MULTILEVEL

ANALYSIS
 2nd Edition

MULTILEVEL
ANALYSIS
An
Introduction to
Basic and Advanced
Multilevel Modeling

Tom A B SNIJDERS
Roel J BOSKER
© Tom A B Snijders and Roel J Bosker 2012

First edition published 1999

Reprinted 2002, 2003, 2004

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as
permitted under the Copyright, Designs and Patents Act, 1988, this publication may be reproduced,
stored or transmitted in any form, or by any means, only with the prior permission in writing of the
publishers, or in the case of reprographic reproduction, in accordance with the terms of licences
issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside those terms
should be sent to the publishers.

SAGE Publications Ltd

1 Oliver’s Yard
55 City Road
London EC1Y 1SP

SAGE Publications Inc.

2455 Teller Road
Thousand Oaks, California 91320

SAGE Publications India Pvt Ltd

B 1/I 1 Mohan Cooperative Industrial Area
Mathura Road
New Delhi 110 044

SAGE Publications Asia-Pacific Pte Ltd

33 Pekin Street #02-01
Far East Square
Singapore 048763

Library of Congress Control Number: 2011926498

British Library Cataloguing in Publication data

A catalogue record for this book is available from the British Library

ISBN 978-1-84920-200-8
ISBN 978-1-84920-201-5

Typeset by C&M Digitals (P) Ltd, Chennai, India

Printed by MPG Books Group, Bodmin, Cornwall
Printed on paper from sustainable resources
Contents

Preface to the Second Edition

Preface to the First Edition

1 Introduction
1.1 Multilevel analysis
1.1.1 Probability models
1.2 This book
1.2.1 Prerequisites
1.2.2 Notation

2 Multilevel Theories, Multistage Sampling, and Multilevel Models

2.1 Dependence as a nuisance
2.2 Dependence as an interesting phenomenon
2.3 Macro-level, micro-level, and cross-level relations
2.4 Glommary

3 Statistical Treatment of Clustered Data

3.1 Aggregation
3.2 Disaggregation
3.3 The intraclass correlation
3.3.1 Within-group and between-group variance
3.3.2 Testing for group differences
3.4 Design effects in two-stage samples
3.5 Reliability of aggregated variables
3.6 Within- and between-group relations
3.6.1 Regressions
 3.6.2 Correlations
3.6.3 Estimation of within- and between-group correlations
3.7 Combination of within-group evidence
3.8 Glommary

4 The Random Intercept Model

4.1 Terminology and notation
4.2 A regression model: fixed effects only
4.3 Variable intercepts: fixed or random parameters?
4.3.1 When to use random coefficient models
4.4 Definition of the random intercept model
4.5 More explanatory variables
4.6 Within- and between-group regressions
4.7 Parameter estimation
4.8 ‘Estimating’ random group effects: posterior means
4.8.1 Posterior confidence intervals
4.9 Three-level random intercept models
4.10 Glommary

5 The Hierarchical Linear Model

5.1 Random slopes
5.1.1 Heteroscedasticity
5.1.2 Do not force τ01 to be 0!
5.1.3 Interpretation of random slope variances
5.2 Explanation of random intercepts and slopes
5.2.1 Cross-level interaction effects
5.2.2 A general formulation of fixed and random parts
5.3 Specification of random slope models
5.3.1 Centering variables with random slopes?
5.4 Estimation
5.5 Three or more levels
5.6 Glommary

6 Testing and Model Specification

 6.1 Tests for fixed parameters
6.1.1 Multiparameter tests for fixed effects
6.2 Deviance tests
6.2.1 More powerful tests for variance parameters
6.3 Other tests for parameters in the random part
6.3.1 Confidence intervals for parameters in the random part
6.4 Model specification
6.4.1 Working upward from level one
6.4.2 Joint consideration of level-one and level-two variables
6.4.3 Concluding remarks on model specification
6.5 Glommary

7 How Much Does the Model Explain?

7.1 Explained variance
7.1.1 Negative values of R2?
7.1.2 Definitions of the proportion of explained variance in two-
level models
7.1.3 Explained variance in three-level models
7.1.4 Explained variance in models with random slopes
7.2 Components of variance
7.2.1 Random intercept models
7.2.2 Random slope models
7.3 Glommary

8 Heteroscedasticity
8.1 Heteroscedasticity at level one
8.1.1 Linear variance functions
8.1.2 Quadratic variance functions
8.2 Heteroscedasticity at level two
8.3 Glommary

9 Missing Data
9.1 General issues for missing data
9.1.1 Implications for design
 9.2 Missing values of the dependent variable
9.3 Full maximum likelihood
9.4 Imputation
9.4.1 The imputation method
9.4.2 Putting together the multiple results
9.5 Multiple imputations by chained equations
9.6 Choice of the imputation model
9.7 Glommary

10 Assumptions of the Hierarchical Linear Model

10.1 Assumptions of the hierarchical linear model
10.2 Following the logic of the hierarchical linear model
10.2.1 Include contextual effects
10.2.2 Check whether variables have random effects
10.2.3 Explained variance
10.3 Specification of the fixed part
10.4 Specification of the random part
10.4.1 Testing for heteroscedasticity
10.4.2 What to do in case of heteroscedasticity
10.5 Inspection of level-one residuals
10.6 Residuals at level two
10.7 Influence of level-two units
10.8 More general distributional assumptions
10.9 Glommary

11 Designing Multilevel Studies

11.1 Some introductory notes on power
11.2 Estimating a population mean
11.3 Measurement of subjects
11.4 Estimating association between variables
11.4.1 Cross-level interaction effects
11.5 Allocating treatment to groups or individuals
11.6 Exploring the variance structure
11.6.1 The intraclass correlation
 11.6.2 Variance parameters
11.7 Glommary

12 Other Methods and Models

12.1 Bayesian inference
12.2 Sandwich estimators for standard errors
12.3 Latent class models
12.4 Glommary

13 Imperfect Hierarchies
13.1 A two-level model with a crossed random factor
13.2 Crossed random effects in three-level models
13.3 Multiple membership models
13.4 Multiple membership multiple classification models
13.5 Glommary

14 Survey Weights
14.1 Model-based and design-based inference
14.1.1 Descriptive and analytic use of surveys
14.2 Two kinds of weights
14.3 Choosing between model-based and design-based analysis
14.3.1 Inclusion probabilities and two-level weights
14.3.2 Exploring the informativeness of the sampling design
14.4 Example: Metacognitive strategies as measured in the PISA study
14.4.1 Sampling design
14.4.2 Model-based analysis of data divided into parts
14.4.3 Inclusion of weights in the model
14.5 How to assign weights in multilevel models
14.6 Appendix. Matrix expressions for the single-level estimators
14.7 Glommary

15 Longitudinal Data
15.1 Fixed occasions
15.1.1 The compound symmetry model
15.1.2 Random slopes
 15.1.3 The fully multivariate model
15.1.4 Multivariate regression analysis
15.1.5 Explained variance
15.2 Variable occasion designs
15.2.1 Populations of curves
15.2.2 Random functions
15.2.3 Explaining the functions
15.2.4 Changing covariates
15.3 Autocorrelated residuals
15.4 Glommary

16 Multivariate Multilevel Models

16.1 Why analyze multiple dependent variables simultaneously?
16.2 The multivariate random intercept model
16.3 Multivariate random slope models
16.4 Glommary

17 Discrete Dependent Variables

17.1 Hierarchical generalized linear models
17.2 Introduction to multilevel logistic regression
17.2.1 Heterogeneous proportions
17.2.2 The logit function: Log-odds
17.2.3 The empty model
17.2.4 The random intercept model
17.2.5 Estimation
17.2.6 Aggregation
17.3 Further topics on multilevel logistic regression
17.3.1 Random slope model
17.3.2 Representation as a threshold model
17.3.3 Residual intraclass correlation coefficient
17.3.4 Explained variance
17.3.5 Consequences of adding effects to the model
17.4 Ordered categorical variables
 17.5 Multilevel event history analysis
17.6 Multilevel Poisson regression
17.7 Glommary

18 Software
18.1 Special software for multilevel modeling
18.1.1 HLM
18.1.2 MLwiN
18.1.3 The MIXOR suite and SuperMix
18.2 Modules in general-purpose software packages
18.2.1 SAS procedures VARCOMP, MIXED, GLIMMIX, and
NLMIXED
18.2.2 R
18.2.3 Stata
18.2.4 SPSS, commands VARCOMP and MIXED
18.3 Other multilevel software
18.3.1 PinT
18.3.2 Optimal Design
18.3.3 MLPowSim
18.3.4 Mplus
18.3.5 Latent Gold
18.3.6 REALCOM
18.3.7 WinBUGS

References

Index
Preface to the Second Edition

Multilevel analysis is a domain of data analysis that has been developing

strongly before as well as after the publication of the first edition of our
book in 1999. This second edition has been seriously revised. It contains
more material, it has been updated and corrected, and a number of
explanations were clarified.
The main new material is in three new chapters. A chapter was added
about missing data, and another about the use of multilevel modeling for
surveys with nonconstant inclusion probabilities (‘survey weights’). Also a
chapter was added in which three special techniques are briefly treated:
Bayesian estimation, cluster-robust standard errors (the socalled sandwich
standard error), and latent class models. The topics of these new chapters all
belong to the ‘advanced’ part of the title. Among what was not covered in
the first edition, these are some of the topics which we believe are most
frequently encountered in the practice of multilevel research.
New material has been added also to existing chapters. The main
example (starting in Chapter 4) has been renewed because the treatment of
missing data in the old version was inadequate. Various other new examples
also have been added. Further, there now is a more elaborate treatment of
the combination of within-group evidence without using full-blown
multilevel modeling (Section 3.7); more detailed considerations are
discussed for choosing between fixed and random effects models (Section
4.3); diagnostic and comparative standard errors of posterior means are
explained (Section 4.8.1); the treatment of tests for parameters of the
random part was corrected and confidence intervals for these parameters are
discussed (Sections 6.2 and 6.3); multiple membership models are treated in
Chapter 13; and there has been an overhaul of the treatment of estimation
methods for hierarchical generalized linear models in Chapter 17. Chapter
18 about multilevel software was totally rewritten, keeping it relatively
short because this is the part of any textbook that ages most rapidly.
Throughout all chapters new developments have been mentioned, pointers
are given to the recent literature, various difficulties now are explained in
more elaborate ways, and errors have been corrected.
All chapters (from the second on) now start by an overview, and are
concluded (except for the last) by a ‘glommary’. As every reader will know
after reading this book, this is a summary of the main concepts treated in
the chapter in a form akin to a glossary. Our intention is that these new
elements will improve the didactical qualities of this textbook. Having said
this, we think that the understanding of the book, or parts of it, may be
further enhanced by going through the examples using the data that we
made available (as far as this was allowed) at http://www.stats.ox.ac.uk/
∼snijders/mlbook.htm. This website will also contain our comments on
remarks made on the book by industrious readers, as well as our corrections
for errors if any will be discovered.
We are very grateful for stimulating discussions (over the years or in the
recent period of revising the text), comments on drafts of chapters, and help
with software, to many people: Marnie Bertolet, sir David Cox, Roel de
Jong, Jon Fahlander, Mark Huisman, Johan Koskinen, Catalina Lomos,
Mayra Mascareño, Paulina Preciado, Roy Stewart, Anneke Timmermans,
and Marijtje van Duijn. For help with new data sets we are grateful to some
of these people and also to Hennie Brandsma, Simone Doolaard, Sonja
Drobnič, Anja Knuver, Hans Kuyper, Sascha Peter, Stijn Ruiter, Greetje van
der Werf, and Frank van Tubergen.

Tom Snijders
Roel Bosker
March,2011
Preface to the First Edition

This book grew out of our teaching and consultation activities in the
domain of multilevel analysis. It is intended for the absolute beginner in this
field as well as for those who have already mastered the fundamentals and
are now entering more complicated areas of application. The reader is
referred to Section 1.2 for an overview of this book and for some reading
guidelines.
We are grateful to various people from whom we got reactions on
earlier parts of this manuscript and also to the students who were exposed to
it and helped us realize what was unclear. We received useful comments
from, and benefited from discussions about parts of the manuscript with,
among others, Joerg Blasius, Marijtje van Duijn, Wolfgang Langer, Ralf
Maslowski, and Ian Plewis. Moreover we would like to thank Hennie
Brandsma, Mieke Brekelmans, Jan van Damme, Hetty Dekkers, Miranda
Lubbers, Lyset Rekers-Mombarg and Jan Maarten Wit, Carolina de Weerth,
Beate Völker, Ger van der Werf, and the Zentral Archiv (Cologne) who
kindly permitted us to use data from their respective research projects as
illustrative material for this book. We would also like to thank Annelies
Verstappen-Remmers for her unfailing secretarial assistance.

Tom Snijders
Roel Bosker
June, 1999
 1
Introduction

1.1 Multilevel analysis

Multilevel analysis is a methodology for the analysis of data with complex
patterns of variability, with a focus on nested sources of such variability –
pupils in classes, employees in firms, suspects tried by judges in courts,
animals in litters, longitudinal measurements of subjects, etc. In the analysis
of such data, it is usually illuminating to take account of the fact that each
level of nesting is associated with variability that has a distinct
interpretation. There is variability, for example, between pupils but also
between classes, and one may draw incorrect conclusions if no distinction is
made between these different sources of variability. Multilevel analysis is
an approach to the analysis of such data, including the statistical techniques
as well as the methodology for their use. The term ‘multilevel analysis’ is
used mainly in the social sciences (in the wide sense: sociology, education,
psychology, economics, criminology, etc.), but also in other fields such as
the biomedical sciences. Our focus will be on the social sciences. Other
terms, referring to the technical aspects, are hierarchical linear models,
mixed models, and random coefficient models.
In its present form, multilevel analysis is a stream which has two
tributaries: contextual analysis and mixed effects models. Contextual
analysis is a development in the social sciences which has focused on the
effects of the social context on individual behavior. Some landmarks before
1980 are the paper by Robinson (1950) which discussed the ecological
fallacy (which refers to confusion between aggregate and individual
effects), the paper by Davis et al. (1961) on the distinction between within-
group and between-group regression, the volume edited by Dogan and
Rokkan (1969), and the paper by Burstein et al. (1978) on treating
regression intercepts and slopes on one level as outcomes on the higher
level.
Mixed effects models are statistical models in the analysis of variance
(ANOVA) and in regression analysis where it is assumed that some of the
coefficients are fixed and others are random. This subject is too vast even to
mention some landmarks. A standard reference book on random effects
models and mixed effects models is Searle et al. (1992), Chapter 2 of which
gives an extensive historical overview. The name ‘mixed model’ seems to
have been used first by Eisenhart (1947).
Contextual modeling until about 1980 focused on the definition of
appropriate variables to be used in ordinary least squares regression
analysis. Until the 1980s the main focus in the development of statistical
procedures for mixed models was on random effects (i.e., random
differences between categories in some classification system) more than on
random coefficients (i.e., random effects of numerical variables). Multilevel
analysis as we now know it was formed by these two streams coming
together. It was realized that, in contextual modeling, the individual and the
context are distinct sources of variability, which should both be modeled as
random influences. On the other hand, statistical methods and algorithms
were developed that allowed the practical use of regression-type models
with nested random coefficients. There was a cascade of statistical papers:
Aitkin et al. (1981), Laird and Ware (1982), Mason et al. (1983), Goldstein
(1986), Aitkin and Longford (1986), Raudenbush and Bryk (1986), de
Leeuw and Kreft (1986), and Longford (1987) proposed and developed
techniques for calculating estimates for mixed models with nested
coefficients. These techniques, together with the programs implementing
them which were developed by a number of these researchers or under their
supervision, allowed the practical use of models of which until that moment
only special cases were accessible for practical use. By 1986 the basis of
multilevel analysis was established. The first textbook appeared (by
Goldstein, now in its fourth edition) and was followed by a few others in
the 1990s and many more in the 2000s. The methodology has been further
elaborated since then, and has proved to be quite fruitful in applications in
many fields. On the organizational side, there are stimulating centers such
as the ‘Multilevel Models Project’ in Bristol with its Newsletter and its
website http://www.mlwin.com/, and there is an active internet discussion
group at http://www.jiscmail.ac.uk/lists/multilevel.html.
In the biomedical sciences mixed models were proposed especially for
longitudinal data; in economics mainly for panel data (Swamy, 1971), the
most common longitudinal data in economics. One of the issues treated in
the economics literature was the pooling of cross-sectional and time series
data (e.g., Maddala, 1971; Hausman and Taylor, 1981), which is closely
related to the difference between within-group and between-group
regressions. Overviews are given by Chow (1984) and Baltagi (2008).
A more elaborate history of multilevel analysis is presented in the
bibliographical sections of Longford (1993) and in Kreft and de Leeuw
(1998). For an extensive bibliography of the older literature, see HÜttner
and van den Eeden (1995). A more recent overview of much statistical
work in this area can be found in the handbook by de Leeuw and Meijer
(2008a).

