unions (also
referred to by a variety of other labels, including cohabiting unions, living together
relationships, de facto marriages and de facto relationships). These are unions
of cohabiting heterosexual partners in which formalities signifying marriage in the
society in question have not been completed. They tend to vary greatly in the degree
of mutual commitment of the cohabiting parties to one another, and therefore in
the stability of the unions themselves. At one extreme they can be marriages in all
but law. Sometimes one or both parties is unable to marry because of inability to
legally dissolve a previous marriage, while on other occasions formal marriage is
viewed as adding nothing to the relationship, as complicating its dissolution should
the need arise, as attaching to it attributes (e.g., sex role stereotypes; a subservience
of the female to the male partner) that are viewed negatively, as something over
which there is no particular urgency, etc. At the other extreme consensual unions
are characterized by levels of mutual commitment that are decidedly less than
marriage-like. At the lowest level they may be little more than convenient settings
for regular sexual gratification, exploitation of a woman by a man for the domestic
services she can provide, and/or minimising living costs. At a somewhat higher
level they may be essentially ‘going steady’ relationships in which couples enjoy
one another’s company and intimacy (and the optimal conditions for this afforded
by cohabitation), but are either not seriously interested in long-term commitment
at that time in their lives or intent on thoroughly testing the relationship before
Defining Marriage 215

contemplating making such a commitment. Consensual unions of this type have

a proneness to dissolution that derives from inherent risks that either they will
ultimately be deemed to have failed the test by one or both parties, or the parties
will not progress to a desire for longer term commitment at the same pace.
Consensual unions are not infrequently precursors to marriage, but as just
intimated may also be alternatives to marriage or devoid of marital character.
Individual unions may merit classification into more than one of these categories
over time, or in the perceptions of the two parties to them at a point in time, and
in circumstances of ready access to fertility control this type of categorization is
likely to have implications for levels of childbearing. In consensual unions that
are perceived to be informal marriages the propensity to bear children may differ
little from that within formal marriages. In contemporary Sweden, for example,
formal marriage has virtually no role as a legitimator of parenthood, and a high
proportion of essentially marital childbearing takes place within consensual unions.
One does, however, need to remember that in some countries significant proportions
of informal marriages are associated with repartnering after marriage breakdown
and with decisions to remain voluntarily childless, both of which tend to depress
fertility. Where consensual partnering is a precursor to marriage the propensity
to have children may also be quite high, and indeed pregnancy may tend to be
associated with the transition to marriage, either as a trigger or as part of a planned
dual change of status. On the other hand, those in relatively uncommitted consensual
unions probably tend to view that phase of their lives as ideally childless, although
clearly women have more incentive to ensure that it turns out that way than do men.
Recent decades have also seen recognition accorded to LAT relationships.
LAT stands for living apart together, and describes couples who are in intimate
relationships that are not co-residential. Reasons for this can be varied. They may be
pragmatic, as when differing work locations force residential separation, or a couple
(perhaps both living with parents) can’t afford to co-reside, or a single mother would
jeopardize her entitlement to a welfare benefit were she to co-reside with a boyfriend
who was earning an income. They may be ideologically based, as when couples
or individual members thereof value the autonomy separate residence provides in
conjunction with the intimacy they enjoy. For others they may represent a stage in
the process of developing a relationship – a precursor to possible cohabitation and
subsequently even marriage. Then again, a party to a broken relationship who has
custody of its children may wish to keep a subsequent romantic interest at arms
length from those children, or a previous unpleasant experience of cohabitation
may promote wariness, at least in the meantime, of establishing co-residence with a
new intimate partner. Motivations for LAT relationships can be complex, and LAT
couples can be variously gladly, regretfully or ambivalently living apart. The fact
that these relationships are intimate, however, obviously gives them potential to
contribute to fertility, although in most societies that potential is probably less likely
to be realized than is the potential in co-residential intimate relationships.
A particular variant of LAT relationships which is of less recent origin and in
which childbearing is common is what are known as visiting unions (sometimes
also termed extra-residential unions). They are widespread in Caribbean soci-
216 5 Marriage, Marital Status and Relationships

eties among majority groups of African slave (as opposed to Indian or Chinese)
descent. Visiting unions exist alongside common-law (consensual) unions and
formal marriages, and relationships occasionally proceed through all three forms as
a man’s capacity to economically support his family increases. They are, however,
especially common at lower socio-economic levels, are associated with matrifocal
family structures (female breadwinners who often join forces with female kin,
especially mothers and sisters, to share income generation, childcare and domestic
duties), and have helped generate an extensive literature seeking to understand
and explain Caribbean mating and family formation patterns (Barrow 1996, 1998).
Characteristics attributed to visiting unions include: (i) youthful commencement;
(ii) variable, but commonly short-lived durations; (iii) a depressant effect on fertility
(earlier, but more widely spaced and rapidly ended childbearing compared to marital
unions); (iv) lack of social obligations, which facilitates easy termination by either
party; (v) conducive to frequent partner changing and multiple partnering (often,
from women’s perspective, from necessity rather than desire); and (vi) the emphasis
they give to mother-child, and especially mother-daughter, relationships. Among
rationales offered for the prominence of visiting unions in the Caribbean, where
males have considerable sexual freedom and all except young teenage girls can
also participate in premarital mating without community censure, have been: (i)
the ease with which they can be managed because a lack of formal kinship ties
and expectations maximizes personal choice; (ii) their being a rational adaptive
response to chronic poverty, unemployment and underemployment that renders
fulfilling conventional marital economic roles impossible for lower class males;
(iii) the flexibility they offer both sexes to respond quickly to both socio-economic
constraints and economic opportunities (often through migration) as and when they
arise; (iv) their offering women personal autonomy, shelter from conjugal violence
and scope for ‘child shifting’ (i.e., transferring dependent children to households
not including a natural parent, usually in the interest of economic survival); and
(v) the attraction of childbearing within one’s mother’s home before tackling the
potential difficulties of a co-residential union. In short, Caribbean visiting unions
are in varied ways a coping mechanism for those of African descent at lower socio-
economic levels. Rubenstein (1980: 333) writes of them, as observed in a village on
the island of St Vincent, as follows:
The female is expected to be sexually faithful to her partner, who, in turn, is obliged to
reciprocate with regular or periodic gifts. Beyond this, little is specified about the way the
union should be acted out : : : . Various types of behaviour are therefore permitted. Some
women regularly wash and cook for their non-resident boyfriends, while others never do
so. Some unions are very intimate, while others are very formal and confined to sexual
release. Some unions involve constant visits between households, while others are based
on brief encounters in deserted spots or in the home of a third party. Some unions require
considerable economic assistance from the male, while others are characterized by small or
infrequent gifts. There is a measure of male sexual exclusiveness in certain unions; others
represent an expression of the male desire to have concurrent affairs with several women.
Some show promise of being a prelude to marriage; others are simply fleeting affairs soon
to be terminated.
Defining Marriage 217

The phrase conjugal unions is sometimes used to refer to the combination of

formal marriages and consensual unions; i.e., to all stable co-residential sexual
unions. The term marital unions can have the same meaning, but may also be used
to refer only to formal marriages, so when encountered its precise meaning should
be checked. Demographers ideally prefer to conduct analyses of all conjugal unions,
but data sources available to them often preclude this. It is important to realize also
that different data sources may define ‘marriage’ differently.
In many countries the marriages of couples passing through recognized religious
or civil marriage ceremonies are required to be registered with a government agency,
and the regularly released registration data yielded by this administrative procedure
are widely used in the analysis of marriage patterns and trends. These data concern
marriage events, and adopt a particular, legalistic definition of marriage. Their
availability also means that marriages are more easily studied than are consensual
unions, data on which tend to have to be obtained from special surveys. Census data
on marriage do not pertain to marriage events. They typically focus on the status of
individuals as married, or not married, at the time of the census count, and rely on
respondents’ perceptions of which category they fit into. There is scope for persons
who are legally married to deny that they are, but more importantly there is scope
for persons to claim to be married who have not taken the steps that would result
in their being registered as such. This can occur because ritual and the registration
process are seen as irrelevant to the nature and quality of the union, or because
the notion of registered marriage is culturally alien. If the latter situation prevails
for the bulk of a population there are unlikely to be any registration data available,
and census data should reflect prevailing cultural norms and practices as to what
constitutes marriage. Sometimes, though, registered marriage is the standard model
within a majority culture, but is eschewed by a minority subculture. New Zealand’s
Maori and Australia’s Indigenous minorities are two examples. Both, in the eyes of
dominant European cultures, are particularly prone to form consensual unions, yet
many of these unions are marriages according to Maori or Indigenous custom, and
are likely to be returned as such at a census. The result is that the census records
significantly more married persons than registration data would lead one to expect,
and a blurring of the distinction between marriages and consensual unions occurs.
Should a union which is a marriage according to the cultural customs of the parties
involved be labelled ‘consensual’ because those customs happen not to require
registration? Mind you, with the proliferation of consensual unions in recent years
some censuses have made attempts to enumerate such unions. The 2011 Australian
census, for example, gathered such data through questions on the relationships
between household members which distinguish between ‘husband/wife’ and ‘de
facto partner’ relationships. In New Zealand’s 2006 census individuals were asked
to indicate who lived in the same household as they did, with two options being
‘my legal husband or wife’ and ‘my opposite-sex partner or de facto, boyfriend or
girlfriend’.
A third source of data on marriage is sample surveys. If utilising a similar
question focused on current status, these may impose the same definition of
marriage as a census. However, demographic sample surveys often are at pains to
218 5 Marriage, Marital Status and Relationships

date events like marriages, and this tends to facilitate the separation of registered
marriages from informal unions that, given the chance, respondents would claim to
be ‘marriages without the piece of paper’. Whether it also facilitates the separation
of registered marriages from unions that are marriages by some other cultural
prescription depends on whether that prescription lends itself to specification of
a precise date of marriage. If it does, no distinction is likely to be possible unless
the survey includes other questions for the purpose. However, if marriage is a state
into which couples drift as consensual unions gradually are accorded community
recognition as marriages, fixing a date of marriage is likely to be problematic and
unions are likely to be deemed consensual. Such a procedure is open to the criticism
that it demeans a particular cultural perspective on marriage.

Some Other Features of Marriage Systems

In many cultures marriage is not only a major social and demographic event, but a
significant economic one too. In addition to religious and/or customary ritual it may
entail non-trivial economic transactions. Dowry systems require economic transfers
from the family of the bride to the groom and possibly also his family; bridewealth
or bride-price systems require the groom to make payment to the family of the bride.
Where law or custom prescribe monogamy, no person may have more than
one husband or wife at any given time. The term serial monogamy is applied to
situations where individuals experience a succession of monogamous relationships.
Monogamous marriage systems make life much easier for the demographic analyst
than the alternative polygamous systems. Polygamy is the social institution that
permits a person to have multiple spouses at one time. Where men are permitted
to have multiple wives the more specific social institution that is said to exist is
polygyny. The much rarer institution that allows women to have multiple husbands
is polyandry. Polygynous marriage systems, common in Africa and formerly
widespread in Asia as well, have tended to be sustainable under conditions of high
fertility, and hence very youthful age structures, in conjunction with marked age
differences between husbands and their younger wives. Their viability is seriously
threatened by marked fertility decline.
Demographers sometimes are interested in the extent to which particular groups
of individuals tend to marry across social boundaries. Endogamy is said to exist
when both parties to a marriage are from the same a tribe, clan or other similar social
group; exogamy is the term used to describe marriages across such boundaries.
Another similar distinction is that between homogamy and heterogamy. Marriages
are said to be homogamous when contracted between parties with similar social,
physical or mental characteristics. Thus there might be interest in the extent to which
marriages were religiously (social), racially (physical) or educationally (mental)
homogamous. Marriages which entail intermarriage between members of different
religious, racial, educational, etc. groups are said to be heterogamous.
Marital Status 219

Marital Status

‘Marital status’ is the term used to describe the categories into which the processes
of marriage and marriage dissolution, collectively referred to as the processes of
nuptiality, transfer individuals. In all societies where formal marriage in some guise
takes place there are at least three categories of marital status – never married,
married and widowed – the latter comprising persons whose marriages ended
because their spouses died and who have not remarried. Where it is possible under
civil, religious or customary law to formally terminate a marriage other than by one
spouse dying there is also a category divorced. It may not, however, because of legal
requirements, be possible to enter this category immediately upon deciding to end a
marriage. Thus a category separated, or sometimes legally separated, may exist to
capture persons in transition from being married to being divorced. Members of this
category remain legally married, and may reconcile with their estranged spouses
and rejoin the ‘married’ group. It is sometimes, though, useful to be able to exclude
them from the married population. In studying fertility, for example, unless already
pregnant at the time of separation they are more likely to contribute to non-marital
(or in the official Australian terminology, ‘ex-nuptial’) than to marital (or ‘nuptial’)
childbearing, through having formed sexual relationships with men other than their
husbands. Indeed, in a proportion of cases the formation of such relationships will
have precipitated separation. In recent times census and survey takers have also, in
recognition of widespread marital disruption and repartnering, sometimes sought
to split the ‘married’ group into those in first marriages and remarriages. Then
again, for some purposes it is common for all groups except the never married to be
aggregated as those who have ever married.
It has already been noted that the ‘marriage’ data censuses typically gather are
data on a population’s marital status, and that they generally accept respondents’
perceptions of the marital status categories to which they belong. Surveys may take
the same approach, but have the capacity to probe more deeply and need not be as
problematic. It is important to realize that the marital status distributions that result
from asking ‘What is your marital status?’ and inviting respondents to tick a box
corresponding to the category they belong to need not result in a classification
based on a strictly legal definition of marriage. Some cohabiting individuals may
consider their relationships to be marriages according to another set of cultural
prescriptions besides that enshrined in law, or simply on the basis of the levels
of mutual commitment they feature. They may thus claim to be married even
though, legally, they are not. The same can happen with the statuses ‘separated’,
‘divorced’ and ‘widowed’, with which persons might identify on the basis that a
longstanding consensual union rather than a legal marriage had dissolved. In other
words, census and survey marital status distributions may reflect social, as well as
legal, definitions of marital status. Not even specifically asking for legal marital
status guarantees that that is what respondents will give, and providing additional
categories for the consensually partnered and those who have had consensual unions
dissolve is no solution either. Those who might identify with such categories also
220 5 Marriage, Marital Status and Relationships

have legal marital statuses with which they might identify, so to what extent do
they opt for each alternative? The best that can usually be done is to ask separately
about legal marital status and about living situation (where two response options
might be ‘Living with legal husband/wife’ and ‘Living with a partner unmarried’),
but even this is not foolproof. If respondents wish to represent consensual unions as
marriages, they will do so.
The issue just discussed can pose real problems for demographic analysis
when ratio-type measures refined by marital status draw their numerators and
denominators from different data sources. The classic example is the calculation of
marital and non-marital fertility rates. The numerators of these rates normally come
from birth registration systems, which classify as ‘marital’ only births to women
legally married to the fathers of their children; i.e., the mother and father must be
able to provide their date of marriage. The denominators, however, typically come
from census data, and to the extent that consensually partnered women claim to
be ‘married’ they contribute to the denominator of the marital fertility rate, whereas
births they have contribute to the numerator of the non-marital fertility rate. One can
argue over whether the births or the women are misclassified, but clearly both should
be reflected in one measure; they should not be split between the two. The result
is that the non-marital fertility rate is inflated (because its denominator excludes
some of the women contributing, or potentially contributing, to its numerator) while
the marital fertility rate is artificially depressed (because some of the women in
its denominator cannot contribute births to its numerator). The picture can become
really distorted if one is dealing with a group such as the Maori of New Zealand,
to whom the system of registered marriage underpinning ‘legal’ marriage in the
country is culturally alien. Many women who claim to be married at a census
because they are married according to Maori custom may have their births registered
as ‘ex-nuptial’ because Maori marriage does not yield a precise date of marriage.
The result can be inflated non-marital fertility rates, and misleadingly low marital
fertility rates.
The demographer’s main interest in marital status derives from the different
levels of opportunity for, and social acceptability of, childbearing associated with
different categories of marital status. Changes in sex-age-specific distributions of the
population by marital status may signal, or explain, changes in fertility. For example,
age-specific marital fertility rates may remain unchanged over a period, but age-
specific proportions married might fall; the result would be a decline in overall
age-specific fertility rates, and therefore a decline in total fertility. A fundamental
question that should always be asked when faced with a rise or fall in fertility is
‘How much of the change is due to change in age-specific proportions married, and
how much is due to change in fertility within marriage?’
General Measures of the Marriage Process 221

General Measures of the Marriage Process

Predictably, the most basic measure of the intensity of the process of marriage is a
crude rate, the crude marriage rate. This was previously (Chap. 1, Eq. 1.8) defined
as follows:

CMR D .M=P/ 1;000 (5.1)

Where M D marriages during year y; P D mean (or mid-year) total population in

year y.
Like other crude rates the CMR has major deficiencies, and is for cautious
use when nothing more refined can be calculated. There is an argument that its
numerator should be doubled to recognize the fact that there are two parties to each
marriage, but this makes no difference to the relative levels of its value for different
populations. The real problem is that it is not a true rate; its denominator includes
individuals who cannot possibly contribute to its numerator – they are too young or,
assuming the marriage system is monogamous, already married.
An initial refinement is to calculate a general marriage rate in which, as was
the case with the general fertility rate, the denominator is a measure of the true
population at risk, or the marriageable population. Such a measure excludes persons
under the legal minimum age for marriage and those older persons who are already
married (and not either widowed or divorced). It is usually suggested that the GMR
should be calculated separately for males and females, but it is not a measure that is
in common use. Its equation is:

GMR D .M=Pm / 1;000 (5.2)

Where M D marriages during year y; Pm D the mean (or mid-year) marriageable

population of a given sex.
What are in common use are sex-age-specific marriage rates. Using life table-
type notation these are given by:

n SASMRx D .n Mx =n Px / 1;000 (5.3)

Where x D the lower limit of the age group for which the marriage rate is being
calculated; n D the width of that age group in years; n Mx D the number of
marriages of males or females aged between x and x C n years (exact ages)
during year y; n Px D the mean (or mid-year) male or female population aged
between x and x C n years (exact ages) in year y.
These rates, however, are in the same category as the CMR in having denomi-
nators which include persons who are already married and therefore not at risk of
marrying during year y. To refine them to overcome this deficiency we can calculate
sex-age-marital status-specific marriage rates. These are given by:
222 5 Marriage, Marital Status and Relationships

n SAMSMRx D .n M.m/x =n P.m/x / 1;000 (5.4)

Where x D the lower limit of the age group for which the marriage rate is being
calculated; n D the width of that age group in years; n M(m)x D the number of
marriages of males or females of marital status m aged between x and x C n
years (exact ages) during year y; n P(m)x D the mean (or mid-year) male or female
population of marital status m aged between x and x C n years (exact ages) in
year y.
Where the marital status m is ‘never married’, Eq. 5.4 yields sex-age-specific
first marriage rates. Where m is ‘widowed’ or ‘divorced’ it yields sex-age-specific
remarriage rates. The latter should be calculated separately for widowed and
divorced persons, not for the two marital statuses combined, because levels of
remarriage tend to be quite different depending on which mechanism (death or
divorce) terminated the previous marriage. Should the data on marriages classified
by marital status required to apply Eq. 5.4 not be available, one option to improve
on SASMRs as defined in Eq. 5.3 is to restrict their denominators n Px to the mean
(or mid-year) populations who were never married, widowed or divorced and aged
between x and x C n years (exact ages). This yields ‘true’ rates by eliminating
persons not at risk from the denominators. One lesson from having presented
this option is that you should always, when encountering SASMRs, check what
denominator was used in their calculation – was it a total population denominator
or a ‘not currently married’ population denominator?

First Marriage

Reference to first marriage rates and remarriage rates introduces a fundamental

refinement of any detailed analysis of marriage patterns and trends. You may
recall that in Chap. 1 we distinguished between renewable and non-renewable
events. In Chap. 3 we noted that renewable events (like marriage) posed problems
for demographic analysis, especially in the construction of probabilistic types of
measures. However, these problems could be overcome by subdividing a renewable
event process into several non-renewable processes. Any detailed study of marriage
MUST look separately at first marriages (which are non-renewable events), and
ideally should also break subsequent marriages down by order (second, third, etc.
marriages) and whether remarriage followed widowhood or divorce.

Measures of the Intensity of the First Marriage Process

One of the more commonly encountered measures of the intensity of the process of
first marriage these days, especially in the work of European demographers, is the
First Marriage 223

total first marriage rate, or TFMR. As its name suggests, it is directly analagous
to the more familiar total fertility rate, and is constructed in a similar way. It is a
cross-sectional, or synthetic cohort, measure based on what is known as the reduced
events approach to demographic analysis. This approach was developed by the
French demographer Louis Henry, and is described by Wunsch and Termote (1978:
Chapters 1 and 2). It is applicable to demographic processes (such as fertility, first
marriage and divorce) which ‘do not exclude members of [a] cohort from [future]
observation’ (as mortality and emigration do – they cause individuals to disappear
from a population) (Wunsch and Termote 1978: 45). At the core of the reduced
events approach are annual ratios of events at each single-year duration d since
‘event-origin’ (the event initiating exposure to risk – attainment of the minimum
legal age for marriage in a study of first marriage) to the population at duration
d who were at risk at that earlier life cycle stage corresponding to event-origin
(the total population at duration d in a study of the process of first marriage, since
everyone surviving at that duration was at risk of marrying for the first time when
they reached the minimum legal age for marriage).
In an analysis of the first marriage process, which because its tempo typically
differs for women and men (women generally marry younger) should be refined by
sex, these ratios are ratios of first marriages of females or males aged x in year y
(symbolized by F(s,x), where s stands for ‘sex’) to the size of the mid-year female or
male population aged x in year y (P(s,x)). In case you are puzzled by reference here
to ages, x, instead of durations, d, the former are simply another way of expressing
the latter. Each duration, d, corresponds to an age, x, such that x is the sum of d
and the legal minimum age for marriage. Summing the ratios (reduced events) just
defined over all ages, x, commencing at the minimum legal age for marriage is the
same as summing them over all durations, d, beyond that legal minimum age, and
yields the TFMR for sex s in year y. Thus the TFMRs for females and males are
given by:

TFMRf D †xDl;¨ .F .f; x/ =P .f; x// 1;000 (5.5)

TFMRm D †xDl;¨ .F .m; x/ =P .m; x// 1;000 (5.6)

Where TFMRf and TFMRm are the female and male total first marriage rates; F(f,x)
and F(m,x) D first marriages of females aged x and males aged x during year y;
P(f,x) and P(m,x) D the mid-year female and male populations aged x in year y;
l D the legal minimum age for marriage for the relevant sex; ¨ D the oldest age
at which any first marriage is recorded by the relevant sex in year y.
The TFMR indicates the proportion of a female or male birth cohort who
would EVER marry IF that cohort experienced the age-specific first marriage
ratios prevailing among women/men in the year for which it was calculated. In
other words, as already intimated, it is a synthetic cohort measure; a period, or
cross-sectional, measure for which there is a cohort type of interpretation. You can
probably appreciate the analogy to the total fertility rate as well. The TFR measures
224 5 Marriage, Marital Status and Relationships

the average number of children a female birth cohort would have if it experienced
through life the age-specific fertility rates prevailing in the year for which it was
calculated.
You should recognize that the TFMR can take on theoretically impossible
values; values greater than 1,000, which imply greater than universal marriage. This
phenomenon aroused considerable debate when first noted some years ago, but is
no cause for alarm. It merely indicates that one or more of the forces that were
seen in Chap. 3 to cause period measures to take on exaggerated values compared
to equivalent cohort measures is at work. Either some historical circumstance has
caused an abnormally large number of first marriages to take place during the year
in question or a trend to earlier marriage is in progress, or both. A heaping of first
marriages has occurred cross-sectionally, much the same as a heaping of births may
occur cross-sectionally and result in an exaggerated peak in the trend of the total
fertility rate compared to that in the trend of the cohort completed fertility rate (refer
Chap. 3, Fig. 3.8). First marriage has been such a widespread, or near-universal,
experience over long periods of demographic history that heaping of this kind
easily can produce the ‘theoretically impossible’ result described. Its occurrence,
however, merely implies that a compensating downturn in annual (cross-sectional)
first marriages occurred earlier or is to be expected in future. Moreover, it merely
reflects a reality that there is no good reason to suppress – the reality that first
marriages can heap in individual years at levels unsustainable in the longer term.
Just as trends in the TFMR can peak at levels that are artificially high, so, too,
they can trough at levels decidedly lower than any real cohort will ever experience.
Instead of first marriages heaping cross-sectionally in response to historical events
favourable to marriage and/or a trend to earlier marriage, substantial cross-sectional
deficits can occur as a result of unfavourable historical events and/or a trend to later
marriage.
It is possible to compute real cohort equivalents of the TFMR, based on
first marriage events distributed diagonally through a Lexis diagram rather than
vertically up its columns. However, because we generally can’t afford to wait around
while a birth cohort passes through all possible marriageable ages, the tendency is to
cumulate age-specific first marriage ratios to some upper limit which is beyond the
main marrying ages, but well below the oldest age at which first marriages occur.
Choice of this upper limit is arbitrary, but the aim is to capture the bulk of first
marriage activity and ignore only ages at which few first marriages occur. Putting
it another way, the aim is to facilitate production of a trend line whose shape is
reliable, even though it may be displaced slightly downward from its position were it
possible to cumulate over all marriageable ages. Suitable upper limits when dealing
with a population like Australia’s might be exact ages 35 for females and 40 for
males, although with more recent trends to later marriage a case might be made for
exact ages 40 and 45 respectively.
When we cumulate age-specific first marriage ratios for real cohorts of women
or men, starting at the youngest age at which marriages occur and working up, the
measure we obtain is a cohort cumulative first marriage rate (CFMR) to whatever
upper age limit we have specified. This tells us the proportion of the cohort who had
First Marriage 225

married by the time it reached that age limit, and because we are now dealing with
the experience of a real cohort we should not obtain results which imply more than
universal marriage. What was only a theoretical impossibility in the context of the
synthetic cohort measure (TFMR) is an actual impossibility in the context of the
equivalent real cohort measure (CFMR). The relevant equations are:

CFMRf;u D †xDl;u1 .F .f; x/ =P .f; x// 1;000 (5.7)

CFMRm;u D †xDl;u1 .F .m; x/ =P .m; x// 1;000 (5.8)

Where CFMRf,u and CFMRm,u are the female and male cohort cumulative first
marriage rates to exact age u; F(f,x) and F(m,x) D first marriages at age x of
females or males who were members of the relevant birth cohort; P(f,x) and
P(m,x) D the size of the female or male birth cohort at the one point in time
when all its members were aged x; l D the legal minimum age for marriage for
the relevant sex; u D the exact age defining the upper limit to which age-specific
first marriage ratios are being cumulated.
By way of example you may like to refer back to Fig. 3.11 in Chap. 3. This shows
for twentieth century Australian female birth cohorts CFMRs to exact age 35. The
equivalent graph for males is presented as Fig. 5.1. Both graphs use first marriage
data up until 2002, beyond which the necessary single-year-of-age data ceased being
published. The prevalence of female marriage by exact age 35 is shown to have
risen from about 82 % for the 1899–1900 birth cohort to well over 95 % for cohorts
born in the early and mid-1930s, before falling sharply again, especially among
cohorts born in the 1950s. The prevalence of male marriage by exact age 40 rose
from a similar level (82 %) for the 1899–1900 birth cohort to a lower peak (a little
over 90 %), before also latterly declining again. In other words, marriage became
progressively more universal among birth cohorts of the first thirty or so years of the
twentieth century, especially female birth cohorts, but became less universal again
among cohorts born following the Second World War.
You will note in both Fig. 3.11 and Fig. 5.1 that the processes of cumulating age-
specific first marriage ratios have been carried out in stages, yielding intermediate
plots of cumulative first marriage rates to various exact ages below the respective
upper limits of 35 and 40 years. These can be useful adjuncts to plots of CFMRs as
defined by Eqs. 5.7 and 5.8. For example, it is interesting to compare, between
the 1:1:46 and 1:1:73 time lines, female plots to exact ages 20 and 22 in Fig.
3.11 with the male plots to exact ages 22 and 25 in Fig. 5.1. The former rise
steeply then plateau while the latter rise more steadily throughout the 27-year
period. Exact ages 20 and 22 for females, and 22 and 25 for males identify roughly
equivalent stages in the marriage process, and this differential pattern of change
reflects a steeper, quicker decline in ages at first marriage among females in post-
war Australia in contrast to a more gradual decline among males. A major factor in
this difference was a marked gender imbalance in the marriage market through the
1950s. Males at this time faced what demographers term a marriage squeeze. Partly
226 5 Marriage, Marital Status and Relationships

Fig. 5.1 Cumulative first marriage rates to selected exact ages for Australian male birth cohorts
(Source: Adapted from Carmichael (1988: Figure 5))

because of depressed fertility during the early 1930s but mainly because of heavy
male dominance of the large volume of early post-war immigration to Australia,
marriageable men faced a shortage of potential marriage partners. The competition
for brides this produced exerted strong pressure on women to marry younger, but
left some men unable to marry as early in life as they might have wished until,
with the first baby boom cohorts entering it, the marriage market improved for them
during the 1960s. This phenomenon also largely explains why the male CFMR to
exact age 40 peaks at a lower level than the female CFMR to exact age 35. Over a
period when almost everyone wanted to marry, some men missed out. There were
too few women to go around.

Measures of the Timing of the First Marriage Process

We have discussed the TFMR and CFMR as measures of the intensity of the
process of first marriage for synthetic and real birth cohorts respectively. The other
First Marriage 227

dimension of the first marriage process in which demographers are interested is its
tempo, or the timing of first marriage. They want to know, in general terms, whether
ages at first marriage are relatively young or relatively old, and whether they are
rising, falling or remaining constant over time.
A commonly used period measure of marriage timing is the median age at first
marriage. The median is a preferable measure of the ‘average’ age at which first
marriages occur to the mean age at first marriage, because distributions of first
marriages by age tend to be heavily skewed to the left. That is, if the number of
first marriages by age is plotted, the resulting graph generally has a pronounced
peak towards the lower end of the age range over which first marriages occur, and
a long ‘tail’ extending over the middle and upper end of that range. The few first
marriages that take place at more advanced ages have an inordinate influence on the
mean age at first marriage, tending to push it above the age at which first marriage
is in fact most common. The median is not subject to this bias, is likely to coincide
better with the graphical peak of the distribution of first marriages by age, and is
thus a more reliable measure of central tendency, or ‘average’ behaviour.
As with any median, to obtain the median age at first marriage one simply
takes an annual distribution of first marriages by age and determines the age above
and below which exactly half of first marriages occurred. It should be calculated
separately for females and males, and where it falls within a single-year age group
rather than neatly at the boundary between two successive age groups (as it almost
always will) its precise value is obtained by linear interpolation within that age
group. The relevant equation is:

Md D X C .T=2  †iDl;x1 Mi / =Mx (5.9)

Where T D the total number of first marriages in year y; Mi D first marriages at age
i in year y; l D the lowest, or youngest, age at which first marriages occurred
in year y; X D the single-year age group within which P the median age at first
marriage lies (i.e., the youngest age group for which iDl, x Mi equals or exceeds
T/2).
The median age at first marriage is in common use. It has a limitation, though,
in not taking account of the varying sizes of single-year age groups. Other things
being equal, the larger the age group the more first marriages it will produce. Trends
over time in the median age at first marriage can thus be affected by shifting age
structures, especially over the ages at which first marriage is most common. A better
procedure for obtaining period trends in marriage timing, and one which controls for
this source of change, is to divide the total first marriage rate into quartiles and
ascertain the ages at which each quartile is reached. To do this:
1. Find the values of quarter, half and three-quarters of the TFMR. These corre-
spond with the first, second and third quartiles of the age distribution of first
marriages respectively.
2. Begin cumulating the F(f,x) / P(f,x) or F(m,x) / P(m,x) ratios from the equation
for the TFMR (Eq. 5.5 or Eq. 5.6 as appropriate) from x D l, the lowest or
228 5 Marriage, Marital Status and Relationships

Fig. 5.2 Quartiles of the age distributions of first marriages for Australian females and males
married between 1921 and 2002 (Source: Adapted from Carmichael (1988: Figure 3))

youngest age at which first marriages occur, on upwards. In doing so, identify
the exact ages at which the three quartiles are reached.
3. Usually you will find that by cumulating to a value of x D m you don’t quite
reach a quartile, whereas by cumulating to x D m C 1 you have gone past it. This
means that the quartile occurs at an exact age that is greater than m C 1, but less
than m C 2 years (x, remember, is a single-year age group, or an age in completed
years). To find the exact age corresponding to a quartile, use:

QA D m C 1 C ŒQ  †xDl;m .F .s; x/ =P .s; x// = .F .s; m C 1/ =P .s; m C 1//

(5.10)

Where QA stands for ‘quartile age’; Q D the relevant quartile value of the TFMR
(i.e., quarter, half or three-quarters of the TFMR, depending on whether
one is evaluating the first, second or third quartile of the age distribution of
first marriages); l D the youngest age at which first marriages occur; m D the
highest value of x (age) for which the TFMR summation on the righthand side
of Eq. 5.5 or Eq. 5.6 lies below Q; F(s,x) and P(s,x) have the same meanings
as in Eqs. 5.5 and 5.6, with s denoting ‘sex’ (either f (female) or m (male)).
First Marriage 229

Quartile ages at first marriage for Australian females and males marrying
between 1921 and 2002 are shown in Fig. 5.2. Single-year-of-age first marriage data
required to calculate them have not been published since 2002. Naturally the lowest
trend line for each sex plots the age by which 25 % of first marriages occurred;
the middle line the age by which half occurred; and the highest line the age by
which 75 % occurred. Ages at first marriage declined through the 1920s, rose again
from the Great Depression to plateaux in the middle and late 1930s, then followed
protracted downward courses from the outbreak of World War 2 until the early
1970s. Declines during the War itself were steep as couples shortly to be separated
when young men left for war service took a ‘now or never’ attitude to marriage and
because of their imminent separation disregarded its usual economic prerequisites.
There were also 12–15 thousand marriages involving visiting American servicemen
during the War.
Following the War continued downward trends were initially steeper for females,
reflecting pressure to marry exerted by the previously noted excess of marriageable
males in the population. Female ages at first marriage in particular looked to have
stabilized during the early and mid-1960s, but the late 1960s and early 1970s
brought further declines as oral contraception permitted very early marriage on
the basis that parenthood could confidently be deferred thereafter. At this time
peer pressure on the young to be sexually active had become intense, but many
parents still preached premarital chastity, and early marriage coupled with reliable
contraception was for some an ideal way of resolving these conflicting pressures.
After the early 1970s the trend to earlier marriage dramatically reversed, and by
the mid-1990s most quartile ages at first marriage had regained and surpassed their
1921 levels. From the pattern of very early marriage followed by a childless, dual
income start to married life that the pill fostered in the late 1960s and early 1970s
it was a short step to couples simply cohabiting without getting married, especially
with a rising divorce rate recommending a more cautious approach to marriage. The
growing acceptability of consensual unions has been one factor behind the trend
to later marriage, but both trends have also been products of the change feminism
induced in young women’s priorities in late adolescence and early adulthood, and
of increased emphasis by the young of both sexes on maintaining their individual
autonomy. Feminism saw education, establishing careers, travel and having a good
time assume higher priorities and finding a husband a lower priority; the quest for
more autonomy overlapped with feminism, but also made young men reluctant to
commit to marriage. A rapid decline in the incidence of ‘shotgun’ marriage over the
period 1971–1976, largely due to freer access to abortion and aided by increased
consensual partnering (hence more reliable contraception) and the introduction of a
welfare benefit for single mothers, was also something of a trigger to the new trend.
It eliminated many marriages at the very youngest ages in which brides had been
especially likely to be already pregnant, and saw the Judaeo-Christian view that
marriage was the honourable response to unplanned premarital pregnancy widely
rejected almost overnight.
You should note that the measures plotted in Fig. 5.2 measure marriage timing
among those who did marry. Beware of measures which are claimed to be, for
230 5 Marriage, Marital Status and Relationships

example, median ages at first marriage, but which actually indicate the age by
which 50 % of all people (not 50 % of those who married) married. There is
sometimes a particular temptation to present this type of measure when examining
the first marriage experience of real birth cohorts, because part of their first marriage
experience still lies in the future and this makes it impossible to calculate a genuine
median. If everyone marries, the age by which 50 % have married is indeed the
median (the halfway point in the age distribution of first marriages), but if, say,
only 80 % of a cohort marries, it is five-eighths of the way up the age distribution
of first marriages. Thus this measure is not a ‘pure’ measure of marriage timing;
its meaning depends on the intensity of the first marriage process and, critically,
changes as intensity changes. Where only 80 % of a cohort ever married the median
age at first marriage would be the age by which 40 % married (i.e., half of those who
ultimately ever married). Failure to marry at all is a different phenomenon from the
timing of marriage among those who do marry.
The procedure for obtaining cross-sectional quartiles of the age distribution of
first marriages summarized in Eq. 5.10 can be extended to studying the first marriage
experience of real birth cohorts by focusing on first marriages of those marrying by
some arbitrarily selected exact age u (standing for ‘upper limit’). Proceed as follows:
1. Find the values of quarter, half and three-quarters of the cohort CFMR to exact
age u. These correspond with the first, second and third quartiles of the age
distribution of first marriages taking place before exact age u respectively.
2. Begin cumulating the F(f,x) / P(f,x) or F(m,x) / P(m,x) ratios from the equation
for the CFMR (Eq. 5.7 or Eq. 5.8 as appropriate) from x D l, the lowest or
youngest age at which first marriages occur, on upwards. In doing so, identify
the exact ages at which the three quartiles are reached.
3. Do this by again applying Eq. 5.10, with obvious modifications to the meaning
of terms in that equation. Q D the relevant quartile value of the cohort CFMR to
exact age u; m D the highest value of x (age) for which the CFMR summation
on the righthand side of Eq. 5.7 or Eq. 5.8 lies below Q; F(s,x) and P(s,x) have
the same meanings as in Eqs. 5.7 and 5.8, with s denoting ‘sex’ (either f (female)
or m (male)).

Nuptiality Tables

The life table principles discussed in relation to the study of mortality in Chap. 4
may also be applied to the study of first marriage. Consider a cohort of never married
persons of one sex attaining the minimum age for marriage, and then passing
through the marriageable ages. Individuals are removed from that population at risk
of first marriage by one of two processes – either they marry, or they die without
having married. On experiencing either of those events a person ceases to be at risk
of marrying for the first time.
Nuptiality Tables 231

If we had all of the information about marriages and deaths occurring to our
cohort we could calculate numbers ‘surviving’ never married at each birthday and
probabilities of marrying and of dying without having married between successive
birthdays. These probabilities would enable us to construct a double decrement life
(or attrition) table for the cohort; i.e., a life table in which there were two modes of
depletion of the initial radix population. This table would allow us to estimate for
persons surviving never married at each birthday the average number of years left
in the never married state before either marrying or dying.
We have talked here of tracing the depletion through first marriage and mortality
of a real cohort of never married males or females. With the conventional life table
we didn’t trace the mortality of real cohorts because they take so long to die out.
Rather we traced the mortality of cross-sectional synthetic cohorts. We do the same
when dealing with first marriage. We construct what is called a net nuptiality table
on the basis of sex-age-specific first marriage and death rates in a calendar year.
As with the conventional life table the key to constructing a net nuptiality table is
the calculation of attrition probabilities – probabilities of first marriage (vx ) and of
death (qx ). Before progressing further with net nuptiality tables, however, there is a
simpler type of nuptiality table to consider – the gross nuptiality table.

Gross Nuptiality Tables

A gross nuptiality table ignores death as a mode of attrition from the radix of never
married persons. It is therefore a single decrement table directly analagous to the
standard life table. The rationale for ignoring mortality is that attrition from this
source is generally so low compared to attrition due to marriage over the main
marrying ages that it is inconsequential. Generating a gross nuptiality table requires
probabilities of first marriage (vx ) to be obtained, with remaining columns of the
table then derived exactly as they would be in a standard single-year-of-age life
table (writing vx for qx in the relevant equations).
The literature provides several equations for obtaining vx values. Recalling the
equations for qx in terms of the age-specific death rate Mx which were derived for
the standard life table in Chap. 4 (Eqs. 4.7 and 4.8), the analagous equations for vx
in terms of age-specific first marriage rates nx are given by:
 
vx D nx = 1 C 1=2 nx D 2nx =.2 C nx / (5.11)

Where nx is first marriages of females or males aged x in year y divided by the

mid-year population of never married females or males aged x in year y.
232 5 Marriage, Marital Status and Relationships

Equation 5.11 is for use when marriage registration data are available to provide
numerators for nx calculations. There is also a procedure available which allows vx
values to be estimated from census data giving sex-age-specific proportions never
married. It is a relatively crude approach, and should only be used when no other
(registration-based) option is open. The relevant equation is:

vx D 1=2 Œ..sx1  sx / =sx1 / C ..sx  sxC1 / =sx / (5.12)

Where sx D the proportion of females or males never married at age x.

An example of a gross nuptiality table is shown as Table 5.1. Several features of
this table should be noted. First, apart from age 0 the first age shown is age 14, the
youngest age at which any marriage occurs in the population. There is no point in
showing intermediate ages because no first marriage activity is taking place, and the
table assumes zero mortality. Note, though, that 14 L0 has a value equal to 14 times
the radix (1,400,000), since everybody is assumed to survive the first 14 years of
their lives never married, and therefore to contribute 14 person-years of exposure to
risk. This exposure must be taken into account in calculating eo 0 , which is perhaps
the most important number in a gross nuptiality table (its interpretation is discussed
below).
Second, Table 5.1 is restricted to ages below 50, so that the Tx column gives total
person-years lived in the never married state between exact age x and exact age
50 (not at all ages beyond exact age x) and eo x gives the average number of years
remaining in the never married state before attaining exact age 50 for persons still
never married at exact age x. Why, you might ask, do we restrict the range of ages
covered by a gross nuptiality table in this way? One reason is that our single mode
of attrition, first marriage, is not a universal experience as death is, so that our radix
of 100,000 never married persons never completely ‘dies out’ (in the sense of all of
them getting married). Eventually a point (an age) is reached at which the great bulk
of the first marriage activity that will ever take place has taken place, and there is
little to be gained from considering what happens at older ages. A second reason for
truncating at age 50 is that the basis upon which we ignore mortality in generating
a gross nuptiality table – that it is an inconsequential mode of decrement compared
to first marriage – is only sustainable over ages where most first marriages occur. It
certainly is not sustainable at older ages, where death takes over as the main force of
attrition among the few remaining survivors never married. Thus we truncate before
reaching the point at which our having disregarded mortality becomes a problem.
A third point that may have occurred to you is that, while mortality may generally
be very low through the main marrying ages, it may not be so low at ages below the
minimum age for marriage, and especially in infancy. The gross nuptiality table
assumes zero mortality at these ages as well, and this influences the value of eo 0 ,
which as already mentioned is a particularly crucial number in a gross nuptiality
table. The quantity eo 0 gives the average number of years spent in the never married
state before age 50 on the assumption that a person remains alive for at least
Nuptiality Tables 233

Table 5.1 Gross nuptiality table for Australian women, 1980

Never married Female first
Age (x) females marriages vx lx n Lx Tx eo x
0 100; 000 1; 400; 000 2; 719; 169 27:2
14 119,633 4 0.000033 100; 000 99; 998 1; 319; 169 13:2
15 121,024 31 0.000256 99; 997 99; 984 1; 219; 171 12:2
16 125,445 787 0.006254 99; 971 99; 658 1; 119; 187 11:2
17 125,266 2144 0.016970 99; 346 98; 503 1; 019; 529 10:3
18 122,915 6379 0.050585 97; 660 95; 190 921; 026 9:4
19 113,468 10,102 0.085235 92; 720 88; 768 825; 836 8:9
20 95,051 12,081 0.119505 84; 817 79; 749 737; 068 8:7
21 80,574 11,962 0.138201 74; 681 69; 520 657; 319 8:8
22 63,022 10,111 0.148521 64; 360 59; 580 587; 799 9:1
23 49,906 7480 0.139432 54; 801 50; 980 528; 219 9:6
24 40,798 5664 0.129818 47; 160 44; 099 477; 238 10:1
25 30,914 4214 0.127615 41; 038 38; 419 433; 140 10:6
26 24,029 3037 0.118876 35; 801 33; 673 394; 720 11:0
27 20,025 2311 0.109109 31; 545 29; 824 361; 048 11:4
28 16,873 1825 0.102611 28; 103 26; 661 331; 224 11:8
29 13,783 1337 0.092516 25; 219 24; 053 304; 563 12:1
30 11,555 1021 0.084621 22; 886 21; 918 280; 510 12:3
31 9518 822 0.082787 20; 949 20; 082 258; 593 12:3
32 8570 638 0.071774 19; 215 18; 525 238; 510 12:4
33 7883 503 0.061835 17; 836 17; 284 219; 985 12:3
34 6111 413 0.065373 16; 733 16; 186 202; 701 12:1
35 6090 297 0.047607 15; 639 15; 267 186; 515 11:9
36 5790 224 0.037953 14; 895 14; 612 171; 248 11:5
37 4423 180 0.039884 14; 329 14; 043 156; 636 10:9
38 4389 161 0.036021 13; 758 13; 510 142; 592 10:4
39 3985 116 0.028691 13; 262 13; 072 129; 082 9:7
40 3737 109 0.028748 12; 882 12; 696 116; 010 9:0
41 3578 100 0.027563 12; 511 12; 339 103; 314 8:3
42 3436 74 0.021307 12; 166 12; 037 90; 975 7:5
43 3570 64 0.017767 11; 907 11; 801 78; 938 6:6
44 3193 55 0.017078 11; 696 11; 596 67; 137 5:7
45 3017 49 0.016110 11; 496 11; 403 55; 541 4:8
46 2991 34 0.011303 11; 311 11; 247 44; 138 3:9
47 3050 42 0.013676 11; 183 11; 106 32; 891 2:9
48 2759 35 0.012605 11; 030 10; 960 21; 785 2:0
49 3082 38 0.012254 10; 891 10; 824 10; 824 1:0
50 3135 38 0.012048 10; 757
234 5 Marriage, Marital Status and Relationships

50 years. It is a useful index of the urgency with which a population marries, and
changes over time in its level can be plotted. It is important to realise, however, that
such changes can be produced either by changes in marriage timing or by changes in
the propensity to get married at all. The larger the proportion of the radix population
that reaches age 50 still never married the higher eo 0 will be, other things (i.e.,
the timing of marriage among those who do marry) remaining equal. Thus eo 0 is
not, strictly speaking, a measure of marriage timing akin to the median age at first
marriage.
The lx column of a gross nuptiality table is of particular interest. Dividing lx
by the radix l0 tells us the proportion of a cohort experiencing the cross-sectional
probabilities of first marriage underlying the table who would remain never married
at exact age x. Subtracting from 1.0 then gives us the proportion ever married by
exact age x, and proportions of this type can be plotted for successive synthetic
cohorts (each the subject of a separate gross nuptiality table) to establish trends over
time.

Net Nuptiality Tables

We can obtain the two sets of probabilities qx and vx needed for the construction
of a net nuptiality table from age-specific death rates, Mx , and age-specific first
marriage rates, nx , using equations already given. To find qx values use equation
4.7 or equation 4.8 from Chap. 4; to find vx values use Eq. 5.11 above. Strictly
speaking these equations are slightly imprecise vehicles for obtaining qx and vx
values for use in a double decrement context. But they are satisfactory, and we
can get away with using them, because of the overwhelming importance of first
marriage, and the minimal importance of death, as modes of attrition from the
never married population through the main marrying ages. More refined procedures
are available, but little is gained from using them in constructing net nuptiality
tables. Note, though, that this is not the case with all double decrement life tables.
For other types of double decrement tables in which both modes of attrition are
reasonably important it is not acceptable to obtain attrition probabilities using
equations modelled directly on Eqs. 4.7 and 4.8 from Chap. 4 and Eq. 5.11 above;
a more refined procedure must be used. One way of thinking of our ability to get
away with a relatively simple approach in the case of net nuptiality tables is to note
that, because of the minimal importance of death, we are very close to really only
having one mode of attrition. That, of course, is also the rationale for often making
do with a gross nuptiality table.
Another point to note is that, strictly speaking, the mortality probabilities in
a net nuptiality table should pertain to never married persons only, not the total
population. Once again, though, because mortality is minimal over the prime
marrying ages there is nothing of consequence to be gained by being rigorous in
Nuptiality Tables 235

this matter, and total population probabilities, which are likely to be readily available
from a standard life table, generally are used.
A net nuptiality table for White females in the United States in the late 1950s is
presented as Table 5.2. Its first two columns give the probabilities of first marriage
and of dying at age 0 and at all ages from 14 to 65 and over. The minimal importance
of mortality from the late teens to the early thirties, when first marriage is most
0 0
prevalent, is clear. Columns 4 and 5, which give values of dx and fx , the numbers of
deaths and first marriages, respectively, at age x, reinforce this. These are obtained
from the equations:

d0x D qx .1  vx =2/ l0x (5.13)

f0x D vx .1  qx =2/ l0x (5.14)

0
The lx column, column 3 of Table 5.2, is the net nuptiality table equivalent of
the lx column of the conventional life table; it gives the number of members of the
radix population (i) still alive and (ii) still never married at exact age x. It is obtained
using the equation:

l0x D l0x1  d0x1  f0x1 (5.15)

0
It may seem strange that Eqs. 5.13 and 5.14 require a value of lx to obtain values
0 0 0
of dx and fx , and Eq. 5.15 requires values of d0x1 and f0x1 to obtain values of lx .
‘Where do I start?’ you might ask. The answer is, by evaluating Eqs. 5.13 and 5.14
0
for x D 0, since l0 is known (the nuptiality table radix D 100,000). It is then possible
0 0 0
to obtain l1 using Eq. 5.15, then d1 and f1 using Eqs. 5.13 and 5.14, and so on.
We simply shuffle between Eqs. 5.13/5.14 and 5.15, working our way down the net
nuptiality table from x D 0, increasing x by 1 each time.
Two idiosyncrasies of Table 5.2 should be noted. First, entries for values of x
in the range 1–13 are excluded because no first marriages take place at those ages.
The same is true at age 0, but this age is represented because (i) we need to know
0
l0 , the radix of the table, and (ii) e0o x , the average number of years of single life
remaining
 0 at birth, is an important feature of a net nuptiality table. Deaths at age 0
d0 D 1; 964 are shown, and any attrition between exact ages 1 and 14 is also due
to mortality. We can ascertain its extent from:
0
 
13 d1 D l01  l014 D l00  d00  l014 D 100;000  1;964  97;411 D 625

The second idiosyncrasy pertains to the bottom of the nuptiality table where the
terminal age group is 65 and over. Here Eqs. 5.13 and 5.14 are not used to obtain
0 0
dx and fx (which are more correctly specified as d0xC and f0xC ). The sum of the
probabilities of first marriage (v65C ) and death while still never married (q65C )
is 1.00000 (since everyone who doesn’t eventually marry ultimately dies never
236 5 Marriage, Marital Status and Relationships

Table 5.2 Net nuptiality table for white females in the United States, 1958–1960
vx qx l0 x d0 x f0 x N0 x %N0 x L0 x T0 x e0o x
Age (x) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
0 0.00000 0.01964 100,000 1,964 0 94,937 94.9 98,319 2,157,010 21.57
14 0.01167 0.00036 97,411 35 1,137 94,937 97.5 96,948 789,410 8.10
15 0.03072 0.00041 96,239 39 2,956 93,800 97.5 94,942 692,462 7.20
16 0.06361 0.00047 93,244 42 5,930 90,844 97.4 90,532 597,520 6.41
17 0.10919 0.00051 87,272 42 9,627 84,914 97.3 82,839 506,988 5.81
18 0.18499 0.00054 77,703 38 14,370 75,387 97.0 70,660 424,149 5.46
19 0.21168 0.00055 63,295 31 13,395 61,017 96.4 56,466 353,489 5.58
20 0.23286 0.00056 49,869 25 11,609 47,622 95.5 43,907 297,023 5.96
21 0.25988 0.00058 38,235 19 9,934 36,013 94.2 33,093 253,116 6.62
22 0.27009 0.00060 28,282 15 7,636 26,079 92.2 24,250 220,023 7.78
23 0.24127 0.00062 20,631 11 4,976 18,443 89.4 17,955 195,773 9.49
24 0.20919 0.00063 15,644 9 3,272 13,467 86.1 13,894 177,818 11.37
25 0.19091 0.00065 12,363 7 2,359 10,195 82.5 11,112 163,924 13.26
26 0.16344 0.00068 9,997 6 1,633 7,836 78.4 9,126 152,812 15.29
27 0.13507 0.00071 8,358 6 1,129 6,203 74.2 7,763 143,686 17.19
28 0.13374 0.00074 7,223 5 966 5,074 70.2 6,721 135,923 18.82
29 0.11805 0.00079 6,252 5 738 4,108 65.7 5,862 129,202 20.67
30 0.09403 0.00085 5,509 4 518 3,370 61.2 5,235 123,340 22.39
31 0.08532 0.00091 4,987 4 425 2,852 57.2 4,765 118,105 23.68
32 0.07274 0.00097 4,558 4 331 2,427 53.2 4,385 113,340 24.87
33 0.06827 0.00105 4,223 4 288 2,096 49.6 4,072 108,955 25.80
34 0.05659 0.00113 3,931 4 222 1,808 46.0 3,814 104,883 26.68
35 0.04851 0.00122 3,705 4 180 1,586 42.8 3,610 101,069 27.28
36 0.04359 0.00133 3,521 5 153 1,406 39.9 3,440 97,459 27.68
37 0.04047 0.00145 3,363 5 136 1,253 37.3 3,291 94,019 27.96
38 0.03740 0.00158 3,222 5 120 1,117 34.7 3,158 90,728 28.16
39 0.03445 0.00174 3,097 5 107 997 32.2 3,040 87,570 28.28
40 0.03166 0.00190 2,985 6 94 890 29.8 2,934 84,530 28.32
41 0.02902 0.00209 2,885 6 84 796 27.6 2,839 81,596 28.28
42 0.02660 0.00229 2,795 6 74 712 25.5 2,754 78,757 28.18
43 0.02438 0.00252 2,715 7 66 638 23.5 2,678 76,003 27.99
44 0.02237 0.00276 2,642 7 59 572 21.6 2,608 73,325 27.75
45 0.02028 0.00303 2,576 8 52 513 19.9 2,546 70,717 27.45
46 0.01886 0.00331 2,516 8 47 461 18.3 2,488 68,171 27.09
47 0.01768 0.00362 2,461 9 43 414 16.8 2,435 65,683 26.69
48 0.01648 0.00396 2,409 9 40 371 15.4 2,384 63,248 26.25
49 0.01530 0.00432 2,360 10 36 331 14.0 2,337 60,864 25.79
50 0.01415 0.00473 2,314 11 33 295 12.7 2,292 58,527 25.29
51 0.01300 0.00517 2,270 12 29 262 11.5 2,249 56,235 24.77
52 0.01189 0.00560 2,229 12 26 233 10.5 2,210 53,986 24.22
53 0.01085 0.00601 2,191 13 24 207 9.4 2,172 51,776 23.63
(continued)
Nuptiality Tables 237

Table 5.2 (continued)

vx qx l0 x d0 x f0 x N0 x %N0 x L0 x T0 x e0o x
Age (x) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
54 0.00984 0.00642 2,154 14 21 183 8.5 2,136 49,604 23.03
55 0.00888 0.00687 2,119 14 19 162 7.6 2,102 47,468 22.40
56 0.00799 0.00740 2,086 15 17 143 6.9 2,070 45,366 21.75
57 0.00718 0.00805 2,054 16 15 126 6.1 2,038 43,296 21.08
58 0.00650 0.00886 2,023 18 13 111 5.5 2,008 41,258 20.39
59 0.00590 0.00981 1,992 19 12 98 4.9 1,977 39,250 19.70
60 0.00540 0.01088 1,961 21 11 86 4.4 1,945 37,273 19.01
61 0.00496 0.01203 1,929 23 10 75 3.9 1,913 35,328 18.31
62 0.00458 0.01325 1,896 25 9 65 3.4 1,879 33,415 17.62
63 0.00432 0.01454 1,862 27 8 56 3.0 1,845 31,536 16.94
64 0.00413 0.01592 1,827 29 7 48 2.6 1,809 29,691 16.25
65C 0.02289 0.97711 1,791 1,750 41 41 2.3 27,882 27,882 15.57
Source: Adapted from Shryock, Siegel and Associates (1973: 561)

0
married), and d065C and f065C are found by simply apportioning l65 according to the
probabilities (proportions) q65C and v65C .
Column 6 of Table 5.2 gives the number of persons from the radix population
0
who marry for the first time at age x and all older ages. Denoted by Nx it has no
counterpart in a standard life table, and is calculated from:

N0x D †iDx;65C f0i (5.16)

0
Column 7, labelled % Nx , indicates the percentage of persons ‘surviving’, alive and
never married, at exact age x who subsequently marry (when aged x or older).
It follows that the difference between this value and 100 gives the percentage of
persons surviving unmarried at exact age x who subsequently die without ever
marrying. The relevant calculating equation is:
 
%N0x D N0x =l0x 100 (5.17)

In Table 5.2, White females had a 94.9 % chance of ever marrying at birth,
and a 5.1 % chance of dying without ever marrying. If they remained alive and
unmarried at exact age 30, for example, they had a 61.2 % chance of ever marrying
subsequently, and a 38.8 % chance of dying as a spinster.
Column 8 of Table 5.2 is the equivalent of the Lx column of the standard life
table. It indicates the number of person-years spent alive and never married at age x
(i.e., between exact ages x and x C 1) by the nuptiality table population. Values of
L0 x are calculated using:
   
L0x D 1=2 l0x C l0xC1 C 1=24 d0xC1 C f0xC1  d0x1  f0x1 (5.18)
238 5 Marriage, Marital Status and Relationships

Where the expression 1/24 .d0xC1 C f0xC1  d0x1  f0x1 / is an adjustment for non-
linearity of the survivorship function between exact ages x and x C 1.
Column 9 is the equivalent of the Tx column of the standard life table, indicating
person-years lived alive and never married at all ages beyond exact age x. We have
that:

T0x D †iDx;65C L0i D T0xC1 C L0x (5.19)

Finally, column 10 of Table 5.2 gives the average number of years of single life
remaining before marrying or dying for persons surviving alive and never married
at exact age x. In a manner analagous to that used to obtain eo x for a standard life
table it is given by:

e0 D T0x =l0x
o
x (5.20)

A net nuptiality table yields various measures which can be plotted for successive
0
synthetic cohorts (each having its own net nuptiality table). Values of lx at key
0
ages can be plotted; the Nx column can be used to divide the age distribution of
0
first marriages into quartiles; N0 gives the proportion of the synthetic cohort who
0
ever marry; % Ny , where y is the youngest age at which marriages occur, gives the
proportion of persons who survive to marriageable age who ever marry; and trends
in e0o x for certain values of x may be of special interest. Changes in these measures
may reflect changes in either first marriage or mortality levels and patterns, but once
again the importance of mortality is likely to pale into insignificance compared to
that of first marriage.
Two issues to conclude with. First, you may wonder whether there is any such
thing as an abridged nuptiality table (net or gross). There tends not to be much point
in constructing such tables because first marriage activity is heavily concentrated
over a very few single years of age. Without a single-years-of-age approach over this
age range a nuptiality table tells us very little. Moreover, if abridged nuptiality tables
were to be constructed, very complex issues of separation, for which satisfactory
generalized solutions such as exist for abridged life tables are not yet readily
available, would arise.
Finally, you should be aware that while gross and net nuptiality tables are often
used in the study of marriage trends and patterns they can introduce distortions into
the age distribution of first marriages. We discussed earlier how to obtain quartiles of
the age distribution of first marriages by finding the values of quarter, half and three-
quarters of the TFMR, then determining the exact ages at which the summation of
age-specific first marriage ratios which produces the TFMR reached these quartile
values. We can similarly use gross and net nuptiality tables to obtain quartiles of the
age distribution of first marriages, but these quartiles will rarely be identical to those
yielded by the TFMR-based procedure. Why?
You will recall that the TFMR can take on ‘impossibly’ high values – values
which imply that if a birth cohort were to experience the age-specific first marriage
Nuptiality Tables 239

ratios recorded in the relevant year, more than 100 % of the cohort would ever
marry. This phenomenon, we noted, occurred because of cross-sectional heaping
of first marriages, due to circumstances being temporarily particularly conducive
to marriage, a trend to earlier marriage being in progress, or both. Logically, it
clearly is perfectly possible for first marriages to heap cross-sectionally at levels
unsustainable in the longer term. In theory, for example, all members of a real birth
cohort could marry at the minimum legal age for marriage. This is an extreme and
highly improbable scenario, but if it happened all of that cohort’s first marriages
would take place during two consecutive calendar years during which it was passing
through the one-year ‘minimum age for marriage’ age group. There would be huge
numbers of first marriages at that age during those two years, causing the TFMR to
‘go through the roof’ (assuming older real cohorts were marrying ‘conventionally’
and contributing first marriages at older ages in the relevant two years), but at the
cost of there being no first marriages contributed by the cohort at older ages in
subsequent years (i.e., there would be compensating adjustments in cross-sectional
data for later years).
Nuptiality tables do not recognise the possibility that first marriages might
heap cross-sectionally at levels consistent with greater than universal marriage.
They force the percentage ever marrying in the synthetic cohort never to exceed
100. In doing so they can distort age distributions of first marriage when, for some
reason, the cross-sectional incidence of first marriage is especially high. Basically
the problem is that in a nuptiality table the number of first marriages a given age-
specific probability of first marriage implies for the distribution of first marriages
by age depends on probabilities at younger ages. These determine the number of
0
survivors (lx ) in the nuptiality table stationary population the probability is applied
to. In the extreme example cited in the previous paragraph, which would lead to
a probability of first marriage of 1.0 at the minimum age for marriage in one
synthetic cohort, the nuptiality table radix population for this synthetic cohort would
be totally depleted after the first marriageable age (i.e. l0yC1 would be zero, where
y was the youngest age for marriage). Probabilities of first marriage at older ages
would all be applied to zero survivors, producing zero first marriages. Despite this
example being implausible to the point of being ridiculous, the principle underlying
it is valid – in a nuptiality table, probabilities of first marriage at younger ages
0
get ‘first bite of the cherry’, and those at older ages are left to operate on the lx
residue that is left. The effect, when first marriages heap cross-sectionally, can
be likened to putting the lid on an overflowing rubbish bin. For the lid to fit the
rubbish must be compressed to the level of the top of the bin, but as one pushes the
rubbish down the degree of compaction is greater at the top of the bin, where the
pressure is being applied, than it is in the middle and at the bottom. In a situation
where cross-sectional first marriage activity is consistent with greater than universal
marriage, nuptiality tables compress activity at older ages in order that the ‘lid can
be put on the bin’ at 100 % ever marrying. The result is that numbers of first
marriages at older ages are understated relative to numbers at younger ages, and
measures like the quartiles of the age distribution of first marriages are distorted. The
beauty of the TFMR-based procedure for obtaining quartiles is that the number of
240 5 Marriage, Marital Status and Relationships

marriages at each age (the ratio of first marriages to total population) is determined
independently of what has happened at younger ages. Each one-year interval of
marriageable age is treated on an identical basis, there is no artificial lid on the
‘bin’, and undistorted quartiles of the age distribution of first marriages result.
‘What about when, instead of first marriages heaping cross-sectionally, there
is a marked cross-sectional deficit?’ you may ask. ‘Do nuptiality tables also yield
a distorted picture of marriage timing in that circumstance?’ The answer is ‘yes’.
In this situation, limited marriage activity at younger ages tends to lead to first
marriage probabilities at older ages being applied to abnormally large numbers of
0
survivors (lx ). The consequence is that numbers of first marriages at those ages are
boosted as a direct result of what is happening at younger ages, and comparatively
speaking the age distribution of first marriages overemphasizes older marriageable
ages. Again, the TFMR-based approach to measuring quartile ages at first marriage
has the advantage that the importance of any single-year marriageable age group in
no way depends on what happens at younger ages in the year to which the relevant
synthetic cohort pertains.

The Singulate Mean Age at Marriage

The measures and techniques for studying the process of first marriage dealt with
so far have presupposed the availability of reliable marriage registration data. In
many countries, and most notably in many of those collectively described as ‘less
developed’ or ‘developing’, such data are not available, and frequent use is made of
the singulate mean age at marriage, or SMAM. This is a measure of marriage timing
whose calculation requires only census or survey data showing the distribution
of members of a population of a given sex by age and marital status. More
specifically, the SMAM requires as input data census or survey percentages of
males/females never married in five-year age groups.
The SMAM gives the average number of person-years lived in the single (never
married) state by persons of a given sex who marry before age 50. The following
is a recipe for its calculation which explains the rationale behind each step in the
calculation process. The recipe assumes one has, as a point of departure, a set of
percentages of females or males never married in five-year age groups. As with other
methods for analysing marriage patterns and trends, it is vital to deal separately with
the two sexes.
Step 1: Assuming no marriages take place before age 15, add the percentages never
married for ages 15–19 to 45–49. If marriages take place before age 15, see the
note after Step 8 below.
Step 2: Multiply this figure by 5. This gives the number of person-years lived in the
never married state between exact ages 15 and 50 by a synthetic cohort of 100
people to whom the observed age-specific percentages never married applied.
The Singulate Mean Age at Marriage 241

Step 3: Add 1500 to the answer at Step 2. This number is the number of person-
years lived never married by the synthetic cohort before age 15 (100  15). We
thus now have total person-years lived never married before age 50.
Step 4: Average the percentages never married at ages 45–49 and 50–54. This
yields an estimate of the percentage still single at exact age 50.
Step 5: Multiply the figure obtained at Step 4 by 50. This gives person-years lived
never married up to exact age 50 by members of our synthetic cohort still single
at that age.
Step 6: Subtract the answer at Step 5 from that at Step 3. This gives person-years
lived never married up to exact age 50 by members of our synthetic cohort who
married before that age.
Step 7: Subtract the answer at Step 4 from 100. This gives the number out of our
synthetic cohort who married by exact age 50.
Step 8: Divide the answer at Step 6 by that at Step 7. This apportions person-
years lived never married before age 50 by members of our synthetic cohort
who married before age 50 among cohort members who married by exact age
50. It is, in other words, the average person-years lived never married by these
individuals, or the singulate mean age at marriage.
Note, however, that if you are dealing with a population in which some first
marriages take place at ages younger than 15 (i.e., the percentage of 10–14 year-
olds never married is less than 100), two modifications must be made to this recipe.
First, at Step 1, add in the percentage never married at ages 10–14 as well (i.e.,
add percentages never married at ages 10–14 to 45–49, rather than at ages 15–19 to
45–49). Second, at Step 3 add 1000, not 1500, to your answer at Step 2. There are
now only 10 years of life, not 15, through which our synthetic cohort lives without
being at risk of marrying. We take age group 10–14 into account at Step 1, where
we deal with ages at which first marriage does occur, rather than at Step 3, where
we deal with ages at which it does not occur.
As an example, consider Table 5.3, which shows the population of Bangladesh in
1974 by marital status, age and sex. We can use these data to obtain the age-specific
percentages never married for females and males that are given following Table 5.3,
which we will use to calculate singulate mean ages at marriage for women and men
using the recipe outlined above.
Because in Bangladesh in 1974 some marriages of both females and males had
taken place at ages 10–14 we must modify step 1 in accordance with the note given
following the SMAM recipe, and add, for each sex, percentages never married for
ages 10–14 to 45–49.
The singulate mean ages at marriage for Bangladeshi women and men in 1974
were 15.93 years and 23.96 years, suggesting an average age difference between
husbands and wives of around eight years. Obviously the SMAM for females at this
time was especially young, and conducive to high fertility. Subsequent calculations
show the female SMAM rising to 17.90 by 1991 and to 19.34 by 2011, with the male
SMAM rising to 24.85 and then 25.01 at these dates. So female ages at first marriage
have risen more rapidly than have those for men and the age difference between
242 5 Marriage, Marital Status and Relationships

Table 5.3 Marital status by age and sex, Bangladesh 1974

Age group Never married Married Widowed Divorced Total
Females
Under 10 12; 576; 557 0 0 0 12; 576; 557
10–14 3; 794; 840 370; 578 5; 125 23; 743 4; 194; 286
15–19 676; 761 1; 983; 740 26; 560 77; 496 2; 764; 557
20–24 80; 869 2; 320; 267 43; 982 50; 404 2; 495; 522
25–29 21; 929 2; 391; 576 67; 516 31; 047 2; 512; 059
30–34 11; 267 1; 892; 513 107; 461 16; 188 2; 027; 429
35–39 7; 648 1; 599; 096 161; 732 11; 204 1; 779; 680
40–44 6; 872 1; 232; 000 266; 850 8; 239 1; 513; 961
45–49 3; 639 824; 457 264; 319 5; 046 1; 097; 461
50–54 3; 659 665; 865 431; 272 4; 173 1; 104; 969
55–59 1; 853 305; 543 265; 882 2; 568 575; 846
60–64 2; 251 272; 159 486; 487 2; 463 763; 360
65 and over 4; 916 208; 707 785; 640 2; 077 1; 001; 340
Males
Under 10 12; 614; 426 0 0 0 12; 614; 426
10–14 4; 952; 787 32; 817 745 180 4; 986; 529
15–19 2; 912; 110 234; 682 5; 052 1; 909 3; 153; 753
20–24 1; 451; 164 944; 778 15; 409 4; 827 2; 416; 178
25–29 529; 113 1; 795; 054 23; 614 5; 656 2; 353; 437
30–34 115; 533 1; 893; 830 22; 347 4; 074 2; 035; 784
35–39 44; 098 1; 965; 393 22; 336 2; 960 2; 034; 787
40–44 26; 095 1; 688; 583 27; 303 2; 563 1; 744; 543
45–49 15; 178 1; 334; 676 27; 717 1; 723 1; 379; 294
50–54 12; 913 1; 230; 463 38; 627 1; 694 1; 283; 697
55–59 6; 298 739; 760 29; 211 758 776; 027
60–64 7; 846 855; 691 54; 593 1; 154 919; 284
65 and over 12; 345 1; 211; 675 147; 891 1; 076 1; 372; 987

Age group Percentage of females never married Percentage of males never married
Under 10 100.000 100.000
10–14 90.476 99.323
15–19 24.480 92.338
20–24 3.241 60.060
25–29 0.873 22.483
30–34 0.556 5.675
35–39 0.430 2.167
40–44 0.454 1.496
45–49 0.332 1.100
50–54 0.331 1.006
55–59 0.322 0.812
60–64 0.295 0.853
65 and over 0.491 0.899
(continued)
Consensual Partnering 243

Table 5.3 (continued)

Step number Females Males
1 Add percentages 120.842 284.642
never married at
ages 10–14 to 45–49
2 Multiply answer at 120.842  5 D 604.210 284.642  5 D 1,423.210
step 1 by 5
3 Add 1,000 to answer 604.210 C 1,000 D 1,604.210 1,423.210 C 1,000 D 2,423.210
at step 2 (in
accordance with the
note given following
the SMAM recipe)
4 Average percentages (0.332 C 0.331) / 2 D 0.3315 (1.100 C 1.006) / 2 D 1.0530
never married at ages
45–49 and 50–54
5 Multiply step 4 0.3315  50 D 16.575 1.0530  50 D 52.650
answer by 50
6 Subtract answer at 1,604.210  16.575 D 1,587.635 2,423.210  52.650 D 2,370.560
step 5 from answer
at step 3
7 Subtract answer at 100  0.3315 D 99.6685 100  1.0530 D 98.9470
step 4 from 100
8 Divide answer at 1,587.635 / 99.6685 2,370.560 / 98.9470
step 6 by answer at
step 7
SMAM D 15.93 SMAM D 23.96

husbands and wives has narrowed. Teenage marriage for females is, however, still
common in Bangladesh and significant scope still exists for further lowering fertility
by increasing female ages at marriage.
It is sometimes claimed that the SMAM and eo 0 in a gross nuptiality table are
equivalent. This, however, is only true if marriage is universal by exact age 50.
The SMAM is a measure of exposure in the single state by persons who marry
before they turn 50. By contrast eo 0 also takes into account exposure in the single
state, up to exact age 50, by those who fail to marry by that point in their lives
(i.e., by (i) those who marry for the first time after they turn 50 and (ii) those who
never marry). Thus, except where everyone marries before their 50th birthday, eo 0
is biased upward in comparison to the SMAM and should be higher than it.

Consensual Partnering

The rise of consensual partnering in developed populations as an integral component

of the ‘second demographic transition’ has greatly complicated the demographic
analysis of ‘partnering’ in the context of understanding fertility change. Where
244 5 Marriage, Marital Status and Relationships

formerly in most countries the commencement of co-residence and regular sexual

relations, and hence susceptibility to conception, largely coincided with formal
marriage in a climate devoid of other than relatively unreliable methods of con-
traception, nowadays consensual partnering and the capacity modern contraceptive
methods (often backed by ready access to induced abortion) have provided to avoid
parenthood for extended periods (and even completely) have substantially loosened
the link between marriage and becoming parents. The picture is further complicated
by the range of perceptions parties to consensual unions have of their unions,
both cross-sectionally and over time. Along both dimensions they range in their
levels of commitment from convenient arrangements for regular sexual gratification
and/or securing domestic services and/or reducing living costs through modern
approaches to ‘going steady’ and forms of engagement to marriage substitutes, with
no guarantee that both parties to a union share the same perception, and hence degree
of commitment, at a given point in time.
Consensual unions are generally not registered in the way marriages often are,
so cannot be studied via registration data. Indeed while marriages can usually be
dated because of the ceremony that surrounds them it can be difficult to date the
commencement of consensual unions, which often evolve from couples initially
staying the night with one another, then doing this with increasing regularity until
the point of continuing to maintain two residences is questioned. Some censuses
nowadays collect data on cohabitation that allows cohabitors and married persons
to be compared. Ideally separate questions are asked on legal marital status and
living arrangement (whether living with a spouse, living with a cohabiting partner,
or not in a coresidential relationship), so that the extent to which cohabitors have
been never and ever previously married can be examined, but with the pressure to
limit the number of questions in censuses a single question that compromises such
an investigation is often used.
More detailed studies of cohabitation have been based on survey data, and par-
ticularly on what are known as event history data. These provide detailed histories
of dates/ages of entry to and exit from cohabiting unions and marriages (as well
as of other events like births of children, migratory moves, changes of employment
status, etc.). They are analysed using event history analysis techniques, a battery of
regression models that overcome shortcomings (centred on censoring and the reality
that explanatory variables can be time-varying) of standard multiple regression
for analysing event history data (Allison 1984; Yamaguchi 1991; Blossfeld and
Gotz, 2001; Blossfeld et al. 2007; Brostrom 2011; Mills, 2011). It is not proposed
to explore these techniques here, but they include such methodologies as logit
regression, proportional hazards modelling and accelerated failure time models.
Analyses of partnering that use event history data focus on what is commonly
termed union formation – that is, entry into either cohabiting unions or marriage.
This can have the unfortunate effect of clouding the complexity of cohabitation
as an institution, tending to equate commencements of cohabiting unions with
boyfriends or girlfriends with those of marriages as life cycle events of comparable
moment and commitment, which in many instances they are not. Equally the
References 245

dissolution of cohabiting unions and marriages may be of dissimilar moment –

the ending of boyfriend-girlfriend relationships of perhaps limited duration that
were never more than tentatively committed as opposed to separations that presage
divorce. Demography still has a way to go to deal adequately with the inherent
complexity of consensual partnering. It serves different purposes for different
participants, and those purposes may differ for the parties to individual relationships
and may change over time.

References

Allison, P. D. (1984). Event history analysis: Regression for longitudinal event data. Newbury
Park/London/New Delhi: Sage.
Barrow, C. (1996). Family in the Caribbean: Themes and perspectives. Kingston: Ian Randle
Publishers.
Barrow, C. (1998). Caribbean masculinity and family: Revisiting ‘marginality’ and ‘reputation’.
In C. Barrow (Ed.), Caribbean portraits: Gender ideologies and identities (pp. 339–358).
Kingston: Ian Randle Publishers and Centre for Gender and Development Studies, University
of the West Indies.
Blossfeld, H.-P., & Rohwer, G. (2001). Techniques of event history modelling: New approaches to
causal analysis. Mahwah: Lawrence Erlbaum.
Blossfeld, H.-P., Golsch, K., & Gotz, R. (2007). Event history analysis with stata. Mahwah:
Lawrence Erlbaum.
Brostrom, G. (2011). Event history analysis with R. Bosa Roca: Taylor & Francis.
Carmichael, G.A. (1988). With this ring: First marriage patterns, trends and prospects in Australia.
Canberra: Department of Demography, Australian National University and Australian Institute
of Family Studies.
International Union for the Scientific Study of Population. (1958). Multilingual demographic
dictionary: English section. New York: United Nations Department of Economic and Social
Affairs.
Mills, M. (2010). Introducing survival and event history analysis. London: Sage.
Rubenstein, H. (1980). Conjugal behaviour and parental role flexibility in an Afro-Caribbean
village. Canadian Review of Anthropology and Sociology, 17(4), 330–337.
Shryock, Henry, S., Siegel, Jacob S. & Associates (1973) The methods and materials of
demography. 2nd Printing (revised). Washington, DC: U.S. Bureau of the Census.
Van de Kaa, D. (1987). Europe’s second demographic transition. Population Bulletin, 42(1), 1–57.
Wunsch, G. J., & Termote, M. G. (1978). Introduction to demographic analysis principles and
methods. New York: Plenum Press.
Yamaguchi, K. (1991). Event history analysis. Thousand Oaks: Sage.
Chapter 6
Analysis of Fertility

The Importance of Fertility Analysis in Demography

The analysis of human fertility is of central importance in demographic analysis.

In the first place, births are a vital component of the population balancing equation
introduced in Chap. 1, and thus a vital element in the study of population change.
More particularly they are, on a world scale, the single growth element in the
balancing equation, and concern over population growth arguably has been the
main stimulus to demography’s development as a discipline. Efforts to slow pop-
ulation growth typically concentrate on attempting to reduce fertility. Deliberately
increasing mortality is not an acceptable option, and while restricting immigration
sometimes offers short-term potential, promoting emigration usually is neither a
feasible nor a long-term solution either. The need to monitor efforts to reduce
population growth by reducing fertility highlights the importance of being able to
measure fertility, fertility change, and the sources of that change.
Fertility change also is intimately linked to the processes of social and economic
development. The movement of women away from traditional roles which see
them regarded more or less exclusively as bearers and nurturers of children, for
example, generally is reflected in fertility decline and change in the age pattern of
childbearing. The provision of health services, especially those directly linked to
reproduction, and development-induced changes to household economies may have
similar impacts. Debate over the causes and consequences of fertility change, and
in particular fertility transition from high to low levels, has generated a substantial,
hotly contested theoretical literature. Indices and approaches to analysis discussed
in this chapter are among the tools that have facilitated this debate, in some instances
having been developed specifically for it.

© Springer International Publishing Switzerland 2016 247

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_6
248 6 Analysis of Fertility

Some Terminology

It is important initially to become familiar with certain terms used in the fertility
analysis literature. The first is the word fertility itself, which in everyday, including
medical, usage refers to a person’s biological capacity to bear or father children. In
demography, however, a person’s fertility is their actual reproductive performance,
and it is this that we are seeking to measure when we talk about ‘measuring fertility’.
Demographers’ term for the biological capacity to bear children is fecundity, and
a woman capable of bearing a child is said to be fecund. The biological inability
to bear children is referred to as infecundity or sterility, and a woman unable to
bear a child is said to be infecund or sterile (the latter word, but not the former,
also being used to describe men who are biologically unable to become fathers).
In everyday usage ‘fertility’ is the opposite of ‘sterility’, but in demography only
the latter word retains its everyday meaning. Arguably if referring to the ‘fertility’
of men there could be doubt as to whether one meant reproductive performance
or biological capacity to reproduce (since one would not refer to the ‘fecundity’
of men), but while some would have it otherwise, fertility analysis in demography
conventionally focuses exclusively on women, so this tends to be a non-issue. That
said there has, of late, been interest in measuring male involvement in reproduction,
and some of this literature has used the term ‘paternity’ as the male equivalent of
female ‘fertility’ (Carmichael 2013).
If all this seems confusing, note that the French word for ‘fertility’ as understood
by demographers is ‘fécondité’, and that for ‘fecundity’ is ‘fertilité’. Be wary,
therefore, should you have occasion to read demographic literature written in French
(as much important demographic literature is).
Used alone the word ‘sterility’ usually is assumed to refer to an irreversible
condition, but a distinction is sometimes made between temporary sterility and
permanent sterility. Among women a distinction may also be made between
primary sterility (where a woman has never been able to have a child) and
secondary sterility (where the biological inability to bear children develops after
a woman has already had one or more children).
Reference was made in Chap. 1 to menarche and menopause as respectively
the points in the life cycle at which a woman first acquires, and finally loses,
the biological capacity to conceive and bear children. More specifically, menarche
marks the onset, and menopause the final cessation, of menstruation, a woman’s
experience of menstrual periods. Menarche is a discrete event, occurring with the
occurrence of the first menstrual period, but ‘menopause’ often refers to an ill-
defined period of the life cycle over which menstruation gradually ceases.
Another word commonly encountered in analyses of fertility is amenorrhoea.
It means a temporary absence of menstruation. This occurs, of course, during a
pregnancy (pregnancy amenorrhoea), but of greatest interest to demographers is the
extent to which it occurs following a confinement (the process of actually giving
birth), when it is known as post-partum amenorrhoea. The length of the period of
post-partum amenorrhoea is highly variable, in particular as a function of whether
Basic Fertility Measures 249

or not, and for how long, a baby is breastfed. This variability and its relation to
breastfeeding are of interest to demographers in populations practising little or
no efficient contraception, because they affect how quickly after a birth a woman
is at risk of becoming pregnant again, and therefore the interval between births.
Promoting breastfeeding is one way of attempting to reduce fertility (as well as of
improving infant and maternal health and survival chances) in such populations,
because in increasing the period of post-partum amenorrhoea it lengthens the
average interval between births, thereby reducing the number of children a woman
bears over a given life cycle phase.
Fertility data distinguish between live births and stillbirths. The World Health
Organization defines a live birth as occurring when a foetus, whatever its gestational
age, exits the mother’s body and subsequently breathes or shows any sign of life,
such as voluntary movement, a heartbeat or pulsation of the umbilical cord, for
however brief a time and regardless of whether the umbilical cord and placenta are
intact (WHO 1993). A stillbirth occurs when a foetus has died in the uterus. The
basic WHO definition of a foetal death is the intrauterine death of any conceptus
at any time during pregnancy, but legal definitions often require the foetus to have
attained a prescribed gestational age or birth weight to distinguish foetal deaths
(stillbirths) from miscarriages. Variability in the minimum gestational age or birth
weight prescribed in such definitions means you should always check the precise
definition in use when dealing with data on foetal deaths/stillbirths.

Basic Fertility Measures

In keeping with the discussion of Chap. 3, fertility may be measured and analysed
in either period (cross-sectional) or cohort terms. In further keeping with that
discussion, the former approach deals with births arranged vertically through the
Lexis diagram (see Fig. 3.7), and the latter with births arranged diagonally through
it and occurring normally to birth cohorts of women (although analysis in terms
of other types of cohorts – e.g., marriage cohorts – is also conceivable). Period
analyses of fertility have the advantage of being up to date, with a new year’s
measures added as each year’s data are released, and are useful for forecasting
purposes. Cohort analyses reflect the experience of actual birth cohorts, and are
not distorted by transient period, or cross-sectional, effects. The more common and
readily calculated fertility measures tend to be period measures, several of which
were introduced in Chap. 1 in illustrating the more general discussion there of types
of demographic measures.
Recapping, the crude birth rate (CBR) is the most basic relative measure of
fertility. Defined as the ratio of live births during year y to the mean (or mid-year)
total population during year y, multiplied by 1,000 (see Chap. 1, Eq. 1.7), it is not a
‘true’ rate, because its denominator is not even a moderately accurate measure of the
population at risk of giving birth. It includes men, and also women who are both too
young (yet to reach menarche) and too old (post-menopausal) to be able biologically
250 6 Analysis of Fertility

to bear children. Thus the CBR is for use when no more refined data are available,
typically for broad international comparisons. It is, however, a better summary
measure of fertility than is the crude death rate of mortality, despite the latter being
a true rate. The proportion of total population who are females of childbearing age
tends to be less variable than does the proportion who are in the high mortality
age groups. This means that the sort of situation encountered in comparing CDRs
for Australia and Malaysia in Chaps. 1 and 2, where the underlying differential in
mortality levels was so distorted by differences in age structure as to be reversed, is
less likely to occur.
A brief digression is in order while we are discussing the relative merits of the
CBR and the CDR. You will recall from the discussion of the population balancing
equation in Chap. 1 that population change over time is the sum of two components:
natural increase and net migration. The difference between the CBR and the CDR
(i.e., CBR – CDR) gives a measure known as the crude rate of natural increase.
Let us, though, return to our consideration of basic fertility measures.
In endeavouring to obtain measures of fertility more refined than the CBR the
broad refinement principle outlined in Chap. 1 is applied. We seek to close in on
the population at risk – by eliminating from the denominator those who could not
possibly give birth, recognizing that certain personal characteristics make a woman
more, or less, likely to give birth, or both. As previously discussed (see Chap. 1,
Eq. 1.10), a first refinement of the CBR is to calculate the general fertility rate
(GFR) by restricting the denominator to the female population aged 15–49 (i.e.,
to those persons who, broadly, are biologically capable of bearing children). This
denominator is still only a rough approximation of the population at risk, since
females may be fecund before age 15 and beyond age 49, and some of those aged
15–49 will be infecund (either involuntarily through having yet to reach menarche,
already passed menopause or suffering from some dysfunction of the reproductive
organs, or voluntarily through having been surgically sterilized). It could also be
argued that some fecund women aged 15–49 are at no risk of giving birth because
they are not sexually active or their sexual partners are sterile. Then, too, risk occurs
in varying degrees among those actually exposed to it. Some women are more
fecund or have sexual partners with more potent semen than others, and so conceive
more readily; differing patterns of partner absence and frequencies of coitus affect
risk, as do varying tendencies to miscarry, have stillbirths and have menstruation
interrupted by breastfeeding or nutritional deficiencies; and increasingly, use of
contraception has facilitated self-definition of risk. The concept ‘population at risk’
thus becomes very complicated when thought about in detail. But the denominator
of the GFR (sometimes restricted to females aged 15–44 instead of 15–49) clearly
represents a much more reasonable population at risk of giving birth than does the
total population.
A second refinement discussed in Chap. 1 was to recognize that married women
are more likely to be sexually active and to be consciously trying to have children
than are unmarried women, so that there was a case for splitting the GFR by marital
status into a (general) marital fertility rate (see Chap. 1, Eq. 1.11) and a (general)
non-marital fertility rate (sometimes also called the illegitimacy rate, although this
Basic Fertility Measures 251

terminology is generally frowned upon these days) (see Chap. 1, Eq. 1.12). A more
fundamental refinement, however, is by age, since within the reproductive age range
(15–49 years) a variety of both biological and social factors make childbearing more
likely at some ages than at others. Comparisons of GFRs thus are susceptible to
distortion due to differences in age structure within the reproductive range. The
calculation of age-specific fertility rates (or birth rates) (ASFRs) is one way of
focusing on this age variability. Techniques of standardization and decomposition
can be applied to the GFR as to other demographic indices to control for differences
in age structure, but in fertility analysis the detail offered by age-specific rates
often is inherently of interest and is not too difficult to digest given that the female
reproductive ages span only seven five-year age groups. Age-specific rates are also
the building blocks of the most widely used fertility measure – the total fertility rate
(or ratio) (TFR – see below).
If data on births by age of father are available, it is possible to calculate ASFRs
(or age-specific paternity rates – ASPRs) for males. Usually, however, they are
computed only for females. The equation for a (female) ASFR is:

ASFRx D .Bx =Px / 1;000 (6.1)

Where x denotes an age group; Bx D live births to women aged x during year y;
Px D the mean (or mid-year) female population aged x in year y.
While in theory the age range 15–49 years can be divided up in many ways for
the purpose of calculating ASFRs, in practice rates most commonly are calculated
for five-year age groups, sometimes for single-year age groups. ASFRs for five-
year age groups 15–19 to 45–49 years often are used to contrast fertility patterns
of discrete populations (Does fertility peak at different ages? How do relativities in
fertility levels at different ages compare?), and to compare patterns for individual
populations over time (Was fertility increase or decline concentrated at particular
ages?).
Both five-year and single-year ASFRs also are used to obtain the TFR. Relevant
equations were given as Eqs. 1.19 and 1.20, respectively, in Chap. 1. As indicated
at that point, the TFR measures the number of children the average woman would
have during her life assuming she experienced at each reproductive age the ASFR
for that age recorded during the year for which the index is calculated. It is a very
widely used measure of fertility. Demographers often summarize fertility change,
especially fertility decline, by quoting before/after TFRs (e.g., the TFR fell from
6.4 to 3.7), and frequently judge the fertility level in a population with reference
to a TFR of from about 2.1 (for low mortality populations) to about 2.4 (for high
mortality populations). These are approximate levels of fertility required for long-
term replacement – i.e., levels that, maintained indefinitely along with existing
levels of mortality, would see a population exactly replace itself, the new generation
being exactly the same size as the older, parental generation. The figures of 2.1–
2.4 comprise one child to replace the woman, one to replace her male partner,
and 0.1–0.4 as a contribution to the additional fertility required to replace those
252 6 Analysis of Fertility

unable to participate, or participate fully, in the replacement process because of

premature death. Clearly the lower the mortality in a population the lower this ‘per
woman’ contribution needs to be. In Australia, which has low mortality, the precise
value of the replacement level TFR currently is about 2.06, but 2.1 is a commonly
used general guideline for developed country populations. A TFR below long-term
replacement, if maintained, signals eventual population decline, although the onset
of decline invariably is delayed for quite some time by momentum for increase
built into the population’s age structure by past, above replacement, fertility; a TFR
above long-term replacement level implies a growing population.
While specific fertility rates are most commonly computed specific for age, there
are other possibilities. Marital status already has been mentioned. Fertility levels,
differentials and trends for the married and unmarried components of populations
often are very different, and changing or different marital status compositions can be
important explanations of fertility trends or differentials as indexed by measures like
the CBR or GFR. In addition to the (general) marital and non-marital fertility rates
already discussed, age-specific marital fertility rates (ASMFR) and age-specific
non-marital fertility rates (ASNMFR) can be valuable in examining such issues.
Their equations are:
 
ASMFRx D B.m;x/ =P.f;m;x/ 1;000 (6.2)

Where x denotes an age group; B(m,x) D marital live births to women aged x during
year y; P(f,m,x) D the mean (or mid-year) married female population aged x in
year y.
And:
 
ASNMFRx D B.n;x/ =P.f;u;x/ 1;000 (6.3)

Where x denotes an age group; B(n,x) D non-marital live births to women aged x
during year y; P(f,u,x) D the mean (or mid-year) unmarried female population aged
x in year y.
Of course the contemporary trend to cohabitation rather than marriage for
significant chunks of the life course in developed populations raises the issue of
exposure to the risk of bearing children while cohabiting. Cohabitors of childbearing
age are obviously likely to be regularly sexually active as married couples are, and so
exposed to the risk of pregnancy. They nonetheless tend to be treated as unmarried,
and their fertility is captured by the various non-marital fertility rates. In theory,
if mid-year estimates of population cohabiting and annual counts of non-marital
births to cohabiting couples were available it would be possible to compute separate
specific fertility rates for cohabiting (as opposed to non-coresident) unmarried
women, but thus far requisite data are rarely available.
Order-specific fertility rates (OSFR) and age-order-specific fertility rates
(AOSFR) are another two types of specific fertility rates which pay attention to
birth order (whether a birth is the mother’s first, second, third, etc. live birth).
Basic Fertility Measures 253

Birth order and its companion attribute, a woman’s parity, are important variables
in fertility analysis, since as the number of children a woman has increases so do
the likelihoods (a) that she will wish to take steps to avoid further births and (b)
of her physiological capacity to have further children being impaired by previous
childbearing and/or advancing age. The relevant equations are:
 
OSFRi D Bi =P.f;1549/ 1;000 (6.4)

Where i denotes a birth order; Bi D live births of order i during year y;

P(f,1549) D the mean (or mid-year) female population aged 15–49 in year y.

 
AOSFR.x;i/ D B.i;x/ =P.f;x/ 1;000 (6.5)

Where x denotes an age group; i denotes a birth order; B(i,x) D live births of order i to
women aged x during year y; P(f,x) D the mean (or mid-year) female population
aged x in year y.
Note that in these two equations the denominators are respectively all women
of reproductive age and all women aged x; they are not women of parity i-1 and
the relevant age who, ignoring the complication of multiple births, are the women
strictly at risk of producing births of order i. It follows that the sum of OSFR over
all birth orders is the GFR, while the sum of AOSFR over all birth orders for age
group x is the ASFR for age group x. The value of these indices is thus that they
allow the GFR and ASFRs to be partitioned by birth order. Such exercises may
aid understanding of the mechanics of fertility change over time – for example,
the extent to which overall change reflects change in the frequency of higher order
births.
You do need to be wary of registration data on births by birth order (or
mother’s parity). They can be seriously inaccurate. In Australian birth registration
data, for example, births by mother’s parity have historically been based on
questions on birth registration forms (the precise wording of which varies across
the States and Territories that have jurisdiction for vital registration) asking for
details of previous children of this marriage or, more recently, of this relationship
(Carmichael 1986; Corr and Kippen 2006). So they exclude children of previous
marriages/relationships, and when ‘this marriage’ was the referent, previous issue
of the mothers of non-marital births, and hence the parities of those births, were
ignored altogether. Parities have thus been systematically understated through such
things as a previous non-marital birth by another man being ignored, or a woman
divorcing after having children, remarrying, and then returning to parity one when
bearing a first child by her second husband. The Australian Bureau of Statistics
(ABS) has been trying for some years to persuade State/Territory registrars of births
to collect parity data that reflect a woman’s complete fertility history, but the alleged
sensitivity of requiring disclosure of the details of previous births a current partner
may be unaware of has prevented this happening in some States. As of 2007 the
254 6 Analysis of Fertility

ABS began publishing parity data of this type for States/Territories that collected it,
but this excluded two of the largest States, Queensland and Victoria, for which such
data continue to be impaired. Current national data are thus hybrid.
Fertility rates specific for age, marital status and birth order are but three
examples of many specific rates that might be useful in fertility analysis. Rates
specific for variables such as educational attainment, ethnic group, occupation,
contraceptive use, income, etc. are also conceivable. Their equations take the form
‘live births to women with the relevant attribute divided by the mean, or mid-year,
female population with that attribute, multiplied by 1,000’, and invariably they could
usefully be calculated specific also for age.

Fertility Measures from Censuses and Surveys

Measures discussed in the previous section require two types of data. They require
data on population, classified by age, sex and possibly other variables, and data
on live births, classified by age of mother and possibly other variables. The former
typically derive from a census or from census data adjusted to the date of interest
for intervening vital events; the latter normally derive from vital registration. Many
countries, especially developing countries, do not, however, have reliable vital
registration data, and in this circumstance fertility measures which require only
census or survey data are widely used. The focus in this section is on measures
which can be obtained directly from such sources, as opposed to those obtained
from them using what are known as the methods of indirect estimation (or indirect
methods) (United Nations 1983).
Two types of fertility data commonly are collected by censuses: Data on
children ever born, and on births within the previous 12 months (so-called ‘current
fertility’). These types of data may also be gathered through surveys, although
surveys may collect other types of fertility data as well, the most detailed of all
being full pregnancy histories.

Measures Based on Children Ever Born

Data on children ever born (CEB) can be used to construct estimates of standard
fertility measures (like the TFR) using indirect methods. However, these methods
require either data from two successive censuses or acceptance of an assumption
that fertility has not changed in the recent past. Assumptions of lack of change in
demographic parameters often are made in demographic analysis for populations
for which data are poor, and many techniques of analysis are only modestly
sensitive to this type of assumption and therefore quite robust. However, methods
of fertility analysis which assume lack of change usually are VERY sensitive to
that assumption, and should be used with extreme caution. Methods of indirect
Fertility Measures from Censuses and Surveys 255

estimation do not, though, fall within the scope of the present discussion. Data on
CEB are mentioned because they also yield other useful summary measures.
One such measure is the proportion of women childless at each age. This, at
younger ages, provides information about the speed with which childbearing begins
in a population, but a more common use is as a measure of the extent of infecundity
at ages above exact age 40. When used for this purpose in relation to populations in
which marriage is not universal, calculations obviously should be restricted to ever
married women.
A more widely used measure using CEB data is the age-specific mean number
of children ever born. For any age group this mean summarizes the cohort’s fertility
history, and its main use is in differential fertility analysis – i.e., in comparisons of
the fertility experience of subgroups of women within the population under study.
Provided they are reliable, mean CEB data are likely to facilitate a much wider
range of fertility comparisons than are ‘standard’ measures like age-specific and
total fertility rates. The detailed cross-tabulations of births needed to calculate the
latter for many subgroups frequently are not available in registration data, whereas
mean CEB data, deriving from a single data source and one apt to include a wide
range of personal attributes besides CEB, generally offer a much wider range of
comparison options (e.g., by education, labour force status, rural-urban residence,
ethnicity, religion).
The qualification ‘provided they are reliable’ made in respect of mean CEB data
must not, however, be lightly brushed aside. Regrettably, CEB data frequently have
deficiencies. At younger ages it is not uncommon to encounter large numbers of
‘not stated’ responses because of failure to obtain information for young unmarried
women. Census enumerators may have felt uncomfortable asking the question of
such women or have met frequent refusals to answer it, or in self-enumerations
householders may have deemed the question irrelevant or inappropriate in respect of
their unmarried daughters. A mean CEB calculation which ignores ‘not stated’ cases
in this circumstance generally has a serious upward bias, since the great majority of
those cases are likely to be childless. There are analytical techniques which attempt
to compensate for this bias, but it is safest to work on the basis that the calculation
and comparison of mean numbers of CEB should only be undertaken where the ‘not
stated’ category is small. Often, in fact, CEB questions are not even asked of the
never married, in which case mean CEB data obviously cannot be interpreted as if
they applied to all women.
It is also sometimes found that mean CEB for older age groups (beyond, say,
exact age 35 or 40) are less reliable than those for younger age groups. Possible
reasons for this include the exclusion of deceased children and, in societies in
which adoption is common, confusion between biological and social parenthood.
Again, caution is called for. Age-specific mean numbers of CEB are nonetheless
an extremely valuable type of demographic measure, especially for analyses of
differential fertility focused on age group 30–34 (and perhaps the two adjacent age
groups), at which data defects just canvassed tend to be least in evidence.
256 6 Analysis of Fertility

‘Current Fertility’ Measures

Asking about live births in the preceding 12 months is most common in censuses
and surveys undertaken in less developed countries, because it yields data almost
equivalent to birth registration data (and these are, of course, the types of countries
least likely to have such data). In theory ‘current fertility’ data are in one sense
superior to registration data; they facilitate calculation of standard fertility measures
for the year preceding the census using a single data source, thereby eliminating
problems of differential levels of enumeration in registration and census data.
However, they have other problems of their own.
There are often problems of coverage. For example, respondents may not observe
a precise 12-month reference period, births ending as infant deaths may be ignored,
and distinguishing stillbirths from early neonatal deaths (and hence whether a live
birth occurred) may be more problematic than ever. Also, births are classified by
the mother’s age at enumeration, which may be different from her age at the date of
the birth. On average births occurred 6 months prior to the census or survey, so that
when classified by mother’s age in five-year groups, these age groups effectively are
displaced 6 months. This need not deter direct calculation of a TFR, but this is not
often done. Instead, current fertility data more typically are used as input for indirect
methods – not always with certainty that the outcome will be an improvement on a
direct calculation.

The Child-Woman Ratio

The measures discussed so far in this section often are regarded as part of the body
of methods of indirect estimation, although there would seem to be nothing indirect
about calculating the mean number of CEB for a female age cohort. But another type
of fertility measure that undeniably is calculated directly from census data (though
how direct a measure of fertility it is is another question) is the child-woman ratio.
The child-woman ratio is different in not deriving from any type of fertility data
as such. It requires only a census age-sex distribution, and is the ratio of children
aged 0–4 to females aged 15–49, multiplied by 1,000 (see Chap. 1, Eq. 1.18). It
is a somewhat rough measure, making no allowance, in a comparative context,
for differences in female age structure across the reproductive ages, for differently
shaped age-specific fertility distributions or for different levels of child survival. It
is perhaps best regarded as giving a reasonable approximate guide to differences in
fertility levels in situations where its limited data requirements commend it.
Measures of Reproduction 257

Measures of Reproduction

The concept of reproduction, or reproductivity, as conventionally used in demog-

raphy, has to do with the extent to which a cohort of women (either real or
synthetic) – a generation of women – replaces, or reproduces, itself (by having
daughters). Its analysis led Alfred Lotka in the early twentieth century (Leonard
Euler independently pre-empted Lotka as early as 1760) to the concept of a stable
population, which underpins several techniques of demographic analysis today.
Euler and Lotka demonstrated that a closed population subjected indefinitely to
fixed schedules of age-specific fertility and mortality rates ultimately assumed a
fixed, or stable, age distribution (i.e., an age distribution with a constant shape)
and a constant rate of natural increase. Stable population theory will be addressed
in Chap. 8. Reproduction rates can in theory be calculated for males or for both
sexes combined as well as for females, but because the female reproductive span is
shorter and better defined, and because mothers’ ages are more likely to be known
than fathers’ in the case of non-marital births, rates linking daughters to mothers are
the norm.