1.1.1 Probability models

The main statistical model of multilevel analysis is the hierarchical linear
model, an extension of multiple linear regression to a model that includes
nested random coefficients. This model is explained in Chapter 5 and forms
the basis of most of this book.
There are several ways to argue why it makes sense to use a probability
model for data analysis. In sampling theory a distinction is made between
design-based inference and model-based inference. This is discussed further
in Chapter 14. The former means that the researcher draws a probability
sample from some finite population, and wishes to make inferences from
the sample to this finite population. The probability model then follows
from how the sample is drawn by the researcher. Model-based inference
means that the researcher postulates a probability model, usually aiming at
inference to some large and sometimes hypothetical population such as all
English primary school pupils in the 2000s or all human adults living in a
present-day industrialized culture. If the probability model is adequate then
so are the inferences based on it, but checking this adequacy is possible
only to a limited extent.
It is possible to apply model-based inference to data collected by
investigating some entire research population, such as all 12-year-old pupils
in Amsterdam at a given moment. Sometimes the question arises as to why
one should use a probability model if no sample is drawn but an entire
population is observed. Using a probability model that assumes statistical
variability, even though an entire research population was investigated, can
be justified by realizing that conclusions are sought which apply not only to
the investigated research population but also to a wider population. The
investigated research population is assumed to be representative of this
wider population – for pupils who are older or younger, in other towns,
perhaps in other countries. This is called a superpopulation in Chapter 14,
where the relation between model-based and design-based inference is
further discussed. Applicability of the statistical inference to such a wider
population is not automatic, but has to be carefully argued by considering
whether indeed the research population may be considered to be
representative of the larger (often vaguely outlined) population. This is the
‘second span of the bridge of statistical inference’ as discussed by Cornfield
and Tukey (1956).1 The inference then is not primarily about a given
delimited set of individuals but about social, behavioral, biological, etc.,
mechanisms and processes. The random effects, or residuals, playing a role
in such probability models can be regarded as resulting from the factors that
are not included in the explanatory variables used. They reflect the
approximative nature of the model used. The model-based inference will be
adequate to the extent that the assumptions of the probability model are an
adequate reflection of the effects that are not explicitly included by means
of observed variables.
As we shall see in Chapters 3–5, the basic idea of multilevel analysis is
that data sets with a nesting structure that includes unexplained variability
at each level of nesting, such as pupils in classes or employees in firms, are
usually not adequately represented by the probability model of multiple
linear regression analysis, but are often adequately represented by the
hierarchical linear model. Thus, the use of the hierarchical linear model in
multilevel analysis is in the tradition of model-based inference.

1.2 This book

This book is intended as an introductory textbook and as a reference book
for practical users of multilevel analysis. We have tried to include all the
main points that come up when applying multilevel analysis. Most of the
data sets used in the examples, and corresponding commands to run the
examples in various computer programs (see Chapter 18), are available on
the website http://www.stats.ox.ac.uk/~snijders/mlbook.htm.
After this introductory chapter, the book proceeds with a conceptual
chapter about multilevel questions and a chapter on ways to treat multilevel
data that are not based on the hierarchical linear model. Chapters 4–6 treat
the basic conceptual ideas of the hierarchical linear model, and how to work
with it in practice. Chapter 4 introduces the random intercept model as the
primary example of the hierarchical linear model. This is extended in
Chapter 5 to random slope models. Chapters 4 and 5 focus on
understanding the hierarchical linear model and its parameters, paying only
very limited attention to procedures and algorithms for parameter
estimation (estimation being work that most researchers delegate to the
computer). Chapter 6 is concerned with testing parameters and specifying a
multilevel model.
An introductory course on multilevel analysis could cover Chapters 1–6
and Section 7.1, with selected material from other chapters. A minimal
course would focus on Chapters 4–6. The later chapters cover topics that
are more specialized or more advanced, but important in the practice of
multilevel analysis.
The text of this book is not based on a particular computer program for
multilevel analysis. The last chapter, 18, gives a brief review of computer
programs that can be used for multilevel analysis.
Chapters 7 (on the explanatory power of the model) and 10 (on model
assumptions) are important for the interpretation of the results of statistical
analyses using the hierarchical linear model. Researchers who have data
sets with many missing values, or who plan to collect data sets that may run
this type of risk, will profit from reading Chapter 9. Chapter 11 helps the
researcher in setting up a multilevel study, and in choosing sample sizes at
the various levels.
Some multilevel data sets come from surveys done according to a
complex design, associated with survey weights reflecting the
undersampling and oversampling of parts of the population. Ways to
analyze such data sets using the hierarchical linear model are covered in
Chapter 14.
 Several methods and models have been developed that can sometimes
be useful as additions or alternatives to the more commonly used methods
for the hierarchical linear model. Chapter 12 briefly presents three of these:
Bayesian procedures, the sandwich estimator for standard errors, and latent
class models.
Chapters 8 and 13–17 treat various extensions of the basic hierarchical
linear model that are useful in practical research. The topic of Chapter 8,
heteroscedasticity (nonconstant residual variances), may seem rather
specialized. Modeling heteroscedasticity, however, is easily done within the
hierarchical linear model and can be very useful. It also allows model
checks and model modifications that are used in Chapter 10. Chapter 13
treats data structures where the different sources of variability, the ‘levels’
of the multilevel analysis, are not nested but related in different ways:
crossed classifications and multiple memberships. Chapter 15 is about
longitudinal data, with a fixed occasion design (i.e., repeated measures
data) as well as those with a variable occasion design, where the time
moments of measurement may differ arbitrarily between subjects. This
chapter indicates how the flexibility of the multilevel model gives important
opportunities for data analysis (e.g., for incomplete multivariate or
longitudinal data) that were unavailable earlier. Chapter 16 is about
multilevel analysis for multivariate dependent variables. Chapter 17
describes the multilevel modeling of dichotomous, ordinal, and frequency
data.
Each chapter starts with an overview and finishes with a summarizing
glossary, which we have called a glommary. The glommaries can be
consulted to gain rapid overviews of what is treated in the various chapters.
If additional textbooks are sought, one could consider the excellent
introductions by Hox (2010) and Gelman and Hill (2007); Raudenbush and
Bryk (2002), for an elaborate treatment of the hierarchical linear model;
Longford (1993), Goldstein (2011), Demidenko (2004), and de Leeuw and
Meijer (2008a) for more detailed mathematical background; and Skrondal
and Rabe-Hesketh (2004) for further modeling, especially latent variable
models.

1.2.1 Prerequisites
In order to read this textbook, a good working knowledge of statistics is
required. It is assumed that you know the concepts of probability, random
variable, probability distribution, population, sample, statistical
independence, expectation (population mean), variance, covariance,
correlation, standard deviation, and standard error. Furthermore, it is
assumed that you know the basics of hypothesis testing and multiple
regression analysis, and that you can understand formulas of the kind that
occur in the explanation of regression analysis.
Matrix notation is used only in a few more advanced sections. These
sections can be skipped without loss of understanding of other parts of the
book.

1.2.2 Notation
The main notational conventions are as follows. Abstract variables and
random variables are denoted by italicized capital letters, such as X or Y.
Outcomes of random variables and other fixed values are denoted by
italicized lower-case letters, such as x or z. Thus we speak about the
variable X, but in formulas where the value of this variable is considered as
a fixed, nonrandom value, it will be denoted by x. There are some
exceptions to this, for example in Chapter 2 and in the use of the letter N for
the number of groups (‘level-two units’) in the data.
The letter ε is used to denote the expected value, or population average,
of a random variable. Thus, εY and ε(Y) denote the expected value of Y. For
example, if Pn is the fraction of tails obtained in n coin flips, and the coin is
fair, then the expected value is
Statistical parameters are indicated by Greek letters. Examples are μ, σ2,
and β. The following Greek letters are used.

α alpha
β beta
γ gamma
δ delta
η eta
θ theta
λ lambda
 μ mu
π pi
ρ rho
σ sigma
τ tau
φ phi
χ chi
ω omega
∆ capital Delta
Σ capital Sigma
τ capital Tau
χ capital Chi

1
We are indebted to Ivo Molenaar for this reference.
 2
Multilevel Theories, Multistage
Sampling, and Multilevel Models

Phenomena and data sets in the social sciences often have a multilevel
structure. This may be reflected in the design of data collection: simple
random sampling is often not a very cost-efficient strategy, and multistage
samples may be more efficient instead. This chapter is concerned with the
reasons why it is important to take account of the clustering of the data, also
called their multilevel structure, in the data analysis phase.

OVERVIEW OF THE CHAPTER

First we discuss how methods of inference failing to take into account the
multilevel data structure may lead to erroneous conclusions, because
independence assumptions are likely to be violated. The next two sections
sketch the reasons for interest in a multilevel approach from the
applications point of view. In many cases the multilevel data structure
reflects essential aspects of the social (biological, etc.) world, and important
research questions can be formulated about relations between variables at
different layers in a hierarchical system. In this case the dependency of
observations within clusters is of focal interest, because it reflects the fact
that clusters differ in certain respects. In either case, the use of single-level
statistical models is no longer valid. The fallacies to which their use can
lead are described in the next chapter.

2.1 Dependence as a nuisance

Textbooks on statistics tell us that observations should be sampled
independently of each other as standard. Thus the standard sampling design
on which statistical models are based is simple random sampling with
replacement from an infinite population: the result of one selection is
independent of the result of any other selection, and all single units in the
population have the same chances of being selected into the sample.
Textbooks on sampling, however, make it clear that there are more cost-
efficient sampling designs, based on the idea that probabilities of selection
should be known but do not have to be constant. One of those cost-efficient
sampling designs is the multistage sample: the population of interest
consists of subpopulations, also called clusters, and selection takes place
via those subpopulations.
If there is only one subpopulation level, the design is a two-stage
sample. Pupils, for instance, are grouped in schools, so the population of
pupils consists of subpopulations of schools that contain pupils. Other
examples are: families in neighborhoods, teeth in jawbones, animals in
litters, employees in firms, and children in families. In a random two-stage
sample, a random sample of the primary units (schools, neighborhoods,
jawbones, litters, firms, families) is taken in the first stage, and then the
secondary units (pupils, families, teeth, animals, employees, children) are
sampled at random from the selected primary units in the second stage. A
common mistake in research is to ignore the fact that the sampling scheme
was a two-stage one, and to pretend that the secondary units were selected
independently. The mistake in this case would be that the researcher is
overlooking the fact that the secondary units were not sampled
independently of each other: having selected a primary unit (e.g., a school)
increases the chances of selection of secondary units (e.g., pupils) from that
primary unit. In other words, the multistage sampling design leads to
dependent observations, and failing to deal with this properly in the
statistical analysis may lead to erroneous inferences. An example of the
grossly inflated type I error rates that may then occur is given by Dorman
(2008).
The multistage sampling design can be depicted graphically as in Figure
2.1. This shows a population that consists of 10 subpopulations, each
containing 10 micro-units. A sample of 25% is taken by randomly selecting
5 out of 10 subpopulations and within these – again at random of course – 5
out of 10 micro-units.
 Figure 2.1: Multistage sample.

Multistage samples are preferred in practice, because the costs of

interviewing or testing persons are reduced enormously if these persons are
geographically or organizationally grouped. Such sample designs
correspond to the organization of the social world. It is cheaper to travel to
100 neighbourhoods and interview 10 persons per neighbourhood on their
political preferences than to travel to 1,000 neighbourhoods and interview
one person per neighbourhood. In the next chapters we will see how we can
make adjustments to deal with these dependencies.

2.2 Dependence as an interesting phenomenon

The previous section implies that, if we want to make inferences on, for
example, the earnings of employees in the for-profit sector, it is cost-
efficient to use a multistage sampling design in which employees are
selected via the firms in which they work. A common feature in social
research, however, is that in many cases we wish to make inferences on the
firms as well as on the employees. Questions that we seek to answer may
be: Do employees in multinationals earn more than employees in other
firms? Is there a relation between the performance of pupils and the
experience of their teacher? Is the sentence differential between black and
white suspects different between judges, and if so, can we find
characteristics of judges to which this sentence differential is related? In
this case a variable is defined at the primary unit level (firms, teachers,
judges) as well as at the secondary unit level (employees, pupils, cases).
Henceforth we will refer to primary units as macro-level units (or macro-
units for short) and to secondary units as micro-level units (or micro-units
for short). The micro level is called the lower level (first) and the macro
level is called the higher level (second). For the time being, we will restrict
ourselves to the two-level case, and thus to two-stage samples only. Table
2.1 gives a summary of the terminology.

Table 2.1: Summary of terms to describe units at either level in the two-
level case.

macro-level units micro-level units

macro-units micro-units
primary units secondary units
clusters elementary units
level-two units level-one units

Examples of macro-units and the micro-units nested within them are

presented in Table 2.2. Most of the examples presented in the table have
been dealt with in the text already. It is important to note that what is
defined as a macro-unit or a micro-unit depends on the theory at hand.
Teachers are nested within schools, if we study organizational effects on
teacher burn-out then teachers are the micro-units and schools the macro-
units. But when studying teacher effects on student achievement, teachers
are the macro-units and students the micro-units. The same goes, mutatis
mutandis, for neighborhoods and families (e.g., when studying the effects of
housing conditions on marital problems), and for families and children
(e.g., when studying effects of income on educational performance of
siblings).
In all these instances the dependency of the observations on the micro-
units within the macro-units is of focal interest. If we stick to the example
of schools and pupils, then the dependency (e.g., in mathematics
achievement of pupils within a school) may stem from:

Table 2.2: Some examples of units at the macro and micro level.
 Macro level Micro level
schools teachers
classes pupils
neighbourhoods families
firms employees
jawbones teeth
families children
litters animals
doctors patients
subjects measurements
interviewers respondents
judges suspects

1. pupils within a school sharing the same school environment;

2. pupils within a school sharing the same teachers;
pupils within a school affecting each other by direct communication
3.
or shared group norms;
4. pupils within a school coming from the same neighborhood.