The Gross and Net Reproduction Rates

In distinguishing in Chap. 3 between the tempo and intensity of (real) cohort

demographic processes, equations for two generalized measures of intensity were
presented (as Eqs. 3.1 and 3.2), and it was noted that for the process of women
giving birth to female children these equations yielded measures known as the net
reproduction rate (NRR) and gross reproduction rate (GRR). These are the core
measures of reproduction, and it is appropriate to give here explicit equations for
them (in the real cohort context).
 
GRR D †xDl;¨ b.f;x/ =1 Fx (6.6)

Where b(f,x) D the number of female births occurring at age x to a female birth
cohort; 1 Fx D the number of survivors in the female birth cohort at age x (i.e.,
at the point in time when all cohort members are aged between exact ages x and
x C 1); l D the youngest age at which any cohort member gave birth to a female
child; ¨ D the oldest age at which any cohort member gave birth to a female
child.
 
NRR D †xDl;¨ b.f;x/ =N0 (6.7)

Where b(f,x) D the number of female births occurring at age x to a female birth
cohort; N0 D the original size of the female birth cohort at exact age 0; l D the
258 6 Analysis of Fertility

youngest age at which any cohort member gave birth to a female child; ¨ D the
oldest age at which any cohort member gave birth to a female child.
If we consider these equations applied in respect of a closed population it is
easy to appreciate that the NRR measures the actual extent to which a cohort of
women replaces itself with daughters, while the GRR measures the extent to which
the cohort would replace itself if all members survived to the end of the reproductive
period.
Equations 6.6 and 6.7 and the real cohort, closed population context in which they
apply are helpful in understanding the GRR and NRR. There are, however, period,
or synthetic cohort equivalents which use cross-sectional data, and these tend to be
the GRRs and NRRs demographers ordinarily calculate. Thus, with single-year-of-
age data the usual calculating equations for the GRR and NRR are:
 
GRR D †xDl;¨ ASFFRx D †xDl;¨ b.f;x/ =Fx (6.8)

Where ASFFRx stands for the age-specific fertility rate at age x based on female
births only; b(f,x) D the number of female births occurring to women aged x in
year y; Fx D the mean (or mid-year) female population aged x in year y; l D the
youngest age at which any woman gave birth to a female child in year y; ¨ D the
oldest age at which any woman gave birth to a female child in year y.
And:
 
NRR D †xDl;¨ ASFFRx : Lx =l0 D †xDl;¨ b.f;x/ =Fx .Lx =l0 / (6.9)

Where ASFFRx stands for the age-specific fertility rate at age x based on female
births only; Lx / l0 is the life table survival ratio from birth to age x from a single-
year-of-age life table for the female population in question; b(f,x) D the number of
female births occurring to women aged x in year y; Fx D the mean (or mid-year)
female population aged x in year y; l D the youngest age at which any woman
gave birth to a female child in year y; ¨ D the oldest age at which any woman
gave birth to a female child in year y.
Two difficulties may arise in applying Eqs. 6.8 and 6.9. First, data on births
may not be refined by sex. Second, single-year-of-age data and/or a single-year-
of-age life table may not be available. If data on female births are not available,
approximations of values of ASFFRx can be obtained by multiplying values of
ASFRx (age-specific fertility rates based on total births) by 100/205. This procedure
makes use of the empirical regularity that a sex ratio of approximately 105 male
births per 100 female births is observed in most populations, and if a more precise
sex ratio at birth is available the denominator 205 can be adjusted accordingly.
Where single-year-of-age data or, in the case of the NRR, a single-year-of-age
life table are not available there are modifications of Eqs. 6.8 and 6.9 which allow
calculations to be based on data for five-year age groups and an abridged life table.
Indeed, if undertaking calculations manually, these equations have considerable
Measures of Reproduction 259

appeal even if their use is not essential, as they greatly reduce the quantity of
mechanical computation. One minor drawback is that calculations normally are
limited to the age range 15–49 years. Small numbers of female children born outside
that range sometimes are assigned to the adjacent five-year age group (15–19 or
45–49), or if numbers are non-negligible it is possible to extend the summations
in Eqs. 6.10 and 6.11 over age groups 10–14 and/or 50–54. Extension to include
age group 10–14 is in particular not uncommon. The relevant equations based on
five-year age groups are:
 
GRR D 5†xD1519;4549 b.f;x/ =Fx (6.10)

Where b(f,x) D the number of female births occurring to women aged x in year y;
Fx D the mean (or mid-year) female population aged x in year y.
And:
   
NRR D 5†xD1519;4549 b.f;x/ =Fx .5 Li =5l0 / D †xD1519;4549 b.f;x/ =Fx .5 Li =l0 /
(6.11)

Where i D the lower limit of age group x; 5 Li / 5l0 is the life table survival ratio
from birth to the five-year age group whose lower limit is exact age i from an
abridged life table for the female population in question; b(f,x) D the number of
female births occurring to women aged x in year y; Fx D the mean (or mid-year)
female population aged x in year y.
With Eqs. 6.10 and 6.11, as with Eqs. 6.8 and 6.9, if data on female births are
not available the ASFFRx element b(f,x) / Fx can be replaced by values of ASFRx
multiplied by 100/205 (or a more precise alternative sex ratio).
By now something about the GRR may have struck you. It is the TFR based on
female rather than total births. It measures the number of daughters (as opposed to
children) a woman would have through her life if experiencing the age-specific rates
of having female (as opposed to total) births prevailing in the year for which it is
calculated. This link between the GRR and the TFR allows a good approximation
of either to be obtained from the other, simply by using the appropriate multiplier
based on the sex ratio at birth. Most commonly the TFR is used to estimate the GRR,
the former being multiplied by 100/205 (or thereabouts). To convert in the opposite
direction one would use the inverse of this multiplier (205/100).
Brief reference was made earlier to the concept of long-term replacement.
Strictly speaking, replacement level fertility occurs when age-specific fertility rates
based on female births and age-specific female mortality rates combine to produce
a net reproduction rate of 1.0. Thus achievement of replacement level fertility is
slightly dependent on mortality conditions, and as previously suggested tends to
coincide with a TFR of from about 2.1 in low mortality populations to about 2.4
in high mortality populations. In the latter, the survival of proportionately fewer
women to the end of the reproductive period means that the average woman who
does survive needs to make a larger contribution beyond the two children to replace
260 6 Analysis of Fertility

herself and her male partner to compensate for the lost fertility of women who die
prematurely. A NRR of 1.0 corresponds also with a GRR slightly higher than 1.0,
and hence replacement level fertility can coincide with a GRR of from about 1.01 to
about 1.20. The GRR is always higher than the NRR because it assumes all women
survive through the reproductive period to experience the ASFFRs recorded during
the year in question. The difference between the two measures tends to be slight
when female mortality before age 50 is low, and larger where female mortality at
those ages is higher. The ratio of the NRR to the GRR yields the reproduction-
survival ratio; the proportion of potential reproductivity that survives the effect of
mortality. With a NRR of 1.0 implying exact replacement; anything lower indicates
sub-replacement fertility, and a higher value above replacement fertility.
As a measure of how large the generation produced by its predecessor will be in
proportionate terms, the NRR has a readily understandable meaning. However, just
as a below replacement TFR does not signal an immediate decline in population
numbers, neither does a NRR below 1.0 give this indication. Normally, because
of momentum for growth built into a population’s age structure, NRRs below 1.0
would need to persist for some time before decline began. The NRR was a popular
measure in the 1920s and 1930s, when low values for industrialized countries
were interpreted as indicating levels to which fertility would ultimately fall, but
it was labelled misleading when values suddenly rose again with the baby boom
that occurred following the Second World War. In both instances the measure was
being interpreted naively in real, rather than synthetic, cohort terms, but it has never
recovered its earlier popularity (for a fuller discussion see Shryock, Seigel and
Associates (1973: 534–535)). Requiring fewer data, the TFR has become by far
the preferred and more widely used measure.
Another important point to grasp about NRRs is that differences between them
do not necessarily measure differences in rates of population growth, even if
mortality conditions in populations being compared are identical. The reason is that
net reproduction can take different lengths of time, depending on the mean age
of childbearing. A population that has high age-specific fertility at younger ages
reproduces itself (creates the next generation) more quickly than one with the same
TFR and NRR, but with age-specific fertility distributed more to older reproductive
ages. The latter population is said to have a longer length of generation.

Mean Length of Generation

The mean length of generation for a population is defined as the mean age of
mothers at the birth of their daughters. It is an index designed to measure the
speed with which a group of mothers replaces itself with another group of potential
mothers. With single-year-of-age data and a relevant single-year-of-age life table
the relevant equation is:
Describing the Age Pattern of Fertility 261

       
MLG D †xDl;¨ b.f;x/ =Fx .Lx =l0 / .x C 0:5/ = †xDl;¨ b.f;x/ =Fx .Lx =l0 /
(6.12)

Where all elements mean the same as in Eq. 6.9.

With data for five-year age groups and a relevant abridged life table the relevant
equation is:
       
MLGD †xD1519;4549 b.f;x/ =Fx .5 Li =l0 / .mx/ = †xD1519;4549 b.f;x/ =Fx .5 Li =l0 /
(6.13)

Where mx is the midpoint of age group x; other elements mean the same as in
Eq. 6.11.
As with Eqs. 6.8, 6.9, 6.10, and 6.11, if data on female births are not available
the ASFFRx element b(f,x) / Fx in Eqs. 6.12 and 6.13 can be replaced by values of
ASFRx multiplied by 100/205 (or a more precise alternative sex ratio at birth). And
in Eq. 6.13, summations can be extended over age groups 10–14 and/or 50–54 if
non-negligible childbearing occurs at either or both of those ages.
Note that in Eqs. 6.12 and 6.13 the denominators are the NRR as given by
Eqs. 6.9 and 6.11 respectively. MLG is a weighted mean of the ages at which births
of daughters occur, ages being represented by the mid-points of the single-year or
five-year age groups within which births occur, and the weights being, in effect,
numbers of female births at those ages in the life table stationary population.
Values of MLG tend to fall in the range 25–30 years, and results radically outside
this range should be carefully checked, although values over 30 for developed
populations widely delaying childbearing into their 30s are not uncommon. A
calculation of the MLG for Australia in 2011 using five-year age group data (which
incorporates calculations of the GRR and NRR) is presented in Table 6.1. This table
allocates known total births by age to female births by age using an observed overall
ratio of female births to total births of 146,621/301,617D 0.486116.
Thus Australia in 2011 had sub-replacement fertility. With the NRR at 0.922,
replacement of females of reproductive age by daughters was 92.2 %. The mean
length of generation was 30.5 years, having risen from 28.4 years in 1991. Thus
the deferment of fertility to older ages over this 20-year period as many women
prioritized other things (education, careers, travel, their social lives) through their
teens and twenties had increased the MLG by 2.1 years.

Describing the Age Pattern of Fertility

In comparing the fertility of different populations or studying fertility change,

demographers are interested in differences/changes in the tempo of fertility as well
as in differences/changes in its level, or intensity. One way in which something of
a focus on differences/changes in tempo is facilitated is through an examination of
262 6 Analysis of Fertility

Table 6.1 Calculation of gross reproduction rate, net reproduction rate and mean length of
generation for Australia, 2011
Age group mx b(f,x) Fxb(f,x) / Fx 5 Li / l0 (5)  (6) (2)  (7)
(1) (2) (3) (4)
(5) (6) (7) (8)
15–19 17.5 5,515 706,860
0.007802 4.97293 0.03880 0.67900
20–24 22.5 20,305 788,193
0.025761 4.96647 0.12794 2.87865
25–29 27.5 40,912 817,086
0.050071 4.95958 0.24833 6.82908
30–34 32.5 46,516 766,950
0.060651 4.95101 0.30028 9.75910
35–39 37.5 27,042 791,706
0.034157 4.93878 0.16869 6.32588
40–44 42.5 5,903 800,496
0.007374 4.92008 0.03628 1.54190
45–49 47.5 296 777,690
0.000381 4.89109 0.00186 0.08835
P P
D 0.186197 NRR D 0.92218 D 28.10196
P
GRR D 5 (5) D 5  0.186197 D 0.930985
P P
MLG D (8) / (7) D 28.10196 / 0.92218 D 30.5 years

age patterns of fertility. The age pattern of a population’s fertility is its distribution
by maternal age, after controlling for differences by age in numbers of potential
mothers.

Graphs of Age-Specific Fertility Rates

Age patterns of fertility are commonly examined by graphing age-specific fertility

rates, usually for five-year age groups. When compared such graphs can very
effectively highlight differences in the ages at which fertility peaks, the prominence
of such peaks, and ages at which fertility differences or change are concentrated.
They also provide a strong visual impression of differences/changes in the intensity
of fertility.
Figure 6.1 plots age-specific fertility rates for several countries in the early 1990s
chosen to illustrate the diversity of age patterns of fertility one can encounter.
Bangladesh (TFR 4.45 children per woman) exhibits clearly the highest ASFRs at
all ages, with peak fertility straddling the 20–24 and 25–29 age groups. Childbearing
begins early, and fertility limitation receives scant attention while women are still
in their twenties. Next, the former Soviet republic of Kazakhstan (TFR 2.66) is an
example of a population where fertility is heavily concentrated at younger ages,
with a distinct peak at ages 20–24 and fertility limitation practiced increasingly
thereafter. The former German Democratic Republic (GDR), or East Germany,
and the Federal Republic of Germany (FRG), or West Germany, then provide an
interesting contrast. With TFRs of 1.59 and 1.41, respectively, their fertility levels
are reasonably similar, but their age patterns of fertility are very different. The GDR
pattern is similar to that of Kazakhstan, but at a lower level. Fertility is much higher
at ages 15–19 and 20–24 than in the FRG, and peaks in the second of these age
groups. By contrast FRG fertility occurs decidedly later, peaking at ages 25–29 at a
Describing the Age Pattern of Fertility 263

Fig. 6.1 Age-specific fertility rates for selected countries in the early 1990s (Source: United
Nations demographic yearbook, 1994)

level only marginally above that observed in the GDR, but then being distinctly
higher than in that country at ages 30–34 and 35–39. These differences reflect
different options open to young adults in the two countries in the early 1990s. The
FRG pattern is typical of more developed countries in recent times as the young
have focused through their twenties on acquiring education, establishing careers,
travelling, experimenting with relationships and having a good time. Australia’s
pattern in the early 1990s was similar, although its ASFRs were a little higher
than those for the FRG. In the GDR prior to re-unification early marriage and
childbearing were a more attractive path to independence given more restricted
alternatives, but parlous economic circumstances then recommended a high level
of fertility control beyond age 30. The Netherlands (TFR 1.58), finally, exhibits a
very delayed age pattern of childbearing. Its flat peak across ages 25–29 and 30–
34 was quite unusual in the early 1990s, although Australia would have a similar
pattern 10 years later, and by 2011 age group 30–34 was its peak fertility age group.
This can be seen in Fig. 6.2, which using Australia as an example, demonstrates
change over time (since 1921) in the age pattern of fertility. TFRs represented by
the fertility patterns shown are indicated in parentheses in the graph legends.
In 1921, with a TFR of 3.12 children per woman, fertility in Australia was
highest at ages 25–29, but also very much higher at older ages than it would ever
subsequently be, reflecting relatively late marriage and limited control over fertility
within marriage. The 1934 pattern then captures the impact on fertility of the Great
Depression, with fertility still highest at ages 25–29, but the TFR reduced by one-
third and all age-specific fertility rates much lower. Clearly couples managed to
control fertility to a substantial extent when economic circumstances demanded it,
partly through deferring marriages (see Fig. 5.1 in Chap. 5). By 1945 fertility at
264 6 Analysis of Fertility

Fig. 6.2 Age-specific fertility rates: Australia 1921–2011 (Source: Australian Bureau of Statistics)
Describing the Age Pattern of Fertility 265

ages 15–19 to 25–29 had recovered to 1921 levels, but while that at older ages had
also rebounded, it had not done so to earlier levels. Six years later the post-war
baby boom was well under way, fertility having risen appreciably at ages 20–24 and
substantially also at ages 15–19 and 25–29, and the TFR once again back above
three children per woman. These trends continued through the 1950s until 1961,
when the peak post-war TFR of 3.55 children per woman was recorded and age
group 20–24 had taken over as the peak fertility age group. Couples were marrying
earlier than ever before, not infrequently with the first child having already been
conceived as a sexual revolution unfolded, and for most young women their twenties
were heavily focused on having and raising children.
The year 1961 was the year oral contraception (the pill) first became available in
Australia. By 1971 fertility had declined again (TFR 2.95) at all ages except 15–19
as the pill facilitated deferment of childbearing following marriage and avoidance
of unwanted higher parity births by women in their thirties and forties, but was
largely denied to young unmarried women other than in the shadow of planned
weddings. By 1981 fertility decline had continued apace, with 15–19 year-olds by
now incorporated into it, the TFR below replacement level at 1.94 children per
woman, and with sterilization procedures now also firmly part of the arsenal for
preventing unwanted higher parity births, fertility at ages 30–34 and 35–39 at lowest
ever levels. The incorporation of teenagers into the trend reflected rising consensual
partnering and use of modern contraception, and the opening up following a
landmark 1971 legal ruling of access to induced abortion as the ultimate protection
against unwanted childbearing. Beyond 1981 there has been limited change in the
TFR, but a major change in the timing of fertility toward older ages. Fertility
continued to decline until 2001 at ages 15–29, but rose after 1981 at ages 30–34 and
older. By 2001 fertility at ages 30–34 had moved marginally ahead of that at ages
25–29, and by 2011 age group 30–34 was clearly Australia’s peak childbearing age
group. Fertility at this and older ages was not dissimilar to levels that had prevailed at
those ages in 1961, but it was the fertility of women who had deferred childbearing
into their thirties and forties, not unplanned higher parity fertility attributable to
circumscribed ability to avoid such births.
Comparative observations of the type just made on the basis of inspection of
Figs. 6.1 and 6.2 can be supplemented by simple calculations giving ratios of
ASFRs in particular age groups. Thus, for example, taking ratios of 2011 to 1981
ASFRs in Fig. 6.2 yields values of 0.57, 0.49, 0.71, 1.61, 2.87, 3.38 and 3.00 at ages
15–19 through 45–49. These numbers quantify the extent of the decline in fertility
at younger reproductive ages and the increase at older ones over the 30 year period
in question as the age pattern of childbearing changed radically whilst the quantum
of fertility, as measured by the TFR, barely changed at all.
266 6 Analysis of Fertility

Measures of the Average Age of Fertility

Differences and changes in age patterns of fertility are also frequently indexed by
some sort of measure of what can loosely be termed the ‘average age of fertility’. We
have already discussed, in the context of measuring reproductivity, the mean length
of generation (MLG). This was defined as ‘the mean age of mothers at the birth of
their daughters’, and is an example of one of these measures. There are, however,
others, and because standard names and algebraic symbols are not used, the various
alternatives are easily confused. Two types of measures should be distinguished:
measures of the mean age of childbearing, of which the MLG is an example,
and measures of the mean age of the fertility schedule (MAFS) (a population’s
fertility schedule is its set of ASFRs or ASFFRs). An example of the latter could
be calculated from Table 6.1 by multiplying columns (2) and (5) (the age group
midpoints and the ASFFR-values), summing the results, and dividing by the sum of
column (5). That is:
     
MAFS D †xD1519;4549 b.f;x/ =Fx .mx / =†xD1519;4549 b.f;x/ =Fx (6.14)

Where all elements mean the same as in Eq. 6.13.

Calculated using the data of Table 6.1, the MAFS has a value of 30.5 years,
identical to the value obtained for the MLG, although slightly higher when
calculated to more decimal places. These two indices are related to one another in
the same way as the NRR and GRR are related. Just as the NRR takes mortality into
account and the GRR does not, so, too, the MLG takes mortality into consideration
and the MAFS does not. This means that the MLG is invariably lower than the
MAFS, because the control for mortality reduces the weights applied in calculation
to the midpoints of the reproductive age groups more substantially at the older
reproductive ages (since mortality increases with age). However, in a low mortality
population the difference, as here, may be slight.
A value for the MAFS could also be computed using ASFR-values in place of
ASFFR-values (i.e., b(f,x) in Eq. 6.14 becomes b(x) , meaning total births to women
aged x in year y); if there was no variation in the sex ratio at birth by age of mother
the answer would be the same. But the key point is that when dealing with ‘average
age of fertility’ indices one should be careful to be sure which measure one is
dealing with, especially if a comparison is being made. Since different measures
can take on different values for a population, if measures calculated in different ways
are compared significance could be read into trends or differences that are wholly
or partly spurious.
Two other ‘average age of fertility’ measures that you may encounter are the
median age of the fertility schedule and the mean age of mothers giving birth. The
latter has a calculating equation similar in form to Eqs. 6.13 and 6.14, except that
the ASFFR-element b(f,x) / Fx is replaced by the number of women giving birth at
age x in year y. Once again, be careful to be sure which measure you are dealing
with and to only compare like measures.
Natural Fertility and Associated Fertility Models 267

Natural Fertility and Associated Fertility Models

While graphs of the age pattern of fertility can be instructive, their interpretation in
the manner illustrated above lacks precision. Another approach to the description
of fertility schedules is to compare them with a standard schedule representing
what is known as ‘natural fertility’. This term was coined by French demographer
Louis Henry, who juxtaposed it with the term ‘controlled fertility’. Natural fertility
refers to the fertility of a population which does not consciously practice any
fertility restraint. Henry (1961) distinguished between fertility limiting practices
that did not vary according to a woman’s parity and those that did vary by
parity; natural fertility occurred where no parity-specific fertility limitation was
practiced. Thus, cultural practices like prolonged breastfeeding, prescribed periods
of sexual abstinence and customs regulating the frequency of intercourse, together
with other parity-independent fertility inhibitors (e.g., genetic differences affecting
fecundity and health conditions influencing pregnancy wastage through miscarriage
and stillbirth) are ignored. It follows, of course, that different populations can have
different levels of natural fertility, depending on the practices and circumstances that
produce involuntary fertility limitation. ‘Controlled fertility’ was defined by Henry
(1976: 90) as ‘the fertility of populations which practice birth control effectively’.
Thus defined it is an even less precise concept than ‘natural fertility’ – ‘effectively’
presumably means that only planned births occur, but the fertility desires, and
hence levels of planned fertility, of populations vary. This is not, however, of
major consequence, since it is the concept of natural fertility that is of analytical
importance.
A standard that demographers often use to represent unrestrained, or natural
fertility is the marital fertility schedule achieved by Hutterite women during 1921–
1930. The Hutterites are a small Anabaptist sect living communally in parts of
the western plains of the United States and Canada. Their religion encourages
large families, they breastfeed for relatively short periods, and they practice no
contraception. The observed marital fertility schedule yields a total marital fertility
rate (TMFR) of 14.44 (indicating the number of children the average woman
would have if she survived and was married throughout the reproductive period –
essentially it is treated as the TFR attainable by a female population that is
universally sexually active throughout this period). The observed schedule, however,
includes a marital fertility rate of 0.700 children per woman at ages 15–19. Based
on few cases, analysts have deemed this implausibly high and arbitrarily reduced
it to 0.300, dropping the TMFR to 12.44. This adjusted Hutterite marital fertility
schedule is used in constructing the shortly to be discussed Coale fertility indices
(sometimes also known as the Princeton fertility indices).
268 6 Analysis of Fertility

The Coale-Trussell Fertility Model

First, though, we shall consider a way of describing the deviation of a marital

fertility schedule from a natural shape developed by Ansley Coale and James
Trussell. This uses another standard schedule of marital fertility which averages
ten survey-based schedules described by Henry (1961) as ‘natural’ and which Coale
and Trussell (1974) judged to be not significantly affected by age misreporting. The
basis of their model is their empirical observation that (1974: 187) ‘marital fertility
either follows natural fertility : : : or departs from natural fertility in a way that
increases with age according to a typical pattern.’ The model then specifies that in a
population in which there is voluntary control of fertility, the ratio of marital fertility
at each age, r(a), to the marital fertility rate from the standard schedule, n(a), is given
by:

r .a/ =n .a/ D M : em : v.a/ (6.15)

Whence:

r .a/ D M : n .a/ : em : v.a/ (6.16)

Where r(a) D the observed marital fertility rate for the five-year age group a;
n(a) D the marital fertility rate for age group a in the natural fertility schedule;
v(a) D an empirically derived value expressing the typical departure from natural
fertility at age a due to voluntary fertility control; m D a measure of the extent of
fertility control within marriage; M D a scale factor.
The model is applied over the age range 20–24 to 45–49. Age group 15–19
is omitted because, in the words of Coale and Trussell (1974: 188), ‘premarital
conceptions have a large and irregular effect on teenage marital fertility.’ The
standard schedules of n(a) and v(a) values are given in Table 6.2. The source of
the n(a) values was explained above. The v(a) values are also empirically derived,
as the means of values obtained for 43 marital fertility schedules published in the
1965 United Nations Demographic Yearbook.

Table 6.2 Standard schedules of n(a) and v(a) for the Coale-Trussell fertility model
Age group
20–24 25–29 30–34 35–39 40–44 45–49
n(a) 0.460 0.431 0.395 0.322 0.167 0.024
v(a) 0.000 0.279 0.677 1.042 1.414 1.671
Source: Coale and Trussell (1975)
Note: Values in this table differ from those published in Coale and Trussell’s original (1974) paper.
They were corrected in an erratum published in Population Index, 41(4), 1975
Natural Fertility and Associated Fertility Models 269

Table 6.3 Calculation of m in the Coale-Trussell Model of Marital Fertility, Philippines 1970
Age group r(a) n(a) r(a)/M.n(a) loge (4) v(a) m(a) D (5)/(6)
(1) (2) (3) (4) (5) (6) (7)
20–24 0.327 0.460 1.000 0.000 0.000
25–29 0.256 0.431 0.835 0.180 0.279 0.645
30–34 0.203 0.395 0.723 0.325 0.677 0.479
35–39 0.146 0.322 0.638 0.450 1.042 0.432
40–44 0.068 0.167 0.573 0.557 1.414 0.394
45–49 n.a. 0.024 n.a. n.a. 1.671 n.a.
M D r(20  24) / n(20  24) D 0.711
m D mean of values in column (7) D 0.488

The scale factor M is of limited interest since it captures the level of marital
fertility, not its age pattern, which is what we are interested in. It is often estimated
as:

M D r .20  24/ =n .20  24/ (6.17)

The parameter m, measuring the extent of fertility control indicated by comparison

of the schedule of r(a) values with the natural fertility schedule of n(a) values, is
the value we are most intent on estimating. According to Newell (1988), values of
m tend to range from 0.0 to about 2.5. A value of 0.0 implies no fertility control
(Coale and Trussell actually suggest that anything below 0.2 is evidence of virtually
no control); a value above 2.0 is indicative of extremely strong control. Values of
m can be used to group populations according to the extent to which they control
fertility within marriage, and can also be used to index trends in marital fertility
control over time for individual populations.
With M estimated using Eq. 6.17, values of m (i.e., m(a)) can be found for each
age group 25–29 to 45–49 (or 25–29 to 40–44 if the marital fertility rate for age
group 45–49 is zero or missing) using the rearranged version of Eq. 6.15 given as
Eq. 6.18 below. This rearrangement entails first taking logarithms of both sides of
Eq. 6.15. The family of m-values thus obtained can then be averaged to yield a
single value for the population. Note that there is no m-value for age group 20–24,
the method assuming that marital fertility at those ages is essentially uncontrolled
(or as uncontrolled as it gets in the population in question – M, in fact, picks up
any deviation from natural fertility at ages 20–24). This is a dubious assumption for
many contemporary developed populations, which commonly use highly reliable
contraceptive methods to defer childbearing following marriage until economic and
career goals have been attained. An example of the calculation of m using this
method is presented in Table 6.3.

m .a/ D loge Œr .a/ =M : n .a/ =v .a/ (6.18)

270 6 Analysis of Fertility

Table 6.4 Schedule of Age group

Hutterite marital fertility
15–19 20–24 25–29 30–34 35–39 40–44 45–49
rates, 1921–1930
0.300 0.550 0.502 0.447 0.406 0.222 0.061

Although frequently used, the approaches to evaluating M and m just described have
some limitations. The main one is that M is obtained with reference to only one age
group, and thereafter influences all subsequent calculations. An alternative approach
again entails first taking logarithms of both sides of Eq. 6.15 to yield the equation:

loge Œr .a/ =n .a/ D loge M C m : v .a/ (6.19)

If we let y D loge [r(a) / n(a)], c D loge M, and x D v(a), Eq. 6.19 has the form
y D mx C c, which is the equation of a straight line. This can be fitted to pairs of
(x,y) values for the various five-year age groups (which we can obtain since r(a),
n(a) and v(a) are all known) using ordinary least squares regression to yield values
of m and c, and hence values of m and M. Applied to the example in Table 6.3, this
approach gives estimates of m D 0.379 and M D 0.682. Comparing the former value
with that computed in Table 6.3 a somewhat lower level of marital fertility control
is suggested, and it is clear that in using the Coale-Trussell model it is unwise to use
different methods of estimating m-values which are to be compared.

The Coale (or Princeton) Fertility Indices

Ansley Coale’s indices If (overall fertility), Ig (marital fertility), Ih (‘illegitimate’,

or unmarried, fertility) and Im (marriage index) were developed in the context
of a major study of the decline of fertility in Europe during the nineteenth and
twentieth centuries conducted at Princeton University from the late 1960s. Data
on annual births generally were not available by age of mother, so that TFRs
could not be calculated (GFRs could be, but did not control for age). Coale’s
solution was to develop three fertility indices, If , Ig and Ih , which are in fact
indirectly standardized ratios (see Chap. 2). They express the levels of fertility
(total, marital and ‘illegitimate’ (non-marital)) in a population as proportions of
natural fertility as represented by the previously mentioned 1921–1930 Hutterite
marital fertility schedule (adjusted at ages 15–19). The method simply applies the
Hutterite schedule to distributions of total women, married women and unmarried
women by five-year age groups to yield expected numbers of births to each group
assuming natural fertility prevailed, then calculates ratios of observed to expected
numbers of births. The Hutterite marital fertility schedule is given in Table 6.4.
Calculation of the Coale indices requires two types of data: (i) total births in
a year classified by whether or not the mother was married (births may also be
averaged over a short series of consecutive years), and (ii) the distribution of the
female population at mid-year (or the middle of the period over which annual births
Natural Fertility and Associated Fertility Models 271

are averaged) by marital status and five-year age groups. Calculating equations for
the three fertility indices are as follows.
Index of overall fertility:

If D Bt =†iD1519;4549 Wi : Hi (6.20)

Where Bt D total births in year y (or the average annual number of births over a
short period centred on year y); Wi D the mean (or mid-year) total number of
women aged i in year y; Hi D the marital fertility rate of Hutterite women aged i.
Index of marital fertility:

Ig D Bn =†iD1519;4549 Mi : Hi (6.21)

Where Bn D nuptial (marital) births in year y (or the average annual number of
such births over a short period centred on year y); Mi D the mean (or mid-year)
number of married women aged i in year y; Hi D the marital fertility rate of
Hutterite women aged i.
Index of unmarried fertility:

Ih D Be =†iD1519;4549 Ui : Hi (6.22)

Where Be D ex-nuptial (non-marital) births in year y (or the average annual number
of such births over a short period centred on year y); Ui D the mean (or mid-year)
number of unmarried women aged i in year y; Hi D the marital fertility rate of
Hutterite women aged i.
There was, of course, a fourth index listed at the beginning of this section. This
is Im , the marriage index, or index of the proportion married among women in the
reproductive age group. It is treated separately because it is not a fertility index.
Given, however, that in virtually all societies married women are much more likely
to bear children than are unmarried women, clearly a potentially major reason for
differences/changes in overall fertility is differences/ changes in the extent to which
women of reproductive age are married. Im is designed to recognize this reality and
to allow it to be taken into account in interpreting fertility differentials and trends. It
can be calculated as either a weighted average of the proportions of women married
in five-year age groups 15–19 to 45–49 years, with the Hutterite marital fertility
schedule providing the weights, or as the ratio of expected births to married women
to expected births to total women assuming Hutterite marital fertility. Thus, either:

Im D †iD1519;4549 .Mi =Wi / : Hi =†iD1519;4549 Hi (6.23)

Where Mi D the mean (or mid-year) number of married women aged i in year y;
Wi D the mean (or mid-year) total number of women aged i in year y; Hi D the
marital fertility rate of Hutterite women aged i.
272 6 Analysis of Fertility

Table 6.5 Calculation of Coale fertility indices If , Ig , Ih and Im for Bulgaria, 1985
Age group Mi Ui Wi Hi W i . Hi M i . Hi Ui . Hi
15–19 40,168 259,481 299,649 0.300 89,895 12,050 77,844
20–24 195,672 98,129 293,801 0.550 161,591 107,620 53,971
25–29 259,417 50,178 309,595 0.502 155,417 130,227 25,189
30–34 278,240 41,916 320,156 0.447 143,110 124,373 18,736
35–39 297,120 41,258 338,378 0.406 137,381 120,631 16,751
40–44 246,680 34,084 280,764 0.222 62,330 54,763 7,567
45–49 236,105 36,189 272,294 0.061 16,610 14,402 2,208
P P P
D 766,334 D 564,066 D 202,266
Bt D 118,955; Bn D 105,001; Be D 13,954
If D 118,955 / 766,334 D 0.155
Ig D 105,001 / 564,066 D 0.186
Ih D 13,954 / 202,266 D 0.069
Im D 564,066 / 766,334 D 0.736
Source: United Nations demographic yearbook, 1986 and 1990

Or:

Im D †iD1519;4549 Mi : Hi =†iD1519;4549 Wi : Hi (6.24)

Where all elements mean the same as in Eq. 6.23.

Equation 6.24 is more convenient because the two summations it requires are
also required in evaluating If (Eq. 6.20) and Ig (Eq. 6.21), and it is a simple matter to
divide one by the other. An example of the calculation of If , Ig , Ih and Im is presented
in Table 6.5.
An important relationship between the Coale indices is expressed by the equa-
tion:

If D Im : Ig C .1  Im / : Ih (6.25)

And if Ih (unmarried fertility) is negligible, as it often is and/or historically has been

in many populations, this reduces to the approximate relationship:

If D Im : Ig (6.26)

Thus in populations where non-marital fertility is very low (not more than, say, 3–
4%) we have an index of overall fertility that is approximately the product of an
index of marital fertility and an index of the proportion married.
To illustrate the use of the Coale indices, Fig. 6.3 plots values for Australia for
census years between 1861 and 2011. In 1861 overall fertility (If ) was about 50 % of
the Hutterite natural level, while marital fertility (Ig ) was around 70 % of it. Overall
fertility fell between 1861 and 1881, but marital fertility rose to peak at 74 % of
Natural Fertility and Associated Fertility Models 273

Fig. 6.3 Coale fertility indices for Australia, census years 1861 to 2011 (Source: Jones (1971),
author’s calculations)

the Hutterite level. Thus, overall fertility declined solely because the marriage index
(Im ) dropped sharply. Women married later, thereby delaying the commencement of
childbearing, but did not reduce the regularity with which they had children once
married.
The period between the censuses of 1881 and 1933 saw overall fertility continue
to decline, reaching a level by the latter date that was less than 18 % of the Hutterite
level. Over this period the driving force behind overall fertility decline clearly was
decline in marital fertility, which more than halved from 74 % of the Hutterite level
to 32 %. This was the period of Australia’s fertility transition, the decline in marital
fertility accelerating between 1881 and 1901, easing off in the first decade of the
twentieth century as depression conditions of the 1890s eased, then reasserting itself
between 1911 and the Great Depression of the early 1930s. During this period
the marriage index remained quite stable, a slight rise between 1911 and 1921
partly offsetting the decline in marital fertility and moderating the decline in overall
fertility.
Between 1933 and 1954 the marriage index rose rapidly. Indeed, had there been
a census in the late 1930s an even steeper increase would be apparent, detailed
studies showing this ‘marriage boom’ to have really commenced with the outbreak
of the Second World War. This trend reinforced a modest resurgence in marital
fertility between 1933 and 1947, and then more than offset a slight decline between
1947 and 1954, yielding an upward trend in overall fertility that continued until
1961. Subsequently, with oral contraception introduced in Australia that year, the
trends in both overall and marital fertility were generally downward until marital
fertility picked up again slowly through the 1990s and more quickly after 2001. The
274 6 Analysis of Fertility

marriage index also fell sharply again after 1971 as trends to younger and more
universal marriage reversed emphatically with the rise to prominence of consensual
partnering, and this drove overall fertility down more rapidly than marital fertility.
The period since 1951 has also seen unmarried fertility increase in importance as,
first, younger women and men became more premaritally sexually active, and then
as they gained some control over unplanned fertility resulting from this trend, the
practice of couples cohabiting and having children in such relationships became
more widespread. The brief break in this trend between 1971 and 1976 marks the
transition from non-marital fertility rising largely as a result of unintended teenage
and early adult pregnancies generated by the greater sexual freedom of the 1950s
and 1960s, to it rising at more normative reproductive ages due to the spread of
cohabitation and childbearing within it. That five-year period brought a massive
improvement in unmarried teenage and early adult fertility control as access was
finally gained to the pill (unmarried women had largely been denied it through the
1960s) and induced abortion became readily available after a landmark legal case in
New South Wales gave the green light to stand alone abortion clinics.

Other Fertility Models

Aside from fertility models based on notions of natural fertility, various other
types have been proposed. Two will be dealt with here: William Brass’s relational
Gompertz model and John Bongaarts’s model relating the more important of what
are known as the proximate determinants of fertility to the observed fertility level
of a population.

The Brass Relational Gompertz Model

This model aims to describe any schedule of age-specific fertility rates by two
parameters, ’ and “, which capture the difference between that fertility schedule
expressed in cumulative proportional form and a standard fertility schedule
expressed in similar form. To express a fertility schedule in cumulative proportional
form one cumulates ASFRs for single-year or five-year age groups successively
from the youngest to the oldest reproductive age (multiplying the latter by 5 in
recognition that a woman takes five years to pass through any five-year age group),
assigns the result after each new ASFR is added to the exact age defining the upper
bound of the age group in question, and divides each of these cumulative fertility
rates by the TFR. The outcome is a schedule of cumulative fertility to successive
exact ages expressed as a proportion of ultimate completed fertility (i.e., values rise
from zero at the exact age marking the lower bound of the youngest age group in
which births occur to 1.0 at that marking the upper bound of the oldest age group in
which they occur).
Other Fertility Models 275

The standard fertility schedule used in this model was designed to reflect
the typical pattern of fertility in high fertility populations. Table 6.6 shows both
cumulative proportional and Gompertz-transformed values from this schedule for
one-year intervals of exact age. For use with five-year age group data simply
extract Gompertz-transformed values for exact ages 20, 25, 30, 35, 40 and 45.
The basis of the Brass relational Gompertz approach is the empirical observation
that applying a Gompertz transformation to any fertility schedule expressed in
cumulative proportional form yields approximately a straight line when plotted
against exact age, and thus approximately a straight line when plotted against
any other, or a standard, Gompertz-transformed fertility schedule as well. The
method focuses on the straight line resulting from the plot of Gompertz-transformed
observed and standard fertility schedules against one another. The parameters ’ and
“ are the y-intercept and slope, respectively, of this straight line, which has the
equation:

Gx D ’ C “Gs;x (6.27)

Where Gx D the Gompertz-transformed observed cumulative proportional fertility

rate to exact age x; Gs,x D the Gompertz-transformed standard cumulative
proportional fertility rate to exact age x.
If the observed and standard fertility schedules were identical the straight line
represented by Eq. 6.27 obviously would intercept at the origin and have a slope
of 1.0; hence ’ D 0 and “ D 1.0 are the ‘no difference’ benchmarks against which
computed values of ’ and “ are interpreted. According to Newell (1988: 177–178), ’
‘changes the age-location of the model’ while “ ‘may be interpreted as determining
the spread, or degree of concentration, of the schedule.’ A value of ’ D 0 indicates
that, as in the standard schedule, about half of total childbearing occurs by age
27. Positive values of ’ indicate earlier fertility than the standard; negative values
indicate later fertility. Values of C0.5 and 0.5 respectively suggest very early and
very late age patterns of fertility. Values of “ lower than 1.0 (down to perhaps 0.65)
indicate a wider spread of fertility by age than in the standard population (as, in the
limiting case, might occur in natural fertility populations which also marry early),
while values higher than 1.0 indicate a narrower spread of fertility (values above 1.5
suggest a late start, and early finish, to childbearing, as occurs in some contemporary
European populations).
Having estimated ’ and “ in a manner described below it is possible to work
backwards from Eq. 6.27 and ‘fit’ ASFR values that would exactly satisfy the
equation for each of the standard Gs,x values (remember that we referred above to
the Gompertz-transformed observed and standard cumulative proportional fertility
schedules approximating a straight line relationship). If these ‘fitted’ values were
judged, in comparison with the original ASFRs, to be a ‘good’ fit, one practical
application of Eq. 6.27 would be the estimation of a schedule of single-year
ASFRs from the observed five-year values. Having used the latter to specify the
relevant form of Eq. 6.27, one would work backwards from it, varying x in single-
276 6 Analysis of Fertility

Table 6.6 Standard cumulative proportional and Gompertz-transformed fertility rates for the
Brass relational Gompertz fertility model
Cumulative Cumulative
Exact proportional Gompertz-transformed Exact proportional Gompertz-transformed
age fertility rate rate age fertility rate rate
11 0.00000 3.18852 31 0.65016 0.84272
12 0.00000 2.70008 32 0.68968 0.99014
13 0.00002 2.37295 33 0.72722 1.14407
14 0.00035 2.07262 34 0.76275 1.30627
15 0.00277 1.77306 35 0.79618 1.47872
16 0.01168 1.49286 36 0.82751 1.66426
17 0.03043 1.25061 37 0.85663 1.86597
18 0.05826 1.04479 38 0.88354 2.08894
19 0.09428 0.85927 39 0.90816 2.33192
20 0.13584 0.69130 40 0.93019 2.62602
21 0.18187 0.53325 41 0.94925 2.95500
22 0.22993 0.38524 42 0.96480 3.32873
23 0.27897 0.24423 43 0.97698 3.75984
24 0.32829 0.10783 44 0.98591 4.25499
25 0.37731 0.02564 45 0.99188 4.80970
26 0.42597 0.15853 46 0.99555 5.41311
27 0.47371 0.29147 47 0.99782 6.12864
28 0.52013 0.42515 48 0.99915 7.07022
29 0.56517 0.56101 49 0.99982 8.64839
30 0.60861 0.70000 50 1.00000 C1

Table 6.7 Brass relational Gompertz fertility model fitted to data for Ecuador, 1993
Fitted Fitted Fitted Fitted
Age group ASFR Fx Fx / TFR Gx Gs,x Gx (Gf,x ) Fx / TFR Fx ASFR
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
15–19 109.4 0.5470 0.1511 0.6365 0.6913 0.6026 0.1609 0.5823 116.5
20–24 191.6 1.5050 0.4158 0.1306 0.0256 0.1424 0.4201 1.5206 187.6
25–29 178.9 2.3995 0.6629 0.8888 0.7000 0.8431 0.6503 2.3538 166.6
30–34 133.1 3.0650 0.8468 1.7940 1.4787 1.6523 0.8256 2.9883 126.9
35–39 77.9 3.4545 0.9544 3.0646 2.6260 2.8445 0.9435 3.4150 85.3
40–44 26.8 3.5885 0.9914 4.7517 4.8097 5.1136 0.9940 3.5978 36.6
45–49 6.2 3.6195 1.0000 1.0000 3.6195 4.3
’ D 0.1158; “ D 1.0391

year intervals and substituting single-year-of-age Gompertz-transformed cumulative

proportional fertility rates for the standard population from Table 6.6.
An example should now help to bring the Brass relational Gompertz procedure to
life. In column (2) of Table 6.7 are presented five-year ASFRs for Ecuador in 1993.
The subsequent calculations involve the following steps.
Other Fertility Models 277

Step 1: In column (3), x is the exact age marking the upper bound of each age group
(i.e., 20, 25, : : : , 50) and Fx is the cumulative fertility rate per woman to exact
age x. Thus:

F20 D 5 : ASFR1519 =1; 000

F25 D 5 : .ASFR1519 C ASFR2024 / =1; 000
F30 D 5 : .ASFR1519 C ASFR2024 C ASFR2529 / =1; 000
:::::::::::::
F50 D 5 : .ASFR1519 C    C ASFR4549 / =1; 000

Step 2: Column (4) is obtained by dividing each value in column (3) by the TFR
(i.e., by F50 ) to yield the cumulative proportional fertility rate to exact age x.
Step 3: In column (5) the cumulative proportional fertility rates of column (4) are
Gompertz-transformed using the equation:

Gx D loge .loge .Fx =TFR// (6.28)

Step 4: Column (6) gives Gompertz-transformed cumulative proportional fertility

rates for the standard fertility schedule extracted from Table 6.6.
Step 5: We now need to solve for ’ and “ in Eq. 6.27 using the values of Gx and Gs,x
in Table 6.7. One way to do so would be by ordinary least squares regression,
but a simpler approach commonly is used. The six pairs of Gx , Gs,x values in
Table 6.7 are divided into two groups of three (x D 20, 25 and 30; x D 35, 40 and
45), and for each group the three equations of the form of Eq. 6.27 created by
substituting Gx , Gs,x pairs are summed to yield two simultaneous equations in ’
and “. Thus:

G20 C G25 C G30 D .’ C Gs;20 : “/ C .’ C Gs;25 : “/ C .’ C Gs;30 : “/

i:e:; 0:6365  0:1306  0:8888 D 3’ C .0:6913  0:0256  0:7000/ “
or  0:3829 D 3’  0:0343“
And W G35 C G40 C G45 D .’ C Gs;35 : “/ C .’ C Gs;40 : “/ C .’ C Gs;45 : “/
i:e:; 1:7940  3:0646  4:7517 D 3’ C .1:4787  2:6260  4:8097/ “
or  9:6103 D 3’  8:9144“

Solving these two simultaneous equations gives ’ D 0.1158 and “ D 1.0391.

Step 6: To obtain the fitted values of Gx shown in column (7) of Table 6.7 we simply
apply Eq. 6.27 treating Gx as the unknown (call it Gf,x if you like, where f stands
for ‘fitted’) and substituting the values of ’ and “ just calculated together with
the relevant value of Gs,x . The result is a set of Gx values that exactly satisfy for
each value of Gs,x the straight line whose equation we found by estimating ’ and
“ at step 5.
Step 7: The fitted values of Gx are converted to fitted ASFRs by reversing the
procedure that was followed in generating column (5) from column (2). We first
obtain, at column (8), fitted values of Fx / TFR (i.e., fitted cumulative proportional
278 6 Analysis of Fertility

fertility rates to exact age x) by ‘undoing’ the Gompertz transformations of the

fitted Gx -values. To do this use:

fitted .Fx =TFR/ D exponent .exponent .fitted Gx // (6.29)

Next, calculate fitted Fx values (column (9)) by multiplying the fitted Fx / TFR
values by the original TFR (i.e., by F50 from column (3)). You can perhaps
appreciate at this point that the fitting procedure is effectively redistributing
recorded total fertility among the reproductive age groups.
The fitted Fx values give cumulative (total) fertility to exact ages 20, 25, : : : ,
50. For each five-year age group (except 15–19), find the fitted portion of total
fertility associated with the age group by subtracting the fitted Fx associated with
its lower bound (remember x is an exact age) from that associated with its upper
bound (e.g., for age group 20–24 find fitted F25 - fitted F20 ) (the fitted portion for
age group 15–19 is simply fitted F20 ). Then divide each fitted portion by 5 and
multiply by 1,000 to yield the fitted ASFRs of column (10). Remember, we did
the reverse of this (multiplied by 5 and divided by 1,000) to our original ASFRs
at step 1.
The parameters ’ and “ estimated in this example suggest that fertility in Ecuador
in 1993 occurred a little later than in the standard population and tended to be
marginally more compressed in terms of the ages over which it was spread. The
match between the fitted ASFRs and the original ASFRs is not especially good, and
so we would probably hesitate to use the particular form of Eq. 6.27 obtained using
our estimates of ’ and “ to generate single-year-of-age ASFRs.