The more the achievement levels of pupils within a school are alike (as
compared to pupils from other schools), the more likely it is that causes of
the achievement have to do with the organizational unit (in this case, the
school). Absence of dependency in this case implies absence of institutional
effects on individual performance.
A special kind of nesting is defined by longitudinal data, represented in
Table 2.2 as ‘measurements within subjects’. The measurement occasions
here are the micro-units and the subjects the macro-units. The dependence
of the different measurements for a given subject is of primary importance
in longitudinal data, but the following section on relations between
variables defined at either level is not directly intended for the nesting
structure defined by longitudinal data. Because of the special nature of this
nesting structure, Chapter 15 is specifically devoted to it.
The models treated in this book are for situations where the dependent
variable is at the lowest level. For models with nested data sets where the
dependent variable is defined at a higher level one may consult Croon and
van Veldhoven (2007), LÜdtke et al. (2008), and van Mierlo et al. (2009).

2.3 Macro-level, micro-level, and cross-level

relations
In the study of hierarchical or multilevel systems, Lazarsfeld and Menzel
(1971) made important distinctions between properties and propositions
connected to the different levels. In his summary of this work, Tacq (1986)
distinguished between three kinds of propositions: on micro-units (e.g.,
‘employees have on average 4 effective working hours per day’; ‘boys lag
behind girls in reading comprehension’), on macro-units (e.g., ‘schools
have on average a budget of $20,000 to spend on resources’; ‘in
neighborhoods with bad housing conditions crime rates are above
average’), or on macro–micro relations (e.g., ‘if firms have a salary bonus
system, the productivity of employees will be greater’; ‘a child suffering
from a broken family situation will affect the climate in the classroom’).
Multilevel statistical models are always1 called for if we are interested
in propositions that connect variables defined at different levels, the micro
and the macro, and also if a multistage sample design has been employed.
The use of such a sampling design is quite obvious if we are interested in
macro–micro relations, less obvious (but often necessary from a cost-
effectiveness point of view) if micro-level propositions are our primary
concern, and hardly obvious at all (but sometimes still applicable) if macro-
level propositions are what we are focusing on. These three instances will
be discussed below. To facilitate comprehension, following Tacq (1986) we
use figures with the following conventions: a dotted line indicates that there
are two levels; below the line is the micro level; above the line is the macro
level; macro-level variables are denoted by capitals; micro-level variables
are denoted by lower-case letters; and arrows denote presumed causal
relations.

Multilevel propositions
Multilevel propositions can be represented as in Figure 2.2. In this example
we are interested in the effect of the macro-level variable Z (e.g., teacher
efficacy) on the micro-level variable y (e.g., pupil motivation), controlling
for the micro-level variable x (e.g., pupil aptitude).

Figure 2.2: The structure of a multilevel proposition.

Micro-level propositions
Micro-level propositions are of the form indicated in Figure 2.3. In this case
the line indicates that there is a macro level which is not referred to in the
hypothesis that is put to the test, but which is used in the sampling design in
the first stage. In assessing the strength of the relation between occupational
status and income, for instance, respondents may have been selected for
face-to-face interviews by zip-code area. This then may cause dependency
(as a nuisance) in the data.

Figure 2.3: The structure of a micro-level proposition.

Macro-level propositions
Macro-level propositions are of the form of Figure 2.4. The line separating
the macro level from the micro level seems superfluous here. When
investigating the relation between the long-range strategic planning policy
of firms and their profits, there is no multilevel situation, and a simple
random sample may have been taken. When either or both variables are not
directly observable, however, and have to be measured at the micro level
(e.g., organizational climate measured as the average satisfaction of
employees), then a two-stage sample is needed nevertheless. This is the
case a fortiori for variables defined as aggregates of micro-level variables
(e.g., the crime rate in a neighborhood).

Figure 2.4: The structure of a macro-level proposition.

Macro–micro relations
The most common situation in social research is that macro-level variables
are supposed to have a relation with micro-level variables. There are three
obvious instances of macro-to-micro relations, all of which are typical
examples of the multilevel situation (see Figure 2.5). The first case is the
macro-to-micro proposition. The more explicit the religious norms in social
networks, for example, the more conservative the views that individuals
have on contraception. The second proposition is a special case of this. It
refers to the case where there is a relation between Z and y, given that the
effect of x on y is taken into account. The example given may be modified
to: ‘for individuals of a given educational level’. The last case in the figure
is the macro–micro-interaction, also known as the cross-level interaction:
the relation between x and y is dependent on Z. To put it another way, the
relation between Z and y is dependent on x. The effect of aptitude on
achievement, for instance, may be small in case of ability grouping of
pupils within classrooms but large in ungrouped classrooms.
Next to these three situations there is the so-called emergent, or micro–
macro, proposition (Figure 2.6). In this case, a micro-level variable x affects
a macro-level variable Z (student achievement may affect teachers’
experience of stress).
 Figure 2.5: The structure of macro–micro propositions.

Figure 2.6: The structure of a micro–macro proposition.

It is of course possible to form combinations of the various examples

given. Figure 2.7 contains a causal chain that explains through which
micro-variables there is an association between the macro-level variables W
and Z (cf. Coleman, 1990). As an example of this chain, we may be
interested in why the qualities of a football coach affect his social prestige.
The reason is that good coaches are capable of motivating their players,
thus leading the players to good performance, thus to winning games, and
this of course leads to more social prestige for the coach. Another instance
of a complex multilevel proposition is the contextual effects proposition.
For example, low socio-economic status pupils achieve less in classrooms
with a low average aptitude. This is also a cross-level interaction effect, but
the macro-level variable, average aptitude in the classroom, is now an
aggregate of a micro-level variable.

Figure 2.7: A causal macro–micro–micro–macro chain.

The methodological advances in multilevel modeling are now also

leading to theoretical advances in contextual research: suitable definitions
of context and ‘levels’, meaningful ways of aggregating variables to higher
levels, conceptualizing and analyzing the interplay between characteristics
of lower- and higher-level units. Some examples in various disciplines are
the following. Following up on the initial work of Hauser (1970, 1974), in
which he stated that group composition effects may be artifacts of
underspecification of the micro-level model, Harker and Tymms (2004)
discuss the issue of group composition effects in education. Sampson et al.
(2002) give a review of theoretical work in the analysis of neighborhood
effects. Diez-Roux (2000), Blakely and Woodward (2000), and O’Campo
(2003) comment on advances along these lines in epidemiology and public
health.
In the next chapters the statistical tools to handle multilevel structures
will be introduced for outcome variables defined at the micro level.

2.4 Glommary
Multilevel data structures. Many data sets in the social sciences have a
multilevel structure, that is, they constitute hierarchically nested
systems with multiple levels. Much of our discussion focuses on two-
level structures, but this can be generalized to three or more nested
levels.

Sampling design. Often the multilevel nature of the social world leads to
the practical efficiency of multistage samples. The population then
consists of a nested system of subpopulations, and a nested sample is
drawn accordingly. For example, when employing a random two-stage
sample design, in the first stage a random sample of the primary units
is taken, and in the second stage the secondary units are sampled at
random from the selected primary units.

Levels. The levels are numbered such that the most detailed level is the
first. For example, in a two-level structure of individuals nested in
groups the individuals are called level-one units and the groups level-
two units. (Note the different terminology compared to the words used
in theories of survey sampling: in the preceding example, the primary
units are the level-two units and the secondary units the level-one
units.)

Units. The elements of a level are called units. Higher-level units are also
called clusters. We talk about level-one units, level-two units, etc.
Dependence as a nuisance. Not taking account of the multilevel data
structure, or of the multistage sampling design, is likely to lead to the
use of statistical procedures in which independence assumptions are
violated so that conclusions may be unfounded.

Dependence as an interesting phenomenon. The importance of the

multilevel structure of social (biological, etc.) reality implies that
research can often become more interesting when it takes account of
the multilevel structure.

Multilevel propositions. Illustrations were given of scientific propositions

involving multiple levels: micro-level propositions, macro-level
propositions, macro–micro relations, cross-level interaction, and
emergent propositions or micro–macro relations.

1
As with any rule, there are exceptions. If the data set is such that for each macro-unit only one
micro-unit is included in the sample, single-level methods still can be used.
 3
Statistical Treatment of Clustered
Data

Before proceeding in the next chapters to the explanation of the hierarchical

linear model, the main statistical model for multilevel analysis, this chapter
looks at approaches to analyzing multilevel data sets that are more
elementary and do not use this model.

OVERVIEW OF THE CHAPTER

The chapter starts by considering what will happen if we ignore the
multilevel structure of the data. Are there any instances where one may
proceed with single-level statistical models although the data stem from a
multistage sampling design? What kind of errors – so-called ecological
fallacies – may occur when this is done? The rest of the chapter is devoted
to some statistical methods for multilevel data that attempt to uncover the
role played by the various levels without fitting a full-blown hierarchical
linear model. First, we describe the intraclass correlation coefficient, a basic
measure for the degree of dependency in clustered observations. Second,
some simple statistics (mean, standard error of the mean, variance,
correlation, reliability of aggregates) are treated for two-stage sampling
designs. To avoid ecological fallacies it is essential to distinguish within-
group from between-group regressions. These concepts are explained, and
the relations are spelled out between within-group, between-group, and total
regressions, and similarly for correlations. Finally, we mention some simple
methods for combining evidence from several independent studies, or
groups, in a combined test or a combined estimate.
 3.1 Aggregation
A common procedure in social research with two-level data is to aggregate
the micro-level data to the macro level. The simplest way to do this is to
work with the averages for each macro-unit.
There is nothing wrong with aggregation in cases where the researcher
is only interested in macro-level propositions, although it should be borne
in mind that the reliability of an aggregated variable depends, inter alia, on
the number of micro-level units in a macro-level unit (see later in this
chapter), and thus will be larger for the larger macro-units than for the
smaller ones. In cases where the researcher is interested in macro-micro or
micro-level propositions, however, aggregation may result in gross errors.
The first potential error is the ‘shift of meaning’ (cf. Firebaugh, 1978;
Hüttner, 1981). A variable that is aggregated to the macro level refers to the
macro-units, not directly to the micro-units. The firm average of an
employee rating of working conditions, for example, may be used as an
index for ‘organizational climate’. This variable refers to the firm, not
directly to the employees.
The second potential error with aggregation is the ecological fallacy
(Robinson, 1950). A correlation between macro-level variables cannot be
used to make assertions about micro-level relations. The percentage of
black inhabitants in a neighborhood could be related to average political
views in the neighborhood – for example, the higher the percentage of
blacks in a neighborhood the higher might be the proportion of people with
extreme right wing political views. This, however, does not give us any clue
about the micro-level relation between race and political conviction. (The
shift of meaning plays a role here, too. The percentage of black inhabitants
is a variable that means something for the neighborhood and this meaning is
distinct from the meaning of ethnicity as an individual-level variable.) The
ecological and other related fallacies are extensively discussed by Alker
(1969), Diez-Roux (1998), and Blakely and Woodward (2000). King
(1997), originally focusing on deriving correlates of individual voting
behavior from aggregate data, describes a method for making inferences –
within certain boundaries – at the micro level, when data are only available
in aggregate form at the macro level.
The third potential error is the neglect of the original data structure,
especially when some kind of analysis of covariance is to be used. Suppose
one is interested in assessing between-school differences in pupil
achievement after correcting for intake differences, and that Figure 3.1
depicts the true situation. The figure depicts the situation for five groups,
for each of which we have five observations. The groups are indicated by
the symbols and *. The five group means are indicated by •.

Figure 3.1: Micro-level versus macro-level adjustments. (X, Y) values for

five groups indicated by group averages by •.

Now suppose the question is whether the differences between the

groups on the variable Y, after adjusting for differences on the variable X,
are substantial. The micro-level approach, which adjusts for the within-
group regression of Y on X, will lead to the regression line with positive
slope. In this picture, the micro-units from the group that have the symbol
are all above the line, whereas those from the group with the * symbol are
all below the regression line. The micro-level regression approach will thus
lead us to conclude that the five groups do differ given that an adjustment
for X has been made.
Now suppose we were to aggregate the data, and regress the average Ȳ
on the average . The averages are depicted by •. This situation is
represented in the graph by the regression line with negative slope. The
averages of all groups are almost exactly on the regression line (the
observed average Ȳ can be almost perfectly predicted from the observed
average ), thus leading us to the conclusion that there are almost no
differences between the five groups after adjusting for the average .
Although the situation depicted in the graph is an idealized example, it
clearly shows that working with aggregate data ‘is dangerous at best, and
disastrous at worst’ (Aitkin and Longford 1986, p. 42). When analyzing
multilevel data, without aggregation, the problem described in this section
can be dealt with by distinguishing between the within-group and the
between-group regressions. This is worked out in Sections 3.6, 4.6, and
10.2.1.
The last objection to aggregation is that it prevents us from examining
potential cross-level interaction effects of a specified micro-level variable
with an as yet unspecified macro-level variable. Having aggregated the data
to the macro level one cannot examine relations such as whether the
sentence differential between black and white suspects is different between
judges, when allowance is made for differences in seriousness of crimes.
Or, to give another example, whether the effect of aptitude on achievement,
present in the case of whole-class instruction, is smaller or even absent in
the case of ability grouping of pupils within classrooms.

3.2 Disaggregation
Now suppose that we treat our data at the micro level. There are two
situations:

we also have a measure of a variable at the macro level, next to the

1.
measures at the micro level;
2. we only have measures of micro-level variables.
In situation (1), disaggregation leads to ‘the miraculous multiplication of
the number of units’. To illustrate, suppose a researcher is interested in the
question of whether older judges hand down more lenient sentences than
younger judges. A two-stage sample is taken: in the first stage ten judges
are sampled, and in the second stage for each judge ten trials are sampled
(in total there are thus 10 × 10 = 100 trials). One might disaggregate the
data to the level of the trials and estimate the relation between the
experience of the judge and the length of the sentence, without taking into
account that some trials involve the same judge. This is like pretending that
there are 100 independent observations, whereas in actual fact there are
only 10 independent observations (the 10 judges). This shows that
disaggregation and treating the data as if they are independent implies that
the sample size is dramatically exaggerated. For the study of between-group
differences, disaggregation often leads to serious risks of committing type I
errors (asserting on the basis of the observations that there is a difference
between older and younger judges whereas in the population of judges there
is no such relation). On the other hand, when studying within-group
differences, disaggregation often leads to unnecessarily conservative tests
(i.e., type I error probabilities that are too low); this is discussed in detail in
Moerbeek et al. (2003) and Berkhof and Kampen (2004).
If measures are taken only at the micro level, analyzing the data at the
micro level is a correct way to proceed, as long as one takes into account
that observations within a macro-unit may be correlated. In sampling
theory, this phenomenon is known as the design effect for two-stage
samples. If one wants to estimate the average management capability of
young managers, while in the first stage a limited number of organizations
(say, 10) are selected and within each organization five managers are
sampled, one runs the risk of making an error if (as is usually the case)
there are systematic differences between organizations. In general, two-
stage sampling leads to the situation that the ‘effective’ sample size that
should be used to calculate standard errors is less than the total number of
cases, the latter being given here by the 50 managers. The formula will be
presented in one of the next sections.
Starting with Robinson’s (1950) paper on the ecological fallacy, many
papers have been written about the possibilities and dangers of cross-level
inference, that is, methods to conclude something about relations between
micro-units on the basis of relations between data at the aggregate level, or
conclude something about relations between macro-units on the basis of
relations between disaggregated data. Discussions and many references are
given by Pedhazur (1982, Chapter 13), Aitkin and Longford (1986), and
Diez-Roux (1998). Our conclusion is that if the macro-units have any
meaningful relation to the phenomenon under study, analyzing only
aggregated or only disaggregated data is apt to lead to misleading and
erroneous conclusions. A multilevel approach, in which within-group and
between-group relations are combined, is more difficult but much more
productive. This approach requires, however, assumptions to be specified
about the way in which macro- and micro-effects are put together. The
present chapter presents some multilevel procedures that are based on only
a minimum of such assumptions (e.g., the additive model of equation (3.1)).
Later chapters in this book are based on a more elaborate model, the so-
called hierarchical linear model, which since about 1990 has been the most
widely accepted basis for multilevel analysis.