The Bongaarts Fertility Model

In 1956 the American demographers Kingsley Davis and Judith Blake published a
ground-breaking paper in which they proposed that social, economic and cultural
conditions affected a population’s fertility through a set of 11 intermediate fertility
variables. These were the variables which directly influenced fertility, and through
which any other factor (e.g., the level of female education) had to operate. Davis and
Blake divided these intermediate variables into three subgroups: Factors affecting
exposure to intercourse (age of entry into sexual unions, extent of permanent
female celibacy, amount of the reproductive period lost to marital disruption or
widowhood, extent of voluntary abstinence, extent of involuntary abstinence, and
coital frequency when not abstaining); factors affecting exposure to conception
(natural forces affecting fecundity, extent of contraception, and voluntary practices
affecting fecundity); and factors affecting gestation and successful parturition
(childbirth) (foetal mortality from involuntary causes, and foetal mortality from
voluntary causes).
Other Fertility Models 279

The Davis-Blake framework found wide acceptance in broad concept, but in its
specifics proved difficult to incorporate into quantitative models. For quite some
time those models that were developed were extremely complex. John Bongaarts’s
contribution has been to greatly simplify the modelling of fertility in terms of the
intermediate fertility variables or, as they have come more commonly to be known
these days, the proximate determinants of fertility.
The basis of this simplification is (i) a respecification of the proximate determi-
nants to a list of seven (the female marriage pattern, the extent and effectiveness
of contraception, the prevalence of induced abortion, the duration of post-partum
infecundablility, natural fecundability (indexed by frequency of intercourse), the
prevalence of spontaneous intrauterine mortality (miscarriage), and the extent of
permanent sterility), and (ii) a further collapsing of this list to four ‘principal
proximate determinants’ by eliminating the last three as not sufficiently variable
from population to population to be important in accounting for fertility differentials
and trends. Thus, in the Bongaarts fertility model the four principal proximate
determinants are considered the prime potential inhibitors of fertility to below its
theoretical maximum level – delayed or prematurely ended marriage, contraception,
induced abortion, and post-partum infecundability induced by breastfeeding or
abstinence.
The model proposes four levels of fertility from which, if known, impacts of the
four principal proximate determinants could be derived. With the inhibiting effects
of all four determinants operative a population’s fertility level is measured by its
TFR (based on age-specific ratios of marital births only to total women). If the
fertility-inhibiting effect of delayed and disrupted marriage is eliminated, fertility
rises to a level TM, the total marital fertility rate (based on age-specific ratios
of marital births to married women). If the effects of contraception and induced
abortion are then also eliminated it rises further to a level TN, the total natural
marital fertility rate. And finally, if breastfeeding and post-partum abstinence are
non-existent it rises to a level TF, the total fecundity rate. This last rate is set by the
model at 15.3 children per woman. TFs vary from population to population because
of the effects of the three proximate determinants disregarded by the model, but do
so over a fairly narrow range. The value 15.3 is an average of this narrow range of
empirical values.
Against this theoretical background Bongaarts defines four indices which mea-
sure the levels of the principal proximate determinants on scales from 0 to 1, where
0 corresponds with maximum inhibition of fertility and 1 with zero inhibition. They
are:
Cm D index of marriage (the proportion of women of reproductive age married) D 1
where marriage is universal and 0 where there is no marriage.
Cc D index of contraception D 1 where contraception is totally absent and 0 where
all fecund women use 100 % effective contraception.
Ca D index of induced abortion D 1 where there is no induced abortion and 0 where
all pregnancies are aborted.
280 6 Analysis of Fertility

Ci D index of post-partum infecundability D 1 where there is no breastfeeding or

post-partum abstinence and 0 where the duration of infecundability from these
practices is infinite.
These indices can, by definition, be expressed in terms of the four fertility levels
as follows:

Cm D TFR=TM (6.30)

Cc : Ca D TM=TN (6.31)

Ci D TN=TF (6.32)

From these equations it follows that:

TFR D Cm : .Cc : Ca / : Ci : TF (6.33)

Or

TFR D Cm : .Cc : Ca / : Ci : 15:3 (6.34)

Equation 6.34 is the one that specifies the basic Bongaarts model. Obviously Cm ,
(Cc . Ca ) and Ci can be calculated from Eqs. 6.32, 6.33, and 6.34 if the fertility
rates TFR, TM, TN and TF are known, but it is rare for all to be known (TN and
population-specific values of TF, in particular, are hard to obtain). Thus Bongaarts
provides a series of alternative equations for estimating his indexes of the principal
proximate determinants.
The index of marriage is given by:

Cm D †xD1519;4549 mx : gx =†xD1519;4549 gx (6.35)

Where mx D the age-specific proportion of women aged x currently married (or in a

consensual union); gx D the marital fertility rate at age x.
Equation 6.35 indicates Cm to be a weighted average of age-specific proportions
of women currently married (or in a consensual union), with age-specific marital
fertility rates providing the weights. It is not merely the proportion of all women of
reproductive age who are married (or in a consensual union), because the fertility
impact of marriage depends partly on the age distribution of married women (marital
fertility rates varying with age).
If values of gx are not directly available they can be estimated as gx D fx / mx ,
where fx is the age-specific fertility rate at age x based on marital births only.
The frequent parenthetic reference to ‘or in a consensual union’ is indicative that
measurement problems can arise for populations where such unions are sufficiently
common that they cannot be ignored. Note in particular that births to consensually
Other Fertility Models 281

partnered women should be regarded as ‘marital’ births, but they are frequently
tabulated as, and impossible to distinguish from among, non-marital births. Bon-
gaarts also highlights possible difficulty in estimating g1519 satisfactorily using
the relationship gx D fx / mx , advocating that in this circumstance a reasonable
approximation is to set g1519 D 0.75 g2024 .
The index of contraception is given by:

Cc D 1  1:08 : u : e (6.36)

Where u D the proportion of married women of reproductive age currently using

contraception (including male methods, sterilization and abstinence other than
post-partum); e D the average use-effectiveness of the methods being used.
Estimates of u mostly are obtained from fertility or contraceptive prevalence
surveys. Values of e are computed from data on methods of contraception used and
standard method-specific values of use-effectiveness using:

e D †em : um =u (6.37)

Where em D the use-effectiveness level for method m; umPD the proportion of

married women of reproductive age using method m; u D um .
When writing in the early 1980s, Bongaarts had ‘reasonably reliable’ method-
specific use-effectiveness data only for the Philippines (sterilization D 1.00,
IUD D 0.95, pill D 0.90, other methods D 0.70) and the United States (steriliza-
tion D 1.00, pill D 0.99, IUD D 0.97, condom D 0.94, diaphragm D 0.92, foams,
creams, jellies D 0.91, rhythm D 0.87, other D 0.93). Obviously population-specific
schedules of use-effectiveness are the ideal, but in their absence one’s options are
to select the most plausible schedule for another population available or to decide
that the evaluation of Eq. 6.36 cannot reliably be completed.
The coefficient 1.08 in Eq. 6.36 is a correction factor for the reality that couples
do not use contraception if they know or believe their unions to be involuntarily ster-
ile. This empirically-based coefficient in effect inflates u to yield the proportion of
married women of reproductive age in non-sterile unions currently contracepting.
The idea is that what is needed is a measure of contraceptive practice in situations
where it has fertility-inhibiting potential.
Bongaarts’s equation for his index of induced abortion is empirically based and
designed to reflect research findings (i) that on average an induced abortion averts
less than one birth (because miscarriage or stillbirth might independently have
prevented a live birth, and because a woman resumes ovulation and hence exposure
to the risk of a further pregnancy sooner after induced abortion than if she has a
live birth), and (ii) that the number of births averted per induced abortion varies
depending on the extent to which contraception is practiced following the induced
abortion. The equation is:
282 6 Analysis of Fertility

Ca D TFR= ŒTFR C 0:4 : .1 C u/ : TA (6.38)

Where TFR D total fertility rate (again based on marital births only); u means the
same as in Eq. 6.36; TA D the total abortion rate (the sum over the reproductive
ages of single-year age-specific abortion rates, or five times the sum of five-year
age-specific rates, excluding abortions to unmarried women).
This equation is the ratio of the observed TFR to the estimated TFR in the
absence of induced abortion. If induced abortion is so infrequent as to be virtually
non-existent, Ca D 1.0.
Finally, the index of post-partum infecundability is given by:

Ci D 20= .18:5 C i/ (6.39)

Where i D the average duration of post-partum infecundability caused by breast-

feeding or post-partum abstinence in months.
Values of i are generally not readily available, but can be estimated from survey
data on the duration of breastfeeding using:
2
i D 1:753 : e0:1396B0:001872B (6.40)

Where B D mean duration of breastfeeding in months.

The Bongaarts model can, of course, be used to estimate the value of any of the
principal proximate determinant indices given values of the other indices, and use
can also be made of the equations linking the indices to the fertility levels TFR,
TM, TN and TR to evaluate individual indices or fertility levels. It can be applied
most straightforwardly to populations in which sexual activity is more or less
confined to marriage (whence a ‘regular’ TFR will reflect only marital births) and
consensual partnering is virtually non-existent whence there are no complications in
distinguishing ‘married’ women and ‘marital’ (births). Where these conditions are
not met particular care is needed to ensure that variables are measured appropriately,
and the scope for meeting insurmountable measurement problems increases.
There are also a variety of applications of the Bongaarts model besides the simple
calculation of individual indices and fertility levels. Bongaarts and Potter (1983)
discuss several. While data are rarely if ever available to enable fertility transitions
for individual populations to be traced, those for a range of populations recently at
different stages of transition have been used to construct a synthetic profile of the
‘typical’ transition from natural to controlled fertility. This suggests that transition
is associated with a shortening of post-partum infecundability, a large rise in contra-
ception, and a decline in the proportion married. Actual transitions for contemporary
developing populations are claimed to closely resemble this pattern, but historical
European transitions to differ in that proportions of women married rose and mean
ages at marriage declined. Other applications discussed are the decomposition
of change in fertility (to assess the proximate determinants most responsible),
Other Fertility Models 283

assessment of the proximate determinants underpinning socioeconomic differences

in marital fertility, the construction of models linking age-specific fertility rates to
the proximate determinants, detailed analysis of an empirical relationship between
the CBR and contraceptive prevalence, projection of the fertility consequences of
a nominated change in contraceptive practice, and assessment of the contraceptive
prevalence level required to meet a fertility target.
We will conclude this section by looking briefly at the penultimate of these
applications. Suppose that we are interested in the potential effect of a family
planning program. If TFR1 refers to the total fertility rate before the program and
TFR2 to its value at a later point after the program has had a chance to operate, the
proportionate change in the TFR can be written:

.TFR1  TFR2 / =TFR1 D 1  .TFR2 =TFR1 /

Which in terms of the Bongaarts model (Eq. 6.34) is:

.TFR1  TFR2 / =TFR1

D 1  .Cm;2 : Cc;2 : Ca;2 : Ci;2 : 15:3/ = .Cm;1 : Cc;1 : Ca;1 : Ci;1 : 15:3/

Now if it is reasonable to assume that the female marriage pattern, the abortion
level among married women and the average period of post-partum infecundability
remain constant over the program period, this reduces to:

.TFR1  TFR2 / =TFR1 D 1  Cc;2 =Cc;1

Substituting using Eq. 6.36 and simplifying then yields:

TFR2 D TFR1 .1  1:08 : u2 : e2 / = .1  1:08 : u1 : e1 / (6.41)

By substituting the pre-program TFR, known pre-program contraceptive use (u1 )

and use-effectiveness (e1 ) values, and target post-program use (u2 ) and effectiveness
(e2 ) values it is possible to estimate a post-program TFR, and hence the reduction
in TFR achievable if target use and effectiveness levels are reached. It might be that
one would wish to assume no change in use-effectiveness, in which case e2 D e1 , and
the assessment that is made is an assessment of the fertility reduction achievable by
increasing contraceptive use to level u2 .
This type of practical application to which the Bongaarts fertility model lends
itself makes it especially valuable. It is appropriate to end this discussion of fertility
models, however, with a warning that all fertility models should be used very
cautiously for purposes other than summary description. They are only models, and
placing blind faith in them can be perilous.
284 6 Analysis of Fertility

Analysis of Birth Intervals

Generally speaking, a birth interval is the length of time between successive live
births, and is usually measured in months. Stillbirths and abortions are ignored.
Two birth intervals, however, do not conform to this definition. The first birth
interval, or protogenetic interval, is the interval from marriage to first birth. It
may be negative if the first birth precedes marriage, and non-existent for a woman
who has had one or more children without ever marrying (she may, though, have
second, third, etc. birth intervals). Some populations even record negative average
first birth intervals, but there is really not much point to analysis of first birth
intervals in such circumstances. The concept of the first birth interval presupposes
that marriage marks the commencement of exposure to the risk of conception, or
at least that fertility, by cultural convention, takes place within marriage (so that if
first conception precedes marriage there is strong social pressure to marry before the
child is born). Where these conditions more or less prevail, change in the average
length of the first birth interval for successive cohorts of women can provide useful
information about change in fertility within marriage, and more particularly about
change in the onset of the entire process of family formation. However, where large
proportions of first births occur to unmarried women, or even where premarital
cohabitation is common and marriage then is widely deliberately timed to coincide
with the transition to parenthood, some other reference point than marriage (first
coitus or entry into a first union) seems more relevant. A more general definition of
the first birth interval as the interval from first exposure to the risk of conception
(defined in one of the ways just suggested) to the first live birth sometimes is
encountered, but unless a source clearly states that it has been used you should
assume first marriage as the reference point.
The second birth interval not conforming to the general definition is the open
birth interval. This is the interval, again in months, from the most recent birth, or
from marriage in the case of childless married women, to the date of data collection.
Birth intervals which do conform to the general definition are labelled according
to the order of the birth marking the end of the interval; thus the interval from the
first to the second birth is the second birth interval, etc. These birth intervals are
known as the intergenetic intervals. Like the protogenetic interval they are closed
intervals – closed by the birth that ended the interval.
You will also encounter in the demographic literature various other intervals
closely related to birth intervals as just defined. Sometimes both live births and
stillbirths (i.e., total births) are taken into account. There can in particular be some
justification for this in populations with high levels of infant mortality, because
differentiating stillbirths from very early neonatal deaths may be both problematic
and somewhat artificial. More generally it can be argued that having a stillbirth
indicates intent to have a live birth, and that explaining birth intervals that have
been lengthened by the unforeseen occurrence of a stillbirth is difficult (although
miscarriages raise similar issues). Other intervals that may be focused on are the
interval from first exposure to first conception, intervals from one live birth to
Analysis of Birth Intervals 285

conception of the next, and inter-pregnancy intervals. The last-mentioned can

be (i) the interval from the end of one pregnancy (however terminated) to the
beginning of the next (again, regardless of whether it results in a live birth), or (ii) the
interval between the events (whatever they may be) marking the end of successive
pregnancies. Clearly one should be certain to check in any practical situation which
definition has been used. But these three types of intervals tend to be of interest
when the aim is to measure exposure to the risk of conception when studying issues
such as contraceptive effectiveness. We will not, however, concern ourselves further
with them here.
Data on birth and related intervals generally derive from pregnancy or birth
histories, which date key events in women’s reproductive careers. Intervals are
computed by subtracting the date marking the beginning of an interval from that
marking its end. Pregnancy or birth histories typically are obtained through surveys
rather than censuses, because they entail sequences of questions which are complex,
variable from respondent to respondent, and potentially lengthy. Some data may also
be available from vital registration systems where these, in the course of registering
a birth, gather dates of previous live births and of marriage.
Analysis of birth intervals is most simply accomplished by calculating means
or medians, the latter having the advantage of being less affected by a few very
long intervals. It is, however, significantly complicated by three methodological
problems. The first is data quality, which can be adversely affected by the misdating
of births or failure to even remember their occurrence. The latter may in particular
be a problem where children had died by the time of data collection, and especially
where they had died very young. The second problem is censoring, the reality
that at a date of data collection there are always open birth intervals that may
later become closed. Censored intervals that ultimately will be closed tend to be
longer than already closed intervals of the same birth order, because the longer a
birth interval is, the greater is the chance of it still being open at any given date of
data collection. This means that for cohorts (age groups) of women yet to complete
childbearing, mean or median birth intervals invariably are underestimates of their
eventual values. In turn this obviously poses problems for comparison of cohorts
that have, and have not completed their fertility. But even among the latter group
the prospects of eventual closure of open intervals vary with age (a 20–24 year-old
woman with an open third birth interval is more likely to close it than is a 45–
49 year-old). One way of tackling this problem is to ignore more recent data for
older cohorts so that cohorts effectively are compared at equivalent points in the life
cycle. The other solution is to use life table techniques in which time elapsed since
attaining parity n is the equivalent of age, and attaining parity n C 1 is the equivalent
of death, in a conventional life table. This allows proportions attaining parity n C 1
within nominated birth intervals to be compared. The third methodological problem
is selectivity of the birth intervals available for analysis in survey data. Because
surveys typically target respondents in a restricted age range, intervals begun at any
single year of age by persons outside the survey age range (often those no longer of
reproductive age who have completed their fertility) are excluded from analysis.
286 6 Analysis of Fertility

In populations in which the date of first marriage is a good indicator of the

commencement of exposure to the risk of conception, and in which practically
no contraception is practiced until at least one live birth occurs, the first birth
interval is useful for studying fecundability. Higher order intervals are less useful
because couples with lower fecundability are selected out of them. Populations
conforming to this model also tend to have shorter first than subsequent mean birth
intervals, principally because after a birth the risk of pregnancy is reduced by post-
partum amenorrhoea and the effect thereon of breastfeeding, customs prescribing
post-partum abstinence (possibly), and circumstances such as maternal and infant
health problems, reduced coital frequency, adoption of contraception and increased
subfecundity.
Analysis of birth intervals often focuses on how they vary by characteristics
of women, both social (e.g., education and labour force status) and biological or
physiological (e.g., duration of breastfeeding and use of contraception). Obviously
the more closely spaced children are the larger is the number who can be borne in
a given period. The length of birth intervals thus affects the potential intensity of
fertility, although this relationship is of limited practical moment in populations
where even moderately high levels of fertility control are practiced. Research
generally shows that the spacing of births has little impact on fertility levels, the key
determinant of which is the parity at which women decide to cease childbearing.
Women can proceed quite rapidly to low parities then simply stop having children.
Analysis of birth intervals by social characteristics thus may say more about the
priorities of different groups of women through the life cycle than about their likely
completed family sizes.
Birth intervals are also of interest in studies of child survival and maternal health.
Very close spacing of births generally is found to be injurious to the health and
survival chances of both mothers and their babies, and family planning programs
often seek to promote wider spacing primarily for the health benefits likely to flow
to women and children. The monitoring of family planning programs thus often
entails monitoring their impact on birth intervals.
Whether dependent or independent, lengths of birth intervals have been promi-
nent variables in demographic analysis in recent years. Much analysis has focused
on specific intervals, and in particular the last closed birth interval (i.e., the interval
between a woman’s two most recent live births, or the first birth interval in the case
of women with only one child). This interval yields the most up-to-date data on
fertility and thus permits a focus on ‘current’ fertility.
Bogue (1993) has noted that, arithmetically, mean birth intervals are closely
approximated by inverting marital fertility rates, and proceeds to ask what, in view
of this, birth interval analysis has to offer. He concludes that it has three main
virtues. First, it permits study of person-to-person variation in childbearing and of
personal childbearing patterns, as distinct from the aggregate picture fertility rates
provide. Second, because it is founded in survey research, it allows for the impact
on childbearing of a much wider range of independent variables to be studied. And
third, it facilitates subdivision of the total childbearing span into segments that are
individually worthy of study – the interval from marriage to the first live birth;
Parity Progression 287

intervals between successive live births; and the open interval from the last live
birth to menopause.

Parity Progression

As has been previously noted, parity refers to the number of live births a woman
has had. An alternative term sometimes used for the same concept is issue. As well
as being an attribute of women who have passed menarche, though, parity is also
an attribute of any live birth, being the order of that birth (first, second, etc.) in the
reproductive history of the woman who had it.
Parity is a particularly important variable to consider when analysing the fertility
of populations that exercise significant control over their reproduction. Couples in
such populations tend to have desired, or target, family sizes, sometimes a little
flexible based on the gender mix of early births, but the attainment of which results
in concerted efforts to avoid further births and thus, in the aggregate, in increasing
departure of fertility from a ‘natural’ schedule as female age increases. A focus
on parity allows the researcher to study the pattern produced by these efforts at
fertility limitation, and changes therein as fertility preferences and/or the means of
translating them into achieved family sizes change.
A number of the fertility measures discussed so far have been measures of some
sort of ‘average’ experience. These, however, mask ranges of actual experience. A
focus on parity highlights this variety of experience, and it not infrequently turns out
that only a minority of women actually have approximately the ‘average’ number
of children (to the nearest whole child). McDonald (1990) showed for Australia,
for example, that if the 1988 pattern of fertility by birth order were to apply over
the lifetimes of a cohort of women, only 24 % of those women would have two
children (the 1988 TFR was 1.84), and only 26 % of children would live in two-
child families. Thirty-two percent of women would have three or more children and
61 % of children would have at least two siblings. Contrary to what a TFR of 1.84
might conjure up in one’s mind, while 24 % of women would have one child (20 %
would remain childless), a mere 13 % of children would be only children.
Parity data are gathered by censuses, vital registers and surveys. Census data
usually are obtained from a question like ‘How many children have you ever had?’
or ‘How many children have you ever given birth to?’; i.e., from a question designed
to yield the number of children ever born alive (CEB) to each woman. The question
not infrequently is asked only of currently married or ever married women, since
in many populations to ask young never married women whether they have had
children is considered inappropriate. It often is accompanied by an instruction to
exclude stillbirths, and is susceptible to understating parity unless also accompanied
by supplementary questions asking about children still living and those who have
died (there can be a tendency to overlook children born alive, but now deceased –
especially those who died soon after birth). In the case of parity data gathered
through vital registration it is wise to check the relevant question and instructions
288 6 Analysis of Fertility

on the registration form. As previously noted Australian experience is that these

may be worded with the aim of permitting ‘sensitive’ births (e.g., premarital births
unknown to the father of the current child) to be ignored, perhaps by asking only
about previous children of the current marriage or relationship. The exclusion of
some premarital births obviously introduces a bias here, but the consequences of the
approach are even more serious in a population in which divorce and repartnering
are common. If followed to the letter it results, for example, in a woman registering
a first child by a second husband being classified parity one, when if she also had
children by her first husband she is of higher parity. Parity data gathered in surveys
may derive from census-type questioning or from pregnancy or birth histories.
While demographers have been known to calculate parity-specific (as distinct
from order-specific) fertility rates (ratios of births of parity n during a year to
the mean, or mid-year, female population aged 15–49 of parity n  1), the main
application of parity data is to the calculation of parity progression ratios. A parity
progression ratio (PPR) is the probability that a woman who attains parity n goes
on to attain parity n C 1. Thus we refer to PPRs from parity zero to parity one, from
parity one to parity two, etc. The use of the word ‘probability’ here is worthy of note.
It is the first time it has occurred in this chapter, after having been frequently used in
Chap. 4 in discussing the analysis of mortality. Recall our earlier distinction between
renewable and non-renewable events, and that only the latter are suited to analysis
using probabilistic measures. Childbirth is a renewable event, but in splitting births
by parity we transform them into a series of non-renewable events, thereby opening
them up to probabilistic modes of analysis.
While it is possible to construct period, or synthetic cohort, PPRs (see Pressat
(1972: 228 et seq.), they are more usually calculated for real female birth cohorts
(i.e., age groups) or marriage cohorts (i.e., duration of marriage groups). In the case
of birth cohorts PPRs may be calculated for all women, currently married women
or ever married women, with data availability often dictating choice among these
options. In general terms the PPR for a cohort from parity n to parity n C 1 is given
by:

PPRn;nC1 D an D Women with at least n C 1 CEB=Women with at least n CEB

Given the distribution of a female birth or marriage cohort by number of CEB,

the calculating equation for any PPR from parity n to parity n C 1 is:

an D †iDnC1;¨ Fi =†iDn;¨ Fi (6.42)

Where n D the parity from which women are progressing; ¨ D the highest parity
attained by any member of the cohort; Fi D the number of females of parity i in
the cohort.
An example of the calculation of PPRs from a distribution of the ever married
female birth cohort aged 45–49 by number of CEB at the 1990 Census of Thailand
is presented in Table 6.8. Also shown are PPRs calculated in similar fashion for 45–
Parity Progression 289

Table 6.8 Calculation of parity progression ratios for ever married female cohort aged 45–49
at 1990 census of Thailand, and comparison with ratios for Thai (1960) and ever married New
Zealand (1981) cohorts
an -values Parity distribution per 1,000
PPR calculations Thailand 1990 Thailand NZ Thailand Thailand NZ
P
Parity (n) Fn iDn,¨ Fi an 1960 1981 1990 1960 1981
0 45,045 1,174,224 0.962 0.968 0.941 38 32 59
1 98,777 1,129,179 0.913 0.940 0.935 85 58 61
2 198,459 1,030,402 0.807 0.929 0.756 169 65 215
3 244,703 831,943 0.706 0.913 0.617 208 73 254
4 213,274 587,240 0.637 0.888 0.532 182 87 192
5 154,340 373,966 0.587 0.856 0.538 131 98 101
6 96,515 219,626 0.561 0.819 0.572 82 107 51
7 55,289 123,111 0.551 0.774 0.601 47 108 27
8 31,740 67,822 0.532 0.723 0.599 27 103 16
9 16,212 36,082 0.551 0.663 0.607 14 91 9
10C 19,870 19,870 17 178 15
Total 1,174,224 1,000 1,000 1,000
Sources: 1960 and 1990 Thai censuses; 1981 New Zealand census

49 year-old Thai females in 1960 and ever married New Zealand women of the same
age in 1981, together with parity distributions for radix populations of 1,000 women
in each of the three cohorts. The latter are presented as aids to comparison. They are
obtained by first multiplying the radix l0 by a0 to determine the number of women
who reach parity 1, l1 ; then l1 by a1 to determine l2 ; l2 by a2 to determine l3 ; and
so on. For each parity n we then take ln  lnC1 (for parity 10C treat lnC1 D 0), the
result being the number of women in the radix who finished childbearing at parity
n. The procedure is the equivalent of finding dx values in a standard life table.
The input data to the calculations for Thailand in 1990 are the column of Fn
values (i.e., the census parity distribution for the cohort). In the next column these
values are successively cumulated from the bottom of the column upward (i.e., F10C ,
F10C C F9 , F10C C F9 C F8 , etc.). Values of an (PPRs) then are found in accordance
with Eq. 6.42 by dividing the cumulated value in each row of the table into the
cumulated value in the row below it.
Several points concerning Table 6.8 deserve comment. First, the PPR calculations
for Thai females in 1990 are based on women whose parities were known. Parity
was unknown for a further 50,392 ever married women aged 45–49. Ignoring these
women effectively pro-rates them to the various parity categories, but conceivably
they were disproportionately of zero parity. Assuming them to have all been of
zero parity reduces a0 from 0.962 to 0.922, and increases the number of parity
zero women per 1,000 cohort members from 38 to 78 (with smaller compensating
adjustments to numbers at higher parities). The issue of how ‘parity not stated’
cases are treated is, therefore, potentially a significant one, particularly if the degree
of childlessness in a cohort is of interest. In the results shown for New Zealand
290 6 Analysis of Fertility

in Table 6.8 the effect of ignoring such cases is less serious. Assuming all ‘parity
not stated’ women to have been of zero parity reduces a0 from 0.941 to 0.933, and
increases the number of parity zero women per 1,000 cohort members from 59 to
just 67.
Second, note that the calculations presented are for women aged 45–49; i.e.,
for birth cohorts that had completed their childbearing, or were so close to doing
so that they would have few subsequent births. Parity data for women still of
reproductive age are censored as data on birth intervals are; further progressions
to higher parities will take place in the future, and PPRs are biased by failing to
take them into account (and not necessarily downward – if the earliest members
of a cohort to attain a parity are the most likely to progress beyond it, the bias
could be upward). Thus PPRs are of most use when calculated for cohorts that have
completed their fertility, and if calculated for other cohorts must be interpreted with
great caution. The situation when dealing with birth cohorts is reasonably clearcut –
are cohort members or are they not beyond, or almost beyond, the age at which
births can occur in significant numbers? With marriage cohorts, however, assessing
whether reproductive potential of consequence remains is less straightforward, and
needs to take account of ages at marriage. Except in exceptional circumstances, a
cohort married 30 years or more will be beyond having many further children, but
what about one married 15 years or 20 years? If tending on average to have married
relatively late one could have more confidence in biology limiting further fertility
than if having married early, but then ages at marriage follow a distribution, and
in almost any marriage cohort there would be some early marriers. Sometimes it
is possible to examine parity progression for a marriage or birth cohort still with
reproductive potential on the basis that that potential is likely to be realized mainly
in higher order births, so that PPRs between the major, lower, parities are likely to be
fairly reliable. But the issue of whether there is any censoring of cohort experience
should always be carefully assessed.
Third, the 1960 and 1990 Thai schedules of PPRs (and parity distributions per
1,000 cohort members) in Table 6.8 are very different. The fact that the former
pertains to all women and the latter to ever married women is of little consequence,
since marriage among the cohort aged 45–49 in 1960 was well-nigh universal.
Note that in the 1960 schedule PPRs decline gradually, but consistently, as parity
increases. This tends to be the pattern in populations that exercise limited control
over their fertility and in which biological processes (including deaths of husbands)
play a major role in lowering fertility over the course of the reproductive life cycle.
The parity distribution that results is heavily weighted towards the higher parities,
with a large proportion of women having 10C children. In the 1990 schedule, PPRs
fall more rapidly to a level at about a6 that is well below a9 in the 1960 schedule, and
at which they remain roughly constant thereafter. Clearly between 1960 and 1990
control over higher parity births increased considerably in Thailand, as is reflected
in a parity distribution in which only 187 women per 1,000 had reached parity 6 or
higher, compared to 587 per 1,000 for the equivalent cohort in 1960.
Fourth, the pattern of rapid decline in an with increasing parity to a level beyond
which it becomes more or less constant is even more pronounced in the schedule
Parity Progression 291

of PPRs for ever married 45–49 year-old New Zealand women than in that for ever
married Thai women of that age in 1990. Decline ceases at a4 . Such a pattern tends to
be an empirical regularity in populations that practice high levels of fertility control,
and this regularity has considerable potential for estimating ultimate completed
fertility rates (cohort TFRs) for birth, and more especially marriage, cohorts that
are only part way through their reproductive lives.
If a cohort has completed its childbearing there is a natural relationship between
its PPRs and the cohort TFR. It is given by:

Cohort TFR D a0 C a0 a1 C a0 a1 a2 C a0 a1 a2 a3 C    C a0 a1 a2 : : : a¨1 (6.43)

Where ¨ D the highest parity any cohort member attains.

A related equation allows us to estimate a cohort TFR for a marriage cohort
married for a reasonable length of time (say 15 years or more) or a birth cohort the
great majority of whose members similarly have been married for some time, but
which is yet to complete its childbearing. It is only usable in respect of populations
that exercise considerable control over fertility, but increasingly many developing
country populations are approaching this point. The equation is:

Cohort TFR D a0 C a0 a1 C a0 a1 a2 C .a0 a1 a2 a3 = .1  a4C // (6.44)

Where a4C D the average of a4 , a5 and a6 .

Table 6.9 shows data to which Eq. 6.44 might be applied. One hesitates to apply
it to the cohort married for 10 years because the constancy of PPRs beyond a3 is
problematic, and equilibrium thus seems yet to be established. The basis of the
method is an assumption that after a reasonable period of exposure to the risk of
marital childbearing a cohort will have largely completed its progression through
lower parities, with subsequent childbearing likely to be heavily concentrated at
higher parities at a constant rate of parity progression. Ten years of exposure in an
Australian setting in which couples routinely delay the first birth after marriage to
establish married life on two incomes seems not to be enough to fully embrace this
‘progression’. But 15 years seems adequate, and applied to the cohorts married for
15, 20 and 25 years Eq. 6.44 yields estimated cohort TFRs of 2.69, 2.91 and 2.95
children per woman. This provides solid evidence of quite steep fertility decline
from census data gathered around 15 years after that decline had begun.
PPRs shown in Tables 6.8 and 6.9 have been computed from distributions of
female birth and marriage cohorts by number of CEB obtained from censuses.
Similar distributions might as easily have come from surveys, and these are the
types of sources you are likely to have to rely on in dealing with the populations
of developing countries. However, data from reliable birth registers can also yield
PPRs for birth and marriage cohorts. Without going into detail, we build up parity
distributions for birth and marriage cohorts year by year using annual tabulations of
live births by number of previous issue and age of mother (birth cohorts) or duration
of marriage (marriage cohorts). There are, though, some complications to be aware
292 6 Analysis of Fertility

Table 6.9 Parity progression ratios for cohorts of Australian women married for 10, 15, 20 and
25 years, 1976
Parity progression ratio
Duration of marriage a0 a1 a2 a3 a4 a5 a6 a7
10 years 0.8991 0.8951 0.4722 0.2821 0.2358 0.2475 0.2848 0.3941
15 years 0.9203 0.9247 0.6318 0.4540 0.3598 0.3690 0.3587 0.3953
20 years 0.9105 0.9137 0.6945 0.5495 0.4526 0.4581 0.4625 0.4768
25 years 0.9117 0.8943 0.6880 0.5849 0.5092 0.5088 0.4925 0.5191
Source: Calculated from 1976 Australian census data

of. First, we must obviously obtain estimates of the original sizes of cohorts from
somewhere to be able to estimate a0 , since a birth register gives no direct information
on numbers of women who remain permanently childless. Possible sources are the
birth register itself, a marriage register or a census, and it must be appreciated that
computed cohort levels of childlessness will depend heavily on the accuracy of
these estimates. Second, it may be necessary to make adjustments to allow for the
changing size of cohorts over time due to migration, death and marriage dissolution.
This can require a lot of additional data and be complicated. The beauty of basing
PPRs on a census or survey distribution of a female cohort by number of CEB is that
this sort of problem does not arise, because we are dealing only with ‘survivors’ at
the date of data collection.
What have been covered in this section are the rudiments of parity progression
analysis. With adequate data more complex descriptions of fertility processes are
possible. One extension of the principles covered is to classify parity progression
data by both age and duration of marriage, so that within birth cohorts the behaviour
of different marriage cohorts can be contrasted and vice versa. It is also intuitively
clear that there is a link between parity progression analysis and birth interval
analysis. Together they fully describe the process of fertility – the former addressing
the quantity or intensity dimension and the latter the timing or tempo dimension
(relative to marriage as an initial reference point). Thus parity progression analysis
can be combined with birth interval analysis. Indeed Feeney (1983) has shown how
the two can be fused using parity progression schedules (consisting of the parity
progression ratio and birth interval distribution for the transition from one parity
to the next) and parity cohorts (groups of women attaining a given parity in a
given period) into an approach to fertility analysis that (i) is more natural than
that based on age-specific fertility rates (because it focuses on the whether and
when to have a first or subsequent birth that is at the heart of fertility decision-
making and because the proximate determinants of fertility operate by regulating
parity progression and birth interval length), (ii) is able to overcome the technical
problems of censoring and selection that affect simpler approaches to birth interval
analysis, and (iii) provides a framework for looking at how birth intervals affect
aggregate fertility and population growth trends, in contrast to micro-level analyses
that concern themselves with what happens within birth intervals.
Biological Aspects of Fertility 293

Biological Aspects of Fertility

From earlier discussion of the Bongaarts fertility model it will be clear that several
of the proximate determinants of fertility are biological. In this section we deal
briefly with those proximate determinants.

Fecundability

The term ‘fecundability’ refers to the probability of a fecund woman conceiving

during a single menstrual cycle. It is an important concept in statistical modelling
of the family building process and in studying the effectiveness of family plan-
ning programs, even though Bongaarts sees natural fecundability as insufficiently
variable from population to population to warrant incorporation in his model.
Finding this probability might seem to require medically screening a panel of
women to ensure they were fecund, regularly sexually active, and had no other
impediments to their ability to conceive (e.g., sterility, short-term absence of their
sexual partners, contraception, post-partum amenorrhoea, etc.). In many countries,
however, newly married women who did not give birth before marriage and were
not already pregnant at marriage largely meet these conditions, and one method of
estimating fecundability focuses on this group. Be aware, though, that the method
is inappropriate for populations that delay first births by practicing contraception
immediately after marriage. It assumes sexual activity at an intensity unlikely to be
exceeded subsequently in the marriage, lack of any attempt to prevent conception,
and that both partners physiologically are never in future likely to be more fertile.
Assessing fecundability in populations that practise contraception straight after
marriage is much more difficult, and we will not attempt to deal with it here.
If the probability a newly married fecund woman will conceive during any
menstrual cycle is f, the probabilities of conception during each cycle following
marriage are given by the series:

f; .1  f/ : f; .1  f/2 : f; .1  f/3 : f; .1  f/4 : f; etc:

That is, the probability of conceiving during the first cycle; the probability of not
conceiving during the first multiplied by the probability of conceiving during the
second; the probability of not conceiving during the first two cycles multiplied by
the probability of conceiving during the third; and so on. After some mathematical
manipulation it transpires that the mean number of cycles after marriage before
conception, assuming that marriage coincides with ovulation, is 1/f, while assuming
an even distribution of marriages through the menstrual cycle yields:

i D 1=f  1=2 (6.45)

294 6 Analysis of Fertility

Table 6.10 Calculation of fecundability of women who had given birth within 5 years of first
marriage for respondents to the 1974 Malaysian fertility and family survey, by age at marriage
Mean months to first Mean months to first Mean cycles to first
Age at marriage birth conception conception f
(1) (2) (3) (4) (5)
<15 26.0 16.5 17.9 0.054
15–17 20.1 10.6 11.5 0.083
18–19 17.5 8.0 8.7 0.109
20–21 16.5 7.0 7.6 0.123
22–24 16.5 7.0 7.6 0.123
25–29 18.1 8.6 9.3 0.102
30C 19.0 9.5 10.3 0.092
(3) D (2)  9.5
(4) D i D (3) . 365/336
(5) D 1 / ((4) C 0.5)
Source: Malaysian fertility and family survey 1974, First country report, Table 2.1.1

Where i D mean number of cycles before conception; f D the probability of conceiv-

ing during a single cycle.
It follows that:
 
f D 1= i C 1=2 (6.46)

Use of Eq. 6.46 to estimate fecundability for women who married at different
ages is illustrated in Table 6.10. The input data are mean numbers of months
between marriage and first birth for women who gave birth within five years of first
marriage calculated from the Malaysian World Fertility Survey of 1974. Women
with negative first birth intervals (implying premarital birth) and first birth intervals
of less than 8 months (which imply premarital conception) are disregarded, the latter
by adjusting published means on the assumption that births at marriage durations 0–
7 months occurred on average at exact duration 4 months. Calculations then proceed
through the following steps.
Step 1: Mean first birth intervals are changed to mean intervals to first conception
by subtracting 9.5 months (column (3)).
Step 2: Mean intervals to first conception are converted to mean numbers of com-
pleted cycles before conception (column (4)) by multiplying by 365/336. This is
approximately 13/12, and recognizes that a menstrual cycle lasts approximately
four weeks, so that there are 13 in a year and each month to first conception
embraces 11 /12 cycles).
Step 3: Equation 6.46 is used to generate column (5) from column (4) (i.e., values
of f from values of i).
Biological Aspects of Fertility 295

Note that fecundability rises from a low level for women married when aged
less than 15 to a peak for those married aged 20–24, then drops again for
women married at older ages. This pattern seems intuitively plausible, allowing
for adolescent subfecundity (a commonly observed phenomenon) and for declining
fecundity at older reproductive ages. It is, however, necessary to remember that zero
contraception prior to first conception is assumed. If, for example, women married
at older ages were better educated, more likely to have careers, and consequently
likely to have used contraception following marriage to delay first births, the drop
in fecundability at those ages could, at least in part, be spurious.

Post-partum Amenorrhoea

A definition of amenorrhoea as the temporary absence of menstruation was given

early in this chapter, where a distinction also was made between the condition
during pregnancy (pregnancy amenorrhoea) and following confinement (post-
partum amenorrhoea). The former is strictly biological and a constant in terms of
its inhibiting effect on fertility, and thus is of limited interest to demographers.
The latter, however, has a cultural dimension which makes its effect distinctly
variable from woman to woman and from population to population. It is therefore
of much greater interest both as a potential source of explanation for fertility
trends and differentials, and as something family planning and maternal and infant
health policy should take account of and possibly seek to influence. This cultural
dimension has two components, one of which influences the biological condition
itself and the other of which reinforces that condition, is capable of extending
the period of post-partum non-susceptibility to the risk of conception, and is
difficult to disentangle from the first. The former component focuses on cultural
attitudes to breastfeeding, which tends to extend the base period of post-partum
non-susceptibility (estimated from studies of birth intervals following stillbirths to
be about 3 months). Societies in which women typically breastfeed for extended
periods exhibit longer intergenetic birth intervals, and hence lower completed family
sizes under conditions of natural fertility. Breastfeeding is not a reliable means of
pregnancy prevention at the individual level, but does seem to have some preventive
effect for up to about 12 months following confinement so long as the child
continues to receive breast milk as its only or main source of food (protection
diminishes rapidly once other foods are introduced). Aside from the nutritional
benefits to the child of breast milk, the longer birth intervals extended breastfeeding
promotes are independently beneficial to the health and survival chances of both
mother and child, and thus inherently worthy of encouragement through policy.
They are also of value as a restraining factor on population growth and, insofar
as they enhance child survival and confidence therein, as a factor in encouraging
couples to stop childbearing at moderate parities.
The second component of the cultural dimension to the period of post-partum
amenorrhoea is cultural restrictions on sexual intercourse during the post-partum
296 6 Analysis of Fertility

period. Prescribed periods of abstinence from coitus may extend for fixed lengths
of time, may be linked to the duration of breastfeeding through beliefs, for example,
that semen will pollute the mother’s milk, or both. Strictly speaking amenorrhoea
is a biological condition, which breastfeeding extends through biological processes
but which abstinence does not. Abstaining prevents conception by excluding semen
from the female’s vagina, not by preventing the resumption of ovulation. But
abstention in conjunction with breastfeeding reinforces the biological condition
of post-partum amenorrhoea, and if extending beyond the end of breastfeeding
effectively prolongs the period of post-partum non-susceptibility to conception.
Both together help determine when exposure to the risk of conception resumes;
their individual effects on fertility are difficult to separate; and they thus are apt to
be treated in tandem.
The measurement of post-partum infecundability, or the extent to which breast-
feeding and post-partum abstinence combine with base-level post-partum amenor-
rhoea to defer the resumption of exposure to the risk of conception following a birth,
was covered earlier in discussing Bongaarts’s fertility model. Refer to Eqs. 6.39 and
6.40.

Sterility

Once again as already noted early in this chapter, sterility, or infecundity, are
terms for a biological incapacity to conceive. The most straightforward measure
of the extent of sterility in a population is the proportion of married women who
reach menopause (approximated as ages 45–49) without having had a live birth.
Such women are not necessarily physiologically sterile – they may be capable
of conception but have miscarried due to other physiological problems – but this
sort of measure is generally the only one readily available and is regarded as
reasonable given that the alternative is an expensive special survey incorporating
medical examinations. It is, of course, only of any use for populations where the
desire for children within marriage is universal. If contraception is readily available
and there is reason to believe some married women use it to remain voluntarily
childless, then the measure becomes one of this voluntary childlessness and sterility
combined (excluding, of course, voluntary childlessness associated with decisions
not to marry), but is not a measure of sterility alone.
Sterility has a variety of types and causes, and is not necessarily a (reproductive)
lifelong condition. Aside from lifelong sterility it can be caused by various
diseases, including sexually transmitted diseases (STDs) such as gonorrhoea and
chlamydia which sometimes lead to pelvic inflammatory disease (PID), and thence
to permanent scarring of the fallopian tubes so that the passage of ova is prevented.
Some strains of tuberculosis can have the same effect, and PID can arise from causes
other than gonorrhoea and chlamydia. Other STDs, like syphilis, affect the viability
of the foetus rather than fecundity, and thus increase the probability of miscarriage
or stillbirth. Poor nutrition can induce temporary sterility by causing amenorrhoea
References 297

or other menstrual irregularities, but restoration of appropriate nutrition usually

quickly eliminates such a condition. Lifelong sterility is often referred to as
primary sterility, and that which prevents a woman having further births after
she has already had one or more as secondary sterility. It is, of course, only the
former that the measure discussed in the previous paragraph is able to capture.
Secondary sterility should also be differentiated from sub-fertility, which refers to
the situation of women who have extreme difficulty conceiving from the outset,
although distinguishing between these conditions in practice is not easy and both
may have similar connections to disease patterns in a society.
Finally, sterility can also be medically caused. Sometimes this occurs as a by-
product of a medical procedure carried out for other than contraceptive reasons (e.g.,
a hysterectomy), but increasingly in recent decades couples and individuals have
resorted to medically caused sterility as a terminal method of contraception. Either
female (by tubal ligation) or male (by vasectomy) sterility may be sought to this
end.

References

Bogue, D. J. (1993). Introduction to pregnancy/birth interval analysis. In D. J. Bogue, E. E.

Arriarga, & D. L. Anderton (Eds.), Readings in population research methodology (Vol. 3, pp.
10-59–10-64). Chicago: Social Development Center for UNFPA.
Bongaarts, J., & Potter, R. G. (1983). Fertility, biology, and behavior: An analysis of the proximate
determinants. New York: Academic.
Carmichael, G. A. (1986). Birth order in Australian vital statistics. Journal of the Australian
Population Association, 3(1), 27–39.
Carmichael, G. A. (2013). Estimating paternity in Australia, 1976–2010. Fathering, 11(3), 256–
279.
Coale, A. J., & Trussell, T. J. (1974). Model fertility schedules: Variations in the age structure of
childbearing in human populations. Population Index, 40(2), 185–258.
Coale, A. J., & Trussell, T. J. (1975). Erratum to Coale and Trussell (1974). Population Index,
41(4), 572.
Corr, P., & Kippen, R. (2006). The case for parity and birth order statistics. Australian and New
Zealand Journal of Statistics, 48(2), 171–200.
Davis, K., & Blake, J. (1956). Social structure and fertility: An analytic framework. Economic
Development and Cultural Change, 4(3), 211–235.
Feeney, G. (1983). Population dynamics based on birth intervals and parity progression. Population
Studies, 37(1), 75–89.
Henry, L. (1961). Some data on natural fertility. Eugenics Quarterly, 8, 81–91.
Henry, L. (1976). Population analysis and models. London: Edward Arnold.
Jones, E. F. (1971). Fertility decline in Australia and New Zealand 1861–1936. Population Index,
37(4), 301–338.
McDonald, P. (1990). The 1980s: Social and economic change affecting families. Family Matters,
26, 13–18.
Newell, C. (1988). Methods and models in demography. London: Belhaven Press.
Pressat, R. (1972). Demographic analysis: Methods, results, applications. London: Edward
Arnold.
Shryock, H. S., Siegel, J. S., & Associates. (1973). The methods and materials of demography. 2nd
Printing (revised). Washington, DC: U.S. Bureau of the Census.
298 6 Analysis of Fertility

United Nations. (1983). Manual X. Indirect techniques for demographic estimation.

ST/ESA/SER.A/81. New York: United Nations. Available online at http://www.un.org/esa/
population/techcoop/DemEst/manual10/manual10.html. Accessed 26 July 2013.
World Health Organization. (1993). International statistical classification of diseases and related
health problems (Tenth Revision, Vol. 2). Geneva: WHO.
Chapter 7
Population Distribution, Urbanization
and Migration

Demography and Geography

Certain themes in demographic analysis have a distinctly geographic dimension

to them, and geographers not surprisingly have taken particular interest in them.
This is not the place for a review of the development and contemporary nature of
population geography (Jones 1990; Newbold 2010). Nor does one wish to create
the impression that themes addressed in this chapter represent anything like the
totality of the demographic issues and phenomena in which geographers take an
interest. But the geographic (or ‘spatial’) distribution of population, the movement
of population between geographic locations (variously mobility and migration),
and the process of urbanization, to which population redistribution by migration
invariably is fundamental, are all inherently geographic, and hence clearly of central
concern to geographers.

Population Distribution

Thus far in this book populations focused on in illustrating demographic techniques

and methodological principles have been national populations, or subgroups thereof
defined by variables like ethnicity. For many purposes, however, national-level
data are too crude. The geographic distribution of national populations typically
is very uneven, with large numbers of people living in relatively small areas in
urban centres and sometimes, at the other extreme, vast areas containing almost
nobody. Australia is a classic example, with its population heavily concentrated in
cities and towns along the eastern, southeastern and southwestern coastal areas,
and huge inland tracts that are virtually uninhabited. Then again, demographic
characteristics and processes rarely occur uniformly over a country; subnational

© Springer International Publishing Switzerland 2016 299

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_7
300 7 Population Distribution, Urbanization and Migration

political, administrative and planning bodies have responsibility for, and therefore
an interest in, restricted geographic areas; and these same bodies and national
governments are concerned to know about population movement within as well
as across national borders. All of these considerations argue for demographic data
tabulated for discrete geographic areas within a country. Such data normally derive
from censuses, and to a lesser extent from vital registers, population registers and
administrative sources. They are less likely to come from sample surveys, although
national-level sample surveys can sometimes yield broad regional subsamples and
frequently allow urban/rural and/or metropolitan/non-metropolitan distinctions to
be made.

Levels at Which Data Are Tabulated

Most countries of any size tabulate demographic, and especially census, data at
a range of different geographic levels. Statistical agencies usually must recognize
as a priority in this process a country’s political or administrative subdivisions,
which may range from primary units such as the states of Australia, the USA
or India and the provinces of Canada, Indonesia or Thailand through a hierarchy
of subsidiary levels to local administrative units which may exist at the village
or sub-village level in rural areas, and at the suburban or neighbourhood level
within cities. In many countries simply counting numbers of persons, or of persons
with particular attributes, resident within these subdivisions is the major reason for
conducting a census at all, such counts determining the geographic allocation of
certain resources and apportionment of parliamentary representation. Population
counts for political or administrative units also are necessary because those units
are natural foci for planning, social and economic policy development, and studies
of internal migration.
Hierarchies of administrative units vary in the number of levels they have
and in the names given to units at each level, and often bifurcate along urban-
rural lines. Thus cities, towns, villages and various administrative subdivisions
thereof (boroughs, wards, arondissements (France), ku (Japan, Korea), etc.) often
are recognized alongside a system of counties, shires, etc. which cover rural areas
(and may overlap with the urban system at lower levels). Statistical agencies may
also, however, create geographical units for purely statistical purposes. Primary
administrative units sometimes are aggregated into broader regions for statistical
purposes, as occurs, for example, with the publication by Indonesia of data for island
groupings (which are aggregates of provinces). Regions may also be designed which
are perceived to be functional economic or cultural areas. These may be groupings
of secondary or tertiary administrative units, may cut across the boundaries of
primary administrative units, or may bear no relation to administrative boundaries
at all. Census authorities also are often keen to draw and periodically revise their
own boundaries around urban agglomerations so as to reflect functional rather
than political or administrative reality, and may create areas which facilitate data
Population Distribution 301

collection, cater to the demand for small-area data (see below), are relatively homo-
geneous (particularly socio-economically or culturally), impart flexibility to the
system of geographical classification, and facilitate the presentation of data at levels
of geographic disaggregation perceived to be useful across both urban and rural
landscapes. Thus in Australia, for example, under the new Australian Statistical
Geography Standard (ASGS) introduced in 2011, the basic spatial building block
for release of census data is the Mesh Block, for which little more than population
and dwelling counts are produced. Around 350,000 of these aggregate to 55,000
Level 1 Statistical Areas (SA1s) with populations in the range 200–800 people (the
smallest areas for which more extensive census data are released); these aggregate to
2,200 Level 2 Statistical Areas (SA2s) with populations between 3,000 and 25,000;
these in turn aggregate to around 350 Level 3 Statistical Areas (SA3s) of 30,000–
130,000 people; then there are just over 100 Level 4 Statistical Areas (SA4s) with
populations between 100,000 and 500,000 which aggregate to the nine States and
Territories and then to Australia. Various levels in this hierarchy also aggregate to
a range of other regional subdivisions, including such things as Postal Areas, State
and Commonwealth Electoral Divisions, Tourism Regions, Significant Urban Areas,
and Greater Capital City Statistical Areas.
Mention was just made of small-area data. Particularly for urban areas, such
data tend to be in heavy demand, and the smaller the small areas the better so far
as many users are concerned. In some countries the lack of any administrative unit
sufficiently small to meet small-area data requirements has been a major reason
for developing geographic units which are first and foremost statistical entities,
and only incidentally ever administrative ones. Small-area data permit one to focus
on localized communities and to differentiate urban socio-demographic landscapes
in considerable detail. They also offer great flexibility in constructing regions to
personal specification through aggregation. All of these features are very attractive
to planners, businesses making location decisions and marketers. It should be
appreciated, though, that as geographic disaggregation increases, the classificatory
and cross-tabulatory detail in data statistical agencies are prepared to release may
tend to diminish, as the confidentiality of individual respondents becomes more of
an issue.

Urban and Rural Populations

Significant differences long have been recognized to exist between populations that
live in cities and in small villages or the open countryside. Hence differentiating
between urban and rural populations in demographic data collections has a lengthy
history. But if there is broad agreement on the desirability of making an urban-
rural distinction, there is no universally accepted means of making it. The extremes
are not a problem. The hearts of major cities are clearly urban and areas in which
people almost universally reside on and work on farms are rural. But where, along
the continuum in between, does urban stop and rural begin?
302 7 Population Distribution, Urbanization and Migration

The upshot is that criteria used to separate urban from rural populations vary
considerably between countries, and may vary over time within them. There is not
even agreement that a standard definition of ‘urban’ is desirable, it being widely
felt that different types of definitions suit different purposes and types of countries.
Thus, when dealing with urban-rural data routinely acquaint yourself with the
definition of ‘urban’ that underpins them, especially if any comparative analysis
is to be undertaken (note that ‘rural’ generally is not directly defined, but is treated
as a residual after eliminating ‘urban’).
The recommended focus when dividing populations into urban and rural com-
ponents is the locality, defined by the Statistics Division of the United Nations
(SDUN) (2008: 123) as ‘a distinct population cluster in which the inhabitants live
in neighbouring sets of living quarters and that has a name or a locally recognized
status.’ If that is not possible, the smallest administrative unit in a country should
be used. Each locality or administrative unit, and the population residing at or
within it, is determined to be urban or not urban, the populations of non-urban
localities and areas not deemed to be localities being the rural population. After
affirming that cross-national differences in the characteristics that distinguish urban
from rural areas make a universal definition of ‘urban’ impractical, the SDUN
goes on to note that, in industrialized countries, lifestyle and level-of-living lines
of differentiation between urban and rural areas have become blurred, leaving the
degree of concentration of population as the main discriminator. Thus population
size and density generally are key considerations in classifying localities as ‘urban’.
There is, however, no consensus as to appropriate minimum values for these
variables, which tend to vary depending on how populous and densely settled a
country is overall. In less developed countries, where large densely settled areas
still are characterized by ‘a truly rural way of life’, additional criteria, and perhaps
something more elaborate than a simple urban-rural classification, are in order.
Among these extra criteria are the extent of employment in non-agricultural
activities and the presence of urban facilities. The latter include such things as
access to electricity, piped water and a flush toilet in the home, and to health
services, education and a local public transport system.
Other criteria by which localities may be classified ‘urban’ include whether or
not they have been classified ‘urban’ administratively by government and whether
they are subject to some form of local government. Classifications in which these
principles feature have a tendency to be unhelpfully static and outdated. Thus, for
example, significant areas of clearly urban character by objective criteria may be
excluded from the urban population because the wheels of bureaucracy turn slowly
in formally designating them urban or adding them to existing ‘urban’ areas with
which they are contiguous. Thailand is a present day case in point, with scholars
routinely having to adjust official estimates of the urban population upward to
achieve a realistic outcome.
Population Distribution 303

Methods of Analysis

A seemingly very straightforward approach to examining population distribution is

to map the numbers or densities of inhabitants of geographic areas or localities. A
common method of mapping population numbers is to use a dot to represent a fixed
number of people (say 100 or 1,000), dots being located physically within areas
to which they pertain. This method can, however, run into difficulties when both
sparse rural and dense urban concentrations of population need to be portrayed.
Dots representing urban populations can merge into an undifferentiated mass. One
solution is to use dots of a different colour and representing a larger number of
people in urban areas. Another is to represent urban localities by a geometric image
(perhaps a circle) of size proportional to population.
If mapping population densities the usual approach is to develop a set of density
categories, associate with each a distinctive shading in a scheme in which density
of shading increases with density of population, and shade each geographic area
according to the category into which its population density falls. Well designed
and executed, maps of this type can have a striking visual impact, although they
do sacrifice some precision. Population densities are calculated by dividing the
population of an area by its physical area in, for example, square kilometres to yield
a figure of so many persons per square kilometre. They can, however, be deceptive.
If large parts of a country or region are physically inhospitable and therefore largely
uninhabited, densities may bear little relation to actual densities in those areas that
are populated to any significant extent. In this circumstance density calculations may
for some purposes usefully be made with reference to arable agricultural land area
rather than to total land area. Urban population densities can also deceive, depending
on the definition of ‘urban’ employed. The more a city is defined to include urban-
rural fringe areas the lower its population density is likely to be.
Another simple way of summarizing population distribution is to list the geo-
graphic units of a country or region and calculate percentages of total population
living in each. Such tabulations facilitate the measurement of change in the distri-
bution of population through comparison of percentage distributions at discrete time
points. Normally we focus on a level of geographic disaggregation with relatively
few categories (e.g., the States and Territories of Australia), so that individual
percentages, and changes therein, do not become so small as to mask major trends.
Geographic localities may also be rank-ordered by population size, a strategy again
lending itself to examining distributional change over time. Comparison of rank
orderings at two censuses can point to areas that have grown comparatively rapidly
or slowly in the interim, although this procedure may be insensitive at higher ranks
in the population size hierarchy, where rank to rank differences in population are so
large that substantially different intercensal growth rates would produce no change
in ranking.
Statisticians and geographers have from time to time proposed measures of the
average location and degree of dispersion of population distributions, although they
are somewhat peripheral to the technical armoury of the specialist demographer.
304 7 Population Distribution, Urbanization and Migration

Locating the median point of a population distribution involves drawing the east-
west line on a map above and below which exactly half the population is located,
and the north-south line that also divides the geographic distribution in two. The
median point is then the intersection of these two lines. Intersections of east-west
and north-south lines that divide distributions into quartiles, deciles etc. can also be
located. Plotting these points for successive census dates can provide an interesting
perspective on distributional change over time.
A variant on the median point of a population’s geographic distribution is the
centre of population, otherwise known as the mean point of the distribution or
its centre of population gravity. After overlaying an x,y-scaled grid on a map of
population distribution, the (x,y) coordinates of the centre of population are given
by:

x D †iD1;¨ pi : xi =†iD1;¨ pi I y D †iD1;¨ pi : yi =†iD1;¨ pi (7.1)

Where pi D the population of unit of area i; xi and yi D the x and y coordinates of

the centre of area i; ¨ D the number of units of area into which the country or
region being studied is divided.
Unlike the median point, the centre of population is influenced by the distance
of each person from it. Population redistribution within a quadrant formed by
intersecting east-west and north-south median lines has no effect on the median
point, but does alter the centre of population. Again, movement of the centre of
population over time captures change in the distribution of population.
One measure of the dispersion of a population distribution is the standard
distance. This bears the same relationship to the centre of population as the standard
deviation bears to the arithmetic mean in statistics. While in theory it should
be calculated based on individual members of the population and their locations
relative to the centre of population, in practice we normally work, as in Eq. 7.1,
with data giving population counts for (preferably quite small) units of area. Under
this scenario the standard distance is given by:
p nh i o
DD †iD1;¨ fi .xi  x/2 C †iD1;¨ fi .yi  y/2 =n (7.2)

Where fi D the population of unit of area i; xi and yi D the x and y coordinates of

the centre of area i; x and y D the x and y coordinates of the centre of population;
n D the total population of the country or region for which D is sought; ¨ D the
number of units of area into which that country or region is divided.
The answer obtained will be in units of distance corresponding to a value of 1
on the scales of x and y coordinates. It can be converted to a meaningful distance
(e.g., in kilometres) by multiplying by whatever this unit of distance is (e.g., if an
increment of 1 on the x or y scale corresponds to 100 km, we multiply by 100 to
obtain an answer in kilometres).
Population Distribution 305

A graphical technique for depicting the concentration or dispersion of population

is provided by the Lorenz curve. Originally developed for studying inequalities in
the distribution of wealth or income it requires geographic localities within a country
or region, or groups of localities classified by size of population, to be arrayed
from the most to the least populous, and cumulative proportions of localities and
of total population to be calculated down this array. These cumulative proportions
then are plotted against one another (population on the x-axis; localities on the y-
axis) and compared to the x D y diagonal, which represents the situation where all
localities have equal population (maximum dispersion of population). The closer to
this diagonal a Lorenz curve lies the less concentrated (or more evenly dispersed
among localities) a population is. Note that where a proportion of population is
reported as not living in localities it has to be ignored. This is not a problem in
Fig. 7.1, which shows Lorenz curves for Venezuela and Italy in 1981; only 0.5 % of
Venezuela’s population and none of Italy’s was not living in a locality. It can be seen
that the Venezuelan curve hugs the x-axis and righthand border of the graph much
more closely than that for Italy, or in other words is decidedly further displaced from
the reference diagonal. It signifies a population a very high proportion of which is
concentrated in very few localities, and in which a relatively high proportion of
localities are home to a relatively small proportion of population. The pattern for

Fig. 7.1 Lorenz curves showing degrees of population concentration in Venezuela and Italy, 1981
306 7 Population Distribution, Urbanization and Migration

Italy, while also indicating a concentration of population in relatively few, larger

localities, is much less pronounced. Clearly there is a wider geographic spread of
population.
The degree of concentration of population indicated by a Lorenz curve can be
quantified. The Gini concentration ratio measures the proportion of the area under
the reference diagonal that lies between the diagonal and the Lorenz curve. The
higher the ratio the greater is the concentration of population in a few localities. The
relevant equation is:

Gi D †iD1;n1 xi :yiC1  †iD1;n1 xiC1 : yi (7.3)

Where xi D cumulative proportion of population living in localities in size category i

and all larger categories; yi D cumulative proportion of localities in size category
i and all larger categories; n D number of size of locality categories arrayed from
largest (n D 1) to smallest (n D n).
Table 7.1 sets out the calculations for obtaining both the (xi , yi ) Lorenz curve
plots and the Gini concentration ratio for Venezuela. The Gini ratio of 0.8919
compares with a value of 0.7115 for Italy.
Another approach to analysing population concentration using the Lorenz curve
and Gini concentration ratio entails focusing on areas rather than localities. Instead
of arraying localities and their populations from the most to the least populous,
a country’s (or region’s) geographic subareas, together with their land areas and
populations, are arrayed in order of population density from most dense to least
dense. Cumulative proportions of total land area and of total population then are
calculated down this array and plotted against one another as before (population on
the x-axis; land area on the y-axis), and the Gini concentration ratio is calculated
from these sets of (xi , yi ) values as before. One significant advantage of this
approach is that the entire population is covered. There is no problem with the
sometimes sizeable population that is deemed not to be living in a locality and
therefore has to be disregarded; all people reside in one or another geographic
subdivision.
We will also deal here with a measure commonly known as the index of
dissimilarity. This is not a measure specifically developed to aid the study of
population distribution, but it can be used for that purpose. The index of dissimilarity
enables two percentage distributions to be compared, and indicates the percentage
of either of the comparator populations who would have to change categories (be
redistributed) for the two distributions to become identical. Thus its value ranges
from zero, where two distributions are already identical, to 100, where all members
of one population are in different categories from all members of the second. The
word ‘population’ is used here in the statistical sense – the percentage distributions
do not have to be distributions of people (although in demographic applications
they mostly will be). They do, however, have to feature the same categories in any
comparison. You cannot compare a percentage distribution by age groups with one
by, say, categories of religion (not that it would make any sense to try to do this),
Table 7.1 Calculation of Lorenz curve coordinates and Gini concentration ratio for Venezuela, 1981
Population Distribution

Cumulative proportion of
Proportion of Localities Pop’n
Size of locality No. of localities Population Localities Pop’n yi xi xi . yiC1 xiC1 . yi
500,000C 3 2,559,517 0.0002 0.1772 0.0002 0.1772 0.0002 0.0001
100,000–499,999 19 3,688,847 0.0010 0.2554 0.0011 0.4326 0.0011 0.0006
50,000–99,999 28 1,905,009 0.0015 0.1319 0.0026 0.5645 0.0030 0.0018
20,000–49,999 53 1,610,324 0.0028 0.1115 0.0054 0.6760 0.0057 0.0039
10,000–19,999 57 776,551 0.0030 0.0538 0.0084 0.7298 0.0095 0.0065
5,000–9,999 89 598,963 0.0046 0.0415 0.0130 0.7713 0.0158 0.0105
2,000–4,999 143 496,010 0.0075 0.0343 0.0205 0.8056 0.0324 0.0173
1,000–1,999 377 569,642 0.0197 0.0394 0.0402 0.8450 0.0658 0.0354
500–999 722 496,523 0.0377 0.0344 0.0779 0.8794 0.1830 0.0726
200–499 2,493 758,630 0.1302 0.0525 0.2081 0.9320 0.9320 0.2081
<200 15,165 982,782 0.7919 0.0680 1.0000 1.0000
Total 19,149 14,442,798 1.0001 0.9999 1.2486 0.3567
Gini concentration ratio D 1.2486  0.3567 D 0.8919
307
308 7 Population Distribution, Urbanization and Migration

and if comparing two age distributions the age groups would need to be the same in
both distributions.
The index of dissimilarity is given by:

ID D 1=2†iD1;k jxi  yi j (7.4)

Where k D the number of categories in each of the percentage distributions being

compared; xi D the percentage in category i in the first distribution; yi D the
percentage in category i in the second distribution.
If the notation which sees the expression xi  yi enclosed by two vertical lines
in Eq. 7.4 is unfamiliar, it means ‘the absolute value of’. This in turn means that
if the expression enclosed turns out to be a negative number, the minus sign is
disregarded and the number is treated as though it was positive.
One application of Eq. 7.4 to the study of population distribution is to compar-
isons over time of percentage distributions of population among geographic areas.
The index of dissimilarity allows the degree of change between two dates to be
quantified on a 0–100 scale (although realistically values are likely to fall towards
the lower end of this range). The index can, however, also be used to measure
population concentration. To do this a country’s, or a region’s, geographic subareas
are the categories, and percentage distributions among them of total land area
and total population are compared. Used in this way the index of dissimilarity is
known as the index of concentration. A value of zero indicates an absolutely even
distribution of population to geographic subareas in proportion to their proportions
of total land area (zero concentration). Higher values indicate the degree to which
population is distributed at above average densities to some areas, and below average
densities to others. In this application the maximum possible value is uncertain, but
is less than 100, since in the limiting case all population would reside in one subarea.
That subarea, however, could not have zero land area, and thus zero percent of total
land area, so that not all the 100 % of population would need redistributing for the
population and land area distributions to match.

Urbanization

The topics of population distribution and urbanization are closely interlinked. While
definitions of urbanization vary, just as definitions of ‘urban’ do, there is little
dispute that both the process and the level of urbanization in a population have
broadly to do with geographic concentration of population. The level of urbanization
can be defined as the percentage of total population living in areas defined
to be ‘urban’. It is thus susceptible to the vagaries of different definitions of
‘urban’, and these should always be assessed, especially when comparisons are
being made. If population size is a key criterion, are urbanization levels for different
populations based on the same minimum size threshold, and are the criteria used to
Urbanization 309

define the boundaries of urban localities comparable? If the definition emphasizes

administrative units defined to be urban, how up to date is the process of conferring
urban status on areas newly deserving of it? To what extent do definitions for
different populations place different emphases on these criteria and on criteria such
as the prevalence of non-agricultural employment and access to ‘urban’ facilities?
Under a definition of urban that emphasizes the sizes of localities’ populations,
the process of urbanization occurs due to (i) additions to the populations of urban
places at rates in excess of additions to the populations of non-urban places, and
(ii) formerly non-urban places attaining the minimum population threshold to be
deemed urban (at a rate faster than formerly urban places slip below that threshold).
These sorts of changes often occur mainly as a result of internal or international
migration, cities being magnets for migrants. But natural increase can also be
a factor, especially in carrying localities past a minimum population threshold.
The physical growth of cities due to migration and natural increase also increases
urbanization by capturing formerly non-urban areas and their populations. Others,
however, see the process of urbanization being more complex, previously non-urban
areas meriting reclassification not just on the basis of population size but on the
basis of a change from largely agricultural to largely non-agricultural employment,
or through the acquisition of ‘urban’ facilities.

Methods of Analysis

The most commonly encountered measure of the level of urbanization is that

indicated above as defining that level – the percentage of total population living
in urban areas. Thus:

pu D .u=t/ : 100 (7.5)

Where u D urban population; t D total population.

Where international comparisons are to be made the percentage of total popula-
tion living in localities of at least some minimum population size (e.g., 20,000) is
perhaps a preferable measure. This really amounts to imposing a uniform definition
of ‘urban’, at least insofar as size of locality is concerned (the criteria used to
establish locality boundaries are another matter), after which Eq. 7.5 can again
be used. Other measures of the level of urbanization have been proposed which
go beyond the simplicity of these two and try to take the sizes or size classes of
urban localities into account (the extent to which urban population is concentrated
in localities of larger size), but none has been widely adopted.
Besides the level of urbanization, analysts often measure the tempo of urban-
ization. One common approach has been to use the compound interest formula to
calculate the average annual rate of urban population growth between two points
310 7 Population Distribution, Urbanization and Migration

in time at which the size of the urban population is known. The compound interest
formula is:

Q2 D Q1 .1 C r/n

Where Q1 and Q2 are quantities at times 1 and 2, n years apart; r D the annual rate
of growth in Q.
If Q1 and Q2 are urban populations n years apart which we denote as u1 and
u2 , this formula can be modified and rearranged to yield the average annual rate of
urban population growth:
p
ru D n .u2 =u1 /  1 (7.6)

Although widely used, the average annual rate of urban population growth has,
for technical purists, drawbacks which disappear if an instantaneous rate of urban
population growth is calculated. The equation for this growth rate derives from the
exponential formula:

Q2 D Q1 : ein

Where Q1 and Q2 are again quantities at times 1 and 2, n years apart; i D the
instantaneous rate of growth in Q.
If Q1 and Q2 are urban populations n years apart which we again denote as
u1 and u2 , the exponential formula can be modified and rearranged to yield the
instantaneous rate of urban population growth:

iu D ln .u2 =u1 / =n (7.7)

In this equation ‘ln’ means ‘the natural logarithm of’. You may be wondering if
there is a link between ru and iu . There is, and in the notation of the compound
interest and exponential formulae it is provided by the relationship r D ei  1. Using
this relationship with r D ru and i D iu , an alternative equation for ru can be obtained:

ru D eln.u2 =u1 /=n  1 (7.8)

If urban population is increasing, the average annual growth rate ru will be

slightly higher than the instantaneous growth rate iu , the difference increasing as the
growth rates increase. If urban population is declining, the negative instantaneous
growth rate iu will be slightly larger (numerically) than the negative average
annual growth rate ru . Either growth rate will, however, serve the purposes of
distinguishing periods of more rapid and less rapid urban growth for a population,
and differentiating populations with more rapid and less rapid urban growth over a
specified period of time.
Average annual and instantaneous rates of urban population growth are not,
however, strictly measures of change in the level of urbanization. Urban population
Urbanization 311

growth need not mean increasing urbanization if rural population is growing as, or
more, rapidly, and even if urbanization is increasing it is unlikely to be doing so
at the same rate as urban population. Thus, four other measures of the tempo of
urbanization are the urban-rural growth difference, the average annual increment
in the level of urbanization, and the average annual and instantaneous rates of
increase in the level of urbanization.
Average annual and instantaneous rates of rural population growth, rr and ir , can
be found using Eq. 7.6 (or Eq. 7.8) and Eq. 7.7, respectively, by substituting the
rural populations at times 1 and 2, r1 and r2 , for u1 and u2 . The urban-rural growth
difference then is given by:

URGD D ru  rr (7.9)

Or:

URGD D iu  ir (7.10)

The average annual increment in the level of urbanization is given by:

iu D .pu;2  pu;1 / =n (7.11)

Where pu,1 and pu,2 D the percentages of population that were urban at times 1 and
2; n D the interval in years between times 1 and 2.
The average annual and instantaneous rates of increase in the level of urbaniza-
tion, riu and iiu , can be found using Eq. 7.6 (or Eq. 7.8) and Eq. 7.7, respectively,
but writing pu,1 and pu,2 (with the same meanings as in Eq. 7.11) for u1 and u2
respectively. Thus:
p
riu D n .pu;2 =pu;1 /  1 (7.12)

Or:

riu D eln .pu;2 =pu;1 /=n  1 (7.13)

And:

iiu D ln .pu;2 =pu;1 / =n (7.14)

It is worth considering briefly each of these measures of the tempo of urban-

ization in the context of the projection of urbanization levels. The rate of urban
population growth is normally higher than the rate of rural population growth, and
therefore also higher than the rate of total population growth. It thus follows that,
projected sufficiently far into the future, constant rates of urban and total population
growth where the former exceeds the latter will eventually result in the absurdity of
312 7 Population Distribution, Urbanization and Migration

an urban population that exceeds total population. Similarly, projecting an annual

increment in the level of urbanization into the future eventually produces the
absurdity of a level of urbanization in excess of 100 %, while projecting an annual
rate of increment in the level of urbanization leads to this absurdity even faster,
because the annually increasing base level to which the rate is applied means an
annually increasing, as opposed to constant, increment in the level of urbanization.
The one measure which does not inherently lead ultimately to an absurd result is the
urban-rural growth difference, although that is not to say that projections based on
it will be accurate.
Similar levels of urbanization can mask different urban structures. Students of
urbanization are also concerned to measure this dimension of a country’s rural-
urban distribution of population. Is urban population concentrated in a single large,
or primate, city, or does it tend to be distributed in a more balanced way among
several large urban agglomerations and a network of smaller cities and towns? The
former pattern tends to be found more often, but not always, in less developed
regions; the latter in more developed countries.
One approach to studying urban structure, or the urban hierarchy, is to examine
absolute and percentage distributions of urban localities and urban population by
locality size classes. Such distributions may be studied cross-nationally or over
time for individual countries to trace differentials in, or the evolution of, urban
hierarchies. The Lorenz curve and Gini concentration ratio outlined in the previous
section are also devices that permit us to focus on a country’s urban structure.
The Venezuelan example presented in Fig. 7.1 is typical of a strongly primate
urban hierarchy, while the Italian example is illustrative of a much more balanced
hierarchy.
Primacy may be measured by the percentage of urban population living in the
largest city. This measure is sometimes also treated in conjunction with a primacy
index or four-city index, which is the ratio of the population of the largest city to
the sum of the populations of the second to fourth largest cities. Thus:

PI D p1 = .p2 C p3 C p4 / (7.15)

Where p1 , : : : , p4 are the populations of the first through fourth largest cities.
A condition of primacy is said to exist if PI exceeds 1.0. Measures of primacy
should always, however, be analysed in conjunction with measures of the level of
urbanization and city size distributions, since very high primacy can coexist with
quite low levels of urbanization (e.g., Conakry in Guinea was in 1990 home to 76 %
of the nation’s urban population, but only 26 % of the population was urban), and
low primacy does not exclude the existence of very large cities (e.g., China in 1990
had a primacy index of only 0.543, but had two cities with populations over 10
million).
Urban structure can also be examined by dividing the urban population into
‘city’ and ‘town’ components based on size of locality, then treating these two
components of urban population in ways analogous to those used to treat the
Urbanization 313

urban and rural components of total population. Thus, the percentage of urban
population living in ‘cities’ becomes a measure of the level of concentration of
urban population analogous to the percentage of total population that is urban (the
level of urbanization). Likewise measures of the tempo of urban concentration can
be constructed which are analogous to the measures of the tempo of urbanization
discussed above; the city-town growth difference, the average annual increment
in the level of concentration of urban population and the average annual or
instantaneous rate of increase in the level of concentration of urban population.
A benchmark sometimes used for the analysis of urban structures is the rank-
size rule. Although observed empirically by Felix Auerbach as early as 1913, it
was fully developed by George Kingsley Zipf in the early 1940s. It states that the
expected size of a city’s population is inversely proportional to its rank in the urban
hierarchy. That is:

pi D k=ri (7.16)

Where pi D the population of urban locality i; ri D the rank of urban locality i in the
urban hierarchy; k D the population of the largest city in the urban hierarchy.
Thus the second-ranked city in an urban hierarchy is expected to have half
the population of the first-ranked, the third-ranked to have one-third of the first-
ranked city’s population, and so on. The rank-size rule can be used to compute
‘expected’ populations of cities in an urban hierarchy for comparison with observed
populations, but is perhaps more effective as a graphical device, enabling departures
from rank-size regularity to be seen when the rank-size distribution for an urban
system is plotted together with that predicted by the rank-size rule. If wishing to
compare empirical rank-size distributions, whether cross-nationally or over time,
a difficulty arises with variations in the size of the largest city, but this is easily
overcome by indexing the sizes of urban places in each hierarchy to a value of, say,
1,000 for the largest city.

Counterurbanization

It is appropriate to end this discussion of urbanization by noting that the process

can occur in reverse. The phenomenon that has come to be known as counterur-
banization was first noted in a study of early 1970s US data (Beale 1975), and has
subsequently been observed in a range of more developed countries, sometimes to
be supplanted again by conventional urbanization. As its name suggests it refers to a
reversal of longstanding trends towards increased population concentration in urban
areas (i.e., ‘deconcentration’) as the lifestyle and economic attractions of living in
more rural surroundings have come to be appreciated and as technological advances
have increased opportunities to work from home. It is not proposed to do more than
mention the concept here, but it is a fascinating one that has rejuvenated some rural
314 7 Population Distribution, Urbanization and Migration

localities that had hitherto been stagnating because of long-term population drift
to urban areas. Readers interested in pursuing it further might start by consulting
Champion (1989).

Migration

In Chap. 1 the population balancing equation for an open population made explicit
the fact that population change is the product of three processes: fertility (births),
mortality (deaths) and migration (arrivals of in-migrants and departures of out-
migrants). The first two of these processes were dealt with in Chaps. 6 and 4,
respectively. We now turn to the third.

Mobility and Migration

Territorial, or spatial, mobility of population occurs at a variety of scales. At

one extreme it is exceedingly temporary and involves movement over very short
distances (e.g., a trip from one’s home to the local shops and return); at the other
it can entail movement halfway round the world, never to return (e.g., a permanent
change of residence from England to Australia). Distance and permanence are two
attributes that tend to make some territorial moves of greater interest than others.
That is not to say that some moves that occur over relatively short distances and/or
are distinctly temporary are not of interest. Local planning agencies, for example,
often are intensely interested in journey to work patterns; the moves mostly occur
over short distances and do not entail changes of place of residence, but their daily
repetition has major implications for the provision of various services and facilities.
Similarly a country, or a region, may have great interest in temporary population
mobility that constitutes tourism. Demographers, however, are most interested in
territorial moves which are sufficiently sustained to be construed as changing a
person’s place of usual residence, and it is moves of this type that qualify as
migration. This definition excludes certain categories of movement that sometimes
are referred to as ‘migration’; for example, nomadism, seasonal movement of
persons who routinely live in two or more locations during the year, and movement
of persons with no fixed place of residence. It is not, however, a clearcut definition in
that there is no universally accepted way of determining when a change of place of
usual residence has occurred. Sometimes an arbitrary minimum actual or intended
duration of stay at a new locality is the criterion used, an approach which, when later
behaviour fails to match stated intention when moving, generates a headache for
analysts known as category jumping. On other occasions respondents’ perceptions
of geographical moves as changing their own, or others’, places of residence are
accepted at face value. Where movement occurs across an international border
the type of travel document a person carries may determine whether a change
Migration 315

of place of residence is considered to have taken place (although people do not

always behave in accordance with entitlements such documents may give them).
Then again, some approaches to the study of migration focus not on individuals,
but on net population aggregates; if, after taking births and deaths into account, the
population of an area has increased or decreased in size compared to some earlier
point in time, a net rise or fall in the number of persons usually residing in that
area due to migration is assumed to have occurred. It is argued as well that very
short distance changes of place of residence (e.g., between apartments within the
same building or houses in the same neighbourhood or town) should not be labelled
‘migrations’, and that migration is better defined as involving a change in locality
of usual residence.

Data and Terminology

The term ‘locality’ arguably is less precise than ‘place’, but the intent in its use is
clear, and in practice its meaning often is dictated by the units of area for which
government statistical agencies provide data. In practice we rarely have data giving
precise points of origin and destination for territorial moves, and therefore rarely
have the capacity to classify moves with absolute precision by distance. Rather we
have data indicating numbers of changes of residence between units of area, and the
larger those units are the greater the proportion of moves that are too short to cross
a unit boundary and thus qualify as ‘migrations’ (in Australia, for example, there
are fewer inter-State migratory moves than there are inter-SA1 migratory moves,
because inter-SA1 moves within States are not inter-State moves). Idiosyncrasies
also arise from the variable shapes of units of area, the closeness to unit boundaries
of points of origin and destination, and the directions of moves from a point of
origin. If area A is long and narrow, moves in the direction of the narrow dimension
are more likely, and those in the direction of the long dimension are less likely, to
qualify as migratory than those originating in area B, which is similar in area but
roughly circular in shape. Second, the closer to an area boundary a point of origin
is the shorter the move needed to cross that boundary and qualify as migratory. And
finally, where a point of origin is not centrally located in an area, moves in certain
directions qualify as migratory when moves of equal distance in other directions do
not. These elements of imprecision and inequality usually have to be accepted as
flaws in migration data about which nothing can be done.
Migration data normally are compiled with reference to a migration interval.
Migration intervals may be fixed (e.g., 1 year, 5 years, or an intercensal period) or
indefinite (e.g., the lifetime of a population). The point is that to measure migration
effectively requires a reference period within which movement has to have taken
place. Lifetime migration is a concept sometimes studied when data relating to a
fixed migration interval are not available. It uses data on places (or localities) of
current residence and of birth, and defines a lifetime migrant as anyone for whom
the two differ. It has very obvious flaws. Lifetime migration may be the net outcome
316 7 Population Distribution, Urbanization and Migration

of several separate migratory moves, and other things being equal, the older a person
is the larger the number of individual moves that may be obscured. A person may
even be a non-migrant in lifetime terms, yet actually have made several moves; it
just happens that at the time data were gathered he/she had returned to live at his/her
birthplace. Moreover, the timing of migration is unknown; even if the product of a
single move, that move may have occurred anything from days before data were
gathered to several decades previously.
Migrants are, of course, persons who engage in migration. They may do so more
than once within a migration interval, so that if border crossings are being counted,
the number of migrations typically exceeds the number of migrants, and by a greater
amount the longer the migration interval. Other types of migration data, however,
measure net movement of an individual over a migration interval (e.g., census data
based on current locality of residence and locality of residence x years ago), in
which case a count of migrants excludes those who moved away from, but later
returned to, a locality of residence, but counts of migrants and of numbers of net
migrations should be equal. Return migrants are persons who, having migrated,
move back to the localities from which they migrated to live at a later date.
The terms origin and destination have already been used, and their meanings
probably are clear. They are the locations (or localities, or areas) where migratory
moves (or net moves) begin and end, respectively; the points/areas of departure and
arrival. A migration stream is the body of migrants who have common areas of
origin and destination. Data for migratory moves among subareas of any defined
larger area can be organized into a matrix of origin-destination pairs, each cell
of which (except those on the leading diagonal, for which origin and destination
are identical and therefore the frequency of migration is zero) corresponds to a
migration stream. These streams form natural pairs in which origin for one is
destination for the other and vice versa. The larger stream in any pair is often
referred to as the stream, and the smaller stream as the counterstream, while
the sum of stream plus counterstream is the gross migration interchange, or
gross migration (sometimes also referred to as the migration turnover), for a
pair of localities. The difference between flows in opposite directions between two
localities, or between one locality and all, or a specified selection of, other localities,
computed as arrivals minus departures, is net migration. It may be positive, when
arrivals exceed departures, or negative, when departures exceed arrivals, and in the
context of any stream and counterstream is of identical magnitude, but opposite sign,
at each of the two localities in question.
A fundamental distinction routinely made in the study of migration is between
international migration and internal migration. The former entails movement
across national borders, and is distinctive because it normally is highly regulated by
governments. This government involvement needs to be clearly understood in order
to interpret migration flows. It also establishes administrative mechanisms (border
control procedures and requirements for passports, visas, work permits, etc.) which
can be important sources of data, although these can be of varying degrees of
completeness and reliability depending on the scope and incentive for avoiding the
mechanisms and for behaving contrary to intentions expressed when interrogated by
Migration 317

them. Other sources of data include aircraft and ship passenger manifests, census or
survey data which establish residence in another country at birth or at a specified
prior date, and occasionally population registers. Internal migration, in contrast,
typically is a much freer category of population movement, with individuals and
households responding to changing economic, social, environmental and life cycle
circumstances as they see fit. A lack of regulation of movement generally means a
lack of data comparable to border control data, and hence the need to rely on other
data sources. Sometimes administrative systems which record changes of address
for reasons other than keeping track of people’s movements per se (e.g., in Australia,
the medicare system) are potential sources of internal migration data. Whether
they can be accessed for this purpose is, however, another matter, and population
censuses are the major source of data on internal migration, followed by sample
surveys and population registers. The former two sources gather retrospective
information on migration. Censuses do this either directly by asking questions about
places of birth, residence at specific dates prior to the census, current residence,
etc., or indirectly through techniques that employ survival ratios and data on age-
sex composition and births (see Chap. 4). Sample surveys may also ask census-type
questions, but in addition may gather individual migration histories, which date and
give the origin and destination of each move. Population registers, on the other hand,
record changes of place of residence at an individual level on an ongoing basis.
Though of considerable potential, however, they are not widely available and their
potential is largely unrealized.
There are exceptions to the generalization that internal migration is ‘a much freer
category of population movement’, and they tend to be associated with totalitarian
governments. In China, for example, the hukou system of household registration
used to determine quite rigidly where one could live and work, and still has influence
on, in particular, rural-urban migration. Under this system one is registered at one’s
hukou place and must apply to change that registration to obtain legal right of
residency and, more importantly, access to state-subsidized welfare covering food,
housing, education and permanent employment at a desired migration destination.
The system is a de facto internal passport mechanism, and while approvals to change
one’s hukou place are easier to obtain in the post-economic reform era than they
were in the pre-reform era, even today it can be difficult for rural peasants to gain
approval to move to medium-sized and large cities.
The distinction between internal and international migration is matched by
distinct terms for arrivals and departures. Internal arrivals are known as in-migrants
and departures as out-migrants, while the totalities of internal movements into
and out of any given unit of area are known as in-migration and out-migration,
respectively. International arrivals and departures are referred to as immigrants
and emigrants, respectively, the totalities of movements into and out of a country
being immigration and emigration. It does happen that measures of migration
for localities within a country combine (or are unable to separate) internal and
international movement, or that a discussion of migration might be referring to
either category of migration (e.g., the discussion of the I and O elements in
318 7 Population Distribution, Urbanization and Migration

the population balancing equation for an open population in Chap. 1). In such
circumstances the terminology for internal migration prevails.

Demographic Perspectives on Migration

Donald Bogue (1993) has argued that demographers (and those with demographic
interests in kindred disciplines – geography, economics and anthropology) study the
territorial mobility of population from four perspectives. First, as a component of
population change, not only in the balancing equation sense of its being a factor
in population growth or decline, but also in the sense of its altering population
composition through selective movement by persons with particular characteristics.
Second, as a mechanism for socioeconomic adaptation; a means by which individ-
uals, households and even entire communities adjust to changing social, economic,
environmental and political realities which tend to ‘push’ them from one locality and
‘pull’ them towards another. Third, as a routine life-course event associated with
such things as the attainment of adulthood (departure from the parental home to
marry, attend institutions of higher learning, follow desired careers, or just establish
independence), family formation (the desire to obtain housing perceived suited to
the arrival, numbers and ages of children), employment transitions (job transfers,
upward mobility in the labour market, and lapses into periods of unemployment),
and retirement. Finally, demographers examine territorial mobility as rational
entrepreneurship. The focus here is on households or families as small businesses
or enterprises seeking to maximise their wellbeing, an activity which may involve
decisions to move or send individual members to other localities to improve the
return on human and other capital.

Measuring International Migration

Administrative data on international migration are for many countries non-existent

or of poor quality. There are also immense problems for international comparability
of data in countries’ different ways of defining and classifying migrants. Defining a
migrant is less straightforward than defining a birth or a death, and even superficially
similar approaches to the definition and classification exercise may mask subtle
differences that result in quite varied estimates of the sizes of ostensibly the
same migration flows. This was clearly demonstrated, for example, in an analysis
of Australian and New Zealand data on migration between the two countries
(Carmichael 1993), and these are countries with supposedly relatively high quality
international migration data. One approach to filling gaps in administrative data on
international migration, or to attempting to improve on data of suspect quality, is to
seek out data gathered at the other end of migration streams involving the country
Migration 319

of interest, although combining data from several sources obviously brings issues of
comparability into sharp focus.
While in Chap. 1 we noted that absolute measures were of limited value in
demography other than as input into relative measures, they perhaps have more of
a place in analyses of international migration than in any other area of demographic
analysis. There is interest in the sheer volumes of immigration, emigration and
net migration flows, not only overall, but in particular streams and migration
subcategories (family reunion, businesspersons, etc.).

Indirect Estimation of Net Migration

Where counts of immigrants and emigrants are not available, the population
balancing equation in the form presented as Eq. 1.5 in Chap. 1 can be used to
estimate net migration to a country between two censuses held at times 1 and 2.
To recap:

M D .P2  P1 /  .B  D/

Where P1 and P2 D the total populations at times 1 and 2; B D births and D D deaths
between times 1 and 2.
Estimates of intercensal net international migration by age and sex can be
made using another methodology already discussed – forward survival. Known as
the intercensal cohort-component method of estimating net migration, this was
discussed for age groups at a second census which were already alive at an earlier
first census in Chap. 4 (see Eqs. 4.75, 4.76, 4.77 and 4.78). The strategy is to forward
survive age groups (or age cohorts – hence the technique’s name) at a census held at
time 1 to obtain their expected sizes assuming zero net migration at a second census
held at time 2, n years later (when they form age groups n years older). Expected
cohort sizes then are subtracted from those observed at the second census. Resulting
differences are the cohort net population gains or losses due to migration over the
intercensal period, and when the method is applied to national-level data these are
net gains or losses due to international migration (applied at the subnational level
net migration estimates produced are net gains or losses due to both international
and internal migration). Recall that because survival chances between any two
age groups of a given width differ for males and females, net migration estimates
typically are computed separately for the two sexes.
In outlining the intercensal cohort-component method in Chap. 4, separate
equations were presented for (i) estimating expected cohort sizes at time 2 and (ii)
subtracting these from observed cohort sizes to yield net migration. Separate pairs
of equations also were presented to cover single-year and t-year age cohorts. This
was done to break the procedure down and aid understanding of it, but it can be
generalized in a single, much simpler looking equation as follows:
320 7 Population Distribution, Urbanization and Migration

NM D P2;a  s:P1;an (7.17)

Where P2,a D the population in age group a, the age group defining the age cohort, at
the second census (time 2); P1,an D the population in the corresponding n-year
younger age group at the first census (time 1); n D the length of the intercensal
period in years; s D the life table survival ratio from age group a-n to age group a.
If age group a begins at exact age x and is of width t years, then this survival ratio
will be given by s Dt Lx /t Lxn , these quantities being extracted from a relevant
life table.
Equation 7.17 does not, however, deal with any younger age cohort at time 2
which was not alive at time 1, and thus is the product of births during the intercensal
period. Net migration for any such cohort must be estimated by forward surviving
the relevant number of births to obtain the expected size of the cohort at the second
census date, then subtracting this expected size from the observed cohort size at
the census. This procedure, too, was covered in Chap. 4 (Eqs. 4.79 and 4.80 in
combination with Eqs. 4.76 and 4.78, respectively, for the single-year and t-year
cohort situations), but again a single-equation summary is possible:

NM D P2;a  s:B (7.18)

Where P2,a D the population in age group a, the age group defining the age cohort, at
the second census (time 2); B D the number of live births during the period over
which the cohort was born (which will be all or part of the intercensal period);
s D the life table survival ratio from birth to age group a. If age group a begins
at exact age x and is of width t years, then this survival ratio will be given by
s Dt Lx / (t . l0 ), where t Lx and l0 come from a relevant life table. The period over
which the cohort was born will be the t-year period commenced the date of the
second census in year y-x-t, where y is the year in which the second census was
held.
The easiest way to carry out this second part of an intercensal cohort-component
estimation of net migration (which again typically is conducted separately for males
and females) is to define a single cohort covering all ages at the second census which
are the product of intercensal births. That way B in Eq. 7.18 becomes all births (of a
given sex) during the intercensal period. However, if the intercensal period is, say, of
duration 10 years, and migration estimates based on Eq. 7.17 have been calculated
for 5-year age cohorts, we may wish to treat those born during the intercensal period
as two 5-year cohorts (aged 0-4 and 5–9 at the second census) instead of one 10-year
cohort (aged 0-9).
The intercensal cohort-component method as outlined does tend to overstate or
understate the implied number of deaths over an intercensal period, depending on
whether a country, or a cohort, is an ‘emigration’ or an ‘immigration’ country or
cohort. In the ‘emigration’ case the size of a cohort at the first census overstates,
and in the ‘immigration’ case it understates, the average population at risk of
Migration 321

dying during the intercensal period. Should this problem be judged to be severe
one can obtain two estimates of net migration for each age cohort, one using
standard forward survival from the first census (Eq. 7.17) or from intercensal births
(Eq. 7.18), and the other using reverse survival from the second census, then average
these estimates. The reverse survival companion to Eq. 7.17 is:

NM D P2;a =s  P1;an (7.19)

Where all elements have the same meaning as in Eq. 7.17.

The reverse survival companion to Eq. 7.18 is:

NM D P2;a =s  B (7.20)

Where all elements have the same meaning as in Eq. 7.18.

Census-based studies of net international migration for countries which are net
recipients of international migrants may also be carried out by focusing on the
overseas-born. Such studies do not, of course, reveal anything about migration of
the native-born, but a modified version of the population balancing equation gives
net intercensal migration of the overseas-born as:

Mo D Po;2  Po;1 C Do (7.21)

Where Po,1 and Po,2 D the overseas-born populations at censuses held at times 1 and
2; Do D deaths of overseas-born persons during the intercensal period.
There is no births element in this equation because, by definition, no births
of overseas-born occur in a country. The equation is capable of being applied
to overseas-born from particular source countries or regions as well as to total
overseas-born, but does rely on having available data on deaths by birthplace. If
not directly available, these can be estimated by applying total population age-
sex-specific death rates at the midpoint of the intercensal period to means of the
numbers of overseas-born in each sex-age group at the two censuses, summing to
obtain an estimate of annual deaths, then multiplying by the length of the intercensal
period in years. This procedure assumes similar mortality conditions among the
total and overseas-born populations, and should not be used if the validity of such
an assumption is in serious doubt.
Estimates of intercensal net migration of the overseas-born, or of subgroups
thereof defined by birthplace, by age cohorts can also be made using the intercensal
cohort-component procedures discussed above, assuming the availability of suitable
life tables. ‘P’ quantities in the respective equations simply become numbers of
overseas-born (or persons born in specified countries or regions) instead of numbers
of total population. The interesting part of such a cohort-component exercise is that
concerning migration in age cohorts born during the intercensal period. Superficially
it might seem that no calculations are required; no persons in these age cohorts were
322 7 Population Distribution, Urbanization and Migration

native-born, and hence since all were born overseas during the intercensal period,
all must have arrived as immigrants during that period. Net migration might seem
to be simply the numbers of overseas-born enumerated at the second census, and
certainly Eq. 7.18 would suggest this, since B D 0.
We have not, however, allowed that some immigrants during the intercensal
period may have died before the second census. To take this possibility into account
the reverse survival procedure using Eq. 7.20 can be used, with B D 0. This,
however, effectively assumes that all mortality associated with immigrants born
during the intercensal period occurred in the country of destination, which amounts
to assuming that all migrated immediately they were born. This is not a plausible
assumption; in reality migration will have occurred progressively between birth and
the second census date, and in effect some who would have migrated will not have
done so because they died in their countries of birth before being able to migrate.
One compromise would be to average the net migration estimate(s) yielded by
Eq. 7.20 and the observed cohort size(s) at the second census, effectively assuming
that half the mortality occurred after migration. But this is still likely to overestimate
migration because childhood mortality is concentrated in the first year of life, and
within that year very early in the first year of life. This makes it probable that well
under half the mortality component added to net migration by reverse survival post-
dated migration, with well over half of it preventing migration. In other words,
the observed sizes at the second census of overseas-born cohorts born during the
intercensal period might not be all that inaccurate as estimates of net migration after
all! If precision is desirable, the distribution of mortality by age between birth and
cohorts’ ages at the second census should be taken into account in apportioning
the mortality component added by reverse survival between that likely to have
prevented, and that likely to have post-dated, migration (assuming migration to have
been evenly distributed between birth and the second census date).
It is sometimes claimed that intercensal cohort-component procedures for age
cohorts alive at both censuses can use census survival ratios rather than life table
survival ratios. The former are calculated by dividing the number of persons in a sex-
age (or age) group at a second census by the number of persons in the corresponding
younger sex-age (age) group at an earlier census. For age cohorts born during
the intercensal period, equivalent survival ratios divide numbers in an age cohort
at the second census by the number of intercensal births over the period during
which the age cohort was created. One advantage of this approach is claimed to be
that intercensal differences in completeness of enumeration are incorporated in the
survival ratios instead of contaminating net migration estimates. The method is of no
use, however, where cohort-component procedures are applied to data for the total
population, since census survival ratios assume a closed population and therefore
yield zero net migration estimates. It may have some validity in application to data
for the overseas-born, in which case the census survival ratios would be based on
data for the native-born. The assumptions then are (i) that native-born mortality
patterns apply to the overseas-born population (and to subgroups thereof defined by
birthplace if they are the focus of analysis), and (ii) that the native-born population
Migration 323

is closed, so that intercensal changes in the sizes of age cohorts are the product of
mortality alone. Neither assumption may be especially accurate.

Measures Using Migration Flow Data

Few techniques have been specifically developed for analysing immigration and
emigration flow data. Several techniques are shared with studies of internal
migration, and will be dealt with under that heading. Data on population flows
across international borders should always be classified by type (temporariness or
permanence) of movement, one of the functions of this classification being to (rather
arbitrarily) distinguish immigrants and emigrants from other types of movers. Then,
for the two migrant categories of movers information on countries of last and
next ‘permanent’ residence, respectively, should be gathered to facilitate analysis
of migration streams and counterstreams. Data on country of birth and citizenship
may serve as substitutes for, or supplements to, such information. Border crossing
databases can also usefully include information on reasons for moving (including
data on visa categories for inward flows), and on various demographic, social and
economic characteristics of movers in general and migrants in particular (since
both international migration and non-migratory international movement tend to be
selective by age, labour force characteristics, etc.).
Special difficulties attach to the construction of migration rates and ratios, which
often, though not always, make use of flow data. These difficulties emanate chiefly
from the fact that each migration involves both a place of origin and a place of
destination. While emigration (and out-migration) can be regarded as ‘risks’ for
the populations of countries (or localities) of origin in the same sense as dying
and giving birth are risks, immigration (and in-migration) are not risks for the
populations of countries (or localities) of destination, since they involve people
external to those countries/localities. Then again, besides being able to assume either
positive or negative values, measures of net migration, as residuals after subtracting
gross emigration from gross immigration, do not conform to the definition of the
numerator of a demographic rate. Whether a person experiences the event of birth,
death, marriage, emigration, immigration, etc. is clearcut (although in the case of
immigration, as just noted, there is a problem defining an appropriate denominator
for a demographic rate), but one cannot say that a person experiences the event of
net emigration or net immigration.
Also an issue in the measurement of international migration flows is the fact
that emigration and immigration are renewable events. Thus it is important to know
whether numbers of migrations or of migrants are being counted. Typically, as
noted earlier, the former are counted, and the obvious potential for double counting
migrants in any migration interval poses further difficulty for notions of being able
to calculate migration rates.
Studies of international migration combine immigration (I), emigration (E), gross
migration (I C E) and net migration (I  E) in a range of ratios that can reveal
interesting features of and trends in migration flows. The ratio E / I is a measure of
324 7 Population Distribution, Urbanization and Migration

the extent to which immigration is counterbalanced by emigration; a value greater

than 1.0 indicates net emigration, while the further below 1.0 the index gets the more
dominant immigration is. Ratios of net immigration to total immigration ((I  E) / I;
where I > E) and net emigration to total emigration ((E  I) / E; where E > I) give the
proportions of total immigration and total emigration effectively added to and lost
from a population in situations of net immigration and net emigration, respectively.
Calculations of this type allow cognisance to be taken of the reality that similar net
migration flows can be underpinned by very different total flows in the direction of
net movement. Sometimes the bulk of a total flow is converted to net flow because
there is little counterflow, whereas on other occasions a large counterflow means
relatively little of a total flow becomes net flow. This type of measure can be very
effectively applied with respect to individual migration streams as well as with
respect to total flows into and out of a country.
The ratios I / (I C E) and E / (I C E) measure the relative contributions of immi-
gration and emigration to gross migration. One of the more useful and widely used
ratios, however, is the migration effectiveness ratio (MER), defined as the ratio of
net to gross migration or:

MER D Œ.I  E/ = .I C E/ : 100 (7.22)

The MER measures the percentage of gross migration that is converted to population
loss (a negative MER) or gain (a positive MER). Its value ranges from 100 where
gross migration is entirely emigration to C100 where it is entirely immigration.
Once again, it is not a measure applicable only to entire migration flows into and
out of a country. It can be calculated for individual migration streams, and even for
substreams defined by, for example, occupation.
While it was indicated above that several problems attend the construction of
migration rates, there are a number of measures in use in the study of international
migration that are called ‘rates’, despite not satisfying the definition of such
measures presented in Chap. 1. The crude rates of immigration, emigration, net
migration and gross migration for a population are given, respectively, by:

CIR D .I=P/ : 1;000 (7.23)

CER D .E=P/ : 1;000 (7.24)

CNMR D Œ.I  E/ =P : 1;000 (7.25)

CGMR D Œ.I C E/ =P : 1;000 (7.26)

Where I D immigration during a calendar year; E D emigration during that year;

P D the mean or mid-year total population in that year.
Migration 325

These are measures of the extent to which a country’s population was augmented
by immigration, depleted by emigration, and augmented or depleted by the net
effect of both processes, and of the extent of international migration turnover in
a population.
It is also possible to calculate a variety of specific ‘rates’ of international
migration, the most common and useful probably being specific rates of net
migration. For any population subgroup s (which might be an age group, a sex-
age group, a birthplace group, etc.) the subgroup-specific rate of net migration is
given by:

SSRNM D Œ.Is  Es / =Ps  : 1;000 (7.27)

Where Is D immigration into the subgroup during a calendar year; Es D emigration

from the subgroup during that year; Ps D the mean or mid-year population in the
subgroup.
The ‘rates’ calculated using Eqs. 7.23, 7.24, 7.25, 7.26 and 7.27 have been
presented as annual rates. If data on immigration and emigration are only available
for some migration interval longer than a year, the same equations can be used,
with the denominator P, or Ps , becoming the relevant population at the mid-point
of the migration interval. Thus we derive n-year rates, and if we wish to compare
migration intervals all of which are of the same length no further modification is
needed. If, however, migration intervals of varying length need to be compared,
average annual rates can be computed by dividing n-year rates by n (although this
procedure can be problematic where migration is cyclical and a shorter migration
interval might include only an up-cycle or a down-cycle, while a longer interval
might include both).
Equations 7.25 and 7.27 may also be used to calculate rates of net migration
using census-derived estimates of net migration. In this application the (I  E)
numerators are obtained as single values using the population balancing equation or
the intercensal cohort-component method. When Eq. 7.27 is applied to age cohorts
using such data, it normally is applied only in respect of cohorts alive at both
censuses defining the migration interval. Moreover, sometimes no attempt is made
to obtain denominators which are mid-interval cohort sizes. Rather, the size of the
cohort at the second census is used. Given that the ‘rates’ being calculated are not
really rates at all, such latitude is not a major problem. The relative impacts of
migration on different cohorts over a migration interval are adequately revealed.