3.3 The intraclass correlation

The degree of resemblance between micro-units belonging to the same
macro-unit can be expressed by the intraclass correlation coefficient. The
use of the term ‘class’ is conventional here and refers to the macro-units in
the classification system under consideration. There are, however, several
definitions of this coefficient, depending on the assumptions about the
sampling design. In this section we assume a two-stage sampling design
and infinite populations at either level. The macro-units will also be
referred to as groups.
A relevant model here is the random effects ANOVA model.1 Denoting
by Yij the outcome value observed for micro-unit i within macro-unit j, this
model can be expressed as

where μ is the population grand mean, Uj is the specific effect of macro-unit

j, and Rij is the residual effect for micro-unit i within this macro-unit. In
other words, macro–unit j has the ‘true mean’ μ + Uj, and each
measurement of a micro-unit within this macro-unit deviates from this true
mean by some value Rij. Units differ randomly from one another, which is
reflected in the fact that Uj is a random variable and the name ‘random
effects model’. Some units have a high true mean, corresponding to a high
value of Uj, others have a true mean close to average, and still others a low
true mean. It is assumed that all variables are independent, the group effects
Uj having population mean 0 and population variance τ2 (the population
between-group variance), and the residuals having mean 0 and variance σ2
(the population within-group variance). For example, if micro-units are
pupils and macro-units are schools, then the within-group variance is the
variance within the schools about their true means, while the between-group
variance is the variance between the schools’ true means. The total variance
of Yij is then equal to the sum of these two variances,

The number of micro-units within the jth macro-unit is denoted by nj. The
number of macro-units is N, and the total sample size is M = Σjnj.
In this situation, the intraclass correlation coefficient ρI can be defined
as

This is the proportion of variance that is accounted for by the group level.
This parameter is called a correlation coefficient because it is equal to the
correlation between values of two randomly drawn micro-units in the same,
randomly drawn, macro-unit. Hedges and Hedberg (2007) report on a large
variety of studies of educational performance in American schools, and find
that values often range between 0.10 and 0.25.
It is important to note that the population variance between macro-units
is not directly reflected by the observed variance between the means of the
macro-units (the observed between-macro-unit variance). The reason is that
in a two-stage sample, variation between micro-units will also show up as
extra observed variance between macro-units. It is indicated below how the
observed variance between cluster means must be adjusted to yield a good
estimator for the population variance between macro-units.

Example 3.1 Random data.

Suppose we have a series of 100 observations as in the random digits in
Table 3.1. The core part of the table contains the random digits. Now
suppose that each row in the table is a macro-unit, so that for each macro-
unit we have observations on 10 micro-units. The averages of the scores for
each macro-unit are in the last column. There seem to be large differences
between the randomly constructed macro-units, if we look at the variance in
the macro-unit averages (which is 105.7). The total observed variance
between the 100 micro-units is 814.0. Suppose the macro-units were
schools, the micro-units pupils, and the random digits test scores.
According to these two observed variances we might conclude that the
schools differ considerably with respect to their average test scores. We
know in this case, however, that in ‘reality’ the macro-units differ only by
chance.

The following subsections show how the intraclass correlation can be

estimated and tested. For a review of various inference procedures for the
intraclass correlation we refer to Donner (1986). An extensive overview of
many methods for estimating and testing the within-group and between-
group variances is given by McCulloch and Searle (2001).

Table 3.1: Data grouped into macro-units (random digits from Glass and
Stanley, 1970, p. 511).

3.3.1 Within-group and between-group variance

We continue to refer to the macro-units as groups. To disentangle the
information contained in the data about the population between-group
variance and the population within-group variance, we consider the
observed variance between groups and the observed variance within
groups. These are defined as follows. The mean of macro-unit j is denoted
by

and the overall mean is

The observed variance within group j is given by

This number will vary from group to group. To have one parameter that
expresses the within-group variability for all groups jointly, one uses the
observed within-group variance, or pooled within-group variance. This is a
weighted average of the variances within the various macro-units, defined
as

If model (3.1) holds, the expected value of the observed within-group

variance is exactly equal to the population within-group variance:

The situation for the between-group variance is a little more

complicated. For equal group sizes nj, the observed between-group variance
is defined as the variance between the group means,

For unequal group sizes, the contributions of the various groups need to be
weighted. The following formula uses weights that are useful for estimating
the population between-group variance:
In this formula, ñ is defined by

where is the mean sample size and

is the variance of the sample sizes. If all nj have the same value, then ñ also
has this value. In this case, S2between is just the variance of the group means,
given by (3.5).
It can be shown that the total observed variance is a combination of the
within-group and the between-group variances, expressed as follows:

The complications with respect to the between-group variance arise from

the fact that the micro-level residuals Rij also contribute, albeit to a minor
extent, to the observed between-group variance. Statistical theory tells us
that the expected between-group variance is given by

Expected observed variance between

= True variance between + Expected sampling error
variance.

More specifically, the formula is

(cf. Hays (1988, Section 13.3) for the case with constant nj and Searle et al.
(1992, Section 3.6) for the general case), which holds provided that model
(3.1) is valid. The second term in this formula becomes small when ñ
becomes large. Thus for large group sizes, the expected observed between
variance is practically equal to the true between variance. For small group
sizes, however, it tends to be larger than the true between variance due to
the random differences that also exist between the group means.
In practice, we do not know the population values of the between and
within macro-unit variances; these have to be estimated from the data. One
way of estimating these parameters is based on formulas (3.4) and (3.9).
From the former it follows that the population within-group variance, σ 2,
can be estimated without bias by the observed within-group variance:

From the combination of the last two formulas it follows that the population
between-group variance, τ2, can be estimated without bias by taking the
observed between-group variance and subtracting the contribution that the
true within-group variance makes, on average, according to (3.9), to the
observed between-group variance:

(Another expression is given in (3.14).) This expression can take negative

values. This happens when the difference between group means is less than
would be expected on the basis of the within-group variability, even if the
true between-group variance τ2 were 0. In such a case, it is natural to
estimate τ2 as 0.
It can be concluded that the split between the observed within-group
variance and observed between-group variance does not correspond
precisely to the split between the within-group and between-group
variances in the population: the observed between-group variance reflects
the population between-group variance plus a little of the population
within-group variance.
The intraclass correlation is estimated according to formula (3.2) by
(Formula (3.15) gives another, equivalent, expression.) The standard error
of this estimator in the case where all group sizes are constant, nj = n, is
given by

This formula was given by Fisher (1958, Section 39) and by Donner (1986,
equation (6.1)), who also gives the (quite complicated) formula for the
standard error for the case of variable group sizes. Donner and Wells (1986)
compare various ways to construct confidence intervals for the intraclass
correlation coefficient.
The estimators given above are so-called analysis of variance or
ANOVA estimators. They have the advantage that they can be represented
by explicit formulas. Other much used estimators are those produced by the
maximum likelihood (ML) and residual maximum likelihood (REML)
methods (cf. Section 4.7). For equal group sizes, the ANOVA estimators are
the same as the REML estimators (Searle et al., 1992). For unequal group
sizes, the ML and REML estimators are slightly more efficient than the
ANOVA estimators. Multilevel software can be used to calculate the ML
and REML estimates.
Example 3.2 Within- and between-group variability for random data.
For our random digits table of the earlier example the observed between variance is S2between = 105.7.
The observed variance within the macro-units can be computed from formula (3.8). The observed
total variance is known to be 814.0 and the observed between variance is given by 105.7. Solving
(3.8) for the observed within variance yields S2within = (99/90) × (814.0–(10/11)x105.7)= 789.7 Then
the estimated true variance within the macro-units is also The estimate for the true
between macro-units variance is computed from (3.11) as
Finally, the estimate of the intraclass correlation is Its
standard error, computed from (3.13), is 0.06.

3.3.2 Testing for group differences

The intraclass correlation as defined by (3.2) can be zero or positive.2 A
statistical test can be performed to investigate whether a positive value for
this coefficient could be attributed to chance. If it may be assumed that the
within-group deviations Rij are normally distributed one can use an exact
test for the hypothesis that the intraclass correlation is 0, which is the same
as the null hypothesis that there are no group differences, or the true
between-group variance is 0. This is just the F-test for a group effect in the
one-way analysis of variance, which can be found in any textbook on
ANOVA. The test statistic can be written as

and it has an F distribution with N − 1 and M−N degrees of freedom if the

null hypothesis holds.

Example 3.3 The F-test for the random data set.

For the data of Table 3.1, F = (10 × 105.7)/789.7 = 1.34 with 9 and 90
degrees of freedom. This value is far from significant (p > 0.10). Thus,
there is no evidence of true between-group differences.
Statistical computer packages usually give the F-statistic and the within-
group variance, S2within. From this output, the estimated population
between-group variance can be calculated by

and the estimated intraclass correlation coefficient by

where ñ is given by (3.7). If F < 1, it is natural to replace both of these

expressions by 0. These formulas show that a high value for the F-statistic
will lead to large estimates for the between-group variance as well as the
intraclass correlation, but that the group sizes, as expressed by ñ, moderate
the relation between the test statistic and the parameter estimates.
If there are covariates, it often is relevant to test whether there are group
differences in addition to those accounted for by the effect of the covariates.
This is achieved by the usual F-test for the group effect in an analysis of
covariance (ANCOVA). Such a test is relevant because it is possible that the
ANOVA F-test does not demonstrate any group effects, but that such effects
do emerge when controlling for the covariates (or vice versa). Another
check on whether the groups make a difference can be carried out by testing
the group-by-covariate interaction effect. These tests can be found in
textbooks on ANOVA and ANCOVA, and they are contained in the well-
known general-purpose statistical computer programs.
So, to test whether a given nesting structure in a data set calls for
multilevel analysis, one can use standard ANOVA techniques. In addition to
testing for the main group effect, it is also advisable to test for group-by-
covariate interactions. If there is neither evidence for a main effect nor for
interaction effects involving the group structure, then the researcher may
leave aside the nesting structure and analyze the data by single-level
methods such as ordinary least squares (‘OLS’) regression analysis. This
approach to testing for group differences can be employed whenever the
number of groups is not too large for the computer program being used. If
there are too many groups, however, the program will refuse to do the job.
In such a case it will still be possible to carry out the tests for group
differences that are treated in the following chapters, following the logic of
the hierarchical linear model. This will require the use of statistical
multilevel software.

3.4 Design effects in two-stage samples

In the design of empirical investigations, the determination of sample sizes
is an important decision. For two-stage samples, this is more complicated
than for simple (‘one-stage’) random samples. An elaborate treatment of
this question is given in Cochran (1977). This section gives a simple
approach to the precision of estimating a population mean, indicating the
basic role played by the intraclass correlation. We return to this question in
Chapter 11.
Large samples are preferable in order to increase the precision of
parameter estimates, that is, to obtain tight confidence intervals around the
parameter estimates. In a simple random sample the standard error of the
mean is related to the sample size by the formula
This formula can be used to indicate the required sample size (in a simple
random sample) if a given standard error is desired.
When using two-stage samples, however, the clustering of the data
should be taken into account when determining the sample size. Let us
suppose that all group sizes are equal, nj = n for all j. Then the (total)
sample size is Nn. The design effect is a number that indicates how much
the sample size in the denominator of (3.16) is to be adjusted because of the
sampling design used. It is the ratio of the variance of estimation obtained
with the given sampling design to the variance of estimation obtained for a
simple random sample from the same population, supposing that the total
sample size is the same. A large design effect implies a relatively large
variance, which is a disadvantage that may be offset by the cost reductions
implied by the design. The design effect of a two-stage sample with equal
group sizes is given by

This formula expresses the fact that, from a purely statistical point of view,
a two-stage sample becomes less attractive as ρI increases (clusters become
more homogeneous) and as the group size n increases (the two-stage nature
of the sampling design becomes stronger).
Suppose, for example, we were studying the satisfaction of patients with
the treatments provided by their doctors. Furthermore, let us assume that
some doctors have more satisfied patients than others, leading to a ρI of
0.30. The researchers used a two-stage sample, since that is far cheaper than
selecting patients simply at random. They first randomly selected 100
doctors, from each chosen doctor selected five patients at random, and then
interviewed each of these. In this case the design effect is 1 + (5 − 1) × 0.30
= 2.20. When estimating the standard error of the mean, we no longer can
treat the observations as independent of each other. The effective sample
size, that is, the equivalent total sample size that we should use in
estimating the standard error, is equal to

in which N is the number of selected macro-units. For our example we find

Neffective = (100 × 5) /2.20 = 227. So the two-stage sample with a total of
500 patients here is equivalent to a simple random sample of 227 patients.
One can also derive the total sample size using a two-stage sampling
design on the basis of a desired level of precision, assuming that ρ1 is
known, and fixing n because of budgetary or time-related considerations.
The general rule is: this required sample size increases as pI increases and it
increases with the number of micro-units one wishes to select per macro-
unit. Using (3.17) and (3.18), this can be derived numerically from the
formula

The quantity Nts in this formula refers to the total desired sample size when
using a two-stage sample, whereas Nsrs refers to the desired sample size if
one had used a simple random sample.
In practice, ρ1 is unknown. However, it often is possible to make an
educated guess about it on the basis of earlier research.
In Figure 3.2, Nts is graphed as a function of n and ρI (0.1, 0.2, 0.4, and
0.6, respectively), and taking Nsrs = 100 as the desired sample size for an
equally informative simple random sample.

Figure 3.2: The total desired sample size in two-stage sampling.

3.5 Reliability of aggregated variables

Reliability, as conceived in psychological test theory (e.g., Lord and
Novick, 1968) and in generalizability theory (e.g., Shavelson and Webb,
1991), is closely related to clustered data – although this may not be
obvious at first sight. Classical psychological test theory considers a subject
(an individual or other observational unit) with a given true score, of which
imprecise, or unreliable, observations may be made. The observations can
be considered to be nested within the subjects. If there is more than one
observation per subject, the data are clustered. Whether there is only one
observation or several, equation (3.1) is the model for this situation: the true
score of subject j is μ + Uj and the ith observation on this subject is Yij, with
associated measurement error Rij. If several observations are taken, these
can be aggregated to the mean value Ỵ̄j which is then the measurement for
the true score of subject j.
The same idea can be used when it is not an individual subject that is to
be measured, but some collective entity: a school, a firm, or in general any
macro-level unit such as those mentioned in Table 2.2. For example, when
the school climate is measured on the basis of questions posed to pupils of
the school, then Yij could refer to the answer by pupil i in school j to a given
question, and the opinion of the pupils about this school would be measured
by the mean value Ỵ̄j. In terms of psychological test theory, the micro-level
units i are regarded as parallel items for measuring the macro-level unit j.
The reliability of a measurement is defined generally as

It can be proved that this is equal to the correlation between independent

replications of measuring the same subject. (In the mathematical model this
means that the same value Uj is measured, but with independent realizations
of the random errors Rij.) The reliability is indicated by the symbol λj.3
For measurement on the basis of a single observation according to
model (3.1), the reliability is just the intraclass correlation coefficient:
 When several measurements are made for each macro-level unit, these
constitute a cluster or group of measurements which are aggregated to the
group mean Ȳ·j. To apply the general definition of reliability to Ȳ·j, note that
the observed variance is the variance between the observed means Ȳ·j, while
the true variance is the variance between the true scores µ + Uj. Therefore
the reliability of the aggregate is

Example 3.4 Reliability for random data.

If in our previous random digits example the digits represented, for
example, the perceptions by teachers in schools of their working conditions,
then the aggregated variable, an indicator for organizational climate, has an
estimated reliability of 26.7/105.7 = 0.25. (The population value of this
reliability is 0, however, as the data are random, so the true variance is 0.)

It can readily be demonstrated that the reliability of aggregated

variables increases as the number of micro-units per macro-unit increases,
since the true variance of the group mean (with group size nj) is τ2 while the
expected observed variance of the group mean is τ2 + σ2 / nj. Hence the
reliability can be expressed by

It is quite clear that if nj is very large and ρI is positive, then λj is almost 1.

If nj = 1 we are not able to distinguish between within- and between-group
variance. Figure 3.3 presents a graph in which the reliability of an aggregate
is depicted as a function of nj (denoted by n) and ρI (0.1 and 0.4,
respectively).
Much more can be said about constructing higher-level variables by
aggregating lower-level variables; see, for example, Raudenbush and
Sampson (1999b) and van Mierlo et al. (2009).
 3.6 Within- and between-group relations
We saw in Section 3.1 that regressions at the macro level between
aggregated variables and Ȳ can be completely different from the
regressions between the micro-level variables X and Y. This section
considers in more detail the interplay between macro-level and micro-level
relations between two variables. First the focus is on regression of Y on X,
then on the correlation between X and Y.

Figure 3.3: Reliability of aggregated variables.

The main point of this section is that within-group relations can be, in
principle, completely different from between-group relations. This is
natural, because the processes at work within groups may be different from
the processes at work between groups (see Section 3.1). Total relations, that
is, relations at the micro level when the clustering into macro-units is
disregarded, are mostly a kind of average of the within-group and between-
group relations. Therefore it is necessary to consider within- and between-
group relations jointly, whenever the clustering of micro-units in macro-
units is meaningful for the phenomenon being studied.

3.6.1 Regressions
The linear regression of a ‘dependent’ variable Y on an ‘explanatory’ or
‘independent’ variable X is the linear function of X that yields the best4
prediction of Y. When the bivariate distribution of (X, Y) is known and the
data structure has only a single level, the expression for this regression
function is

where the regression coefficients are given by

The constant term β0 is called the intercept, while β1 is called the regression
coefficient. The term R is the residual or error component, and expresses
the part of the dependent variable Y that cannot be approximated by a linear
function of Y. Recall from Section 1.2.2 that ε (X) and ε (Y) denote the
population mean (expected value) of X and Y, respectively.

Table 3.2: Artificial data on five macro-units, each with two micro-units.

In a multilevel data structure, this principle can be applied in various

ways, depending on which population of X and Y values is being
considered.
Let us consider the artificial data set of Table 3.2. The first two columns
in the table contain the identification numbers of the macro-unit (j) and the
micro-unit (i). The other four columns contain the data. Xij is the variable
observed for micro-unit i in macro-unit j, the average of the Xij values for
group j, and similarly for the dependent variable Y.
 One might be interested in the relation between Yij and Xij. The linear
regression of Yij on Xij at the micro level for the total group of 10
observations is

This is the disaggregated relation, since the nesting of micro-units in macro-

units is not taken into account. The regression coefficient is −0.33.
The aggregated relation is the linear regression relationship at the macro
level of the group means Ȳ.j on the group means . This regression line is

The regression coefficient is now −1.00.

A third option is to describe the relation between Yij and Xij within each
single group. Assuming that the regression coefficient has the same value in
each group, this is the same as the regression of the within-group Y-
deviations (Yij − Ȳj) on the X-deviations (Xij − ).

with a regression coefficient of +1.00.

Finally (and this is how the artificial data set was constructed), Yij can
be written as a function of the within-group and between-group relations
between Y and X. This amounts to putting together the between-group and
within-group regression equations. The result is

Figure 3.4 graphically depicts the total, within-group, and between-group

relations between the variables.
 Figure 3.4: Within, between, and total relations.

The five parallel ascending lines represent the within-group relation

between Y and X. The steep descending line represents the relation at the
aggregate level (i.e., between the group means), whereas the almost
horizontal descending line represents the total relationship, that is, the
micro-level relation between X and Y ignoring the hierarchical structure.
The within-group regression coefficient is +1 whereas the between-
group coefficient is −1. The total regression coefficient, −0.33, lies in
between these two. This illustrates that within-group and between-group
relations can be completely different, even have opposite signs. The true
relation between Y and X is revealed only when the within- and between-
group relations are considered jointly, that is, by the multilevel regression.
In the multilevel regression, both the between-group and within-group
regression coefficients play a role. Thus there are many different ways to
describe the data, of which one is the best, because it describes how the data
were generated. In this artificial data set, the residual R is 0; real data, of
course, have nonzero residuals.

A population model
The interplay of within-group and between-group relations can be better
understood on the basis of a population model such as (3.1). Since this
section is about two variables, X and Y, a bivariate version of the model is
needed. In this model, group (macro-unit) j has specific main effects Uxj
and Uyj for variables X and Y, and associated with individual (micro-unit) j
are the within-group deviations Rxij and Ryij. The population means are
denoted by μx and μy and it is assumed that the Us and the Rs have
population means 0. The Us on the one hand and the Rs on the other are
independent. The formulas for X and Y are then

For the formulas that refer to relations between group means and Ȳ.j, it
is assumed that each group has the same size, denoted by n.
The correlation between the group effects is defined as

while the correlation between the individual deviations is defined by

One of the two variables X and Y might have a stronger group nature than
the other, so that the intraclass correlation coefficients for X and Y may be
different. These are denoted by ρIx and ρIy, respectively.
The within-group regression coefficient is the regression coefficient
within each group of Y on X, assumed to be the same for each group. This
coefficient is denoted by βwithin and defined by the within-group regression
equation,

This equation may be regarded as an analysis of covariance model for Y.

Hence the within-group regression coefficient is also the effect of X in the
ANCOVA approach to this multilevel data.
The within-group regression coefficient is also obtained when the Y-
deviation values (Yij − Ȳ.j) are regressed on the X-deviation values (Xij −
). In other words, it is also the regression coefficient obtained in the
disaggregated analysis of the within-group deviation scores.
The population between-group regression coefficient is defined as the
regression coefficient for the group effects Uy on Ux. This coefficient is
denoted by βbetween U and is defined by the regression equation

where R now is the group-level residual.

The total regression coefficient of Y on X is the regression coefficient in
the disaggregated analysis, that is, when the data are treated as single-level
data:

The total regression coefficient can be expressed as a weighted mean of the

within-and between-group coefficients, where the weight for the between-
group coefficient is just the intraclass correlation for X. The formula is

This expression implies that if X is a pure macro-level variable (so that ρIx =
1), the total regression coefficient is equal to the between-group coefficient.
Conversely, if X is a pure micro-level variable we have ρIx = 0, and the total
regression coefficient is just the within-group coefficient. Usually X will
have both a within-group and a between-group component and the total
regression coefficient will be somewhere between the two level-specific
regression coefficients.

Regressions between observed group means5

At the macro level, the regression of the observed group means Ỵ̄j on is
not the same as the regression of the ‘true’ group effects Uy on Ux. This is
because the observed group averages, and Ỵ̄j, can be regarded as the
‘true’ group means to which some error, and has been added.6
Therefore the regression coefficient for the observed group means is not
exactly equal to the (population) between-group regression coefficient, but
it is given by
where λxj is the reliability of the group means for measuring μx + Uxj,
given by equation (3.21) applied to the X-variable. If n is large the
reliability will be close to unity, and the regression coefficient for the group
means will be close to the between-group regression coefficient at the
population level.
Combining equations (3.24) and (3.25) leads to another expression for
the total regression coefficient. This expression uses the correlation ratio η2x
which is defined as the ratio of the intraclass correlation coefficient to the
reliability of the group mean,

For large group sizes the reliability approaches unity, so the correlation ratio
approaches the intraclass correlation.
In the data, the correlation ratio η2x is the same as the proportion of
variance in Xij explained by the group means, and it can be computed as the
ratio of the between-group sum of squares relative to the total sum of
squares in an analysis of variance, that is,

The combined expression indicates how the total regression coefficient

depends on the within-group regression coefficient and the regression
coefficient between the group means:

Expression (3.27) was first given by Duncan et al. (1961) and can also be
found, for example, in Pedhazur (1982, p. 538). A multivariate version was
given by Maddala (1971). To apply this equation to an unbalanced data set,
the regression coefficient between group means must be calculated in a
weighted regression, group j having weight nj.

Example 3.5 Within- and between-group regressions for artificial data.

In the artificial example given, the total sum of squares of Xij as well as Yij
is 30 and the between-group sums of squares for X and Y are 20. Hence the
correlation ratios are = 0.667. If we use this value and
plug it into formula (3.27), we find

which is indeed what we found earlier.

3.6.2 Correlations
The quite extreme nature of the artificial data set of Table 3.2 becomes
apparent when we consider the correlations.
The group means ( , ȲJ) lie on a decreasing straight line, so the
observed between-group correlation, which is defined as the correlation
between the group means, is Rbetween = – 1. The within-group correlation is
defined as the correlation within the groups, assuming that this correlation
is the same within each group. This can be calculated as the correlation
coefficient between the within-group deviation scores and
In this data set the deviation scores are (−1, −1) for
i = 1 and (+1, +1) for i = 2, so the within-group correlation here is Rwithin =
+1. Thus, we see that the within-group as well as the between-group
correlations are perfect, but of opposite signs. The disaggregated
correlation, that is, the correlation computed without taking the nesting
structure into account, is Rtotal = −0.33. (This is the same as the value for
the regression coefficient in the total (disaggregated) regression equation,
because X and Y have the same variance.)

The population model again

Recall that in the population model mentioned above, the correlation
coefficient between the group effects Ux and Uy was defined as ρbetween and
the correlation between the individual deviations Rx and Ry was defined as
ρwithin. The intraclass correlation coefficients for X and Y were denoted by
ρIx and ρIy.
 How do these correlations between unobservable variables relate to
correlations between observables? The population within-group correlation
is also the correlation between the within-group deviation scores

For the between-group coefficient the relation is, as always, a little more
complicated. The correlation coefficient between the group means is equal
to

where λxj and λyj are the reliability coefficients of the group means (see
equations (3.20, 3.21)). For large group sizes the reliabilities will be close to
1 (provided the intraclass correlations are larger than 0), so that the
correlation between the group means will then be close to ρbetween.
The total correlation (i.e., the correlation in the disaggregated analysis)
is a combination of the within-group and between-group correlation
coefficients. The combination depends on the intraclass correlations, as
shown by the formula

If the intraclass correlations are low, then X and Y primarily have the nature
of level-one variables, and the total correlation will be close to the within-
group correlation; on the other hand, if the intraclass correlations are close
to 1, then X and Y almost have the nature of level-two variables and the
total correlation is close to the between-group correlation. As a third
possibility, if one of the intraclass correlations is close to 0 and the other is
close to 1, then one variable is mainly a level-one variable and the other
mainly a level-two variable. Formula (3.30) then implies that the total
correlation coefficient is close to 0, no matter how large the within-group
and between-group correlations. This is intuitively obvious, since a level-
one variable with hardly any between-group variability cannot be
substantially correlated with a variable having hardly any within-group
variability.
 If the intraclass correlations of X and Y are equal and denoted by ρI, then
(3.30) can be formulated more simply as

In this case the weights ρI and (1 − ρI) add up to 1 and the total correlation
coefficient is necessarily between the within-group and the between-group
correlation coefficients. This is not true in general, because the sum of the
weights in (3.30) is smaller than 1 if the intraclass correlations for X and Y
are different.
The reliabilities of the group means then also are equal, λxj = λyj; let us
denote these by λj. The correlation coefficient between the group means
(3.29) then simplifies to

The last two formulae can help in understanding how aggregation changes
correlations – under the continued assumption that the intraclass
correlations of X and Y are the same. The total correlation as well as the
correlation between group means are between the within-group and the
between-group correlations. However, since the reliability coefficient λj is
greater7 than the intraclass correlation ρI, the correlation between the group
means is drawn toward the between-group correlation more strongly than
the total correlation is. Therefore, aggregation will increase correlation if
and only if the between-group correlation coefficient is larger than the
within-group correlation coefficient. Therefore, the fact that correlations
between group means are often higher than correlations between individuals
is not the mathematical consequence of aggregation, but the consequence of
the processes at the group level (determining the value of ρbetween) being
different from the processes at the individual level (which determine the
value of ρwithin).

Correlations between observed group means8

Analogous to the regression coefficients, for the correlation coefficients we
can also combine the equations to see how the total correlation depends on
the within-group correlation and the correlation between the group means.
This yields

This expression was given by Robinson (1950) and can also be found, for
example, in Pedhazur (1982, p. 536). When it is applied to an unbalanced
data set, the correlation between the group means should be calculated with
weights nj.
It may be noted that many texts do not make the explicit distinction
between population and data. If the population and the data are equated,
then the reliabilities are unity, the correlation ratios are the same as the
intraclass correlations, and the population between-group correlation is
equal to the correlation between the group means. The equation for the total
correlation then becomes

When parameter estimation is being considered, however, confusion may be

caused by neglecting this distinction.

Example 3.6 Within- and between-group correlations for artificial data.

The correlation ratios in the artificial data example are ηx2 = ηy2 = 0.667 and
we also saw above that
Rwithin = +1 and Rbetween = –1. Filling in these numbers in formula
(3.32) yields

which indeed is the value found earlier for the total correlation.

3.6.3 Estimation of within- and between-group correlations

There are several ways of obtaining estimates for the correlation parameters
treated in this section.
 A quick method is based on the intraclass correlations, estimated as in
Section 3.3.1 or from the output of a multilevel computer program, and the
observed within-group and total correlations. The observed within-group
correlation is just the ordinary correlation coefficient between the within-
group deviations and the total correlation is the
ordinary correlation coefficient between X and Y in the whole data set. The
quick method is then based on (3.28) and (3.30). This leads to the estimates

and

This is not statistically the most efficient method, but it is straightforward

and leads to good results if sample sizes are not too small.
The ANOVA method (Searle, 1956) goes via the variances and
covariances, based on the definition

Estimating the within- and between-group variances was discussed in

Section 3.3.1. The within- and between-group covariances between X and Y
can be estimated by formulas analogous to (3.3), (3.6), (3.10) and (3.11),
replacing the squares by the cross-products
It is shown in Searle et al.
(1992, Section 11.1.a) how these calculations can be replaced by a
calculation involving only sums of squares.
Finally, the maximum likelihood and residual maximum likelihood
methods can be used. These are the estimation methods most often used (cf.
Section 4.7) and are implemented in multilevel software. Chapter 16
describes multivariate multilevel models; the correlation coefficient
between two variables refers to the simplest multivariate situation, namely,
bivariate data. Formula (16.5) represents the model which allows the ML
and REML estimation of within-group and between-group correlations.
Example 3.7 Within- and between-group correlations for school tests.
Many examples in Chapters 4 and 5 use a data set on the school
performance of 3,758 pupils in N = 211 schools. The present example
considers the correlation between the scores on an arithmetic test (X) and a
language test (Y). We delete only pupils who have missing data on either of
these variables, leaving M = 3,762 pupils. Within-school correlations reflect
the correspondence between the pupils’ language and arithmetic capabilities
(attenuated by the unreliability of the tests). Between-school correlations
reflect the processes that determine the schools’ populations (intake policy,
social segregation of neighborhoods) and the influence of teachers and
school policy. Thus the within-school correlations and the between-school
correlations are caused by quite different processes.
The indices ‘within’, ‘between’, and ‘total’ will be abbreviated to ‘w’,
‘b’, and ‘t’. The observed correlations are Rw = 0.634,Rb = 0.871, and Rt =
0.699.
The ANOVA estimators are calculated in accordance with the principles
of Section 3.3.1. The observed within-group variances and covariance,
which are also the estimated within-group population variances and
covariance (cf. (3.10)), are
Swxy = 28.432.
The observed between-group variances and covariance are
14.254. For this data set, ñ = 17.816.
According to (3.11), the estimated population between-group variances and
covariance are
From these estimated variances, the intraclass correlations are computed
as ρIx = 11.602/ (11.602 + 32.118) = 0.2654 and ρIy = 16.452/(16.452 +
62.607)= 0.2081.
The ‘quick’ method uses the observed within and total correlations and
the intraclass correlations. The resulting estimates are
The ANOVA estimates for the within- and between-group correlations
use the estimated within- and between-group population variances and
covariance. The results are
The ML estimates for the within-group variances and covariance are
obtained in Chapter 16 as For the
between-group variances and covariance they are
 This leads to estimated correlations
It can be concluded that, for this large data set, the three methods all
yield practically the same results. The within-school (pupil-level)
correlation, 0.63, is substantial. Thus pupils’ language and arithmetic
capacities are closely correlated. The between-school correlation, 0.92, is
very high. This demonstrates that the school policies, the teaching quality,
and the processes that determine the composition of the school population
have practically the same effect on the pupils’ language as on their
arithmetic performance. Note that the observed between-school correlation,
0.87, is not quite as high because of the attenuation caused by unreliability
that follows from (3.29).