International Migration and Population Growth

International migration is, of course, also of interest as a factor in national

population growth, as the population balancing equation made clear in Chap. 1. It
was noted that total growth over any period (P2  P1 ) is the sum of net (international)
migration (M) and natural increase (B  D). Various ratios of these quantities are
326 7 Population Distribution, Urbanization and Migration

commonly used in analyses of population growth to index the relative importance

of migration and natural increase in growth. The percentages of total growth due to
natural increase and to net migration are given by:

GNI D Œ.B  D/ = .P2  P1 / : 100 (7.28)

And:

GNM D ŒM= .P2  P1 / : 100 (7.29)

And the ratio of net migration to natural increase, which we might call the
components of growth ratio is given by:

CGR D ŒM= .B  D/ : 100 (7.30)

Assuming positive natural increase, a positive value of the CGR indicates the
percentage by which net migration augmented growth due to natural increase,
while a negative value indicates the percentage by which it offset growth due to
natural increase. If natural increase is negative (i.e., natural decrease), as it has been
in recent years in Japan and a number of European countries (e.g., in 2008–11,
Belarus, Bosnia-Herzegovina, Bulgaria, Croatia, Estonia, Germany, Hungary, Italy,
Latvia, Lithuania, Portugal, Moldova, Romania, the Russian Federation, Serbia and
Ukraine), a positive CGR indicates the percentage by which net migration increased
population loss due to natural decrease, and a negative one indicates the percentage
by which it offset loss due to natural decrease.
Assessments of the contribution of migration to population growth often also
seek to take account not only of migration per se, but of that element in natural
increase that is attributable to migrants and perhaps also to their descendants. In
the broadest view this type of assessment perhaps becomes rather ridiculous for a
country like Australia, whose non-Indigenous population is entirely descended from
migrants. At the other extreme the narrowest approach is to adjust net migration
during a period for the natural increase for which migrants during that period are
responsible. This can be done by forward surviving the population at a first census,
plus births to that population during the intercensal period, to a second census
date, then subtracting the resultant estimate of survivors from the second census
population count. This becomes the migration contribution to intercensal growth,
and the difference between it and total growth is the contribution of natural increase
among the population present at the first census. The tricky part of this exercise
is the estimation of intercensal births, which needs somehow to exclude births to
migrant arrivals during the intercensal period. Otherwise the method becomes a
straightforward series of intercensal cohort-component calculations.
In between these extremes is the type of exercise undertaken for Australia over
many years by C.A. Price, who sought to monitor the contribution to population
growth of post-Second World War migrants and their descendants. Thus he took
the 1947 Census population (itself with a significant immigrant, not to mention
Migration 327

Fig. 7.2 Age-sex pyramid for Australia showing components of population added by post-Second
World War immigration, 1993 (Source: Price 1996: 67–68)

immigrant-descended, element) as a base, thereafter progressively identifying, as

they became features of the demographic landscape, (i) the post-war immigrant
population, (ii) children of those immigrants (including half the children with
one immigrant and one non-immigrant parent), and (iii) the third generation as
components of the immigrant contribution to population growth. The methodology
he used is complicated, summarized in Price (1996), and beyond being detailed here,
but it produced the 1993 age pyramid shown in Fig. 7.2. You can probably imagine
how, using data from which diagrams like Fig. 7.2 were constructed at regular
intervals, population growth over those intervals could be apportioned by simple
subtraction between that due to the 1947 Census population and its descendants,
and that due to post-1947 immigrants and their descendants.

Measuring Internal Migration

Operationally, internal migration normally takes place when a person changes

his/her place of residence from one statistical or civil division within a country
to another. This definition tends to presuppose censuses as the source of data on
internal migration, and they are indeed the major source. They yield data on internal
migration either through direct questions on place of residence, place of birth,
328 7 Population Distribution, Urbanization and Migration

duration of current residence, place of previous residence and place of residence

at specified past dates, or indirectly through comparisons of population counts at
successive censuses in conjunction with data on intercensal births and deaths, or
through providing age distributions from which net migration estimates can be
derived with the aid of survival ratios.

Census Migration Questions

It was noted earlier that place of birth data, in conjunction with place of current
residence data, facilitate the study of lifetime migration. The shortcomings of this
concept also were discussed, but in the absence of better data, and ideally at a fairly
broad level of geographic disaggregation, it provides some sort of indication of
major population movements within and into (though not out of) a country over
preceding decades. Internal lifetime migrants are those for whom place of birth is
a locality within the country being studied which is different from the locality of
current residence. International lifetime migrants are those whose birthplaces are
external to the country being studied. With respect to internal movement, migration
streams and counterstreams may be identified by focusing on place of birth/place
of current residence pairs, always recognizing, of course, that at the individual level
paths travelled between the two may have been indirect, and that migration patterns
emerging may not be particularly representative of the recent past. One way of
tackling the latter problem to a degree is to examine lifetime migration patterns for
age cohorts. Very young cohorts could only have migrated relatively recently, and
cohorts in the prime migration age groups (late teens to early thirties, say) are also
worth focusing upon as likely to have migrated recently if they have migrated at all.
Place of birth data for a common set of spatial units from consecutive censuses
can also be used to make indirect estimates of intercensal net migration for each
unit. The relevant equation is:

NM D .I2  O2 /  .si I1  so O1 / (7.31)

Where I1 and I2 D lifetime in-migrants to an area at the first and second censuses;
O1 and O2 D lifetime out-migrants from that area at the two censuses; si and
so D survival ratios applicable to the lifetime in-migrants and out-migrants,
respectively, present at the first census over the intercensal period.
The main challenge in using Eq. 7.31 is to obtain satisfactory values for the two
survival ratios. Data needed to calculate accurate values are rarely available, and
so approximations of varying degrees of complexity are used. Perhaps the simplest
approach is to set both survival ratios equal to the ratio total population aged n years
and older at the second census divided by total population at the first census (or to
the life table survival ratio Tn / T0 if an appropriate life table is available), where n
is the length of the intercensal period. Other more elaborate procedures are outlined
in United Nations (1970: 8–12).
Migration 329

Data on duration of current residence identify as migrants anyone for whom that
duration is less than his/her current age. This approach has the capacity to identify
as migrants persons who moved away from, then later returned to, their place (or
locality) of birth, thereby providing a more accurate count of lifetime migrants. It
also facilitates the identification of migration cohorts, defined in terms of the timing
of the most recent migratory move, and this allows more of a focus on the recent
migration history of an area. These cohorts are, however, survivors of those who
actually migrated during the periods defining them; survivors both in the sense of
not having subsequently died and in the sense of not having subsequently migrated
again. Other shortcomings are that, depending on the wording of the census question
and the way respondents interpret it, moves which change a person’s address
without changing his/her locality of residence may be accorded the status of
migrations; that there is no information on place, or locality of origin; and that
therefore there is no capacity to focus on out-migration, net migration, or migration
streams, or to distinguish internal from international migrants. Duration of residence
data from successive censuses can, though, be used to examine remigration, by
surviving a duration of residence group from the first census to the second and
comparing expected with observed survivors in the corresponding longer duration
of residence group at the second census.
Many of the limitations for migration analysis of data on duration of current
residence can be overcome if a question on most recent place of previous residence
is also asked. This adds the missing information on localities of origin, and thus
the ability to examine concepts listed in the previous paragraph. It also improves on
place of birth data because it provides information on direct moves, rather than on
moves which may be direct or very indirect. Analysed in isolation, however, most
recent place of previous residence data share with place of birth data the problem
that migrations may have taken place anything from within weeks of the census to
many years previously.
In many respects the preferred census data for studying internal migration are
data on place of residence at a specified past date. Combined with data on current
place of residence they provide a clear migration interval, often 1 year or 5 years, and
where this coincides with the length of the intercensal period, successive censuses
allow a continuous picture of internal migration to be built up. Census authorities
also like the fact that one question classifies persons into migrants and non-migrants,
provides a migration interval, and also yields the places of origin of migrants, in
contrast to the two questions the duration of current residence/most recent place of
previous residence approach requires. The method does revive problems associated
with lifetime migration, in that if several moves take place within the migration
interval only their net effect is measured, and if that net effect is a return to the
place/locality of origin a person is deemed a non-migrant. But the strictly limited
time frame within which these problems apply makes them far less serious. The
capacity of respondents to recall accurately their places of residence at an arbitrary
past date, as compared to their capacity to recall where they last moved from, has
been questioned, and as with all census migration questions the tendency of persons
to respond on behalf of others (e.g., a household head filling out census forms for
330 7 Population Distribution, Urbanization and Migration

all members of the household) can affect data quality through lack of knowledge
of other household members’ migration histories. But the place of residence at a
specified past date approach does provide measures of in-, out- and net migration,
for well-defined areas and migration streams, and over a well-defined migration
interval.

Indirect Estimation of Net Migration

One method of indirectly estimating net migration already has been covered; that
which uses Eq. 7.31 and is based on place of birth data for a common set of
spatial units from consecutive censuses. Where reliable birth and death statistics
are available for statistical or civil geographical divisions within a country on a
place of residence basis, net migration for those divisions can also be estimated
indirectly using the population balancing equation. Sometimes called the vital
statistics method of estimating net migration, this approach can be applied either to
the total population of a division, or to subpopulations defined by attributes which
do not change (sex, birthplace, etc.) or which change predictably (age), provided
births and deaths are available classified by those attributes. In the case of estimating
net migration for an age cohort, the balancing equation M D (P2  P1 )  (B  D)
becomes, for an age cohort initially aged a, and an n-year intercensal period:

Ma D .P2;aCn  P1;a / C Da (7.32)

Where P2,aCn D the population in the age group n years older than age group a at
time 2; P1,a D the population in age group a at time 1; Da D deaths between times
1 and 2 in the cohort aged a at time 1. This number of deaths would need to be
estimated from data on deaths by calendar year and single years of age using
Lexis diagram principles.
Note that in Eq. 7.32 there is no reference to births (B), since intercensal births
cannot possibly add to a cohort aged a (and therefore already born) at time 1.
It is important to appreciate that net migration estimates based on the population
balancing equation are susceptible to error from two sources. First, if levels of
underenumeration at the two censuses defining the migration interval are different,
that difference is incorporated into the net migration estimate. So, for example, an
improvement in census coverage (i.e., a reduction in underenumeration) is treated as
a net migration gain. This can happen either overall, or in the context of individual
age cohorts (e.g., as they age from their teens or early twenties, when their lifestyles
might render them easily missed by census enumerators, to ages at which more
settled lifestyles reduce the risk of this happening). The second source of error
is the vital statistics used. In developing countries in particular, vital data often
are of poor quality, and errors they contain are reflected in migration estimates
(although to some extent undercounts of births and deaths may cancel out). Another
point about net migration estimates calculated for subareas of a country using
Migration 331

the vital statistics method was alluded to earlier; if a population is not closed,
estimates reflect the combined effects of internal and international migration, and
are not pure indicators of internal migration. One partial way around this problem,
assuming availability of the census and vital data by birthplace, is to apply the
balancing equation to the native-born population only. This approach presupposes
little international migration by the native-born, and if this is not the case the
problem remains.
The intercensal cohort-component method of estimating net migration can also
be applied in respect of statistical or civil divisions within a country to derive net
migration estimates for those divisions. As when measuring international migration,
two estimates can be made for any age cohort, one using forward survival and
the other using reverse survival (Eqs. 7.17 and 7.19 for cohorts alive at the initial
census, and Eqs. 7.18 and 7.20 for those born during the intercensal period).
These approaches respectively assume that all migration occurred at the end of
the migration interval, effectively failing to allow for deaths among migrants after
arrival, and at the beginning of the migration interval, effectively overestimating
such deaths, so that a common ploy is to average the results they yield. This gives
a result that assumes that migrations and deaths were evenly distributed over the
migration interval, or alternatively that all migrations occurred in the middle of
the interval. Thus, if NM1 and NM2 are the net migration estimates yielded for a
cohort by forward survival (Eq. 7.17 or Eq. 7.18) and reverse survival (Eq. 7.19 or
Eq. 7.20), respectively, we have:

NM D .NM1 C NM2 / =2 (7.33)

We do not, however, need to evaluate NM2 to obtain this average. It can be shown
that:

NM D Œ.1 C s/ =2s : NM1 (7.34)

Thus we can take a net migration estimate obtained by forward survival using
Eq. 7.17 or Eq. 7.18 and adjust it using the survival ratio, s, from that equation
to yield the average of forward and reverse survival estimates. Note that Eq. 7.34
works equally well whether the age cohort in question was alive at the first census
(when Eq. 7.17 is used) or born during the intercensal period (when Eq. 7.18 is
used). The survival ratio, s, must, however, be the one from whichever of Eq. 7.17
or Eq. 7.18 was used to find NM1 .
When the intercensal cohort-component method is applied to the estimation of
net migration for statistical or civil divisions within a country, the use of census
survival ratios (CSRs) as an alternative to, or in preference to, life table survival
ratios (LTSRs) in Eqs. 7.17, 7.19 and 7.34 becomes more of an issue. Recall that
a CSR is the number of persons in a sex-age (age) group at a second census
divided by the number in the corresponding younger sex-age (age) group at an
earlier census. LTSRs for use with the intercensal cohort-component method can
be obtained from male and female life tables for the area in question around the
332 7 Population Distribution, Urbanization and Migration

midpoint of the intercensal period (averages of LTSRs from tables for the endpoints
of the period can also be used), or from appropriate model life tables. Their use
is questionable, however, where census age data are of dubious quality, and
age distributions consequently are irregular; the combination of an irregular age
distribution with a smooth set of LTSRs tends to produce distorted net migration
estimates which do not sum over all areas within a country to zero. The latter
situation is likely to stem partly from the population not being closed, so that
net migration estimates incorporate international as well as internal migration, and
it would be possible to smooth the census age distributions prior to application
of cohort-component procedures. But the use of CSRs has the advantage that
irregularities in age data and the effects of international migration tend to be
absorbed into the survival ratios themselves, rather than accruing entirely to the net
migration estimates. This can result in CSRs which fluctuate rather than following
the smooth trend with age characteristic of LTSRs. It can even result in CSRs above
unity, which clearly are not accurate measures of survivorship. But the purpose is to
measure net migration, not survivorship, and where internal migration is concerned,
CSRs serve that purpose quite well.
Normal practice is to compute national CSRs, then use these in applying
Eqs. 7.17, 7.19 and 7.34 to population data for subnational areas. One intuitively
attractive feature of this approach is that, provided calculations are made for areas
which together cover the entire country, net migration estimates for all areas sum
to zero. The use of CSRs is not, however, without problems. First, although it
has been stated that CSRs absorb the effects of international migration rather than
have them affect the migration estimates themselves, this is only partly true. The
CSR method works best for a closed population. While for an open population the
effects of net international migration are absorbed into the CSRs, the subsequent
application of these CSRs uniformly across a country assumes an even spatial
impact of international migration, which is unlikely to coincide with reality. In
a net immigration country, for example, areas (like major cities) which receive
disproportionately large shares of net immigration are allocated less than their shares
by this process, effectively converting some international migration gain to internal
migration gain; other areas are allocated more than their shares of international
migration gain, effectively lowering internal gain below its true level. One approach
to overcoming this problem is to attempt to adjust the second census population for
intercensal international migration, thus creating approximate intercensal closure
before calculating CSRs. This, however, requires quality administrative data on
international migration, which often are not available.
The second problem with using CSRs is the assumption that mortality conditions
are uniform across statistical or civil divisions within a country. Especially in
countries with high mortality this is unlikely to be the case, and some correction of
net migration estimates for regional differences in mortality may be called for. Lack
of data on such differences is often a barrier to making these corrections, but if one
has a basic idea of regional variations in expectations of life at birth it is possible
to use ratios of survival ratios in model life tables applicable to those regions to
Migration 333

corresponding survival ratios in a model table applicable to the country as a whole

as a basis for adjusting CSRs before application to regional population data.
A third problem with using CSRs to estimate internal migration via the inter-
censal cohort-component method is that the procedure assumes that, for any age
cohort in any spatial unit, degrees of completeness of enumeration at the first and
second censuses, relative to those applicable at the national level, are the same.
Thus if an age cohort in an area is more underenumerated than the equivalent
national cohort by, say, 2 % at the first census, it is assumed to remain 2 % more
underenumerated than the national cohort at the second census as well. Checking
the validity of this assumption normally is nigh impossible, but the mechanism that
allows CSRs to inherently correct for census enumeration error is effective only to
the extent that it holds. And even if it does hold, if there is overall underenumeration
or overenumeration in the national population the total amount of net migration
(over all ages) will be underestimated or overestimated accordingly. However, a
rate of migration based on the population at the second census will be free of
enumeration error, since both numerator and denominator will feature the same
degrees of such error.
The CSR approach to the intercensal cohort-component method technically
cannot deal with cohorts born during the intercensal period, because, by definition,
no cohort members were present at the first census, and therefore no CSR can be
calculated for the cohort. However, if reliable birth registration data are available
for the intercensal period, survival ratios analagous to CSRs can be calculated
by taking the ratio of the size of the cohort at the second census to the number
of births (of a given sex if calculations are being made for males and females
separately) during the period over which the cohort was created. These survival
ratios can then be used in place of LTSRs in Eqs. 7.18, 7.20 and 7.34.
If reliable data on births are not available, approximations of net migration in
5-year age cohorts born during a 10-year intercensal period can be obtained using
geographic area-specific child-woman ratios from the second census. If we define
two child-woman ratios CWR(1) D children aged 0–4 divided by women aged 15–
44 and CWR(2) D children aged 5–9 divided by women aged 20–49, then:

NM.04/ D 1=4CWR.1/ : NM.f;1544/ (7.35)

NM.59/ D 3=4CWR.2/ : NM.f;2049/ (7.36)

Where NM(04) and NM(59) are net migration for the cohorts aged 0–4 and 5–9 at
the second census; NM(f,1544) and NM(f,2049) are net migration for the female
cohorts aged 15–44 and 20–49 at the second census.
Choice of a method for estimating internal migration often is dictated by the
availability of data. Where options exist, however, direct measures, such as can
be obtained from data on place of residence at a fixed past date, or on duration
of residence by place of last residence, normally are preferable. Among the three
334 7 Population Distribution, Urbanization and Migration

indirect approaches discussed, that using place of birth data (Eq. 7.31), the vital
statistics method and the intercensal cohort-component method, it is impossible to
be categorical as to which is preferable. Generally the intercensal cohort-component
method, ideally executed using CSRs rather than LTSRs, seems to be favoured.
One of its attributes is its inherent provision of age-specific detail, which the vital
statistics method may, but does not necessarily, yield. But there are circumstances,
often linked to the comparative quality of the data it and other approaches require,
when the CSR-based cohort-component method is not the best option. These can be
followed up in United Nations (1970: 35–36).

Migration Rates and Ratios

It has already been noted that the construction of migration rates and ratios is
complicated by the reality that any migration involves both a place of origin and
a place of destination. This difficulty can be set aside if one’s focus is the totality of
internal migration within a country or region during a migration interval. One’s
interest is only in whether individuals moved (internally) during the migration
interval, not where they moved from or to. All members of the population are at
risk of having become internal migrants, and a rate of migration can be calculated
as follows:

M D .m=p/ : 1;000 (7.37)

Where m D the number of internal migrations or the number of persons migrating

internally during the migration interval; p D the population exposed to the risk
of migrating during the migration interval – i.e., the sum of migrants and non-
migrants.
Whether the numerator in this measure is a count of migrations (which may be
multiple for an individual) or of migrants will depend on the nature of available
data and will determine whether M measures the incidence of moves or of persons
moving at least once. A count of migrants is perhaps preferable and tends to be the
type of count yielded by the usual census approaches to obtaining internal migration
data, subject to the limitation that migrants who had returned to their places of birth
(the place of birth, place of current residence approach to lifetime migration) or
to their places of residence at a specified past date cannot be detected. It is also
arguable that p should exclude immigrants (i.e., persons who joined the population
under study by external migration during the migration interval) because they were
not at risk of internal migration throughout the migration interval, and any internal
movements will have been suppressed in favour of their external movements in data
based on place of birth or of residence at a fixed past date. The numerator m, too,
should exclude external migrants, or their migrations (internal as well as external),
as the case may be.
In the relatively rare event of data tabulating migrants by number of migrations
being available for a migration interval, the possibility exists of computing not
Migration 335

only both varieties of migration rate detailed above, but also rates of migrating
only once (as opposed to at least once), more than once, and specific numbers of
times in excess of once. In each case Eq. 7.37 could be used, with m redefined to
be the number of migrants experiencing the relevant number of migrations. The
denominator p would remain the sum of migrants and non-migrants, but could
also be conceived of as the sum of migrants experiencing the relevant number
of migrations, migrants experiencing some other number of migrations, and non-
migrants. A further possible measure with data tabulating migrants by number of
migrations is a rate of remigration among migrants, defined as:

R D .m2C =m/ : 1;000 (7.38)

Where m D the number of persons migrating internally during the migration

interval; m2C D the number of persons migrating internally on two or more
occasions during the migration interval.
When analysis becomes focused on individual migration streams, strictly speak-
ing only rates of out-migration can be calculated, since only persons at places
of origin are at risk of becoming members of a stream (persons at places of
destination are at risk of becoming members of the counterstream, for which a rate
of out-migration could also be calculated). Such rates of out-migration, which can
be construed as measures of the attraction the place of destination holds for the
population at origin, are given by:
 
Mij D mij = .pi  mi C mi / : 1;000 (7.39)

Where i D locality of origin; j D locality of destination; mij D migrations from i

to j during a migration interval; pi D population at locality i at the end of the
migration interval; m*i D in-migrations to locality i from all other localities
during the migration interval; mi* D out-migrations from locality i to all other
localities during the migration interval.
The denominator on the righthand side of this equation is a measure of the
population alive and living in the locality of origin at the beginning of the
migration interval who survived and were still living in the country/region for
which internal migration is being studied at the end of the interval. As they
should be, immigrants (arrivals through external migration) during the migration
interval are excluded (as one component of the m*i that is subtracted from pi ).
However, emigrants, who should have been included (as a component of the mi*
that is added to pi ) because, despite emigrating, they were also at risk of migrating
between localities i and j are also excluded (because the census marking the end of
the migration interval does not cover them).
When Eq. 7.39 is applied in respect of migration data based on residence at
a specified prior date, its denominator is the population of the locality of origin
at that date who survived the migration interval, less those who emigrated during
the interval. When applied in respect of lifetime migration data, mij is persons
336 7 Population Distribution, Urbanization and Migration

born in locality i and resident in locality j at the census, and the denominator is
the population born in locality i who survived to the census date, less those who
were lifetime emigrants. In both cases mij are net moves and not necessarily direct
moves. When applied, finally, in respect of data on duration of present residence
and place of last previous residence, mij is persons whose most recent move within
a selected migration interval (e.g., 5 or 10 years) was from locality i to locality j, and
is a count of direct moves. The denominator, however, does not relate to a particular
date and comprises persons at the census who had lived in locality i throughout the
migration interval (persons enumerated in locality i (pi ) less those who had migrated
there during the migration interval (m*i )), plus those whose most recent move within
the migration interval originated in locality i (although they need not have lived there
at the beginning of the migration interval).
While measures that express a migration stream mij as a ratio of the population at
destination pj are not migration rates in the sense of having ‘at risk’ denominators,
they are calculated, and can be thought of as indexing the impact of particular
migration streams on destination populations. One could adjust the denominator as
before by subtracting all migrations during the migration interval with locality j as
destination (m*j ) and adding all those with it as origin (mj* ), although because doing
so does not yield a purer denominator of the ‘at risk’ type there is less necessity to
do this. The case for adjustment is stronger if the aim is to compare the impacts of
migration from each locality of origin on a range of localities of destination than
if it is to compare the impacts on each locality of destination of migration from a
range of localities of origin.
‘Rates’ of net migration and gross migration (migration turnover) within migra-
tion streams typically have the average of the adjusted census populations of the two
localities as their denominator. Thus:
n    o
Mijji D mij  mji =1=2 .pi  mi C mi / C pj  mj C mj : 1;000
(7.40)

And:
n    o
MijCji D mij C mji =1=2 .pi  mi C mi / C pj  mj C mj : 1;000
(7.41)

Where i, j represent the two localities connected by a migration stream and

counterstream; all elements otherwise have meanings the same as, or patterned
on, those of elements in Eq. 7.39.
Besides focusing on migration streams and counterstreams connecting pairs
of localities, analyses of internal migration may also examine in-migration to,
out-migration from, and net migration between individual localities and all other
localities. Considerable debate has occurred over the most appropriate forms for the
denominators of ‘rates’ of in-migration, out-migration and net migration calculated
Migration 337

on this basis (see United Nations 1970: 41–42), but one view defines these rates for
a locality i, respectively, as:

Mi D .mi =pi / : 1;000 (7.42)

Mi D Œmi = .pi  mi C mi / : 1;000 (7.43)

h  i
Mii D .mi  mi / = pi  1=2 .mi  mi / : 1;000 (7.44)

Where elements on the righthand sides of all three equations have the same
meanings as in Eq. 7.39.
Other approaches argue that the denominator in Eq. 7.44 should also be used in
Eqs. 7.42 and 7.43 (whence M*ii* D M*i  Mi* ), or that all three rates should use
the non-migrant population (pi  M*i ) as denominator.
The foregoing discussion assumes the availability of data on gross migration
flows between origin-destination pairs. However, available migration data often
consist of net migration estimates derived by indirect methods. Such data obviously
lend themselves only to the calculation of ‘rates’ of net migration, and the
question that arises is again one of the most appropriate denominator. Where net
migration estimates are obtained by application of the vital statistics method (i.e.,
the population balancing equation) the appropriate denominator is usually taken to
be the average of the sizes of the population of the locality (area) for which a net
migration rate is being calculated at the two censuses defining the migration interval.
Thus, where net migration for locality i is obtained by the vital statistics method, the
rate of net migration is given by:

Mii D nmi =1=2 .pi;t C pi;tCn / : 1;000 (7.45)

Where nmi D the net migration estimate for locality i; pi,t D the population of
locality i at a census held at time t which marks the start of the migration interval;
n D the length of the intercensal migration interval.
When net migration estimates for localities (areas) are obtained by applying
the intercensal cohort-component method using census survival ratios it transpires
that the same denominator is appropriate for a net migration ‘rate’ whether net
migration is estimated by forward survival, reverse survival, or the average of the
two. Adjustment factors applicable in the reverse survival and average cases are
applicable equally to the numerators of rates, and thus conveniently cancel out.
The relevant denominator is the equivalent of that in Eq. 7.44, except that only the
net difference between flows out of and into a locality, not the separate flows (m*i
and mi* ), is known, and the equation for the rate of net migration takes the form:
 
Mii D nmi = pi  1=2nmi : 1;000 (7.46)
338 7 Population Distribution, Urbanization and Migration

Where nmi D the net migration estimate for locality i; pi D the population of locality
i at the census marking the end of the migration interval.
The calculation, finally, of specific ‘rates’ of migration can generally be under-
taken using whichever of the foregoing equations is appropriate given the nature of
the migration data being used, with the various elements in the equation confined
to members of the population subgroup specific to which the calculation is being
made. Most commonly, as always, rates specific for age cohorts are of interest, and
care must be taken to ensure that all migration flows, net migration estimates and
population counts, as appropriate, pertain to the relevant cohort.
It should be kept in mind that internal migration ‘rates’ of the types just discussed
pertain to particular migration intervals. There is obviously no problem with
comparing rates calculated for different localities (defined at a given geographic
scale) with respect to the same migration interval, or with comparing rates for a
particular locality over different migration intervals provided these are of the same
length. But certain types of comparisons are not valid. You should not compare
migration rates for ‘localities’ that are defined at different geographic scales. In
the Australian context, for example, it makes no sense to compare a rate of internal
migration calculated at the State/Territory level with one calculated at the SA1 level.
Moves which are migrations at the latter level often will be too short to be migrations
at the former level. Second, comparisons of migration rates based on migration
intervals of different lengths are not valid. Some types of demographic comparisons
where data are available for periods of differing length can be legitimized by
calculating annual averages; rates of population growth over intercensal periods of
varying length are the obvious example. However, when dealing with migration data
which are counts of migrants rather than of migrations, as the length of the migration
interval increases so does the opportunity for individuals to have made multiple
moves, and hence so does the degree to which a count of migrants underestimates
the number of migrations. It follows that an annual average migration rate based on
such data is biased downward to a greater extent the longer the migration interval,
compromising the comparability desired. There is arguably less of a problem
associated with annualizing rates of net migration, given that for any migration
interval the net balances of migrants and of migrations are equal, but even then
problems can arise, especially if age-specific rates are being computed. Assuming
age cohorts are defined as at the end of migration intervals, where these intervals
are of differing lengths, so are the ranges of ages over which cohort migration
experience is being annually averaged. Bias is introduced, and annual averages are
rendered non-comparable, because migration levels are highly variable by age. For
example, the migration that is being averaged for a cohort aged 20–24 over a 5-year
migration interval is migration between ages 15–19 and 20–24, a highly mobile
phase of the life cycle. But that being averaged for a cohort aged 20–24 over a
ten-year migration interval is migration between ages 10–14 and 20–24, and the life
cycle phase extending from ages 10–14 to ages 15–19 (when children are mostly still
resident with their parents) tends to be a good deal less mobile than that extending
from ages 15–19 to 20–24. In summary, the calculation of annual averages in an
Migration 339

attempt to render comparable measures for intervals of differing lengths must be

approached with caution, irrespective of the nature of the migration data.

Internal Migration Indices

Besides the various so-called ‘rates’ of internal migration (many of which are ratios,
but not strictly rates – hence the title of the previous subsection) there are a number
of other indices that can be calculated. A measure of the amount of population
redistribution due to internal migration is the sum over the spatial units (localities)
making up a country or region for which internal migration is being studied of all
positive net migration estimates, or half the sum of the absolute values of all net
migration estimates (since the sums of all positive and all negative estimates will
be identical numerically). This can then be used to calculate a rate of population
redistribution due to migration:

Rm D 1=2†iD1;k jmi  mi j =1=2 .pt C ptCn / D †iD1;k jmi  mi j = .pt C ptCn /
(7.47)

Where k D the number of spatial units (localities) in the country or region; pt D its
population at a census held at time t marking the start of the migration interval;
n D the length of the intercensal migration interval; m*i and mi* mean the same
as in Eq. 7.39.
Note that a rate of population redistribution Rm is specific to the particular level of
geographic disaggregation of the country or region on which net migration estimates
used in its calculation are based, and its value will vary from level to level for a given
country or region. Thus in the Australian context, for example, Rm values based
on inter-State/Territory net migration estimates and on inter-SA1-level estimates
would differ, the latter being higher because of the more intricate geographic
disaggregation (and hence greater chance of a move crossing a boundary and
becoming a migration) involved. This sensitivity of Rm to the level of geographic
disaggregation makes its use for comparison inappropriate, except over time for a
given country/region analysed at a given level of geographic disaggregation.
An advantage of migration data obtained from censuses is that because data
on personal attributes gathered for the general population are available also for
migrants, considerable scope exists for analysing migrant selectivity and migration
differentials. Studies of selectivity focus on whether members of certain subgroups
of populations at localities of origin were more likely to become migrants than were
members of others. Studies of migration differentials are concerned with whether
in-migration rates differ among population subgroups at localities of destination.
Suppose that m1 , m2 , : : : , mk represent the distribution of migrantsPat a locality
of destination across the k categories of some characteristic, that m D mi , that n1 ,
n2 , : : : , nk represent the distribution of non-migrants at the same locality across the
340 7 Population Distribution, Urbanization and Migration

P
same k categories, and that n D ni . Two indices of migration differentials (IMD)
can be defined as:

IMDi D fŒ.mi =m/  .ni =n/ = .ni =n/g : 100

And:

IMDi D fŒ.mi =ni /  .m=n/ = .m=n/g : 100

Both of which reduce to:

IMDi D Œ.mi n  mni / =mni  : 100 (7.48)

Positive values of IMDi indicate an overrepresentation of migrants in category i;

negative values indicate an underrepresentation. The numerical value of the index
indicates the percentage by which migrants are overrepresented or underrepresented
in category i compared to the representation of non-migrants in that category. Thus,
for example, IMDi D 29 means that the concentration of migrants in category i is
29 % higher than the concentration of non-migrants in that category (concentration
being measured by the percentage of total migrants or non-migrants who are in the
category); and IMDi D 55 means that the concentration of migrants in category i
is 55 % lower than the concentration of non-migrants in that category.
Indices of migrant selectivity can be constructed in a similar manner. However,
whereas in measuring migration differentials the comparison that is made is
between migrants and non-migrants at a place of destination (the latter representing
the population being joined by the former), in measuring migration selectivity the
comparison made is between migrants and the total population at a place of origin
(since the latter represents the population being left by the former). Thus, suppose
that m1 , m2 , : : : , mk represent the distribution of migrants
P at a locality of origin
across the k categories of some characteristic, that m D mi , that p1 , p2 , : : : , pk
represent the distributionP of total population at the same locality across the same k
categories, and that p D pi . Then the index of migrant selectivity (IMS) for any
category i, following the pattern of Eq. 7.48, is given by:

IMSi D Œ.mi p  mpi / =mpi  : 100 (7.49)

Interpretation of values of IMSi follows the same principles as interpretation of

values of IMDi , except that the comparator is the concentration of total population
(rather than non-migrants) in category i. Positive values indicate the percentage by
which the concentration of migrants in category i exceeds that of the total population
(positive selectivity) and negative values indicate the percentage by which the
concentration of migrants in category i is below that of the total population (negative
selectivity).
As well as being able to be calculated for individual localities (areas) of origin of
internal migrants, indices of migrant selectivity can also be calculated for all internal
Migration 341

migrants in a country or region (whence m1 , m2 , etc. are the sums of all internal
migrants in categories 1, 2, etc. for the country or region and p1 , p2 , etc. are the total
national or regional populations in those categories), for international migrants
(whence m1 , m2 , etc. are the numbers of emigrants from a country in categories 1,
2, etc. and p1 , p2 , etc. are the total national populations in those categories), and
even for international migrants to particular destinations (whence m1 , m2 , etc. are
the numbers of emigrants from a country to a particular destination in categories 1,
2, etc. and p1 , p2 , etc. are again the total national populations in those categories).
In a similar manner indices of migration differentials can be computed for
international migration (whence m1 , m2 , etc. are numbers of immigrants to a
country in categories 1, 2, etc. during a migration interval and n1 , n2 , etc. are
the populations in those categories who did not immigrate during the migration
interval) and for international migration from particular countries of origin
(whence m1 , m2 , etc. are the numbers of immigrants to a country from a particular
origin in categories 1, 2, etc. during a migration interval and n1 , n2 , etc. are again
the populations in those categories who did not immigrate during the migration
interval).
Two other indices already dealt with in discussing population distribution and
international migration, respectively, also have application in studies of internal
migration (and in the former case in studies of international migration as well). The
index of dissimilarity, ID, (see Eq. 7.4) can be used to compare the percentage
distribution of migrants across categories of any variable with an equivalent
distribution of non-migrants or total population as appropriate. The distributions
are more similar the closer ID is to 0, and more dissimilar the higher its value
gets (the theoretical maximum being 100, but being rarely closely approached).
Second, the migration effectiveness ratio, MER, (see Eq. 7.22) can be applied to
net and gross internal migration figures for individual localities within a country
or region, or for locality pairs defining migration streams and counterstreams. In
these applications the quantities I and E, instead of standing for immigrants and
emigrants, stand for in-migrants from, and out-migrants to, all other localities in
the country/region (effectiveness for a particular locality) or a particular locality
(effectiveness for a particular migration stream).
Finally, if m represents the total number of internal migrants or migrations in a
country or region during a migration interval, pk represents the population of locality
k and P represents the total national or regional population, then the proportions of
migrants/migrations, m, expected to originate and terminate in locality k assuming
equal likelihoods of movement from and to every locality are both given by pk / P.
It follows that the expected number of migrants/migrations from a locality i to a
second locality j under the assumption that all localities are equally likely to provide
migrants to and receive them from all other localities is given by:
 
m : .pi =P/ : pj =P

The actual number of migrants/migrations from i to j (mij ) can be compared with

this expected number to yield an index of migration preference for moves from
locality i to locality j:
342 7 Population Distribution, Urbanization and Migration

   
IMPij D mij = m : .pi =P/ : pj =P : 100 (7.50)

The higher this index rises above 100, the more popular the two localities are as
an origin-destination pair; the lower it falls below 100 the less popular they are. A
value of 100 indicates that the scale of migration between the two localities exactly
matches that within the country or region as a whole.

References

Beale, C. L. (1975). The revival of population growth in non-metropolitan America (Economic

research service publication No. 605). Washington, DC: US Department of Agriculture.
Bogue, D. J. (1993). Spatial mobility and migration research: Editor’s introduction. In D. J. Bogue,
E. E. Arriarga, & D. L. Anderton (Eds.), Readings in population research methodology (Vol. 3,
pp. 14-1–14-2). Chicago: Social Development Center for UNFPA.
Carmichael, G. A. (1993). Beware the passenger card! Australian and New Zealand data on
population movement between the two countries. International Migration Review, 27(4), 819–
849.
Champion, A. G. (1989). Counterurbanization: The changing pace and nature of population
deconcentration. London: Edward Arnold.
Jones, H. R. (1990). Population geography. London/Thousand Oaks/New Delhi: Sage.
Newbold, K. B. (2010). Population geography: Tools and issues. Lanham: Rowman & Littlefield.
Price, C. A. (1996). Immigration and ethnicity. Canberra: Commonwealth Department of Immi-
gration and Multicultural Affairs.
Statistics Division of the United Nations. (2008). Principles and recommendations for population
and housing censuses: Revision 2. New York: United Nations.
United Nations. (1970). Manual VI. Methods of measuring internal migration. ST/SOA/SER.A/47.
New York: United Nations. Available online at http://www.un.org/esa/population/techcoop/
IntMig/manual6/manual6.html. Accessed 20th Feb 2014.
Chapter 8
Stable Population Theory

Stationary and Stable Populations

The concept of a stationary population was introduced in Chap. 4. The life table
population, it was noted, was a stationary population. It was closed to migration,
experienced a constant annual
P number of births (given by l0 ), a constant and equal
annual number of deaths ( dx D l0 ), thus had a constant size (T0 ) and a zero growth
rate, and had a constant age structure (given by the Lx -column).
A stationary population is a special case of a stable population. Development
of the stable population concept is, as was noted in Chap. 6, generally attributed to
Alfred Lotka, who developed it in a series of papers published between 1907 and
1925, although the basic idea can be found in the work of Leonard Euler well over
100 years earlier. The discovery these men made was that if a closed population
is subjected to constant schedules of fertility and mortality for a long period
of time, eventually a fixed age structure develops which is independent of the
age structure at the time the constant fertility and mortality schedules were first
established. Thus, a stable population can be defined as the limit to which a closed
population’s age (and sex) structure tends when it is subjected to constant age-
specific schedules of fertility and mortality. The ‘stability’ to which the concept
refers is stability in age-sex structure, by which is meant an unchanging shape to
the age-sex pyramid. It is perfectly possible for the size of a stable population, and
therefore the numbers of people in each age-sex group, to increase or decrease.
But once stability is attained the proportions of total population in each age-sex
group do not change. Moreover, the rate of growth of a stable population, r, usually
referred to as the intrinsic rate of population growth or the intrinsic rate of natural
increase (since stable populations are closed and thus migration is not a factor in
their growth), is also constant or unchanging.
The intrinsic rate of natural increase provides the link between stationary and
stable populations. A stationary population is a stable population whose intrinsic

© Springer International Publishing Switzerland 2016 343

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_8
344 8 Stable Population Theory

rate of natural increase is zero. In other words, the intrinsic birth and death rates
(the constant crude birth and death rates of the stable population) are equal (yielding
equal annual numbers of births and deaths).
The independence of the stable age-sex structures of stable populations from
those that exist when constant schedules of fertility and mortality commence has
been illustrated by Pollard et al. (1990), who take the very different age-sex
pyramids for Sri Lanka and Sweden in 1960 (the former broad-based in the manner
typical of young, high fertility populations and the latter much older) and show how
they evolve subsequently if both populations are assumed to experience Sweden’s
1960 fertility and mortality regimes indefinitely. By 2060 the two age-sex pyramids
are virtually indistinguishable. The United Nations (1968: 6–8) presents similar
evidence for East German and Thai pyramids projected under identical constant
fertility and mortality schedules from the mid-1950s. The determinants of the shape
of a stable population’s age-sex pyramid are thus its constant fertility and mortality
schedules alone, and have nothing to do with pyramid shape when those schedules
first become established.
Because of the nature of the stability of stable populations described above, the
intrinsic rate of natural increase, r, specifies not only the annual rate of population
growth, but the annual rates of growth in numbers in each age group, and in numbers
of both births (total and at each maternal age) and deaths (total and at each age).

Generating Stable Population Measures

For any population it is possible to define a stable equivalent population (age-sex)

structure; the population structure that would eventually emerge if current fertility
and mortality schedules were to prevail indefinitely. Conventionally a one-sex model
is developed first, for females, since details like age-specific birth rates typically
are only available for women and their reproductive ages are more well-defined.
Parameters of the male component of the stable population then are defined in
terms of those for the female component. The procedure requires, first, obtaining
the relevant stationary population structure (the Lx -column of a life table based on
current female age-specific death rates – see Chap. 4), then adjusting it so that the
rate of growth is the intrinsic rate of natural increase r, not the zero growth rate of a
stationary population.

The Intrinsic Rate of Natural Increase

It is important at the outset to differentiate the intrinsic rate of natural increase

from the crude rate of natural increase. The latter applies to the existing population
for which a stable equivalent is sought; the former to the stable equivalent, a
hypothetical population that would evolve well into the future if fertility and
Generating Stable Population Measures 345

mortality remained constant at current levels. Both populations have identical age-
specific fertility and mortality schedules, but unless the original population is
already stable, different age structures and thus different rates of natural increase.
Unless a population is stable at the outset, the stabilising process alters its age
structure, often radically.
The concept of a stable population is intimately related to the notion of reproduc-
tivity discussed in Chap. 6, and a reasonable approximation of the intrinsic rate of
natural increase r can be obtained in terms of two measures introduced there – the
net reproduction rate (NRR) and the mean length of generation (MLG). The NRR,
the extent to which a cohort of women replaces itself with daughters if experiencing
current age-specific female fertility and mortality rates, is a measure whose value is
determined by those rates alone, and is unaffected by age composition. It therefore
has the same value for both an observed population and that population’s stable
equivalent, since the two by definition share the same age-specific female fertility
and mortality schedules, and their invariably different age structures are not a factor
(unlike the situation with their crude birth rates, which are affected by their different
age structures). The NRR (denoted as R0 ) thus can be thought of as a measure of
the extent to which population increases in a generation, and the MLG (denoted
as T) as a measure of the period over which that increase occurs. Then, making
use of the exponential formula (on which the instantaneous rates of urban and rural
population growth were based in Chap. 7 – see discussion preceding Eq. (7.7)), we
have:

R0 D erT (8.1)

Whence loge R0 D rT, and so:

r D loge R0 =T (8.2)

Derivation of a more precise estimate of r entails greater mathematical com-

plexity than some readers will be comfortable with (Impagliazzo 1989), but leads
along the way to another approximate equation for r whose evaluation requires only
a slight extension of the procedure for obtaining the NRR and MLG presented in
Chap. 6 (see Table 6.1). That equation is:
 p 2 
r D ’ C ’ C 2“ loge R0 =“ (8.3)

Where R0 D the net reproduction rate as before; ’ D R1 /R0 ; “ D ’2  R2 /

R0 D (R1 /R0 )2  R2 /R0 ; R1 and R2 are what are known as the first and second
moments of the curve representing the age schedule of net reproductivity.
Using the notation used in calculating the NRR and MLG in Chap. 6, assuming
data for five-year age groups:
 
R0 D NRR D †xD1519;4549 b.f;x/ =Fx .5 Li =l0 / (8.4)
346 8 Stable Population Theory

 
R1 D †xD1519;4549 b.f;x/ =Fx .5 Li =l0 / .mx / (8.5)

   
R2 D †xD1519;4549 b.f;x/ =Fx .5 Li =l0 / mx 2 (8.6)

Where mx is the midpoint of age group x; other elements mean the same as in Eq.
6.11 in Chap. 6.
Besides R0 being simply the NRR, or the sum of column (7) in Table 6.1, Chap.
6, R1 is the numerator of Eq. 6.13 in Chap. 6 (the equation for the MLG), and is
simply the sum of column (8) in Table 6.1. Furthermore, to obtain R2 in Table 6.1
we would merely have to add a column (9) in which we would multiply values in
column (8) by values in column (2). R2 would be the sum of this column (9). Thus,
while the evaluation of Eq. (8.3) above may seem complicated it really involves little
more data manipulation than was required earlier to obtain the NRR and MLG.
If we were to add a column (9) to Table 6.1 we would obtain as its
sum R2 D 888.57045. Thus, with values of R0 D 0.92218 (column (7)) and
R1 D 28.10196 (column (8)), we would have ’ D R1 /R0 D 30.47 and “ D ’2  R2 /R0
D 34.93, whence r D 0.00265.
Should data for single years of age be available, the equations for R0 , R1 and R2
become:
 
R0 D NRR D †xDy;¨ b.f;x/ =Fx .Lx =l0 / (8.7)

Where all elements mean the same as in Eq. 6.9, Chap. 6.

 
R1 D †xDy;¨ b.f;x/ =Fx .Lx =l0 / .x C 0:5/ (8.8)

 
R2 D †xDy;¨ b.f;x/ =Fx .Lx =l0 / .x C 0:5/2 (8.9)

The Intrinsic Birth and Death Rates

The constant crude birth rate which will emerge once a population becomes stable is
given, in terms of the intrinsic rate of natural increase r, by the following equations
for five-year and single-year age group data respectively:

b D 1=†xD04;¨ erm.x/ : .5 Li =l0 / (8.10)

Where ¨ D the oldest five-year age group represented in the population; m(x) D the
midpoint of age group x; i D the lower limit of age group x.
And:
Generating Stable Population Measures 347

b D 1=†xD0;¨ er.xC0:5/ : .Lx =l0 / (8.11)

Where ¨ D the oldest single-year age group represented in the population.

Moreover, once the intrinsic birth rate b has been calculated, the intrinsic death
rate d is simply:

dDbr (8.12)

The assumption at this stage is that, like earlier equations associated with the
calculation of r, Eqs. 8.10 and 8.11 are evaluated using data for females only, giving
rates of female births among the female population. However, whereas the value
of r is the same for both females and males (as it must be for the sex ratio of the
stable population to remain constant), and therefore also for the total population, the
same is not necessarily the case with b and d. Values for the female, male and total
populations may vary, provided that in each case b  d D r (which does not vary).
To calculate b based on male data the numerator in Eq. 8.10 or 8.11 becomes the
sex ratio at birth (sb ) instead of 1 (i.e., male births divided by female births; a value
likely to be in the vicinity of 1.05), and the denominators are (i) evaluated using
values of 5 Li or Lx from the relevant male life table and (ii) each multiplied by sb .
To calculate b based on the total population the numerator becomes 1 C sb and the
denominators become the sums of the denominators of the separate equations for
females and males. In each case Eq. 8.12 can then be used to obtain a value for d.