3.7 Combination of within-group evidence

When research focuses on within-group relations and several groups
(macro-units) have been investigated, it is often desired to combine the
evidence gathered in the various groups. For example, consider a study in
which a relation between work satisfaction (X) and sickness leave (Y) is
studied in several organizations. If the organizations are sufficiently similar,
or if they can be considered as a sample from some population of
organizations, then the data can be analyzed according to the hierarchical
linear model of the next chapters. This also requires that the number of
organizations is large enough (e.g., 20 or more), because this number is the
sample size from the population of organizations. On the other hand, if the
organizations are too diverse and not representative of any population, or if
there are not enough of them, there are still inferential procedures to
combine the results across the organizations and provide a single answer on
the relation between X and Y in which the evidence from all organizations is
combined.
Another example is meta-analysis, the statistically based combination of
several studies. There are many texts about meta-analysis, among them
Hedges and Olkin (1985), Rosenthal (1991), Hedges (1992), Wilson and
Lipsey (2001), and Bohrenstein et al. (2009). A number of publications may
contain information about the same phenomenon, and it may be important
to combine this information in a single test.
 There exist various methods for combining evidence from several
studies that are based only on the assumption that this evidence is
statistically independent. They can be applied if the number of independent
studies is at least two. The least demanding method is Fisher’s combination
of p -values (Fisher, 1958; Hedges and Olkin, 1985). This method assumes
that in each of the N studies a null hypothesis is tested, which results in
independent p -values p1, . . . , pN. The combined null hypothesis is that in
all of the studies the null hypothesis holds; the combined alternative
hypothesis is that in at least one of the studies the alternative hypothesis
holds. It is not required that the N independent studies used the same
operationalizations or methods of analysis, only that it is meaningful to test
this combined null hypothesis. This hypothesis can be tested by minus
twice the sum of the natural logarithms of the p -values,

which under the combined null hypothesis has a chi-squared distribution

with 2N degrees of freedom. Because of the shape of the logarithmic
function, this combined statistic will already have a large value if at least
one of the p -values is very small.
A stronger combination procedure can be achieved if the several studies
all lead to estimates of theoretically the same parameter, denoted here by θ.
Suppose that the jth study yields a parameter estimate with standard error
sj and that all the studies are statistically independent. Then the combined
estimate with smallest standard error is the weighted average with weights
inversely proportional to s2j,

with standard error

For example, if standard errors are inversely proportional to the square root
of sample size, for some value σ, then the weights are directly
proportional to the sample sizes and the standard error of the combined
estimate is If the individual estimates are approximately normally
distributed, or N is large even while the estimates are not nearly normally
distributed, the t-ratio,

can be tested in a standard normal distribution.

A third, intermediate combination procedure is based on the assumption
that the parameters estimated by the several studies are not necessarily the
same, but are independent random draws from the same population. If the
true value of the parameter in population j is denoted by θj, this means that

where θ is the mean parameter in the population of all potential studies, and
Ej is the deviation from this value in study j, dependent on particulars of the
group under study, the measurement instrument used, etc. The estimate can
be represented as

where Rj is a random residual. From each study we know the standard error
Combining these two equations leads to a
representation of the parameter estimates as a mean value plus a double
error term,

If it is reasonable to assume that the studies are a random draw from a

population, then the residuals Ej may be regarded as being independent with
expected value 0 and a common unknown variance σ2. If each of the studies
is large enough and the errors for each study are approximately normally
distributed, then the residuals Rj can be regarded as normally distributed
variables with expected value 0 and known variance s2j. Then (3.39)
represents a heteroscedastic model, with variances σ2 + s2j. If the number of
studies is large enough, then study-level variables may be added, giving, for
one explanatory variable, a regression model such as

Such two-stage models can be estimated with software such as HLM,

MLwiN, or R. This model, and extended versions, are treated for meta-
analysis, for example, by Raudenbush and Bryk (2002, Chapter 7), where
they are called the ‘V-known problem’. The model can be regarded as a
special case of the hierarchical linear model treated in the following
chapters, and as a two-step approach to multilevel analysis. In contrast to
some earlier two-step approaches, it is correct in the sense that it recognizes
the distinct contributions of within-study ‘error’ variability and between-
study ‘true’ variability, a distinction that has permeated this chapter since
Section 3.3.1. Achen (2005) discussed the suitability of this model for
multilevel analysis, focusing on applications in political science.
The choice between these three combination methods can be made as
follows. The combination of estimates, expressed by (3.36), is more
suitable if conceptually it can be argued that the parameter being estimated
in each of the combined studies does indeed have the same theoretical
interpretation and it is plausible that this parameter (the estimated θ) has the
same value in each of the studies. Models (3.39) and (3.40) with a double
error term are applicable if it can be argued that the theoretical
interpretation is the same, but the value of this parameter could differ across
studies. On the other hand, Fisher’s method (3.35) for combining p -values
requires only that the null hypotheses of ‘no effect’ can be regarded as the
same, and it has good power even if the effect sizes are very different
between the N studies. For example, the individual tests could be two-sided
tests and the effects could be of different signs; as long as it is meaningful
to test the combined null hypothesis that the null hypothesis holds in all
individual studies, this method may be applied. More combination methods
can be found in the literature about meta-analysis (e.g., Hedges and Olkin,
1985).
 Sometimes it may be helpful to apply a formal statistical test of the null
hypothesis that all the estimates do indeed have the same expected value:

Assuming that the standard errors were estimated very precisely, this can be
done by a chi-squared test as derived, for example, in Lehmann and
Romano (2005, Section 7.3). The test statistic is

and has, under the null hypothesis and if the estimates hypothesis and if the
estimates are normally distributed with variances s2j, a chi-squared
distribution with N – 1 degrees of freedom.

Example 3.8 Gossip behavior in six organizations.

Wittek and Wielers (1998) investigated effects of informal social network
structures on gossip behavior in six work organizations. One of the
hypotheses tested was that individuals tend to gossip more if they are
involved in more coalition triads. An individual A is involved in a coalition
triad with two others, B and C, if he has a positive relation with B while A
and B both have a negative relation with C. Six organizations were studied
which were so different that an approach following the lines of the
hierarchical linear model was not considered appropriate. For each
organization separately, a multiple regression was carried out to estimate
the effect of the number of coalition triads in which a person was involved
on a measure for gossip behavior, controlling for some relevant other
variables.
The p -values obtained were 0.015, 0.42, 0.19, 0.13, 0.25, and 0.43.
Only one of these is significant (i.e., less than 0.05), and the question is
whether this combination of six p– values would be unlikely under the
combined null hypothesis which states that in all six organizations the
effect of coalition triads on gossip is absent. Equation (3.35) yields the test
statistic χ = 22.00 with 2 × 6 = 12 degrees of freedom, p < 0.05. Thus the
result is significant, which shows that in the combined data there is indeed
evidence of an effect of coalition triads on gossip behavior, even though this
effect is significant in only one of the organizations considered separately.

3.8 Glommary
Aggregation and disaggregation. Multilevel data structures can be
analyzed by aggregating data to the higher level (e.g., by taking the
means) and analyzing these; or by disaggregating to the lower level,
which means that characteristics of higher-level units are used as
characteristics of the lower-level units contained in them, and further
nesting is not used. Both approaches provide only a limited perspective
because they focus on only one level among several.

Shift of meaning. Variables aggregated from a lower to a higher level (e.g.,

the average income in a neighborhood or the proportions of girls in a
classroom) have a theoretically different meaning at the level of
aggregation than at their original level, because of the social processes
taking place at the higher level.

Ecological fallacy. Errors committed by taking a relation between variables

established at one level and transferring it to a different level without
checking its validity for that level.

Intraclass correlation. A measure for the internal homogeneity of level-

two units (also called clusters, classes, or groups) with respect to a
given variable. It can be interpreted in two equivalent ways: the
proportion of total variance of this variable that is accounted for by the
higher level; or the correlation between the variables measured for two
randomly drawn different individuals in the same group.

Design effect of a two-stage sample. The ratio of the variance of

estimation obtained with the given sampling design, to the variance of
estimation obtained for a simple random sample from the same
population, and with the same total sample size.
Reliability of aggregated variables. Cluster means can be used as
measurements for properties of the clusters. Their reliability depends
on the number of units used for averaging and on the intraclass
correlation.

Within- and between-group regressions. In a multilevel structure, the

relation between two variables X and Y will often be different at the
different levels of the hierarchy. The within-group regression
coefficient of Y on X is the expected difference in the values of Y,
when comparing two level-one units in the same group with a unit
difference in the value of the variable X. The between-group regression
coefficient of Y on X is the expected difference in the values of the
group means for Y, when comparing two groups with a unit difference
in the group mean with respect to X. The distinction between between-
group and within-group regression coefficients is discussed further in
Section 4.6. Similarly, within-group correlations are defined as
correlations valid within the groups, and the between-group
correlations as the correlations between the group means. A population
model was introduced to clarify these concepts and show how they
relate to total regressions and total correlations, which are the
regressions and correlations obtained for a simple random sample from
this population.

Meta-analysis. Statistical methods for the combination of several studies.

Combination of tests. A method for combining results from several studies

demanding little in terms of assumptions, only their independence. The
best-known method is Fisher’s combination of independent p-values.

Combination of estimates. If several studies are combined and it is

reasonable to assume that the same parameter is estimated in all of
these studies, then a pooled parameter estimate can be obtained with
smaller variance than the individual estimates.

1
This model is also known in the statistical literature as the one-way random effects ANOVA model
and as Eisen-hart’s Type II ANOVA model. In multilevel modeling it is known as the empty model,
and is treated further in Section 4.4.
2
In a data set it is possible for the estimated intraclass correlation coefficient to be negative. This is
always the case, for example, for group-centered variables. In a population satisfying model (3.1),
however, the population intraclass correlation cannot be negative.
3
In the literature the reliability of a measurement is X frequently denoted by the symbol ρXX, so that
the reliability coefficient λj here could also be denoted by ρYY.
4
‘Best prediction’ means here the prediction that has the smallest mean squared error: the so-called
least squares criterion.
5
The remainder of this subsection may be skipped by the cursory reader.
6
The same phenomenon is at the heart of formulas (3.9) and (3.11).
7
We argue under the assumption that the group size is at least 2, and the common intraclass
correlation is positive.
8
The remainder of Section 3.6.2 may also be skipped by the cursory reader.
 4
The Random Intercept Model

In the preceding chapters it was argued that the best approach to the
analysis of multilevel data is one that represents within-group as well as
between-group relations within a single analysis, where ‘group’ refers to the
units at the higher levels of the nesting hierarchy. Very often it makes sense
to represent the variability within and between groups by a probability
model, in other words, to conceive of the unexplained variation within
groups and that between groups as random variability. For a study of
students within schools, for example, this means that not only unexplained
variation between students, but also unexplained variation between schools
is regarded as random variability. This can be expressed by statistical
models with so-called random coefficients.
The hierarchical linear model is such a random coefficient model for
multilevel, or hierarchically structured, data and has become the main tool
for multilevel analysis. In this chapter and the next the definition of this
model and the interpretation of the model parameters are discussed. This
chapter discusses the simpler case of the random intercept model; Chapter 5
treats the general hierarchical linear model, which also has random slopes.
Chapter 6 is concerned with testing the various components of the model.
Later chapters treat various elaborations and other aspects of the
hierarchical linear model. The focus of this treatment is on the two-level
case, but Chapters 4 and 5 also contain sections on models with more than
two levels of variability.

OVERVIEW OF THE CHAPTER

This chapter starts by discussing regression models that may have different
parameters in different groups, and discusses when it is more appropriate to
have separate parameters in each group (‘fixed effects’), and when to have
random variability between groups (‘random effects’). Then the random
intercept model is defined as a random effects model, with special attention
given to its simplest case, the empty model, serving as a starting point for
multilevel modeling. Here also a measure for heterogeneity between groups
is defined, the intraclass correlation coefficient.
A major issue is treated next: the avoidance of ecological fallacies by
differentiating the between-group regression from the within-group
regression. Then estimation techniques for the parameters are discussed on
an intuitive, nonmathematical level. A special topic is the estimation of the
random coefficients in each group by posterior means, where the
information for each single group is combined with the information for the
whole data set. This is an example of the advantages of the multilevel
approach: for inference about any given group, strength is borrowed from
the data for other groups. The final section of this chapter is about random
intercept models with more than two levels.

4.1 Terminology and notation

For the sake of concreteness, in a two-level model we refer to the level-one
units as ‘individuals’, and to the level-two units as ‘groups’. The reader
with a different application in mind may supply other names for the units;
for example, if the application is to repeated measurements, ‘measurement
occasions’ for level-one units and ‘subjects’ for level-two units. The nesting
situation of ‘measurement occasions within individuals’ is given special
attention in Chapter 15. The groups are also called clusters in the literature.
The number of groups in the data is denoted by N; the number of
individuals in the groups may vary from group to group, and is denoted by
nj for group j (j = 1,2,...,N). The total number of individuals is denoted by M
= Σj nj.
The hierarchical linear model is a type of regression model that is
particularly suitable for multilevel data. It differs from the usual multiple
regression model in that the equation defining the hierarchical linear model
contains more than one error term: one (or more) for each level. As in all
regression models, a distinction is made between dependent and
explanatory variables: the aim is to construct a model that expresses how
the dependent variable depends on, or is explained by, the explanatory
variables. Instead of ‘explanatory variable’, the names ‘predictor variable’
and ‘independent variable’ are also used; and ‘criterion’ is also used for
‘dependent variable’. The dependent variable must be a variable at level
one: the hierarchical linear model is a model for explaining something that
happens at the lowest, most detailed level. Models for multilevel structures
where the dependent variable is defined at a higher level are treated by
Croon and van Veldhoven (2007), Lüdtke et al. (2008), and van Mierlo et
al. (2009).
In this section, we assume that one explanatory variable is available at
either level. In the notation, we distinguish the following types of indices
and variables:

j is the index for the groups (j = 1, . . . ,N);

i is the index for the individuals within the groups (i = 1, . . . , nj).

The indices can be regarded as case numbers; note that the numbering for
the individuals starts again in every group. For example, individual 1 in
group 1 is different from individual 1 in group 2.
For individual i in group j, we have the following variables:

Yij is the dependent variable

xij is the explanatory variable at the individual level;

for group j,

zj is the explanatory variable at the group level.

To understand the notation, it is essential to realize that the indices i and j

indicate precisely what the variables depend on. The notation Yij, for
example, indicates that the value of variable Y depends on group j and also
on individual i. (Since the individuals are nested within groups, the index i
makes sense only if it is accompanied by the index j: to identify individual i
= 1, we must know which group we are referring to!) The notation zj, on the
other hand, indicates that the value of Z depends only on group j, and not on
individual i.
The basic idea of multilevel modeling is that the outcome variable Y has
an individual as well as a group aspect. This also carries through to other
level-one variables. The X-variable, although it is a variable at the
individual level, may also contain a group aspect. The mean of X in one
group may be different from the mean in another group. In other words, X
may (and often will) have a positive between-group variance. To put it more
generally, the compositions of the various groups with respect to X may
differ from one another. It should be kept in mind that explanatory variables
that are defined at the individual level often also contain some information
about the groups, as discussed in Section 3.1.

4.2 A regression model: fixed effects only

The simplest model is one without the random effects that are characteristic
of multilevel models; it is the classical model of multiple regression. This
model states that the dependent variable, Yij, can be written as the sum of a
systematic part (a linear combination of the explanatory variables) and a
random residual,

In this model equation, the βs are the regression parameters: β0 is the

intercept (i.e., the value obtained if both xij and zj are 0), β1 is the coefficient
for the individual variable X, and β2 is the coefficient for the group variable
Z. The variable Rij is the residual (sometimes called error) – an essential
requirement in regression model (4.1) is that all residuals are mutually
independent and have mean 0; a convenient assumption is that in all groups
they have the same variances (the homoscedasticity assumption) and are
normally distributed. This model has a multilevel nature only to the extent
that some explanatory variables may refer to the lower and others to the
higher level.
Model (4.1) can be extended to a regression model containing not only
the main effects of X and Z, but also the cross-level interaction effect. This
type of interaction is discussed in more detail in Chapter 5. It means that the
product variable ZX = Z × X is added to the list of explanatory variables.
The resulting regression equation is

These models pretend, as it were, that all the multilevel structure in the data
is fully explained by the group variable Z and the individual variable X. In
other words, if two individuals are being considered and their X- and Z-
values are given, then for their Y-value it is immaterial whether they belong
to the same or to different groups.
Models of types (4.1) and (4.2), and their extensions to more
explanatory variables at either or both levels, have in the past been widely
used in research on data with a multilevel structure. They are convenient to
handle for anybody who knows multiple regression analysis. Is anything
wrong with them? YES! For data with a meaningful multilevel structure, it
is practically always incorrect to make the a priori assumption that all of
the group structure is represented by the explanatory variables. Given that
there are only N groups, it is unjustified to act as if one has n1 + n2 + . . . +
nN independent replications. There is one exception: when all group sample
sizes nj are equal to 1, the researcher need have no qualms about using these
models because the nesting structure is not present in the relation between
observed variables, even if it may be present in the structure of the
population. Designs with nj = 1 can be used when the explanatory variables
were chosen on the basis of substantive theory, and the focus of the research
is on the regression coefficients rather than on how the variability of Y is
partitioned into within-group and between-group variability.
In designs with group sizes larger than 1, however, the nesting structure
often cannot be represented completely in the regression model by the
explanatory variables. Additional effects of the nesting structure can be
represented by letting the regression coefficients vary from group to group.
Thus, the coefficients β0 and β1 in equation (4.1) must depend on the group,
denoted by j. This is expressed in the formula by an extra index j for these
Groups j can have a higher (or lower) value of β0j, indicating that, for any
given value of X, they tend to have higher (or lower) values of the
dependent variable Y. Groups can also have a higher or lower value of β1j,
which indicates that the effect of X on Y is higher or lower. Since Z is a
group-level variable, it would not make much sense conceptually to let the
coefficient of Z depend on the group. Therefore β2 is left unaltered in this
formula.
The multilevel models treated in the following sections and in Chapter 5
contain diverse specifications of the varying coefficients β0j and β1j. The
simplest version of model (4.3) is that where β0j and β1j are constant (do not
depend on j), that is, the nesting structure has no effect, and we are back at
model (4.1). If this is an appropriate model, which we said above is a
doubtful supposition, then the ordinary least squares (OLS) regression
models of type (4.1) and (4.2) offer a good approach to analyzing the data.
If, on the other hand, the coefficients β0j and β1j do depend on j, then these
regression models may give misleading results. Then it is preferable to take
into account how the nesting structure influences the effects of X and Z on
Y. This can be done using the random coefficient model of this and the
following chapters. This chapter examines the case where the intercept β0j
depends on the group; the next chapter treats the case where the regression
coefficient β1j is also group-dependent.

4.3 Variable intercepts: fixed or random

parameters?
Let us first consider only the regression on the level-one variable X. A first
step toward modeling between-group variability is to let the intercept vary
between groups. This reflects the tendency for some groups to have, on
average, higher responses Y and others to have lower responses. This model
is halfway between (4.1) and (4.3) in the sense that the intercept β0j does
depend on the group but the regression coefficient of X, β1, is constant:
 For simplicity here the effect of Z or other variables is omitted, but the
discussion applies equally to the case with more explanatory variables.
This model is depicted in Figure 4.1. The group-dependent intercept can
be split into an average intercept and the group-dependent deviation:

Figure 4.1: Different parallel regression lines. The point y12 is indicated
with its residual R12.

For reasons that will become clear in Chapter 5, the notation for the
regression coefficients is changed here, and the average intercept is called
γ00 while the regression coefficient for X is called γ10. Substitution now
leads to the model

The values U0j are the main effects of the groups: conditional on an
individual having a given X-value and being in group j, the Y-value is
expected to be U0j higher than in the average group. Model (4.5) can be
understood in two ways:

1. As a model where the U0j are fixed parameters, N in number, of the

statistical model. This is relevant if the groups j refer to categories
each with their own distinct interpretation – for example, a
classification according to gender or religious denomination. In
order to obtain identified parameters, the restriction that ΣjU0j = 0
 can be imposed, so that effectively the groups lead to N − 1
regression parameters. This is the usual analysis of covariance
model, in which the grouping variable is a factor. (Since we prefer
to use Greek letters for statistical parameters and capital letters for
random variables, we would prefer to use a Greek letter rather than
U when we take this view of model (4.5).) In this specification it is
impossible to use a group-level variable Zj as an explanatory
variable because it would be redundant given the fixed group
effects.
2. As a model where the U0j are independent and identically
distributed random variables. Note that the U0j are the unexplained
group effects, which also may be called group residuals,
controlling for the effects of variable X. These residuals are now
assumed to be randomly drawn from a population with zero mean
and an a priori unknown variance. This is relevant if the effects of
the groups j (which may be neighborhoods, schools, companies,
etc.), controlling for the explanatory variables, can be considered
to be exchangeable. There is one parameter associated with the U0j
in this statistical model: their variance. This is the simplest random
coefficient regression model. It is called the random intercept
model because the group-dependent intercept, γ00 + U0j, is a
quantity that varies randomly from group to group. The groups are
regarded as a sample from a population of groups. It is possible
that there are group-level variables Zj that express relevant
attributes of the groups (such variables will be incorporated in the
model in Section 4.5 and, more extensively, in Section 5.2). Then it
is not the groups that are exchangeable but the group-level
residuals U0j conditional on the group variables.

The next section discusses how to determine which of these specifications

of model (4.5) is appropriate in a given situation.
Note that models (4.1) and (4.2) are OLS models, or fixed effects
models, which do not take the nesting structure into account (except
perhaps by the use of a group-level variable Zj), whereas models of type (1)
above are OLS models that do take the nesting structure into account. The
latter kind of OLS model has a much larger number of regression
parameters, since in such models N groups lead to N − 1 regression
coefficients. It is important to distinguish between these two kinds of OLS
models in discussing how to handle data with a nested structure.

4.3.1 When to use random coefficient models

These two different interpretations of equation (4.5) imply that multilevel
data can be approached in two different ways, using models with fixed or
with random coefficients. Which of these two interpretations is the most
appropriate in a given situation depends on the focus of the statistical
inference, the nature of the set of N groups, the magnitudes of the group
sample sizes nj, and the population distributions involved.

1. If the groups are regarded as unique categories and the researcher

wishes primarily to draw conclusions pertaining to each of these N
specific categories, then it is appropriate to use the analysis of
covariance (fixed effects) model. Examples are groups defined by
gender or ethnic background.
2. If the groups are regarded as a sample from a (real or hypothetical)
population and the researcher wishes to draw conclusions
pertaining to this population, then the random coefficient model is
appropriate. Examples are the groupings mentioned in Table 2.2.
3. If the researcher wishes to test effects of group-level variables, the
random coefficient model should be used. The reason is that the
fixed effects model already ‘explains’ all differences between
groups by the fixed effects, and there is no unexplained between-
group variability left that could be explained by group-level
variables. ‘Random effects’ and ‘unexplained variability’ are two
ways of saying the same thing.
4. Especially for relatively small group sizes (in the range from 2 to
50 or 100), the random coefficient model has important advantages
over the analysis of covariance model, provided that the
assumptions about the random coefficients are reasonable. This can
be understood as follows.
 The random coefficient model includes the extra assumption of
independent and identically distributed group effects U0j. To put it
less formally, the unexplained parts of the group differences are
governed by ‘mechanisms’ that are roughly similar from one group
to the next, and operate independently between the groups. The
group-level residuals are then said to be exchangeable. This
assumption helps to counteract the paucity of the data that is
implied by relatively small group sizes nj. Since all group effects
are assumed to come from the same population, the data from each
group also have a bearing on inference with respect to the other
groups, namely, through the information that is provided about the
population of groups.
In the analysis of covariance model, on the other hand, each of the
U0j is estimated as a separate parameter (or fixed effect). If group
sizes are small, then the data do not contain very much information
about the values of the U0j and there will be considerable
overfitting in the analysis of covariance model: many parameters
have large standard errors. This overfitting is avoided by using the
random coefficient model, because the U0j do not figure as
individual parameters. If, on the other hand, the group sizes are
large (say, 100 or more), then in the analysis of covariance the
group-dependent parameters U0j are estimated very precisely (with
small standard errors), and the additional information that they
come from the same population does not add much to this
precision. In such a situation the difference between the results of
the two approaches will be negligible.
5. If the researcher is interested only in within-group differences, and
wishes to control as much as possible for between-group
differences without modeling them as such or testing them, then
the analysis of covariance model is more appropriate. This is
because the only assumption made about the group differences is
that they can be represented by main group effects β0j as in (4.4),
without any assumption about their distribution or about how they
might be associated with the explanatory variables in the model.
Thus, the parameter β1 in the fixed effects interpretation of (4.4)
 represents only the within-group regression parameter: it is the
expected difference between values of the dependent variable for
cases in the same group differing by one unit on variable X, and
having the same values for all other explanatory variables, without
any influence of differences between the group averages.
6. The random coefficient model is mostly used with the additional
assumption that the random coefficients, U0j and Rij in (4.5), are
normally distributed. If this assumption is a very poor
approximation, results obtained with this model may be unreliable.
This can happen, for example, when there are more outlying
groups than can be accounted for by a normally distributed group
effect U0j with a common variance. Models for group residuals
with nonnormal distributions have been developed, however; see
Sections 10.8, 12.2, and 12.3.

Other discussions about the choice between fixed and random

coefficients can be found, for example, in Searle et al. (1992; Section 1.4),
Hsiao (1995, Section 8), and Greene (2008, Chapter 9). An often mentioned
condition for the use of random coefficient models is the restriction that the
random coefficients should be independent of the explanatory variables.
However, if there is a possible correlation between group-dependent
coefficients and explanatory variables, this residual correlation can be
removed, while continuing to use a random coefficient model, by also
including effects of the group means of the explanatory variables.
Therefore, such a correlation does not imply the necessity of using a fixed
effects model instead of a random coefficient model. This is treated in
Sections 4.6 and 10.2.1.
In order to choose between regarding the group-dependent intercepts
U0j as fixed statistical parameters and regarding them as random variables,
a rule of thumb that often works in educational and social research is the
following. This rule mainly depends on N, the number of groups in the data.
If N is small, say, N < 10, then use the analysis of covariance approach: the
problem with viewing the groups as a sample from a population in this case
is that the data will contain only scant information about this population. If
N is not small, say N ≥ 20; if, furthermore, the level-two units are indeed
regarded as a sample of an actual or hypothetical population (see the
discussion on analytic inference in Section 14.1.1) and the researcher
wishes to draw conclusions about differences in this population (such as
regression coefficients for level-two variables or unexplained variability
between level-two units); and if also nj is small or intermediate, say nj <
100 – then use the random coefficient approach. The reason here is that 20
or more groups is usually too large a number to be regarded as unique
categories. If the group sizes nj are large, say all nj ≥ 100, then it does not
matter much which view we take provided that within- and between-group
regression coefficients are appropriately differentiated (Section 4.6). The
situation with 10–20 groups is ambiguous in this respect. However, this rule
of thumb should be taken with a large pinch of salt and serves only to
provide a first hunch, not to determine the choice between fixed and
random effects.1

Populations and populations

Having chosen to work with a random coefficient model, the researcher
must be aware that more than one population is involved in the multilevel
analysis. Each level corresponds to a population. For example, for a study
of students in schools, there is a population of schools and a population of
students; for voters in municipalities, there is a population of municipalities
and a population of voters. Recall that in this book we take a model-based
view (cf. Section 14.1). This implies that populations are infinite
hypothetical entities, which express ‘what could be the case’. The random
residuals and coefficients can be regarded as representing the effects of
unmeasured variables and the approximate nature of the linear model.
Randomness, in this sense, may be interpreted as unexplained variability.

Sometimes a random coefficient model can be used also when the

population idea at the lower level is less natural. For example, in a study of
longitudinal data where respondents are measured repeatedly, a multilevel
model can be used with respondents at the second and measurements at the
first level: measurements are nested within respondents. Then the
population of respondents is an obvious concept. Measurements may be
related to a population of time points. This will sometimes be natural, but
not always. Another way of expressing the idea of random coefficient
models in such a situation is to say that the level-two units are a sample
from a population, residual (unexplained) variability is present at level one
as well as at level two, and this unexplained variability is represented as
random variation in a probability model.

4.4 Definition of the random intercept model

Here we treat the random coefficient view of model (4.5). This model, the
random intercept model, is a simple case of the so-called hierarchical linear
model. We shall not specifically treat the analysis of covariance model, and
refer for this purpose to texts on analysis of variance and covariance
(Shadish et al., 2002; Stevens, 2009). However, we shall encounter a
number of considerations from the analysis of covariance that also play a
role in multilevel modeling.