The Stable Age Distribution

The age structure of a stable population is a function of its intrinsic growth rate r and
its intrinsic birth rate b. The constant proportion of the population in an age group x
is given, for five-year and single-year age groups, respectively, by:

c .x/ D b : erm.x/ : .5 Li =l0 / (8.13)

Where m(x) D the midpoint of age group x; i D the lower limit of age group x.
And:

c .x/ D b : er.xC0:5/ : .Lx =l0 / (8.14)

The life table quantities 5 Li / l0 and Lx / l0 in these two equations are sometimes
generalized as a quantity p(x), meaning the average number of years spent in age
group x by members of the life table stationary population.
You will note that Eqs. 8.10 and 8.11 feature the sums over all five-year and
single-year age groups of values of erm(x) . (5 Li / l0 ) and er(xC0.5) . (Lx / l0 ),
respectively, while Eqs. 8.13 and 8.14 feature these same values for the particular
348 8 Stable Population Theory

age group for which c(x) is being calculated. Thus the normal approach to
calculating c(x) values is to use the same calculating table as is generated to find b,
so that effectively Eqs. 8.13 and 8.14 are rewritten with b replaced by the righthand
sides of Eqs. 8.10 and 8.11, respectively. That is:
h i h i
c .x/ D erm.x/ : .5 Li =l0 / = †xD04;¨ erm.x/ : .5 Li =l0 / (8.15)

And:
h i h i
c .x/ D er.xC0:5/ : .Lx =l0 / = †xD0;¨ er.xC0:5/ : .Lx =l0 / (8.16)

It is also common to inflate values of c(x) by a factor of 1,000 or 10,000, so that

stable age structures are given to either of those bases rather than to a base of 1.
Once again, the equations above are presented in the context of generating a
one-sex, female, stable population. That is, life table functions come from the
relevant female life table, b is the rate of female births among females, and c(x)
is the proportion of females in age group x. The adjustments needed to generate
stable male and total population age structures follow a similar pattern to those
described above for generating values of b for stable male and total populations.
In Eqs. 8.13 and 8.14 the b-elements obviously are adjusted as already described.
Then, to generate a stable male age structure the life table function 5 Li or Lx is
taken from the relevant male life table, and a multiplier of the sex ratio at birth, sb ,
is applied. To generate a stable total age (and sex) structure the life table function
5 Li or Lx is taken from the relevant female life table when calculating c(x) values
for the female component of the population and from the relevant male life table
when calculating them for the male component, with the multiplier sb again being
applied for males. In the total population case the modifications made in calculating
b result in c(x) values that sum to 1.0 over all age groups for both sexes rather than
for one sex.
Table 8.1 carries forward the example dealing with Australia in 2011 which was
originally presented as Table 6.1 in Chap. 6 to illustrate calculation of the gross and
net reproduction rates and the mean length of generation. It has already been shown
that a minor extension of Table 6.1 yields a value of r D 0.00265, and this value
is used in the further calculations presented in Table 8.1, along with the observed
2011 value of sb D 1.05712. Calculations are based on five-year age group data.
The stable age distribution for females is given by column (6), and that for males
by column (9). In each case c(x) values have been inflated to a base of 10,000 rather
than 1.0. Columns (10) and (11) then give the stable age-sex distribution (the stable
age distribution for the total population), again to a base of 10,000 (the numbers of
males and females in the stable age-sex distribution add to 10,000). You will note
that despite the sex ratio at birth favouring males, the stable population features a
slight excess of females. Their superior longevity more than compensates for the
excess of males at birth. Males outnumber females through to age group 60–64, but
females are decisively more numerous from ages 70–74 onwards.
Table 8.1 Generation of intrinsic birth rates and age distributions of stable female, male and total populations for Australia, 2011
P P
P P (5)/(
P (8)/( (5) C
(5) C P
Age Female (5)/ (5) Male 5 Li /l0 (8)/ (8) (8)) (8))
group m(x) erm(x) 5 Li /l0 (3)  (4) (10,000) ( sb D 1.05712) (3)  (7) (10,000) (10,000) (10,000)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
0–4 2.5 1.006647 4.98215 5.015266 529 5.26177 5.296745 560 265 280
5–9 7.5 1.020074 4.97907 5.079019 536 5.25784 5.363386 567 268 283
10–14 12.5 1.033680 4.97708 5.144707 543 5.25541 5.432412 574 272 287
15–19 17.5 1.047467 4.97313 5.209190 550 5.24933 5.498500 581 275 290
20–24 22.5 1.061438 4.96666 5.271804 556 5.23470 5.556309 587 278 293
25–29 27.5 1.075596 4.96011 5.335075 563 5.21729 5.611696 593 282 296
Generating Stable Population Measures

30–34 32.5 1.089943 4.95153 5.396883 569 5.19683 5.664248 598 285 299
35–39 37.5 1.104480 4.93908 5.455117 576 5.17129 5.711586 603 288 301
40–44 42.5 1.119212 4.92031 5.506871 581 5.13746 5.749907 607 291 304
45–49 47.5 1.134140 4.89160 5.547761 585 5.08938 5.772069 611 293 305
50–54 52.5 1.149268 4.84883 5.572604 587 5.01784 5.766843 609 294 304
55–59 57.5 1.164597 4.78532 5.572969 588 4.91076 5.719056 604 294 302
60–64 62.5 1.180130 4.69046 5.535355 584 4.75095 5.606739 592 292 296
65–69 67.5 1.195871 4.54887 5.439863 574 4.51127 5.394897 570 287 285
70–74 72.5 1.211822 4.32732 5.243942 553 4.15216 5.031679 532 277 266
75–79 77.5 1.227985 3.96635 4.870620 514 3.60645 4.428667 468 257 233
80–84 82.5 1.244365 3.37541 4.200241 443 2.80275 3.487644 368 222 184
85–89 87.5 1.260962 2.44110 3.078135 325 1.77486 2.238031 236 162 118
90–94 92.5 1.277781 1.29832 1.658969 175 0.78985 1.009255 107 88 53
95–99 97.5 1.294824 0.41894 0.542454 57 0.20766 0.268883 28 29 14
100–104 102.5 1.312095 0.08550 0.112184 12 0.03387 0.044441 5 6 2
P P P P P P
D 94.789029 D 10,000 D 94.652993 D 10,000 D 5,005 D 4,995
P P P P
bf D 1/ (5) D 0.01055; df D bf  r D .01320; bm D sb / (8) D 0.01117; dm D bm  r D 0.01382; bt D (1 C sb )/( (5) C (8)) D 0.01086;
349

dt D bt  r D 0.01351
350 8 Stable Population Theory

Intrinsic birth and death rates for the female, male and total populations are
computed at the bottom of Table 8.1. As foreshadowed earlier, values for the three
populations differ.

Applications of Stable Population Theory

The primary purposes of this discussion have been to introduce the concept of a
stable population and outline how its basic parameters can be calculated. While it is
not proposed to develop applications of the theory at this juncture, it is appropriate
to conclude with a brief statement indicating what the major applications are.
Stable population models have in the past been widely used in mathematical
demography, demographic estimation and population projection. They can be used
to estimate the age structures that will result from ongoing stability of fertility
and mortality schedules in any nominated configuration. This sort of exercise has
become increasingly relevant in the context of populations which have passed
through the demographic transition and have established new, post-transitional
fertility and mortality equilibria. It can also serve to illustrate what might be
anticipated in a population if stability of fertility and mortality conditions in
conformity with specified schedules were to be achieved.
Another application has been to the estimation of vital rates when an observed
population can be regarded as approximating a stable population but has no, or no
reliable, vital registration data. If relatively good census data exist for the population,
an assumption that the population is closed is reasonable, and an estimate of the
population growth rate is available, this information can be compared with sets
of model stable populations. The objective is to find the best fit, or a series of
reasonable fits, then from the stable population(s) identified to obtain estimates of
vital rates for the observed population.
This type of exercise has of course been of greatest interest among those studying
the populations of developing countries, which were more likely to have approxi-
mated stability in the past than they have been more recently. Prior to major fertility
transitions setting in, many of these populations had relatively stable (high) fertility
in combination with declining mortality. Studies of such populations led to the
development of correction factors which could be applied to the birth and death rates
of stable populations, and to the labelling of these populations as quasi-stable. Use
of these correction factors extended the application of the general approach to demo-
graphic estimation to situations in which it would otherwise have been inappropri-
ate. Discussions of fertility and mortality estimation using model stable age distri-
butions can be found in United Nations (1968: Chapter VII, 1983: Chapter VII).
Stable populations are also useful for understanding population dynamics and
their link to population structure. They can be used, for example, to demonstrate that
fertility is a much stronger determinant of the proportions of total population under
age 15 and over age 65 than is mortality. They are an experimental tool, allowing
demographers to change parameters as they wish and to assess the consequences of
those changes.
References 351

References

Impagliazzo, J. (1989). Stable population theory and applications. In S. A. Levin, T. G. Hallam, &
L. J. Gross (Eds.), Applied mathematical ecology (pp. 408–427). Berlin/Heidelberg: Springer.
Pollard, A. H., Yusuf, F., & Pollard, G. N. (1990). Demographic techniques (3rd ed.).
Sydney/Oxford/New York: Pergamon Press.
United Nations. (1968). The concept of a stable population: Application to the study of populations
of countries with incomplete demographic statistics. ST/SOA/SER.A/39. New York: United
Nations. Available online at http://www.un.org/en/development/desa/population/publications/
mannual/model/stable-population.shtml. Accessed 27 July 2013.
United Nations. (1983). Manual X. Indirect techniques for demographic estimation.
ST/ESA/SER.A/81. New York: United Nations. Available online at http://www.un.org/esa/
population/techcoop/DemEst/manual10/manual10.html. Accessed 26 July 2013.
Chapter 9
Population Projections

The Bread and Butter of Demography

Population projection was mentioned briefly in Chap. 4 when in Fig. 4.8 a Lexis-
type diagram was presented to illustrate how projecting a population n-years into
the future entailed (in part) forward survival of both the population at the beginning
of the projection period and the births that would occur during the projection
period. Population projection is arguably the most marketable skill demographers
have – their bread and butter, widely assumed by potential employers to be their
core business. It is fundamental to what is often termed ‘applied demography’
(Rowland 2003), defined by Siegel (2001: 2) as ‘the sub-field of demography
concerned with the application of the materials and methods of demography to the
analysis and solution of the problems of business, private non-profit organizations,
and governments, at the local, national, and international levels, with a primary
orientation toward particular areas and the present and future.’ National statistical
agencies, businesses and planning agencies at various levels from national through
regional to local government recruit demographers first and foremost with an
expectation that they will be skilled in preparing and/or making intelligent use of
population projections. The provision of all manner of services and facilities is
dependent on quality estimates of future demographic trends at all levels, from
national to local, to ensure to the maximum extent possible that they are provided on
time, in sufficient quantity and where, geographically, they are needed. Businesses
also have a major interest in population projections as they plan the marketing
of their products and the locations of their activities. And government policy
formulation in areas like immigration, housing, education and ageing is underpinned
by population projections.

© Springer International Publishing Switzerland 2016 353

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3_9
354 9 Population Projections

Population Projections and Population Forecasts

The terms ‘projection’ and ‘forecast’ are sometimes used interchangeably, and
therefore loosely, by demographers to describe their efforts to predict future demo-
graphic conditions. It is, however, more correct to distinguish between the two. Until
the mid-1940s official statistical agencies generally put out ‘forecasts’, although
some called their numbers future population ‘estimates’, but the word ‘forecast’ was
considered to imply a modicum of accuracy that caused U.S. agencies to be ‘deeply
embarrassed’ when the unexpected upturn of the birth rate following the Second
World War revealed ‘a conspicuous divergence’ between their population forecasts
and subsequent reality (Keyfitz 1987: 242). Agencies could not, however, cease
publishing the results of these activities, because despite the gross errors demand
for them was strong. It was in this context, according to Keyfitz, that (1987:242)
‘someone came up with a distinction between projections and forecasts. The former
consisted of a non-committal working out of a set of stated assumptions and
did not pretend to be an account of the future. That would protect the agency
from blame for the inevitable errors.’ Thus, in the words of a Keyfitz section
header, projections are marked by ‘professional caution’; associated ‘uncertainty’
is routinely acknowledged (O’Neill et al. 2001: 217).
The standard approach to addressing this uncertainty is to produce a range of
projections. Particularly when preparing projections at national level or above (i.e.,
for major international regions or the world as a whole), demographers are apt to
present series of alternative assumptions concerning the possible future courses
of the components of population growth – fertility, mortality and migration –
then undertake the mathematics that show the future consequences of different
combinations of these assumptions. Commonly there is a medium, a high and
a low variant for each component, although there may be more or fewer and
the number of variants per component may differ. There may, for example, be
more alternative fertility than mortality assumptions made because future fertility
is deemed more unpredictable than future mortality. But projections show the
size and usually also the composition of the population if particular assumptions
were to hold true. They are ‘conditional statements about the future’ and are
‘nonjudgemental’ (Smith et al. 2001: 3). The user is left to choose which projection
seems most realistic for his or her purpose, or to contemplate a range of projections
and the possibility that planning may need to be sufficiently flexible to allow
for different plausible alternative eventualities. Projections create scenarios, which
demographers usually do not claim to be forecasts. Instead they are, to again quote
Keyfitz (1987: 246), ‘ways of focusing discussion and judgement.’ A developing
alternative approach to dealing with uncertainty is probabilistic projections. These
use expert opinion, statistical analysis (the fitting of auto-regressive integrated
moving average, or ARIMA, models) and/or analysis of errors in past projections
to associate probabilities with alternative future regimes of fertility, mortality and
migration (O’Neill et al. 2001; Wilson and Rees 2005).
The overriding message here is that from a predictive perspective projection is
an inexact science, and users need to be cognisant of that. Another indicator of
Some Other Features of Population Projections 355

this reality is the regularity with which statistical agencies update and revise their
population projections. Usually every new census, at least, is a signal to do this for
national-level projections, establishing anew a (hopefully) robust base from which to
project, but given that most countries do not have censuses as frequently as Australia
does (five-yearly), other signals recommending revision may be the intercensal
development of unanticipated trends in particular components of population change
that call into question assumptions built into previous projections. International
projections such as are prepared by agencies like the United Nations, on the other
hand, are subject to constant updating as new census results for particular countries
become available every year and things such as major epidemics emerge or are
brought under a measure of control in different countries and regions. The classic
recent example of a major epidemic has, of course, been HIV/AIDS, which suddenly
dramatically impacted life expectancies in a range of countries after the mid-1980s
and then, more recently, has seen its mortality impact in some of these countries
rapidly reduced through the making available of anti-retroviral drugs.
One of the more widely used texts on demographic methodology in recent
years defines population projections as ‘in general purely formal calculations,
developing the implications of : : : assumptions that are made [about the future
course of fertility, mortality and migration]’ (Preston et al. 2001:117). By contrast a
population forecast is ‘a projection in which the assumptions are considered to yield
a realistic picture of the probable future development of a population’ (Preston et al.
2001: 117). Or to quote Hinde (1998: 198), ‘The term forecast is used to indicate
the actual predictions about which demographers feel reasonably confident.’ By
these definitions the quality of a projection rests on its INTERNAL validity;
the mathematical soundness of the relations among demographic variables that it
models. In the absence of computational error projections are always ‘correct’. The
quality of a forecast, in contrast, is dictated by its EXTERNAL validity; or how
well its predictions correspond with subsequent reality. In Smith et al.’s (2001: 3)
words, ‘forecasts are explicitly judgemental’. They can be proven right or wrong, or
since absolute accuracy is nigh unattainable (whence forecasts are almost inevitably
‘wrong’ to some degree), relatively good or relatively poor, by subsequent events.
Mention was made above of the practice of producing a range of projection variants.
When this happens the tendency tends to be for a central, or medium variant to be
regarded as the ‘best’ assessment of what the future will be like, and the one most
likely to be elevated to the status of a ‘forecast’. Forecasting is by definition future-
focused. So, too, generally is projection, but projection backwards into the past is
also a possibility, although some would call such a process one of estimation rather
than projection (George et al. 2004).

Some Other Features of Population Projections

Assuming a future focus, the predictive reliability of population projections

that are accorded the status of forecasts clearly diminishes as the projection
period lengthens. Common sense says that regardless of the sophistication of one’s
356 9 Population Projections

projection methodology, one has a better chance of being ‘ballpark accurate’ five
years into the future than 20 or 50 years into the future. This is another reality
demographers should constantly stress to users. The further into the future a
projection extends, the less likely it is to merit being treated as producing a reliable
forecast.
Projections produced by official statistical agencies have the advantage of being
dispassionate. That is, they are produced as ends in themselves by professionals
skilled in such work, not with the aim of buttressing a particular business case or
policy argument. Numbers produced will inevitably be inaccurate predictors of the
future to a greater or lesser degree (the future is unknown and, in most respects,
unknowable), but they are produced independently, not fabricated to support an
argument. It cannot be claimed that projections produced by statistical agencies
are beyond being contested, especially where forecast status is claimed for them.
Demographers and statisticians may argue among themselves about the appropriate
assumptions to make in carrying them out. International projections prepared by
the International Institute for Applied Systems Research (IIASA), for example, are
explicitly based on deliberations of expert groups convened to debate appropriate
fertility, mortality and migration scenarios for major world regions (Lutz 1996).
Planning agencies focused on particular geographic areas may, of course, be more
intimately across demographic drivers in their areas of responsibility than their
national statistical agencies, and may in consequence prefer to generate their own
projections. However, one should be especially cautious when projections have been
produced by interest groups or individuals pushing particular policy agendas. It is
preferable that they emanate from an agency that is overtly indifferent to any policy
agenda being spruiked, unless it can be convincingly argued that such independent
projections are flawed and alternatives are warranted.
It is important to appreciate that population projection is usually not just about
projecting population numbers. It is generally also about projecting population
composition. Most fundamentally it is about projecting age-sex composition, but
projections that explicitly focus on other compositional variables – e.g., ethnicity,
household type and labour force status – are also undertaken and are receiving
increasing attention in the literature (Wilson and Rees 2005).
Projections are carried out at a range of different geographic scales. Interna-
tional agencies such as the United Nations and World Bank prepare them for the
world as a whole and for major, often continental or sub-continental, regions. For
a comprehensive review of projections at this level and their history see O’Neill
et al. (2001). National agencies naturally have a primary national focus, although
they may also produce projections for major sub-national regions (e.g., Australia’s
States and Territories) or populations (e.g., Indigenous Australians). Below national
level planning agencies with responsibility for sub-national political units (e.g.,
states, provinces or cities), functional regions or local government areas also
often produce projections of their own. These agencies are likely to be more
knowledgeable about the forces for change operating within their jurisdictions,
which at a local level might, for example, involve, in different areas, (i) conventional
ageing of suburban populations, with housing initially occupied by young families,
then children departing as they reach adolescence and adulthood followed by empty
Some Other Features of Population Projections 357

nesters ageing; (ii) constant turnover of similar types of people (e.g., in areas near
universities with lots of student rental housing); or (iii) marked suburban compo-
sitional change as older residents relinquish housing stock and regeneration and
perhaps also gentrification takes place. The latter occurs when those buying into a
suburb are of higher socio-economic status than those who formerly lived there and
upgrade housing stock. It is often sparked by urban expansion causing a premium
to come to be placed on locational attributes of formerly lower class areas that were
peripheral or otherwise less desirable when the city was smaller, but have acquired
‘inner city’ status as suburban sprawl has redefined the meaning of the term.
Projections at these different geographic levels raise different issues. Projection
time horizons tend to shorten moving from international and national levels to local
levels. Whereas at the former levels they typically extend several decades into the
future and projections often only break population down by age and sex, small area
projections usually have more immediate time horizons because of their short-term
planning importance, a greater likelihood of taking other characteristics besides age
and sex into consideration (e.g., education, labour force composition, rural-urban
residence, household type), and the realization that forecast reliability diminishes
rapidly as the projection horizon lengthens (O’Neill et al. 2001). At international
level variability in the quality of data and missing data from certain countries can be
major problems, requiring considerable adjustment and estimation. At a world scale,
of course, in the absence of interplanetary travel, migration is not a potential source
of population change and the focus is entirely on fertility and mortality. National-
level projection options may benefit from greater data availability and reliability
than exist for sub-national projections. Some data may only be available at national
level; some may be available with greater frequency at that level; and if data are
sample-based, sampling errors are likely to be smaller at national level than at sub-
national levels. Moreover, migration tends to be a more major consideration in
making sub-national, and especially local, population projections. While there
can be exceptions, international migration typically has a relatively modest, if at
times temporally variable, impact on national population change, whereas internal
migration is often the major source of sub-national population change. It is apt to
be more spatially variable and temporally volatile than either fertility or mortality,
making it difficult to project accurately and a major source of uncertainty in sub-
national population projections. Especially at more local levels single occurrences,
such as the closing of a major employer, a new residential development, or the
launching of a major venture with a construction phase followed by an ongoing
permanent boost to employment, can have significant migration-based effects on
small area demographics, sometimes with limited warning. These sorts of issues
make what are often termed small area population projections a specialized field of
demographic endeavour, often inviting different choices concerning data, projection
methodology and assumptions than might be made at other geographic levels
(Wilson 2011, 2014). Choices may also be influenced by differing user needs.
Commercial organizations, for example, may want a single ‘most likely’ short-term
forecast refined by socioeconomic variables as well as age and sex to inform their
marketing. Government planners may be interested in ageing and therefore in longer
term projections that highlight the likely future health status and living arrangements
358 9 Population Projections

of the elderly. Policy makers may prioritize alternative scenarios that attempt to
model the demographic consequences of different policy options.
Projection of populations of subnational areas is not infrequently rendered more
difficult than it would otherwise be by a lack of geographically consistent time
series. In other words subnational geographic boundaries other than basic ones
that might divide a country into states, provinces etc. have a tendency to be
redefined over time as, for example, cities expand into previously rural areas,
perhaps capturing formerly isolated smaller settlements in the process. Boundary
changes may require adjustments to historical data so as to approximate a current
geographic configuration, and subnational projections may also need to anticipate
future settlement expansion likely to generate further boundary revisions.
Population projections serve a range of different purposes. The most obvious
one is the prediction of future population change. Undertaken for this purpose it
is imperative to play close attention to the plausibility of underlying assumptions.
Without plausible assumptions, a projection will not yield a result meriting des-
ignation as a forecast. Second, projections can be used in a ‘What if?’ manner to
study determinants of population change. In this application they have a simulation
role that is illustrative of the effect of hypothetical changes, not predictive. Closely
related to this second purpose is the third – presenting alternative scenarios as a
way of trying to understand the range of possible future outcomes as components
of demographic change are varied across plausible ranges. This type of exercise
facilitates planning for worst-case outcomes. The fourth purpose is to support a
particular political or economic agenda or to sound a warning about a perceived
future threat, the aim in the latter instance being to stimulate preventive action.
And the fifth purpose is to provide a rational basis for decision making. In this
application a population projection with forecast status may be used as a base for
projecting other phenomena – for example, the labour force (through application
of appropriate labour force participation rates), housing demand (by incorporating
assumed future household composition) or demand for educational services (by
applying projected enrolment rates to relevant age groups).
George et al. (2004) observe that population projections take advantage of
two strong features of demography – its accurate recording of demographic
processes over lengthy periods and the momentum that links those processes for
one time period with those for a later time period. Demographic futures, in the
short to medium term at least, are usually intimately tied to demographic pasts, so
projections based on past trends and relationships (p. 562) ‘often serve as forecasts
of population change that are sufficiently accurate to support good decision making.’

Approaches to Population Projection

At one very basic level a distinction can be drawn between subjective and objective
approaches to population projection. Subjective approaches lack clearly defined
processes for analysing data; they rely on general impressions, intuition, analogy
Approaches to Population Projection 359

or even wild guesswork. The nature of the projection process is not clearly specified
so as to be replicable by another analyst. Such approaches have their place when it
comes to forecasting things like political change or technological change, but they
are rarely if ever defensible ways of projecting demographic change.
Respectable approaches to population projection are objective. That is, the
projection process is clearly specified in terms of its assumptions, data sources
and the mathematical relationships employed. It is amenable to being replicated
by someone else. Objective approaches do, though, have subjective elements –
they involve judgement, in respect of assumptions, data sources, key variables,
appropriate time periods and functional forms. In addition it is important to
appreciate that projection methodology is constantly evolving (Willekens 1990;
Wilson and Rees 2005).
The selection of a projection methodology depends on the desired level of detail
in the output (whether the aim is just to project total population or whether elements
of population composition are also important) and the availability of requisite data.
Methodological sophistication is not necessarily advantageous if additional data
it requires are of poor quality. George et al. (2004) list three broad categories of
objective projection methods – trend extrapolation methods, the cohort-component
method and structural models.
Trend extrapolation methods are essentially what Hinde (1998) calls the math-
ematical method. They fit mathematical functions to observed historical data and
use these functions to extrapolate into the future. Such methods are typically used
to project total populations, and while they may be separately applied independently
to population subgroups, do not project population composition as such. They
are mostly quick and simple to apply, and have minimal data requirements. They
may thus be methods of choice when data series are incomplete, and/or time
and/or budget are constrained. However, besides providing little or no information
on the future demographic characteristics of a population these methods have a
number of limitations. With the possible exception of logistic extrapolations, which
imply a Malthusian population growth dynamic (S-shaped, with initial slow growth
accelerating for a period then slowing again), they cannot be related to theories
of population growth and so have limited utility for analysing the determinants
of population growth. They make implicit assumptions about the continuity of
population change according to the chosen mathematical model throughout the
projection period, which can lead to unrealistic and even absurd results, even over
relatively short time horizons. Their failure to explicitly factor in declining fertility
rates has in recent times, for example, made them prone to overestimate future
population. And when applied to population subgroups they are apt to ignore logical
interdependencies among those subgroups, so that subgroup projections do not
necessarily sum to an overall projection.
The cohort-component approach to projection is the most widely used approach
for national and international projections. Indeed Preston et al. (2001: 119)
describe it as ‘now nearly the only method used for [such] population projections,
representing a rare consensus for the social sciences.’ Broadly, it entails dividing a
base population into subgroups assumed to be differentially exposed to the risks of
360 9 Population Projections

fertility, mortality and migration, and separately estimating the changes over time
for each subgroup. At a minimum these subgroups are defined by age and sex, but
more complex divisions that also recognize variables such as race/ethnicity, rural-
urban residence, country of birth (nationality), religion, educational attainment,
etc. are also conceivable. Projection periods (the period between the date of the
base population and the most distant date to which the projection is being carried
out) are typically divided into intervals of length equal to the width of the age
groups adopted, and projections are then carried out one interval at a time so that
projected populations at dates within the projection period are also produced, not
just an end-of-projection-period projected population. So, if single-year-of-age data
are used, projected populations at single-year intervals are produced; if five-year
age group data are used, projected populations at five-year intervals are generated
(unless the methodology incorporates means of converting the five-year age group
input data to single-year data, as, for example, the cohort-component projection
procedure in the widely used United Nations MortPak suite of demographic
computer programs does). The cohort-component method is traced back to Cannan
(1895), was independently developed by Whelpton (1928, 1936), and was first used
to project global population by Notestein (1945). It has become more detailed and
sophisticated over time, not least as computers have eliminated the tedium formerly
associated with its use, but its basic framework has changed little from that outlined
by its pioneers.
Structural models typically come into play when planners and decision-makers
encounter questions projection methods based entirely on demographic factors or
the extrapolation of historical trends are ill-equipped to answer. These are questions
concerning the demographic impact of major changes or developments in an area
that will obviously have implications divergent from what would be expected in
their absence – the opening of a large new industrial plant, or the building of a
major new piece of transportation infrastructure, or the closing down of a major
source of employment, for example. Structural models rely on relationships between
demographic and non-demographic variables, basing projected population changes
on projected changes in the non-demographic variables. These relationships are
usually developed using regression-based techniques. Population projections carried
out in this way allow for factors such as projected changes in the economy, land use,
housing, the transport system and the environment. Two general categories have
been recognized (Smith et al. 2001; George et al. 2004) – economic-demographic
models and urban systems models. The former are mostly used to project population
and economic activities for nations and larger sub-national geographic areas; the
latter are more a tool for small area projections at local geographic levels. Structural
models may contain only a few equations and variables or may be extremely com-
plex, with huge systems of simultaneous equations featuring numerous variables
and parameters.
Trend Extrapolation Projections 361

Trend Extrapolation Projections

Trend extrapolation projection methods fall into three subcategories – simple,

complex and ratio extrapolation methods.

Simple Extrapolation Methods

The simplest of the simple methods assumes linear change; i.e., that annual
absolute population change over a projection period will equal the mean annual
change over the base period – the past period on which the projection is being based.
The mean annual change over the base period is given by:

 D .p2  p1 / =y (9.1)

Where  is the mean annual population change over the base period; p2 D the
population at the end of the base period; p1 D the population at the beginning
of the base period; y D the length of the base period in years.
A projection assuming linear change beyond the base period is then generated
using:

p3 D p2 C z (9.2)

Where p3 D the projected population at the end of a z-year projection period; p2 is

again the population at the end of the base period (i.e., at the beginning of the
projection period); z D the length in years of the interval from the end of the base
period to the date for which a projected population is required.
A second simple method of trend extrapolation assumes geometric change; i.e.,
that the annual average percentage increase in population in future will be the same
as the annual average change during the base period. The average geometric rate of
population change during the base period is given by:
h i
r D .p2 =p1 /.1=y/  1 (9.3)

Where r is the average annual rate of population change over the base period;
p2 D the population at the end of the base period; p1 D the population at the
beginning of the base period; y D the length of the base period in years.
And a projection assuming geometric change beyond the base period is then
generated by:

p3 D p2 .1 C r/z (9.4)
362 9 Population Projections

Where aside from r, all elements have the same meaning as in Eq. 9.2.
The final simple method of trend extrapolation assumes exponential change.
This model is closely related to the geometric one, but views change as occurring
continuously rather than over discrete intervals. The exponential rate of population
change during the base period is given by:

r D Œln .p2 =p1 / =y (9.5)

Where r is on this occasion the average annual exponential rate of population change
over the base period; ln means ‘natural logarithm of’; other elements have the
same meaning as in Eq. 9.3.
So that a projection assuming exponential change in population is generated by:

p3 D p2 : erz (9.6)

Where r is defined by Eq. 9.5 rather than Eq. 9.3; other elements have the same
meanings as in Eq. 9.2.

Complex Extrapolation Methods

Complex trend extrapolation methods are distinguished from simple ones by the fact
that they use base-period data for more than two dates; that is, they don’t just use
data for the two endpoints of the base period. This makes them in theory better suited
to dealing with non-linear population change, although there can be no guarantee
that projections produced will be more accurate than would have been yielded by
a simple extrapolation method. The first step in applying a complex extrapolation
method is to assemble historical population data for different dates during the base
period. These data must be based on consistently defined geographic boundaries
for the area whose population is being projected. Next, parameters of the model
selected to generate the projection must be estimated. And finally, the projection
is generated using this model. It is important that consistent time units are used
when (i) estimating complex extrapolation models and (ii) using them to project
population values. Suppose the base period for a projection extended from 1991
until 2011. Time could be measured on a scale from 1991, 1992, : : : , 2011 or,
equivalently, on a scale from 1, 2, : : : , 21. Either option could be selected, but
once selected the same scale must be used when projecting into the future. You
cannot change scales beyond the date that separates the end of the base period from
the projection period. If the base period is specified as 1991, 1992, : : : , 2011 the
projection period must be specified as 2012, 2013, etc.; if it is specified as 1, 2, : : : ,
21 the projection period must be specified as 22, 23, etc.
The simplest complex extrapolation method is again a linear model, but one
computed differently from the simple extrapolation method already outlined. It uses
Trend Extrapolation Projections 363

a linear regression equation, the parameters of which are first estimated using data
for the base period then used to project the population linearly into the future. The
equation takes the form:

pi D a C bti (9.7)

Where pi is the population at time point ti ; a and b are the intercept and slope
respectively of the linear regression line fitted to (pi , ti ) pairs for values of ti
within the base period.
Having estimated a and b by ordinary least squares regression, Eq. 9.7 can then
be used for projection purposes by substituting those values in conjunction with
values of ti that correspond to time points during the projection period.
A second complex extrapolation approach uses polynomial models. These allow
population change to be non-linear. The general form of a polynomial model relating
population to time is given by:

pi D a C b1 ti C b2 ti 2 C b3 ti 3 C    C bn ti n (9.8)

Where pi is the population at time point ti ; a is the intercept term and b1 , b2 , b3 , : : : ,

bn are the partial slope coefficients of an n-degree polynomial curve that best fits
the (pi , ti ) pairs for values of ti within the base period.
The coefficients a, b1 , b2 , b3 , : : : , bn can again be estimated using ordinary least
squares regression techniques using data for the base period, then substituted in
Eq. 9.8 along with values of ti that correspond to time points during the projection
period to generate the desired population projections. Equation 9.8 is a generalized
equation that looks rather horrendous. In practice population projections normally
make use only of second-degree polynomials, or quadratic functions, whence:

pi D a C b1 ti C b2 ti 2 (9.9)

Quadratic functions can simulate a variety of population growth scenarios, including

an increasing or a decreasing rate of annual growth, and population decline at an
increasing or decreasing rate.
The third complex extrapolation approach models population change logistically.
This means a projection follows an S-shaped trajectory, with initial slow growth
accelerating then tapering back again to approach zero as an upper limit on
population size is approached. As previously noted, the logistic model is consistent
with theories of constrained population growth, the best known of which is that of
Malthus. Several equations are available to fit a logistic model to observed data for a
base period, but the one presented here is a three-parameter model that can be fitted
using NCSS (Number Cruncher Statistical System) software using the equation:
 
pt D a= 1 C bect (9.10)
364 9 Population Projections

Where pt is the projected population at time t; e is the base of the natural logarithm;
a, b and c are the three parameters estimated using the ‘Three-parameter logistic’
routine available under the ‘Curve Fitting – General’ suite of routines in NCSS,
Version 9 (see http://www.ncss.com/).
The final complex extrapolation approach to population projection uses autore-
gressive integrated moving average (ARIMA) models, also referred to as Box-
Jenkins models after Box and Jenkins (1970) who first systematically documented
the approach. It is claimed to be appropriate for projection from base time series
of medium to long length (i.e., at least 50 observations). Procedures used are
complicated and no attempt will be made to outline them here. Mathematically adept
readers are referred for more detail to Smith et al. (2001: 172–176) and George et
al. (2004: 568–570), and further to Box and Jenkins (1970) (revised editions under
the same title were published in 1976 and 1994, the latter with G.C. Reinsel as a
third author) and McCleary and Hay (1980).

Ratio Extrapolation Methods

Ratio extrapolation methods are used where an area containing the population to be
projected is a sub-area of a larger area for which projections already exist. They
are often used where geographic units exist at several levels such that those at each
level aggregate to units at the next higher level, and ultimately all aggregate, with
no omissions, to a single unit. This could be an entire country, a state or province, a
city, etc.
The most straightforward ratio extrapolation method is the constant-share
method. In this approach a smaller area’s share of the larger area’s population is
held constant at a level observed during the base period – typically at the end of the
base period where it transitions into the projection period. The relevant equation is:
 
pit D pir =pjr : pjt (9.11)

Where i and j refer to the smaller sub-area for which a projection is required and the
larger area within which it is located respectively; r and t refer to the reference
year on which the projection is being based and the year for which the projection
is required respectively; values of p are populations for the areas and time points
that are defined by their subscripts.
Application of Eq. 9.11 requires data for only one historical date, so it is
especially useful where changing geographical boundaries or poor records make
constructing longer historical series difficult or impossible. The method’s chief flaw
is that it assumes all smaller areas grow at the same rate as the larger area within
which they are located. This assumption will often not be plausible.
A second ratio extrapolation method is the shift-share method, which is designed
to deal with changes in population shares. The literature offers several variants,
The Cohort-Component Method of Population Projection 365

the most straightforward of which extrapolates population shares linearly over time
from a trend between end points of a base period. The relevant equation for this
variant is:
     
pit D pjt pi2 =pj2 C .z=y/ pi2 =pj2  pi1 =pj1 (9.12)

Where i and j refer to the smaller sub-area for which a projection is required and the
larger area within which it is located respectively; subscripts 1 and 2 respectively
denote the beginning and the end of a base period (the latter doubling as the
beginning of the projection period); y and z denote the lengths of the base period
and of the projection horizon at time t respectively; t refers to the year for which
the projection is required; values of p are populations for the areas and time
points that are defined by their subscripts.
The shift-share method should be used cautiously for longer projection horizons
(e.g., 20 or 30 years). If, during the base period, sub-areas (i) grew very slowly
or declined in population or (ii) grew very rapidly, these scenarios can respectively
lead to substantial projected population losses or absurdly high projected population
increases. These sorts of problems can be dealt with by building constraints into
one’s projections.
The final ratio extrapolation method is the share-of-growth method, which
focuses on shares of population change rather than shares of population size. It
assumes that a smaller area’s share of population change in the larger area over the
projection horizon will be the same as its share of change in the larger area during
the base period. The relevant equation is:
    
pit D pi2 C .pi2  pi1 / = pj2  pj1 pjt  pj2 (9.13)

Where i and j refer to the smaller sub-area for which a projection is required and the
larger area within which it is located respectively; subscripts 1 and 2 respectively
denote the beginning and the end of a base period; t refers to the year for which
the projection is required; values of p are populations for the areas and time
points defined by their subscripts.
This approach often yields more plausible projections than either of the other two
ratio extrapolation methods, but does run into difficulty if a smaller area growth rate
has the opposite sign from that of the larger area. There are ways of dealing with
this, including setting an offending share to zero and not letting it change.

The Cohort-Component Method of Population Projection

If population projections are demographers’ bread and butter the cohort-component

approach to projection is their methodological bread and butter when carrying
out projections. As already noted it is by some margin the most widely used
366 9 Population Projections

approach to projection. Data availability permitting, it can be used at any geographic

level, from the world and major regions thereof down through nations and then
hierarchies of sub-national areas culminating in localized small areas, although
small area projections needing to embrace local idiosyncrasies and data limitations
frequently use other methods. The cohort-component approach derives much of its
importance from the fact that projections of total population alone have limited
utility. For broader planning purposes we are really interested in projecting not just
total population, but population composition, or structure. Education planners,
for example, require projections of the numbers in age groups to which the
services they oversee cater; planning for housing future population requires detailed
projections by age, sex and marital status, among other information. The sorts of
trend extrapolation methods detailed above can, as previously noted, be applied
separately to structural subgroups of populations (defined by age and sex, and by
other structural variables deemed relevant), but such an approach is cumbersome
and has a serious logical flaw in that the separate projection of subgroups in isola-
tion ignores interdependencies between them that cohort-component methodology
inherently embraces. As a consequence subgroup projections tend not to sum to
total population projections.
The name ‘cohort-component’ acknowledges that the method (i) segments a base
population by at least age and sex, and projects age-sex groups (i.e., sex-specific
cohorts) individually, and (ii) projects for those cohorts the individual components
of population change – the types of events through which population change occurs.
When projecting a total population these components are fertility, mortality and
net migration – people are added to the population by births, removed from it by
deaths, and both added to it by in-migration and removed from it by out-migration,
with the net impact of these two processes determining whether population increase
or decrease from migration occurs. If the projection is focused on a subgroup of
the population the number of components may be larger. For example, if the aim
was to project the population of married females aged 30–34, births would not be
a component of change (you can’t be born married and aged 30–34). Deaths and
in- and out-migration of married females aged 30–34 would be, but in addition we
would need to consider (i) numbers of married women celebrating 30th or 35th
birthdays, which would respectively see them join or exit the focal population,
(ii) marriages of women aged 30–34, which would be a way of joining the focal
population, and (iii) divorces of married women aged 30–34 and deaths of husbands
of 30–34 year-old women, each of which would cause members of the focal
population to cease being married and therefore to exit that population. Fortunately
most population projections are projections of total population, so we will ignore
this type of complication.
A cohort-component projection of population is what is known as a discrete-
time model of population dynamics. This means that it only calculates population
characteristics at certain moments of time separated by significant time intervals.
Those intervals are as long as the age intervals employed to define cohorts are
wide. So if single-year age groups are employed the projection period is divided
into one-year intervals and the projection is carried out one year at a time. If
The Cohort-Component Method of Population Projection 367

five-year age groups are employed the projection period is divided into five-year
intervals and the projection is carried out five years at a time. These are not the
only two options, but are by far the most common ones, with national projections
for developed populations typically using single-year-of-age data to produce year
by year projections into the future. Single-year age groups imply the availability of
relevant complete life tables for males and females to use in one’s projection. If only
abridged life tables are available then five-year age groups and projection intervals
will need to be used, unless a method for estimating single-year probabilities of
dying from those for the wider age groups in the abridged life table is used (as,
for example, in the MortPak projection routine marketed by the United Nations).
Age group data, whether single-year or five-year age groups are employed, almost
always terminate with an open age group – something like 70C, 80C, 85C or 90C,
depending on the level of survivorship to older ages in the population.

The Projection Process: A Simplified Model

The first step in a cohort-component projection is to select a base-year population

structure. This will often be a census-based structure, perhaps adjusted slightly
to produce a mid-year structure if the census date deviated from mid-year. For
simplicity we will assume a base population with an age-sex structure only (i.e., no
additional compositional variables involved) and that the age groups are single-year
groups. For now we will also assume a closed population – i.e., we will set aside the
complication of migration and assume the only components of population change
are mortality and fertility. This is the best way to grasp the logic of projection.
Other complexities that make the projection process more realistic will be discussed
subsequently. The projection is undertaken iteratively; that is, sequentially over a
series of separate short projection intervals (single-year intervals when using single-
year age groups) until the date n-years into the future marking the desired outer limit
or horizon of the projection (the end of the projection period) is reached. Along the
way projections to intervening dates are also produced.
With the simplifying assumptions outlined, the second step involves projecting
forward (surviving) the base population in each age-sex group to determine
numbers surviving at the end of the first projection interval, in this case one year
beyond mid-year in the base year. This is a straightforward process using survival
ratios derived from relevant male and female complete life tables (Chap. 4, Eq.
4.65). The equation, applied separately for males and females, is:

sx D px1 : .Lx =Lx1 / (9.14)

Where sx gives survivors aged x at the end of the projection interval; px1 is the base
year population one year younger; Lx and Lx1 are taken from an appropriate
complete life table for the sex (males or females) for which the calculation is
being performed.
368 9 Population Projections

A modified equation is necessary to deal with survivors to the open age interval
that single-year age distributions end with – age group x C (x and over, where x is
usually 70, 80, 85 or 90). It is:
   
sxC D p.x1/C : LxC =L.x1/C D p.x1/C : Tx =T.x1/ (9.15)

Where x is the exact age marking the lower bound of the open age interval; LxC
is the sum of Lx values from the relevant complete life table for all age groups
making up the age interval xC (which is the life table function Tx ).
Equations 9.14 and 9.15 produce sex-specific age distributions of survivors at
ages 1 (last birthday) and older. Survivors aged 0 last birthday, however, have
to be estimated separately, because they were not alive at the date to which the
base population pertains. They were born during the one-year projection interval
separating the base-year population from the population being projected. Hence we
need to estimate births during the projection interval, then survive those births to
the end of the projection interval to obtain the missing survivors aged 0 last birthday
at that time. Figure 9.1 provides a lexis diagram illustration of the projection of
the population (of a given sex) aged 1 and older as survivors from the base-year
population, but the population aged 0 last birthday as survivors from births (of that
sex) during the projection interval (represented by the line AB).
The third step is therefore to estimate the number of births of each sex
occurring during the projection interval. We need births of each sex because we
are projecting the male and female populations separately (since their mortality
conditions, and also their migration patterns, which we are ignoring for the moment,
generally differ), and so need to separately survive male and female births to the end
of the projection interval to obtain male and female survivors aged 0 at that date. In
theory, every birth is produced by two individuals, one of each sex, so that ideally
births would be attributed to sexual unions, whose creation and dissolution would be
treated explicitly in the projection framework. In practice, however, data to facilitate
this are never available and the number of births is usually estimated by applying a
relevant schedule of age-specific fertility rates to averages of the female populations
in reproductive age groups at the beginning and end of the projection interval (as
approximations of person-years lived by women at each reproductive age during the
projection interval), then summing over all reproductive ages. Thus:
 
Bt;tC1 D †xD15;49 F .x/ : 1=2 p.f; x/t C p.f; x/tC1 (9.16)

Where B is births; t and t C 1 are the dates defining the one-year projection interval;
F(x) D the fertility rate for women aged x in the chosen fertility schedule (if rates
are per 1,000, divide by 1,000 to get F(x)); p(f,x) D the female population aged x.
This equation can be rewritten as follows to express births as a function only of
population at the beginning of the projection interval:
The Cohort-Component Method of Population Projection 369

Age last
birthday

A B
Projected population
Base year Base year + 1 after one-year
projection interval AB
Base population

Fig. 9.1 Lexis diagram illustrating projection of a closed population over a one-year projection
interval from a base year

Bt;tC1 D †xD15;49 F .x/ : 1=2 .p.f; x/t C .p.f; x  1/t : .Lx =Lx1 /// (9.17)

Where Lx derives from the complete life table summarizing the female mortality
conditions assumed to prevail during the projection interval; other elements have
the same meanings as in Eq. 9.16.
Total births during the projection interval Bt,tC1 are split into male and female
births by applying an appropriate sex ratio at birth. This is likely to be around 105
males per 100 females (i.e., 105/100), but a more precise ratio based on recent actual
experience of the population being projected is preferable if available. Thus we
have:
370 9 Population Projections

B.f/t;tC1 D .1= .1 C SRB// : Bt;tC1 (9.18)

And

B.m/t;tC1 D .SRB= .1 C SRB// : Bt;tC1 (9.19)

Where B(f)t,tC1 and B(m)t,tC1 are female and male births, respectively, during the
projection interval; SRB stands for ‘sex ratio at birth’.
The fourth step in this simplified cohort-component projection model is to
survive the births during the projection interval for each sex to the end of that
projection interval. This is accomplished by multiplying by L0 / l0 from relevant
male and female life tables (Chap. 4, Eq. 4.67), and fills in the missing projected
numbers of males and females surviving to age 0 at that date. Thus:

s0 D B.s/t;tC1 : .L0 =l0 / (9.20)

Where the equation is applied separately for males and females; s0 means survivors
at age 0; B(s)t,tC1 stands for births of sex s during the projection interval; life table
functions come from the complete life table assumed to summarize mortality
conditions for sex s.
Having projected the base-year population a year into the future, we can treat
that projected population as a new base population and project it one year ahead in
precisely the same way. Thus, in Fig. 9.1 the mid-year population in ‘Base year C 1’
is treated as the new base population from which to obtain a projected population
for ‘Base year C 2’. This process can be repeated for as many additional one-year
projection intervals as we wish (or are game) to add to our overall projection, so
that iteratively a five-year, 10-year or n-year projection is generated. In the n-year
projection, the population aged n years and older will be survivors from the base-
year population, while the younger population aged 0 to n1 years will be survivors
from birth cohorts born during the n-year projection period (Fig. 4.8 in Chap. 4).
The iterative process described does raise issues concerning the life tables and
fertility schedules used. One option is to assume constant mortality and fertility
throughout an n-year projection period, whence the life tables and fertility schedule
don’t change. However, if discernable trends in mortality and/or fertility are under
way in years preceding one’s base year, assumptions of constancy may be dubious.
Other assumptions may be deemed more plausible, and mortality and/or fertility
conditions may be allowed to change in clearly specified ways across the n-
year projection period. Indeed, multiple alternative sets of assumptions may be
adopted, with separate projections based on each set being prepared. This is what
often happens when statistical agencies prepare national population projections
that provide low, medium and high projection variants. What varies between them
is the assumptions made about the future courses of fertility, mortality and also
migration.
The Cohort-Component Method of Population Projection 371

The general approach outlined above has assumed a base population structure
with single-year age-sex groups. The same general approach is adopted when the
base-year age-sex structure has five-year age groups. The projection intervals are
five years rather than one year long; the iterative process produces intermediate
projections at five-year rather than single-year intervals; and the n-year projection
period is such that n is a multiple of 5.
Figure 9.2, modelled on Fig. 9.1, provides a lexis diagram illustration of the
projection of the population (of a given sex) aged 5 and older as survivors from
a base-year population, but the population aged 04 as survivors from births (of
that sex) during the five-year projection interval (again represented by the line
AB). Having chosen a base-year population structure the second step involves
projecting forward (surviving) the base population in each five-year age-sex group
to determine numbers surviving at the end of the first projection interval, this
time five years beyond mid-year in the base year. This again is a straightforward
process using survival ratios derived from relevant male and female abridged life
tables (Chap. 4, Eq. 4.66). The equation, once again applied separately for males
and females and using a standard life table type of notation, is:

5 sx D 5 px5 : .5 Lx =5 Lx5 / (9.21)

Where 5 sx gives survivors at the end of the projection interval in the five-year age
group commencing at exact age x (x D 5, 10, 15, etc.); 5 px5 is the base-year
population in the next younger five-year age group; 5 Lx and 5 Lx5 are taken
from an appropriate abridged life table for the sex (males or females) for which
the calculation is being performed.
As before, a modified equation is necessary to deal with survivors to the open
age interval that five-year age distributions often end with – age group xC (x and
over). It is:
   
sxC D p.x5/C : LxC =L.x5/C D p.x5/C : Tx =T.x5/ (9.22)

Where x is the exact age marking the lower bound of the open age interval; LxC
is the sum of Lx values from the relevant abridged life table for all age groups
making up the age interval xC (which is the life table function Tx ).
Equations 9.21 and 9.22 produce sex-specific age distributions of survivors at
ages 5–9 (last birthday) and older. Survivors aged 0–4 last birthday, however, have
to be estimated separately, because once again they were not alive at the date to
which the base population pertains. They were born during the five-year projection
interval separating the base-year population from the population being projected.
Hence we once more need to estimate births during the projection interval, then sur-
vive those births to the end of that interval to obtain the missing survivors aged 0–4
last birthday at that time. Figure 9.2 provides a lexis diagram illustration of the pro-
jection of the population (of a given sex) aged 5 and older as survivors from the base-
372 9 Population Projections

30

25

20
Exact age

15

10

A B
0
Projected population
after five-year
Mid-year Mid-year in projection interval AB
Base population in base year base year + 5

Fig. 9.2 Lexis diagram illustrating projection of a closed population over a five-year projection
interval from a base year

year population, but the population aged 0–4 last birthday as survivors from births
(of that sex) during the five-year projection interval (represented by the line AB).
The third step is therefore again to estimate the number of births of each sex
occurring during the projection interval. Once more we first estimate total births of
both sexes by applying a relevant schedule of age-specific fertility rates to averages
of the female populations in reproductive age groups at the beginning and end of the
projection interval (as approximations, when multiplied by 5, of person-years lived
by women at each reproductive age during the five-year projection interval), then
summing over all reproductive ages. On this occasion we utilize data for five-year
age groups rather than single-year age groups. Thus:
 
Bt;tC5 D †xD1519;4549 F .x/ : 5 : 1=2 p.f; x/t C p.f; x/tC5 (9.23)
The Cohort-Component Method of Population Projection 373

Where B is births; t and t C 5 are the dates marking the beginning and end of the
five-year projection interval; x is a five-year reproductive age group; F(x) D the
fertility rate for women aged x in the chosen fertility schedule (if rates are per
1,000, divide by 1,000 to get F(x)); p(f,x) D the female population aged x.
As before with Eq. 9.16, this equation can be rewritten to express births as a
function only of population at the beginning of the projection interval:

Bt;tC5 D †xD1519;4549 F .x/ : 5 : 1=2 .p.f; x/t C .p.f; x  1/t : .5 Lx =5 Lx5 ///
(9.24)

Where x1 means the next younger five-year age group to age group x (so if x
is 1519, x1 is 1014); 5 Lx derives from the abridged life table summarizing
the female mortality conditions assumed to prevail during the projection interval;
other elements have the same meanings as in Eq. 9.23.
Total births during the projection interval Bt,tC5 are split into male and female
births by applying an appropriate sex ratio at birth. A precise figure based on recent
actual experience of the population being projected is preferable, but in its absence
a ratio of 105 males per 100 females (105/100) is a reasonable approximation for
most populations. Thus we have:

B.f/t;tC5 D .1= .1 C SRB// : Bt;tC5 (9.25)

And

B.m/t;tC5 D .SRB= .1 C SRB// : Bt;tC5 (9.26)

Where B(f)t,tC5 and B(m)t,tC5 are female and male births, respectively, during the
projection interval; SRB stands for ‘sex ratio at birth’.
The fourth step in this simplified cohort-component projection model based on
five-year age groups is to survive the births during the projection interval for each
sex to the end of that projection interval. This is accomplished by multiplying by
5 L0 / (5.l0 ) from relevant male and female abridged life tables (Chap. 4, Eq. 4.68),
and fills in the missing projected numbers of males and females surviving to ages
04 at that date. Thus:

5 s0 D B.s/t;tC5 : .5 L0 = .5 : l0 // (9.27)

Where the equation is applied separately for males and females; 5 s0 means survivors
aged 04; B(s)t,tC5 stands for births of the relevant sex during the projection
interval; life table functions come from the abridged life table assumed to
summarize mortality conditions for the relevant sex; 5 L0 D1 L0 C4 L1 .
374 9 Population Projections

Having now projected our base-year population five years into the future, we can
treat that projected population as a new base population and project it five years
ahead in precisely the same fashion. Thus, in Fig. 9.2 the mid-year population
in ‘Base year C 5’ is treated as the new base population to project a further five
years ahead and obtain a projected population for ‘Base year C 10’. This process
can then be repeated for as many additional five-year projection intervals as we
wish (or are game) to add to our overall projection, so that iteratively an n-year
projection is generated (where n is some multiple of 5). As with projections over
one-year intervals using single-year-of-age data, life tables and fertility schedules
used may each either change from projection interval to projection interval to allow
for assumed changes in mortality and/or fertility conditions over time, or not change
(constant mortality and/or fertility conditions).

Introducing Migration

While the population of the world is closed, major regional populations, national
populations and sub-national populations at various geographic levels typically are
not, so that projection exercises need also to take account of migration. Migration is
generally harder to forecast accurately than either mortality or fertility, especially
for small areas. Difficulties in forecasting it derive from a number of things: the
greater responsiveness of migration than either fertility or mortality to changing
economic conditions, employment prospects, housing availability, transport options
and neighbourhood conditions; its susceptibility to influence by government policy,
natural disasters and social or cultural conflicts; and overlapping with this its
sensitivity, especially at more local levels, to unpredictable events (like sudden
closures of major sources of employment) and administrative or legislative actions
that introduce or remove what George et al. (2004) call ‘special populations’ – the
likes of refugee groups, university students, military personnel and prison inmates.
For a lot of national populations the net migration component of population
change is typically small compared to the birth and death components. At the same
time, as birth rates have sunk well below replacement level and life expectancies
have risen to new highs in many more developed populations, migration has
become a more significant source of population change. It is also a less predictable
source – considerations such as HIV/AIDS aside, the momentum of fertility
and mortality trends in most parts of the world over recent decades has been
downward and increasingly predictable, whereas migration trends have no such
inherent directionality and the migration contribution to population change can
fluctuate substantially over comparatively short periods. This lack of predictability
may discourage simple extrapolation of past net migration trends and lead to a
conservative widening of the range between ‘high’ and ‘low’ estimates of future
migration in projections.
It is also important to appreciate that migration activity tends to be greatest
among young adults. These, of course, are people of reproductive age, so that
The Cohort-Component Method of Population Projection 375

substantial migration during a projection interval, whether inward or outward in

net terms, has important flow-on implications for projected births during both
the current and subsequent projection intervals. Get your migration assumptions
radically wrong and the natural increase component of your projection is likely to
be compromised as well. Failure to predict migration trends acceptably accurately is
a common reason for revising population projections sooner than originally planned.
A complication with projecting migration is the reality that while national popu-
lation projections need to take account only of international migration, sub-national
projections (e.g., for the States and Territories of Australia) need to incorporate
both international and internal migration. These two types of migration often
draw on different data sources and may need to be separately built into sub-national
projections.
A range of methodologies exist for incorporating migration into population
projections. These variously project either net migration flows or gross migration
flows (i.e., in-migration and out-migration separately). Both approaches are widely
used, the latter being theoretically and computationally more elegant, but requiring
more data and being more complicated to apply. Methods focused on gross
migration flows are often used to deal with internal migration in sub-national
population projections. George et al. (2004) describe two basic techniques used in
preparing State projections for the United States. The first, employing five-year age
groups and projection intervals, uses census data on place of residence five years
previously to compute age-sex-specific out-migration rates for each State (based on
estimated populations in age-sex groups five years before the census). These are then
applied to base-year populations in age-sex groups to yield a projection of the total
pool of interstate out-migrants for all States by age and sex, that pool then being
distributed to receiving States according to the proportions of interstate in-migrants
in age-sex groups that each State received during the five-year period prior to the
census. Note the distinction between using out-migration rates and in-migration
proportions. It reflects the fact that while members of a population are at risk of
migrating from that population, they are not at risk of migrating to it. In-migrants
are sourced from other populations.
The second technique for projecting migration based on gross flows uses what
are known as multiregional models (Rogers 1985, 1995). In these, migration is
treated as part of an integrated system of mortality, fertility and origin-destination
population flows between regions of a larger geographic entity (typically a country)
by age and sex (and sometimes other demographic attributes as well). Thus, inter-
state migration in a multiregional model for the United States can be represented
by a 51  51 matrix (Washington, D.C. is included) showing, by age and sex, the
number of people moving from each State to every other State during the five-year
period preceding a census. Migration rates are calculated for each State by dividing
destination-specific gross migration flows by age and sex by relevant origin State
populations, yielding 50 sets of age-sex-specific out-migration rates, one for each
destination State. These rates can then be assumed to be replicated or modified in
some prescribed way over subsequent five-year projection intervals.
376 9 Population Projections

Projections of net migration use a top-down or bottom-up approach, or both.

The former first projects total net migration based on recent levels, historical trends,
structural models, or some other method, then disaggregates by age and sex based
on past observed distributions. The focus is initially broad, then becomes more
detailed. A bottom-up approach works the other way round. Net migration rates
are developed for each age-sex cohort, then applied to age-sex groups in the base
population. The total volume of net migration projected for a population is then
obtained as the sum of individual values projected for age-sex groups.
A top-down approach to projecting net migration is often taken in national
population projections, where the migration in question is exclusively international
migration. Where net migration models are used in making sub-national projections
they generally combine international and internal migration. Note that because net
migration models do not base migration ‘rates’ on a population at risk, when used
for sub-national projections they are apt to result in inconsistencies in projections
for a group of areas (George et al. 2004). Those for areas with rapidly growing
populations can be too high, and those for areas with slow or negative growth
can be too low. Various constraints can be built into models to prevent this sort
of inconsistency getting out of hand.
Given the variety of different approaches to incorporating migration into pop-
ulation projections (see Wilson and Rees (2005) for a review of more recent
methodological developments) there is obviously diversity in the precise mathemat-
ics employed. Moreover, these days there is rarely a need to carry out projections
laboriously by hand, as there are any number of computer packages that will do
the job for you provided requisite input data are entered in a prescribed manner.
It is therefore not proposed to delve deeply into the mathematics of incorporating
migration into a population projection. However, for illustrative purposes we will
extend the basic methodology already outlined for projection of a closed population
using five-year age groups and five-year projection intervals (Eqs. 9.21, 9.22, 9.23,
9.24, 9.25, 9.26, and 9.27).
It was noted above that while it makes sense to talk of rates of out-migration from
a sub-national population or of emigration from a national population, notions of
rates of in-migration or immigration make no sense because the migrants are drawn
from outside the focal population. Furthermore, immigration is typically affected by
immigration policies which set targets or ceilings in terms of desired or maximum
numbers of migrants, not rates of migration. These sorts of considerations mean
that migration assumptions in population projections, especially national population
projections where internal migration is not an issue, are usually formulated in terms
of absolute numbers rather than rates.
Another key consideration is the reality that migration continuously alters the
populations at risk of dying and giving birth throughout a projection interval. In
national projections emigrants cease to be at risk of dying and female emigrants
cease to be at risk of giving birth as members of the focal population after leaving
it; immigrants may pass away between arrival and the end of a projection interval
while female immigrants may have children during this period who survive to the
end of the projection interval. How do we take this sort of complexity into account?
The Cohort-Component Method of Population Projection 377

One convenient approach is to assume that migration in both directions for any
sub-category of the population is evenly distributed through the projection interval,
which in turn permits an assumption that half moved at the very beginning of the
interval and the other half at its very end. The errors introduced (some migrants
being deemed to have moved earlier than they actually did and others to have moved
later than they did) cancel out.
Suppose, using life table-type notation, we denote net immigration of a given sex
in the five-year age group commencing at exact age x during a five-year projection
interval t to t C 5 by 5 ix . Then to calculate survivors at the end of the projection
interval taking account of migration under the assumption outlined above we need
to modify Eq. 9.21 by (i) adding directly at the end of the interval half of the net
immigration between exact ages x and x C 5 and (ii) adding at the beginning of the
interval half of the net immigration between exact ages x-5 and x and surviving
that immigrant population over the five-year projection interval. Note that we add
immigrants five years younger at the beginning of the projection interval than at
the end because the cohort ages five years during the projection interval. Note also
that net immigration values may be negative if net emigration prevails. Our revised
version of Eq. 9.21 making allowance for net migration becomes:
 
5 sx D 5 px5 C 1=2 : 5 ix5 : .5 Lx =5 Lx5 / C 1=2 : 5 ix (9.28)

Where 5 sx gives survivors at the end of the projection interval in the five-year age
group commencing at exact age x (x D 5, 10, 15, etc.) with migration taken
account of; 5 ix D net immigration during the projection interval between exact
ages x and x C 5; other elements have the same meanings as in Eq. 9.21.
Equation 9.28 is applied separately for males and females. In similar fashion
Eq. 9.22 for survivors in the open age interval needs modification (and also to be
applied separately for each sex).
   
sxC D p.x5/C C 1=2 : i.x5/C : Tx =T.x5/ C 1=2 : ixC (9.29)

Where x is the exact age marking the lower bound of the open age interval;
ixC D net immigration of the relevant sex during the five-year projection interval
at ages x and older; Tx is taken from the abridged life table summarizing
mortality conditions assumed to prevail for that sex during the projection interval;
p(x5)C D the base-year population of the relevant sex aged x5 or older.
The number of births during the projection interval also needs adjusting to take
account of migration. Female migration increments at the end of the projection
interval do not contribute to the number of births during the interval, but those at
the beginning of it are assumed to bear children during the interval at the same rate
as the population they join.
Recall that Eq. 9.24 gives the number of births during the five-year projection
interval as:
378 9 Population Projections

Bt;tC5 D †xD1519;4549 F .x/ : 5 : 1=2 .p.f; x/t C .p.f; x  1/t : .5 Lx =5 Lx5 ///

We adjust for migration by substituting for p(f,x)t in this equation the quantity
p(f,x)t C ½ i(f,x), where i(f,x) is net female immigration during the five-year pro-
jection interval in five-year age group x. Similarly p(f,x1)t becomes p(f,x1)t C ½
i(f,x1), where x1 means the five-year age group immediately younger than age
group x. With these substitutions it transpires that the adjustment to the number
of births to take account of migration (i.e., the quantity to be added to the number
yielded by Eq. 9.24) is given by:

Bt;tC5 D †xD1519;4549 F .x/ : 5 : 1=4 .i .f; x/ C .i .f; x  1/ : .5 Lx =5 Lx5 ///

(9.30)

This quantity will be negative if net emigration is occurring, reflecting a loss of

births due to a loss of potential mothers. The adjusted total number of births is given
by:

ABt;tC5 D Bt;tC5 C Bt;tC5 (9.31)

Where Bt,tC5 and Bt,tC5 are obtained from Eqs. 9.24 and 9.30 respectively.
Adjusted numbers of female births and male births are then obtained from
equations modelled on Eqs. 9.25 and 9.26:

B.f/t;tC5 D .1= .1 C SRB// : ABt;tC5 (9.32)

And

B.m/t;tC5 D .SRB= .1 C SRB// : ABt;tC5 (9.33)

Where B(f)t,tC5 and B(m)t,tC5 are female and male births, respectively, during the
projection interval; SRB stands for ‘sex ratio at birth’.
We then apply Eq. 9.27 twice to survive the adjusted number of births of each
sex during the five-year projection interval to the end of that interval – i.e., to age
group 0–4 years.
What have been developed above are a series of equations for building approx-
imate migration effects into a national population projection using estimates of
net immigration. The methodology should be regarded as basic and illustrative.
Various alternative approaches (e.g., using separate immigration and emigration
flow data) and complications (e.g., different mortality and fertility regimes for
migrants compared to non-migrants) are conceivable. It is also possible to recognize
other structural variables besides age and sex. The cohort-component approach
to population projection can deal with additional structural variables, provided
they pertain to attributes acquired at birth (such as ethnic origin or race, although
Projection Packages 379

assumptions need to be made in allocating births to ethnic or racial categories

given the reality that some will be of mixed parentage). If other structural variables
pertaining to characteristics that change during the life course are of interest,
multistate methodologies that explicitly recognize age-sex patterns of transitions
between states (e.g., labour force status, marital status) and require longitudinal data
are likely to be required.

Projection Packages

As previously noted, in the early years of its development cohort-component pop-

ulation projection was a very tedious undertaking. More recently, however, modern
computing has spawned a considerable array of computer packages for population
projection, most following the cohort-component approach. The following are some
examples.

MortPak for Windows (Version 4.3)

The MortPak suite of demographic procedures was developed by the United Nations
Population Division (UNPD) as an aid to undertaking demographic analyses for
developing country populations. It was first released in 1988 in two versions – Mort-
Pak for use on mainframe computers and MortPak-Lite for use on microcomputers
(United Nations 1988). These initial versions did not include a projection routine,
rather providing 16 programs principally for analysing mortality data, although
two were fertility analysis routines and another estimated the completeness of a
census compared to a second census. The UNPD had separately made available
a population projection program earlier than this (United Nations 1982), but in
introducing version 3.0 of MortPak and MortPak-Lite in 1990 added a projection
routine to the other 16 (United Nations 1990). MortPak for Windows was launched
as version 4.0 in 2003, and the current version, version 4.3, incorporating three
further mortality routines, was released in 2013 (United Nations 2013).
The PROJCT routine in version 4.3 carries out a single-year projection by age
and sex for up to 100 years into the future based on base-year male and female
populations in five-year age groups. These input data are converted to the necessary
single-year age group data using an interpolation method known as Beers multipliers
(Beers 1944, 1945). Thus five-year age group input data do not necessarily mean
that only a projection at five-year intervals can be carried out. The date to which
the base population pertains and the final projection year need to be nominated, as
do the open age group for the two base populations (minimum is 65C; maximum
is 85C) and the sex ratio at birth (minimum 0.75; maximum 1.50; recommended
default in the absence of a trustworthy empirical figure 1.05).
380 9 Population Projections

Other input data required are, first, male and female life expectancies at birth
for at least the base year and the final projection year. Life expectancies at birth
may also be given for intermediate projection years, but in their absence the
routine interpolates linearly between the base and final projection year values.
It also interpolates linearly between successive pairs of life expectancies when
intermediate values are given. The age pattern of mortality is provided as a United
Nations or Coale-Demeny model life table or a user-designated empirical life table
(the UN model life tables accommodate life expectancies at birth up to 92.5 years),
and the MortPak routine UNABR is then used to generate single-year probabilities
of dying (1 qx ) for every projection year from probabilities for standard abridged
life table age groups (exact ages 0–1, 1–5, 5–10, 10–15, etc.). These facilitate the
calculation of single-year survival ratios, and hence survivors to single-year age
group a, at the end of each one-year projection interval.
Fertility assumptions are built in by providing, at a minimum, total fertility rates
and age-specific fertility rates (five-year age groups 15–19 to 45–49) for the base
year and the final projection year. TFRs may also be provided for intermediate
years at the discretion of the user; values for intermediate years for which data
are not provided are calculated by linear interpolation. Age-specific fertility rates
for intermediate years are also calculated by linear interpolation with reference to
corresponding TFR values.
Net male and female migration needs to be given for the initial and final
projection years, and again values may also be given for intermediate years at the
user’s discretion. Linear interpolation is used to establish values for projection years
for which no data are given. Male and female patterns of net migration by age (five-
year age groups up to and including the nominated open age group) also need to be
provided. These patterns are assumed to apply unchanged throughout the projection
period.
PROJCT also requires the user to nominate a print cycle – a value of 1 provides
annual projections as output; a value of 5 provides projections for every fifth year
beyond the base year. For further information see United Nations (2013: 56–60).

DemProj

DemProj is a population projection program produced by the United States Agency

for International Development as part of its Spectrum suite of software to inform
the establishment and monitoring of development programs and policies (USAID
2008). Other programs in the suite (i) project family planning requirements needed
in order to achieve national fertility goals, (ii) compare the costs and benefits of
family planning programs, (iii) project the consequences of an AIDS epidemic, (iv)
project the social and economic consequences of high fertility and rapid population
growth for the labour force, education, health, urbanization and agriculture, (v)
estimate the consequences of adolescent reproductive behaviour, and (vi) examine
Projection Packages 381

the costs and benefits of interventions to prevent the mother-to-child transmission

of HIV.
After choosing a base date and the length of the desired projection period
(maximum 150 years) DemProj’s first input requirement is the base population by
five-year age groups and sex, with a terminal open age group of 80C years. As with
PROJCT in MortPak, these data are converted by DemProj to single-year-of-age
data using Beers multipliers (Beers 1944, 1945), facilitating the generation of male
and female projections at single-year intervals.
The fertility input required is a total fertility rate and an age distribution of
fertility for each year of the projection. TFRs can be input (i) year by individual
year, (ii) as a constant value throughout the projection period or defined portions
thereof, (iii) by nominating base year and end of projection values and interpolating
linearly between them to establish intermediate values, or (iv) by nominating values
for selected years during the projection period and interpolating linearly between
successive pairs of these to establish intermediate values. Percentage distributions of
total fertility can be specified in any of three ways: (i) by selecting one of four United
Nations model fertility schedules (known as ‘UN sub-Saharan Africa’, ‘UN Arab’,
‘UN Asia’ and ‘Average’ (Table 9.1)), (ii) by entering one’s own distributions,
or (iii) by using the Coale-Trussell fertility model. Choosing a United Nations
model is the default option, although as the names of the models suggest they
are geared to projecting developing populations (which DemProj was designed
for) and are unlikely to be suitable if a more developed population was being
projected. The models were originally constructed in 1973 in the context of the
UN’s regular (nowadays biennial) World Population Prospects global projections
exercise, and provide age distributions of fertility for TFRs ranging from 7.0 to 2.0
in decrements of 1.0, with distributions for more precise TFRs established by linear
interpolation. Stover and Kirmeyer (2007) warn that the selection of a UN model
should not be based solely on geography, noting that (p. 21) ‘regions are not as well
associated with fertility patterns as they are with mortality patterns [because fertility
is] subject to relatively more forces that are primarily not biological’. They advise
first checking a country’s or region’s base year age distribution of fertility against
all of those given in Table 9.1 to select the one that is most appropriate, rather than
choosing blindly on the basis of regional geographic location.
The option in DemProj to enter one’s own five-year age group distributions
of fertility has the same input options as were listed above for specifying TFRs.
Separate distributions can be input for each year; a distribution input for a given
year can be kept constant over any subsequent portion of the projection period; or
linear interpolation can be used to generate intermediate distributions between pairs
of distributions specified for years two or more years apart. This option is especially
likely to be attractive if a more developed population is being projected. Table 9.2
shows model age patterns of fertility used by the United Nations Population Division
in carrying out its 2012 revision of World Population Prospects for two groups of
‘low fertility’ countries described as the ‘market economy countries of Europe’ and
the ‘countries with economies in transition’ (United Nations 2014). The former were
used for countries such as the Netherlands and also several low fertility countries
382 9 Population Projections

Table 9.1 United Nations model tables of the age distribution of fertility
Age group
TFR 15–19 20–24 25–29 30–34 35–39 40–44 45–49 Total
Sub-Saharan Africa
2 8.2 35.4 29.9 17.4 7.2 1.7 0.1 100
3 14.0 31.1 24.7 16.6 9.2 3.6 0.6 100
4 14.9 25.9 22.1 17.1 11.7 6.4 1.8 100
5 16.1 25.4 22.0 17.0 11.6 6.2 1.6 100
6 16.4 24.7 22.1 17.3 11.7 6.2 1.5 100
7 14.7 23.5 21.9 17.9 12.8 7.2 2.0 100
Arab Countries
2 7.2 31.1 30.3 19.7 9.0 2.4 0.2 100
3 6.6 29.1 29.8 20.7 10.4 3.2 0.2 100
4 7.6 24.4 26.0 21.1 14.2 6.9 1.4 100
5 8.5 23.1 24.9 21.0 14.2 6.9 1.4 100
6 8.8 21.9 24.3 21.1 14.8 7.5 1.6 100
7 7.8 21.7 25.1 21.9 15.0 7.2 1.4 100
Asia
2 2.8 31.1 38.4 21.1 5.9 0.7 0.0 100
3 2.4 23.5 33.7 25.6 11.9 2.8 0.1 100
4 3.8 20.8 27.9 24.6 15.7 6.3 0.8 100
5 5.6 21.4 26.6 23.3 15.4 6.7 1.0 100
6 7.9 22.8 26.2 22.0 14.2 6.1 0.9 100
7 11.8 24.1 24.1 19.5 13.0 6.3 1.3 100
Average
2 6.1 32.5 32.9 19.4 7.4 1.6 0.1 100
3 7.7 27.9 29.4 21.0 10.5 3.2 0.3 100
4 8.8 23.7 25.3 20.9 13.9 6.5 1.3 100
5 10.1 23.3 24.5 20.4 13.7 6.6 1.3 100
6 11.0 23.1 24.2 20.1 13.6 6.6 1.3 100
7 11.4 23.1 23.7 19.8 13.6 6.9 1.6 100
Source: Stover and Kirmeyer (2007: 19)

outside Europe (almost certainly including the U.S.A., Canada, Australia and New
Zealand) while the latter countries included many of the countries of Eastern Europe
(e.g., Slovenia). A separate set of models was used for ‘high fertility’ and ‘medium
fertility’ populations (not shown; see United Nations (2014: Table II.2)). Note that
the models in Table 9.2 do not differentiate by TFR. All are models applicable to
‘low fertility’ populations, and the skill in using them is to (i) identify the model
that best matches known base-year conditions and (ii) also identify any model(s) to
which one believes the population in question might trend at some nominated future
date(s). Assumptions of constancy or linear change in these selected age patterns
of fertility over prescribed periods can then be built into a projection. The ‘market
economy countries of Europe’ age patterns of fertility tend, of course, to be much
Projection Packages 383

Table 9.2 Model age patterns of fertility used by the United Nations for projecting low fertility
populations in 2012
Age group
Model 15–19 20–24 25–29 30–34 35–39 40–44 45–49 Total Mean age at childbirth
Market economy countries of Europe
1 2.1 22.9 43.2 26.2 5.2 0.2 0.0 100 28.0
2 1.5 17.5 40.4 31.4 8.7 0.6 0.0 100 29.0
3 1.0 13.2 36.3 35.3 13.0 1.3 0.0 100 30.0
4 0.6 9.8 31.6 37.6 17.9 2.5 0.0 100 31.0
5 0.4 7.2 26.7 38.1 23.0 4.5 0.1 100 32.0
Countries with economies in transition
1 7.9 35.3 38.4 15.9 2.4 0.1 0.0 100 26.0
2 5.6 29.5 39.3 21.0 4.4 0.2 0.0 100 27.0
3 4.0 24.1 38.4 25.6 7.3 0.6 0.0 100 28.0
4 2.8 19.4 36.2 29.5 10.8 1.3 0.0 100 29.0
5 2.0 15.4 33.1 32.1 14.8 2.5 0.1 100 30.0
Source: United Nations (2014: Tables II.3 and II.4)

older than either those of the ‘countries with economies in transition’ or those in
Table 9.1 (and also the ‘high fertility’ and ‘medium fertility’ models used by the UN
more recently (United Nations 2014: Table II.2)) – i.e., fertility is less concentrated
below age 25 and distinctly more concentrated above age 30.
The Coale-Trussell option for specifying age patterns of fertility uses the
relational attribute of the Coale-Trussell fertility model discussed in Chap. 6 to
generate future age patterns of fertility from a base year set of age-specific fertility
rates and projected TFRs for following years. It generates patterns that take into
account characteristics and components of the base year age-specific rates that
reflect deviation from an underlying natural pattern of fertility attributable to the
population’s marriage pattern and degree of fertility control. According to Stover
and Kirmeyer (2007: 26), this option ‘is recommended for populations whose
initial fertility distributions do not resemble any regional pattern, or have some
idiosyncrasy.’ It ‘performs best in the medium run, if moderate levels of fertility
are targeted; or in the short run, regardless of levels of fertility.’ However, it is not
recommended for projecting low levels of fertility following a substantial period of
fertility decline, in which circumstance it is apt to yield distributions that are ‘too
peaked’ (Stover and Kirmeyer 2007: 26).
By default DemProj sets the sex ratio at birth to 105 males per 100 females,
but an alternative more precise empirically derived value is also permissible input.
Mortality assumptions require the further input of base year and assumed projection
period male and female life expectancies at birth, and the nomination of a model
life table to generate age patterns of mortality from these. On the former front,
once again separate values can be input for each year, or a value input for a given
year can be kept constant over any subsequent portion of the projection period, or
linear interpolation can be used to generate intermediate values between pairs of life
384 9 Population Projections

expectancies specified for years two or more years apart. Thus, for example, some
sort of linear trend followed by constancy beyond a nominated future date can be
accommodated. The model life tables available are the Coale-Demeny West, North,
East and South alternatives, the United Nations General, Latin American, Chilean,
South Asian or East Asian options, or a customized table supplied by the user.
Migration assumptions require input of the net total number of migrants by sex
for each year and the separate specification of male and female age distributions
of net migration. Net out-migration is input as a negative number; net in-migration
as a positive number. The previously described input options for specifying these
assumptions across the projection period are again available. If all age groups for
a given sex are assumed to contribute net flows in the same direction as the overall
net flow, age distributions of net migration are simple percentage distributions.
However, the DemProj manual (USAID 2008) is silent on how age distributions of
net migration should be specified if there is reason to assume certain age groups
contribute net counterflows to the overall net flow. Presumably such a situation
is dealt with via a mixture of positive and negative age-specific percentages, the
sum of which is C100. Stover and Kirmeyer (2007: 35) make the point that ‘There
are no simple model tables for patterns of migration by age.’ Just as projected net
migration estimates themselves need to be based on historical data for the population
being projected for a period deemed appropriate leading up to the base year (in
conjunction, perhaps, with expressed government policy covering the projection
period or at least the earlier part of it), so, too, appropriate sex-specific assumed
age patterns of net migration need to be based on recent historical experience of that
population.
DemProj also allows for what it calls ‘regional assumptions’. This facility
allows for a binary distinction between sub-populations. Perhaps most commonly
this is a rural-urban distinction, but the term ‘regional’ can apply to any two-
way distinction – e.g., a geographic distinction between highlands and lowlands,
a cultural distinction between the native-born and foreign-born populations or
indigenous and non-indigenous populations, or a political distinction between the
north and south of a country. There are two ways of entering a ‘regional’ assumption.
You can ‘directly choose’ the percentage of the entire population you want in one of
the two subcategories (e.g., urban) in each year of the projection. Or you can base
the sub-population projections on the base year ‘growth rate difference’ between
the two sub-populations. This requires input of the base year growth rates for both
sub-populations.
For further detail on DemProj see the program manual (USAID 2008). In
addition, Stover and Kirmeyer (2007) provide further insight into certain program
functionalities and assumptions.
References 385

References

Beers, H. S. (1944). Six-term formulas for routine actuarial interpolation. The Record of the
American Institute of Actuaries, 33 Part II (68), 245–260.
Beers, H. S. (1945). Discussion of papers presented in the Record, No. 68: ‘Six-term formulas
for routine actuarial interpolation’ by Henry S. Beers. The Record of the American Institute of
Actuaries, 34 Part I (69), 59–60.
Box, G. E. P., & Jenkins, G. M. (1970). Time series analysis: Forecasting and control. San
Francisco: Holden-Day.
Cannan, E. (1895). The probability of the cessation of the growth of population in England and
Wales during the next century. The Economic Journal, 5, 506–515.
George, M. V., Smith, S. K., Swanson, D. A., & Tayman, J. (2004). Population projections. In D.
A. Swanston, J. S. Siegel, & H. S. Shryock (Eds.), The methods and materials of demography
(pp. 561–601). San Diego: Elsevier Academic Press.
Hinde, A. (1998). Demographic methods. London: Arnold.
Keyfitz, N. (1987). The social and political context of population forecasting. In W. Alonso & P.
Starr (Eds.), The politics of numbers (pp. 235–258). New York: Russell Sage.
Lutz, W. (1996). The future population of the world: What can we assume today? London:
Earthscan Publication Limited.
McCleary, R., & Hay, R. A. (1980). Applied time series analysis for the social sciences. Beverly
Hills: Sage.
Notestein, F. W. (1945). Population: The long view. In T. W. Shultz (Ed.), Food for the World (pp.
36–69). Chicago: University of Chicago Press.
O’Neill, B. C., Balk, D., Melanie, B., & Ezra, M. (2001). A guide to global population projections.
Demographic Research, 4(8), 203–288.
Preston, S. H., Heuveline, P., & Guillot, M. (2001). Demography: Measuring and modelling
population processes. Oxford: Blackwell.
Rogers, A. (1985). Regional population projection methods. Beverly Hills: Sage.
Rogers, A. (1995). Multiregional demography: Principles, methods and extensions. Chichester:
Wiley.
Rowland, D. T. (2003). Demographic methods and concepts. Oxford/New York: Oxford University
Press.
Siegel, J. S. (2001). Applied demography. San Diego: Academic.
Smith, S. K., Tayman, J., & Swanson, D. A. (2001). State and local population projections:
Methodology and analysis. New York: Kluwer Academic/Plenum Publishers.
Stover, J. & Kirmeyer, S. (2007). DemProj version 4: A computer program for making population
projections. Washington, DC: USAID. Available online at http://data.unaids.org/pub/Manual/
2007/demproj_2007_en.pdf. Accessed 7 Nov 2014.
United Nations. (1982). A user’s manual to the population projection computer programme of the
population division of the United Nations. ESA/P/WP.77, New York: United Nations.
United Nations. (1988). MortPak: The United Nations software package for mortality measure-
ment: Batch-oriented software for the mainframe computer. ST/ESA/SER.R/78, New York:
United Nations. Available online at http://www.un.org/esa/population/publications/MortPak_
SoftwarePkg/MortPak_SoftwarePkg.htm. Accessed 25 Oct 2014.
United Nations. (1990). MortPak and MortPak-Lite upgrades: Version 3.0 of the United
Nations software packages for mortality measurement. ST/ESA/SER.A/117, New York: United
Nations.
United Nations. (2013). MortPak for Windows (Version 4.3). POP/SW/MORTPAK/2003, New
York: United Nations. Available online at http://www.un.org/en/development/desa/population/
publications/pdf/mortality/mortpak_manual.pdf. Accessed 25 Oct 2014.
386 9 Population Projections

United Nations. (2014). World population prospects – The 2012 revision: Methodology of
the United Nations population estimates and projections. New York: United Nations.
Available online at http://esa.un.org/unpd/wpp/Documentation/pdf/WPP2012_Methodology.
pdf. Accessed 7 Nov 2014.
USAID. (2008). DemProj: A computer program for making population projections. Wash-
ington, DC: USAID. Available online at http://www.healthpolicyinitiative.com/Publications/
Documents/1255_1_DemmanE.pdf. Accessed 16 Aug 2014.
Whelpton, P. (1928). Population of the United States, 1925 to 1975. American Journal of Sociology,
34, 253–270.
Whelpton, P. (1936). An empirical method for calculating future population. Journal of the
American Statistical Association, 31, 457–473.
Willekens, F. J. (1990). Demographic forecasting: State-of-the-art and research needs. In C. Hazeu
& G. Frinking (Eds.), Emerging issues in demographic research (pp. 9–66). Amsterdam:
Elsevier Science.
Wilson, T. (2011). A review of sub-regional population projection methods. Queensland Centre
for Population Research, University of Queensland. Available online at http://gpem.uq.edu.au/
SubRegionalProjectionMethodsReview.pdf. Accessed 20 Jan 2015.
Wilson, T. (2014). New evaluations of simple models for small area population forecasts.
Population, Space and Place. doi:10.1002/psp.1847.
Wilson, T., & Rees, P. (2005). Recent developments in population projection methodology: A
review. Population, Space and Place, 11, 337–360.
Index

A Age-specific fertility rate (ASFR), 251

Abridged life table, 158 Age-specific marital fertility rates
assuming non-linear survivorship beyond (ASMFR), 252
exact age 5, 166 Age-specific mean CEB, 255
generalized equation for n qx values, 166 Age-specific non-marital fertility rates
generating n qx values, 161 (ASNMFR), 252
generating other functions, 172 Age-specific paternity rate (ASPR), 251
mortality probabilities and rates compared Age-specific proportion of women
to those in a single-year-of-age life childless, 255
table, 159 Albert, 202
non-linear survivorship between exact ages Amenorrhoea
1 and 5, 162 post-partum amenorrhoea, 248, 295
notation used for flow functions, 159 pregnancy amenorrhoea, 248
obtained by collapsing a single-year-of-age Applied demography, 6, 353
table, 161 Attrition, 112
stock and flow functions, 158 Auerbach, 313
terminal open age interval, 160 Australian Bureau of Statistics, 15
Administrative data sources, 17 Auto-regressive integrated moving average
border control data, 18 (ARIMA) models, 354
divorce data, 17
medical and welfare records, 19 B
population registers, 17 Barrow, 216
Adolescent subfecundity, 295 Beale, 313
Age Beers, 379, 381
main variable distorting summary measure Beers multipliers, 379, 381
comparisons, 52 Birth interval analysis, 284
Age-age last birthday or in completed years, 30 based on pregnancy/birth histories, 285
Age-exact age, 30 child survival and maternal health, 286
Age pattern of fertility focus on last closed interval, 286
graphs of ASFRs, 262 methodological problems, 285
ratios of ASFRs at different dates, 265 social, biological and physiological
Age-order-specific fertility rates (AOSFR), 252 differentials, 286
Age-sex pyramid, 7 window on women’s life cycle
Age-sex-specific death rates, 33 priorities, 286

© Springer International Publishing Switzerland 2016 387

G.A. Carmichael, Fundamentals of Demographic Analysis: Concepts, Measures
and Methods, The Springer Series on Demographic Methods and Population
Analysis 38, DOI 10.1007/978-3-319-23255-3
388 Index

Birth intervals intensity of, 109

first, or protogenetic interval, 284 tempo of, 109
intergenetic intervals, 284 Cohort expected years of life lost
open birth interval, 284 (CEYLL), 205
other closely related intervals, 284 Cohort-component population projection, 365
Bogue, 1, 2, 286, 318 introducing migration, 374
Bongaarts and Potter, 282 the projection process, 367
Bongaarts fertility model, 278 Comparative mortality factor, 59
Brass relational Gompertz fertility model, 274 Comparison of summary measures, 49
Bruggink, 202 Competing risks, 197
Burden of disease, 202, 203 cause-eliminated life tables, 199
Compositional variable categories, 53
Conjugal union, 217
C Consensual partnering, 213, 243
Cannan, 360 approaches to the study of, 244
Carmichael, 19, 248 link between marriage and parenthood
Category jumping, 19 loosened, 243
Causes of death, 192 Consensual unions, 214
data deficiencies, 193 customary marriages, 217
Cause-specific death rates Counterurbanization, 313
comparisons must standardize for age, 194 Crude birth rate (CBR), 249
CEB data deficiencies, 255 Crude death rate (CDR), 27, 50
Census questions Crude marriage rate, 221
complexity, 5 Crude rate of natural increase, 250
limited range, 5 Crude rates, 26
sensitivity, 5
to study fertility and mortality, 6
Censuses D
administrative utility, 6 Data quality, 20
as a historical record, 7 Davis and Blake, 278
post-enumeration surveys, 13 De facto and de jure censuses, 22
small-area data, 6 Decomposition, 73
Champion, 314 of change over time, 78
Chiang, 200 a complex example, 81
Children ever born (CEB), 254 components of a difference, 75
Child-woman ratio, 46, 256 contrasted with standardization, 73
Cleland and Scott, 13 implicit in direct standardization, 74
Closed and open populations, 21 interaction, 77
Coale (or Princeton) fertility indices, 270 interpretation of results, 76
Coale and Demeny, 144, 206 by reverse subtraction, 78
Coale and Guo, 206 reverse subtraction recipe, 80
Coale and Trussell, 268 Defective data, 20
Coale-Demeny model life tables, 144, 165, Definition of death, 130
206, 380 Demographic analysis, 1, 3
Coale-Trussell fertility model, 268, 383 Demographic data, 3
Cohort, 28 geographic levels at which
real and synthetic, 116 tabulated, 300
Cohort approach to demographic small-area data, 301
analysis, 85 sources of, 3
Cohort cumulative first marriage rate Demographic events, 3
(CFMR), 224 Demographic rates and ratios
Cohort demographic processes denominators of, 7
complexity introduced by renewable Demography, 1
events, 111 Dempsey, 200, 205
Index 389

Dependency ratio, 46 Forward survival

old-age dependency, 46 from birth, 183
youth dependency, 46 and population projection, 183
Differential fertility analysis, 255
Differential mortality, 195
Differential under-enumeration in registration G
and census data, 139 Gardner and Sanborn, 200
Disability-adjusted life year (DALY), 203 GBD cause list, 204
age-weighting, 209 GBD health risk factors, 204
defined, 204 General fertility rate (GFR), 32, 250
discounting future health, 208 General marriage rate, 221
equity considerations, 209 George, 355, 358, 360, 374, 375, 376
social choices built into DALYs, 205 Gini concentration ratio, 306
underpinning general concepts, 204 Global burden of disease (GBD) studies, 203
Disability-free expectation of life, 202 Greville approach to estimating n qx , 167
Duration-duration last anniversary or in Gross reproduction rate (GRR), 257
completed years, 30 Guillard, 1, 2
Duration-exact duration, 30 Guillot and Yu, 202

E H
Early neonatal deaths, 196 Hauser and Duncan, 1, 2
Endogenous and exogenous causes of Health-related quality of life (HRQL)
death, 143 methods, 206
Euler, 257 Henry, 223, 267, 268
Evenly distributed events assumption, 39 Hinde, 2, 355, 359
Exact age, 28 Hutterite marital fertility schedule, 270
Exact duration, 28
Expectation of life at birth, 151
Expectation of life remaining at a I
birthday, 151 Impagliazzo, 345
Impairment, disability and handicap
distinction between, 206
F Incidence measures, 31
Fecundability, 286, 293 Index of concentration, 308
defined, 293 Index of dissimilarity, 306, 341
Fecundity, 248 Indirect methods, 254
Feeney, 292 Infant mortality, 29, 130
Fergany method of estimating n qx , 169 degrees of concentration early in first year
Fertility, 247 of life, 143
definition, 248 lexis representation of, 140
measures based on live births in last 12 uneven distribution of infant deaths
months, 256 between exact ages 0 and 1, 139
Fertility analysis Infant mortality rate (IMR), 28, 99
cohort approach to, 249 Intelligent forms processing, 5
focus on women, 248 Intermediate fertility variables
period approach to, 249 nowadays the proximate determinants of
Fertility transition, 247 fertility, 278
First marriage process Internal migration, 22, 316
intensity, 222 census questions on, 328
need to refine analysis by sex, 223 index of migration preference, 341
timing, 227 indirect estimation of net migration, 330
Foetal deaths, 195 in-migration and out-migration, 317
Forces that discourage census taking, 4 internal migration indices, 339
390 Index

Internal migration (cont.) Life table survival ratios

measurement of, 327 estimation of intercensal migration, 188
migrant selectivity, 339 Lifetime migration, 315
migration differentials, 339 Lim, 203, 204
migration rates and ratios, 334 Linear interpolation
rate of population redistribution, 339 of separation factors, 150
International migration, 22, 316 Linear survivorship, 148
and population growth, 325 Link between TFR and GRR, 259
immigration and emigration, 317 Live birth
indirect estimation of net migration, 319 definition, 249
measurement of, 318 Lopez, 203
measurement of using migration flow Lorenz curve, 305
data, 323 Lotka, 257, 343
migration ‘rates’, 324 Lutz, 356
studied through focus on overses-born, 321
IUSSP, 1, 2
M
Marital fertility, 213
J Marital fertility rate, 250
Jones, 299 Marital fertility rate (MFR), 32, 220
Marital status, 219
legal marital status and living situation, 220
K Marital union, 217
Keyfitz, 354 Marriage, 213
Keyfitz-Frauenthal method of estimating association with fertility, 213
n qx , 169 bridewealth or bride-price systems, 218
Keyfitz separation factor, 139, 143 census data, 217
defined, 214
dissolution, 213
L dowry systems, 218
LAT relationships, 215 endogamy and exogamy, 218
Lexis diagram, 29, 85, 88 first marriages and remarriages, 219, 222
accuracy needed to exploit visual homogamy and heterogamy, 218
potential, 97 market, 225
ages/durations in completed years, 86 monogamy, serial monogamy, 218
even distribution of events/individuals patterns, 213
assumption, 94 polygamy, polygyny and polyandry, 218
exact ages/durations, 86 registration data, 217
interpreting what areas represent, 93 squeeze, 225
life lines, 90 survey data, 217
points, lines and areas thereon, 88 Mathers, 202
representation of cohort demographic McDonald, 287
experience, 91 Mean age of the fertility schedule (MAFS), 266
time and its straight line representation, 85 Mean length of generation (MLG), 260, 345
time instants and time intervals, 86 Mean population at risk, 24
Life cycle phase, 28 Mean population exposed to risk, 44
Life table, 113 Median age at first marriage, 227
abridged, 132 need to check precise nature of measure
constructed from age-sex-specific death used, 229
rates, 130 Menarche, 27, 248
framework for study of survival, 131 Menopause, 27, 248
radix, 134 Micro approach, 14
single-year-of-age, standard, or Mid-year population at risk
complete, 132 estimate of mean population at risk, 25
Index 391

Migration, 314 Parity progression, 287

data and terminology, 315 Parity progression ratio (PPR), 288
demographic perspectives on, 318 Parity progression schedules, 292
distinguished from mobility, 314 Paternity, 248
streams and counterstreams, 316 Perinatal mortality, 196
Migration effectiveness ratio (MER), 324, 341 Period approach to demographic analysis,
Migration turnover, 316, 325 85, 115
Mortality, 129 assumptions and approximations, 125
males and females treated separately, 129 basic question asked, 118
Multiple decrement life tables, 197 cross-sectional, 116
Multiregional migration models, 375 lexis representation of, 116
Multistate, or increment-decrement life Period expected years of life lost
tables, 198 (PEYLL), 205
Murray, 202, 203, 205, 206, 207, 208, 209 Person-years of exposure to risk, 24, 36
Murray and Lopez, 203 Pivotal value - standardized and unstandardized
ratios, 59
Point and period prevalence, 31
N Pollard, 344
Natural fertility Population at risk, 23, 250
approximated by Hutterite marital fertility Population balancing equation, 21
1921–30, 267 Population censuses, 4
defined, 267 disadvantages, 4
juxtaposed with controlled fertility, 267 Population composition
Natural increase, 22 impact on comparison of summary
Neonatal deaths, 196 measures, 49
Net migration, 22, 316 meaning of, 49
intercensal cohort component method, of summary measure denominators, 50
319, 331 Population distribution, 299
vital statistics method, 330 concentration/dispersion of
Net reproduction rate (NRR), 257, 345 population, 305
Newbold, 299 median and mean points of, 304
Non-marital fertility, 213 methods of analysis, 303
Non-marital fertility rate, 250 standard distance of, 304
Non-marital fertility rate (NMFR), 32, 220 Population momentum, 252, 260
Notestein, 360 Population projection packages
Numerator and denominator, 24 DemProj, 380
Numerators and denominators, 28 MortPak for Windows (Version 4.3), 379
Nuptiality, 219 Population projections, 353
Nuptiality tables, 230 approaches to, 359
capacity to distort age distribution of first the cohort-component method, 359
marriages, 237 differentiated from population
gross nuptiality tables, 231 forecasts, 354
net nuptiality tables, 234 general features of, 356
projection packages, 379
structural models, 359
O trend extrapolation methods, 359, 361
O’Neill, 354, 356, 357 Population registers, 4
Order-specific fertility rates (OSFR), 252 Population studies, 3
Potential years of life lost (PYLL), 205
Pressat, 288
P Preston, 355, 359
Parity, 253, 287 Preston, Heuveline and Guillot, 1, 2
data sources, 287 Prevalence measures, 31
registration data, need to be wary of, 253 Price, 326
392 Index

Principle of correspondence, 45 calculation from detailed data on infant

Product-limit life tables, 131 deaths by age, 153
Proximate determinants of fertility, 279 derived from model life tables, 144
used in calculating 4 q1 for abridged life
tables, 164
Q variety of types, 141
Quality of life Series of specific measures, 54
health expectancies, 201 Sex ratio, 46
healthy and health-impaired life, 201 Sex-age-marital status-specific marriage
Quartiles of total first marriage rate rates, 221
ages associated with, 227 Sex-age-specific marriage rates, 221
procedure extended to real birth Shapes of life table functions, 173
cohorts, 230 Shryock, Seigel and Associates, 260
Shryock, Siegel and Associates, 1, 2
Siegel, 353
R Siegel and Swanson, 1, 2, 200
Rates and probabilities Single-year-of-age life table
demographic probabilities, 28 bottom of the life table, 151
demographic rates, 24 core functions, 134
Reduced events approach to demographic generating 1 Lx , Tx and eo x values, 148
analysis, 223 generating 1 qx values, 135
Reed-Merrell approach to estimating n qx , 166 generating lx , 1 dx and 1 px values, 145
Reference event, 29 obtaining age-sex-specific death
Reference to a standard life table method of rates, 133
estimating n qx , 169 other functions, 135
Refinement of rates, 32 recipe for construction of, 157
Renewable and non-renewable events, 36 standard notation for functions, 134
Replacement level fertility, 251, 259 taking migration into account, 152
Reproduction Single-year-of-age life table
concept defined, 257 obtaining 1 q0 , 139
Reproduction-survival ratio, 260 Singulate mean age at marriage (SMAM), 240
Respondent burden, 5 interpretation, 240
Respondents and informants, 14 recipe for calculation, 240
Reverse survival Smith, 200, 354, 355, 360
estimation of births, 186 Specific rates, 32
to obtain information about the past, 186 age-specific rates, 33
Robine, 202 Stable equivalent population structure
Rogers, 375 defined, 344
Rowland, 353 Stable population, 257
Rubenstein, 216 and reproductivity, 345
definition, 343
Stable population theory, 343
S applications of, 350
Salomon, 203, 207, 208 intrinsic birth and death rates, 346
Sample surveys, 13 intrinsic rate of natural increase, 343, 344
cross-sectional or longitudinal, 13 obtaining the stable age distribution, 347
data quality, 13 Standard expected years of life lost
retrospective or prospective, 13 (SEYLL), 206
Sanders, 202 Standardization, 55
Second demographic transition, 213, 243 additional theoretical points, 72
Self-enumeration, 5 direct, 55
Separation factor, 139 direct - data required, 56
adjusts for non-linear survivorship, 149 direct - equation, 57
alternative interpretation, 141 direct - standard population, 56, 66
Index 393

eliminates compositional difference, 56 T

indirect, 55, 68 Total fertility rate (TFR), 47, 251
indirect - data required, 68 estimated from cohort parity progression
indirect - equation, 69 ratios, 291
indirect - population providing standard Total first marriage rate (TFMR), 223
specific measures, 68, 69 capacity to take on theoretically impossible
indirect - technical invalidity of comparing values, 224
multiple standardized ratios, 70 interpretation, 223
Standardized mortality ratio, 70, 194 Trend extrapolation projections
Standardized ratios, 58 complex extrapolation methods, 362
Standardized summary measures ratio extrapolation methods, 364
absolute and relative levels, 56 simple extrapolation methods, 361
hypothetical summary measures, 56 Trude Lappegárde, 18
Stationary population, 132, 343 True rates, 26
life table population, 132, 147, 180 Types of measure
link to stable population concept, 343 absolute measures, 21
Sterility, 248, 296 Types of measures
primary and secondary, 248, 297 relative measures, 23
temporary and permanent, 248
Stillbirths, 195
definition, 249 U
Stover and Kirmeyer, 381, 383, 384 Underlying level
Structural models of demographic phenomenon, 49
economic-demographic models, 360 Underlying level of demographic phenomenon
urban systems models, 360 intensity of phenomenon, 50
Sullivan, 202 Underlying risk, 27
Summary measures Union formation, 244
approaches to more meaningful United nations, 20, 144, 254, 344, 350, 379,
comparisons of, 54 380, 382
composition component, 50 United nations manual x, 20
generalized equation, 51 United Nations model life tables, 144, 165,
specific measures and composition 380
components, 51 United nations statistics division, 4
underlying level, or intensity Urban and rural populations, 301
component, 50 Urban structure
Summary measures at risk of yielding distorted ‘city’ and ‘town’ components, 312
comparisons measures of tempo of urban
definition, 50 concentration, 313
Survival primate cities, primacy index, 312
between birthdays - joint survival, 191 rank-size rule, 313
forward survival, 182 Urbanization, 308
life table survival ratios, 176, 180 level of, 309
population survival ratios, 176 tempo of, 309
relationship between population and life USAID, 380, 384
table survival ratios, 182 Usual resident
survival ratios from birth, 178 definition of, 22
survival ratios multiplicable, 177, 181
Synthetic cohort measures
cross-sectional heaping and deficits of V
events, 119 Van de Kaa, 213
heaping/deficit mechanisms, 120 Verhoef and van de Kaa, 17
period effects, 120 Visiting unions, 215
tempo effects, 122 matrifocal family structures, 216
tendency to exaggerate change, 118 Vital events, 15
394 Index

Vital registration, 7, 15 International Classification of Impairments,

cost, 16 Disabilities and Handicaps, 206
coverage, 16 Wunsch and Termote, 1, 2, 223
Vitiligo, 207
Vos, 203
Y
Years of healthy life lost due to disability
(YLDs), 204
W disability weights, 207
Weeks, 1, 2 Years of life lost due to premature mortality
Weinstein and Pillai, 1, 2 (YLLs), 204
Whelpton, 360 Years of potential life lost (YPLL), 200
Wilson, 357
Wilson and Rees, 354, 376
World health organization Z
International Classification of Diseases, Zhao, 170
Injuries and Causes of Death, 192, 193 Zipf, 313