The empty model

Although this chapter follows an approach along the lines of regression
analysis, the simplest case of the hierarchical linear model is the random
effects analysis of variance model, in which the explanatory variables, X
and Z, do not figure. This model only contains random groups and random
variation within groups. It can be expressed as a model – the same model
encountered before in formula (3.1) – where the dependent variable is the
sum of a general mean, γ00, a random effect at the group level, U0j, and a
random effect at the individual level, Rij:

Groups with a high value of U0j tend to have, on average, high responses
whereas groups with a low value of U0j tend to have, on average, low
responses. The random variables U0j and Rij are assumed to have a mean of
0 (the mean of Yij is already represented by γ00), to be mutually
independent, and to have variances var(Rij) = σ2 and var(U0j) = τ20. In the
context of multilevel modeling (4.6) is called the empty model, because it
contains not a single explanatory variable. It is important because it
provides the basic partition of the variability in the data between the two
levels. Given model (4.6), the total variance of Y can be decomposed as the
sum of the level-two and level-one variances,

The covariance between two individuals (i and i′, with i ≠ i′) in the same
group j is equal to the variance of the contribution U0 j that is shared by
these individuals,

and their correlation is

This parameter is just the intraclass correlation coefficient ρI(Y) which we

encountered in Chapter 3. It can be interpreted in two ways: it is the
correlation between two randomly drawn individuals in one randomly
drawn group, and it is also the fraction of total variability that is due to the
group level.

Example 4.1 Empty model for language scores in elementary schools.

In this example a data set is used that will recur in examples in many
subsequent chapters. The data set is concerned with grade 8 students (age
about 11 years) in elementary schools in the Netherlands. Further
background about the study and primary results can be found in Brandsma
and Knuver (1989), Knuver and Brandsma (1993), and Doolaard (1999).
After deleting 258 students with missing values, the number of students
is M = 3,758, and the number of schools is N = 211. Class sizes in the
original data set range from 5 to 36. In the data set reduced by deleting
cases with missing data, the class sizes range from 4 to 34. The small class
sizes are for classes where grade 8 is combined with grade 7 in a larger
classroom; here only the grade 8 students are studied. One class per school
is included, so the class and the school level are the same in this data set.
The nesting structure is students within classes.2
 The dependent variable is the score on a language test. Most of the
analyses of this data set in this book are concerned with investigating how
the language test score depends on the pupil’s intelligence and his or her
family’s socio-economic status, and on related class variables.
Fitting the empty model yields the parameter estimates presented in
Table 4.1. The ‘deviance’ in this table is given for the sake of completeness
and later reference; this concept is explained in Chapter 6. In the examples
in this and the next chapter, the ML estimation method is used (see Section
4.7) in order to obtain deviances that can be used for comparing all the
models by the deviance (likelihood ratio) test of Section 6.2.

Table 4.1: Estimates for empty model.

The estimates = 62.85 and 18.12 yield an intraclass correlation

coefficient of = 18.12/80.97 = 0.22. This is on the high side but not
unusual, compared to other results in educational research (values between
0.10 and 0.25 are common – see Hedges and Hedberg, 2007). This indicates
that the grouping according to classes leads to an important similarity
between the results of different students in the same class, although (as
practically always) within-class differences are far larger than between-
class differences.
For the overall distribution of the language scores, these estimates
provide a mean of 41.00 and a standard deviation of The
mean of 41.00 should be interpreted as the expected value of the language
score for a random pupil in a randomly drawn class. This is close, but not
identical, to the raw mean 41.41 and standard deviation 8.89 of the sample
of 3,758 students. The reason for this difference is that the estimation of
model (4.6) implies a weighting of the various classes that is not taken into
account in the calculation of the raw mean and standard deviation.

The estimates obtained from multilevel software for the two variance
components, τ20 and σ2, will usually be slightly different from the estimates
obtained from the formulas in Section 3.3.1. The reason is that different
estimation methods are used: multilevel software uses the more efficient
ML or REML method (cf. Section 4.7 below), which in most cases cannot
be expressed in an explicit formula; the formulas of Section 3.3.1 are
explicit but less efficient.

One explanatory variable

The following step is the inclusion of explanatory variables. These are used
to try to explain part of the variability of Y; this refers to variability at level
two as well as level one. With just one explanatory variable X, model (4.5)
is obtained (repeated here for convenience):

Essential assumptions are that all residuals, U0j and Rij, are mutually
independent and have zero means given the values xij of the explanatory
variable. It is assumed that the U0j and Rij are drawn from normally
distributed populations. The population variance of the lower-level
residuals Rij is assumed to be constant across the groups, and is again
denoted by σ2; the population variance of the higher-level residuals U0j is
denoted by τ20. Thus, model (4.5) has four parameters: the regression
coefficients γ00 and γ10 and the variance components σ2 and τ20.
The random variables U0j can be regarded as residuals at the group
level, or group effects that are left unexplained by X. Since residuals, or
random errors, contain those parts of the variability of the dependent
variable that are not modeled explicitly as a function of explanatory
variables, this model contains unexplained variability at two nested levels.
This partition of unexplained variability over the various levels is the
essence of hierarchical random effects models.
 The fixed intercept γ00 is the intercept for the average group. The
regression coefficient γ10 can be interpreted as an unstandardized regression
coefficient in the usual way: a one-unit increase in the value of X is
associated with an average increase in Y of γ10 units. The residual variance
(i.e., the variance conditional on the value of X) is

while the covariance between two different individuals (i and i', with i ≠ i')
in the same group is

The fraction of residual variability that can be ascribed to level one is given
by and to level two by
Part of the covariance or correlation between two individuals in the
same group may be explained by their X-values, and part is unexplained.
This unexplained, or residual, correlation between the X-values of these
individuals is the residual intraclass correlation coefficient,

This parameter is the correlation between the Y-values of two randomly

drawn individuals in one randomly drawn group, controlling for variable X.
It is analogous to the usual intraclass correlation coefficient, but now
controlling for X. The formula for the (nonresidual, or raw) intraclass
correlation coefficient was just the same, but in that case the variance
parameters τ20 and σ2 referred to the variances in the empty model whereas
now they refer to the variances in model (4.5), which includes the effect of
variable X. The values of residual intraclass correlations often are smaller
than those of raw intraclass correlation coefficients, and in educational
research often range between 0.05 and 0.20 (Hedges and Hedberg, 2007).
If model (4.5) is valid while the intraclass correlation coefficient is 0
(U0j = 0 for all groups j), then the grouping is irrelevant for the Y-variable
conditional on X, and one could have used ordinary linear regression, that
is, a model such as (4.1). If the residual intraclass correlation coefficient (or
equivalently τ20) is positive, then the hierarchical linear model is a better
method of analysis than OLS regression analysis because the standard
errors of the estimated coefficients produced by ordinary regression
analysis are not to be trusted. This was discussed in Section 3.2.3
The residual intraclass correlation coefficient is positive if and only if
the intercept variance τ20 is positive. Therefore testing the residual
intraclass correlation coefficient amounts to the same thing as testing the
intercept variance. Tests for this parameter are presented in Sections 6.2 and
6.3.
Example 4.2 Random intercept and one explanatory variable: IQ.
As a variable at the pupil level that is essential for explaining language score, we use the measure for
verbal IQ taken from the ISI test (Snijders and Welten, 1968). The IQ score has been centered, so that
its mean is 0. (The centering took place before taking out the 258 students with some missing values,
so that in this data set the mean is close but not exactly equal to 0.) This facilitates interpretation of
various parameters. Its standard deviation in this data set is 2.04 (this is calculated as a descriptive
statistic, without taking the grouping into account). The results are presented in Table 4.2.
In the model presented in Table 4.2 each class, indexed by the letter j, has its own regression line,
given by

Table 4.2: Estimates for random intercept model with effect for IQ.

The U0j are class-dependent deviations of the intercept and have a mean of 0 and a variance of
(hence, a standard deviation of 9.85 = 3.14). Figure 4.2 depicts 15 such random
regression lines. This figure can be regarded as a random sample from the population of schools
defined by Table 4.2.
The scatter around the regression lines (i.e., the vertical distances Rij between the observations
and the regression line for the class under consideration) has a variance of 40.47 and therefore a
standard deviation of These distances between observations and regression lines
therefore tend to be much larger than the vertical distances between the regression lines. However,
the distances between the regression lines are not negligible.
A school with a typical low average achievement (bottom 2.5%) will have a value of U0j of about
two standard deviations below the expected value of U0j, so that it will have a regression line

whereas a school with a typical high achievement (top 2.5%) will have a regression line

There appears to be a strong effect of IQ. Each additional measurement unit of IQ leads, on average,
to 2.507 additional measurement units of the language score. To obtain a scale for effect that is
independent of the measurement units, one can calculate standardized coefficients, that is,
coefficients expressed in standard deviations as scale units. These are the coefficients that would be
obtained if all variables were rescaled to unit variances. They are given by

in this case estimated by (2.04/9.00) × 2.507 = 0.57. In other words, each additional standard
deviation on IQ leads, on average, to an increase in language score of 0.57 standard deviations.
The residual variance σ2 as well as the random intercept variance τ20 are much lower in this
model than in the empty model (cf. Table 4.1). The residual variance is lower because between-pupil
differences are partially explained. The intercept variance is lower because classes differ in average
 Figure 4.2: Fifteen randomly chosen regression lines according to the
random intercept model of Table 4.2.
IQ score, so that this pupil-level variable also explains part of the differences between classes. The
residual intraclass correlation is estimated by

slightly smaller than the raw intraclass correlation of 0.22 (see Table 4.1).
These results may be compared to those obtained from an OLS regression analysis, in which the
nesting of students in classes is not taken into account. This analysis can be regarded as an analysis
using model (4.5) in which the intercept variance τ20 is constrained to be 0. The results are displayed
in Table 4.3. The parameter estimates for the OLS method seem rather close to those for the random
intercept model. However, the regression coefficient for IQ differs by about three standard errors
between the two models. This implies that, although the numerical values seem similar, they are
nevertheless rather different from a statistical point of view. Further, the standard error of the
intercept is twice as large in the results from the random intercept model as in those from the OLS
analysis. This indicates that the OLS analysis produces an over-optimistic impression of the precision
of this estimate, and illustrates the lack of trustworthiness of OLS estimates for multilevel data.

4.5 More explanatory variables

Just as in multiple regression analysis, more than one explanatory variable
can be used in the random intercept model. When the explanatory variables
at the individual level are denoted by X1, . . . , Xp, and those at the group
level Z1, . . . , Zq, adding their effects to the random intercept model leads to
the formula4

Table 4.3: Estimates for OLS regression.

The regression parameters, γh0 (h = 1, . . . , p) and γ0h (h = 1, . . . , q) for

level-one and level-two explanatory variables, respectively, again have the
same interpretation as unstandardized regression coefficients in multiple
regression models: a one-unit increase in the value of Xh (or Zh) is
associated with an average increase in Y of γh0 (or γ0h) units. Just as in
multiple regression analysis, some of the variables Xh and Zh may be
interaction variables, or nonlinear (e.g., quadratic) transforms of basic
variables.
The first part of the right-hand side of (4.8), incorporating the regression
coefficients,

is called the fixed part of the model, because the coefficients are fixed (i.e.,
nonstochastic). The remainder,

is called the random part of the model.

 It is again assumed that all residuals, U0j and Rij, are mutually
independent and have zero means given the values of the explanatory
variables. A somewhat less crucial assumption is that these residuals are
drawn from normally distributed populations. The population variance of
the level-one residuals Rij is denoted by σ2, while the population variance of
the level-two residuals U0j is denoted by τ20.
Due to the two-level nature of the model under consideration, two
special kinds of variables deserve special attention: group-level variables
that arise as aggregates of individual-level variables; and cross-level
interactions. The latter will be treated in Chapter 5; the former are treated
here and in the next section.
Some level-two variables are directly defined for the units of this level;
others are defined through the subunits (units at level one) that are
comprised in this unit. For example, when the nesting structure refers to
children within families, the type of dwelling is directly a family-dependent
variable, but the average age of the children is based on aggregation of a
variable (age) that is itself a level-one variable. As another example,
referring to a longitudinal study of primary school children, where the
nesting structure refers to repeated measurements within individual
children, the gender of the child is a direct level-two variable, whereas the
average reading score over the age period of 6–12 years, or the slope of the
increase in reading score over this period, is a level-two variable that is an
aggregate of a level-one variable.

Varying group sizes

In most research, the group sizes nj are variable between groups. Even in
situations such as repeated measures designs, where level two corresponds
to individual subjects and level one to time points of measurement of these
subjects, and where it is often the researcher’s intention to measure all
subjects at the same moments, there often occur missing data which again
leads to a data structure with variable nj. This in no way constitutes a
problem for the application of the hierarchical linear model. The
hierarchical linear model can even be applied if some groups have size nj =
1, as long as some other groups have greater sizes. Of course, the smaller
groups will have a smaller influence on the results than the larger groups.
 However, it may be worthwhile to give some thought to the substantive
meaning of the group size. The number of students in a school class may
have effects on teaching and learning processes, the size of a department in
a firm may have influence on the performance of the department and on
interpersonal relations. In such cases it may be advisable to include group
size in the model as an explanatory variable with a fixed effect.

4.6 Within- and between-group regressions

Group means are an especially important type of explanatory variable. A
group mean for a given level-one explanatory variable is defined as the
mean over all individuals, or level-one units, within the given group, or
level-two unit.5 This can be an important contextual variable. The group
mean of a level-one explanatory variable allows the difference between
within-group and between-group regressions to be expressed, as proposed
by Davis et al. (1961) and extended by various authors, including Neuhaus
and Kalbfleisch (1998). We saw in Section 3.6 that the coefficients for these
two types of regression can be completely different. The within-group
regression coefficient expresses the effect of the explanatory variable within
a given group; the between-group regression coefficient expresses the effect
of the group mean of the explanatory variable on the group mean of the
dependent variable. In other words, the between-group regression
coefficient is just the coefficient in a regression analysis for data that are
aggregated (by averaging) to the group level.
Continuing with the example of children within families, suppose that
we are interested in the amount of pocket money that children receive. This
will depend on the child’s age, but it could also depend on the average age
of the children in the family. The within-group regression coefficient
measures the effect of age differences within a given family; the between-
group regression coefficient measures the effect of average age on the
average pocket money received by the children in the family. In a simple
model, not taking other variables into account, denote age (in years) of
child i in family j by xij, and the average age of all children in family j by
In the model
the within-group and between-group regression coefficients are forced to be
equal. If we add the family mean, as an explanatory variable, we
obtain

This model is more flexible in that the within-group regression coefficient

is allowed to differ from the between-group regression coefficient. This can
be seen as follows.
If model (4.9) is considered within a given group, the terms can be
reordered as

The part in parentheses is the (random) intercept for this group, and the
regression coefficient of X within this group is γ10. The systematic
(nonrandom) part for group j is the within-group regression line

On the other hand, taking the group average on both sides of the equality
sign in (4.9) yields the between-group regression model,

The systematic part of this model is represented by the between-group

regression line

This shows that the between-group regression coefficient is γ10 + γ01.

The difference between the within-group and the between-group
regression coefficients can be tested in this model by testing the null
hypothesis that γ01 = 0 by the method of Section 6.1.
 This test is a version of what is known in econometrics as the Hausman
specification test; see Hausman and Taylor (1981) or Baltagi (2008, Section
4.3). Some textbooks suggest the use of this test as the primary means of
choosing between a fixed effects and a random effects approach; see, for
example, Greene (2008, Section 9.5.4). However, if this test is significant, it
is not necessary to conclude that one should abandon the use of a random
effects model; there may be reasons to continue using a random effects
model while including the group mean of X as an explanatory variable. The
choice between fixed and random effects models should be based on other
considerations, as reviewed in Section 4.3. This incorrect interpretation of
the Hausman test is discussed in Fielding (2004a) and Snijders and Berkhof
(2008, p. 147).
Figure 4.3 is a sketch of within-group regression lines that differ from
the between-group regression line in the sense that the between-group
regression is stronger (γ01 > 0). The reverse, where the between-group
regression line is less steep than the within-group regression lines, can also
occur.

Figure 4.3: Different between-group and within-group regression lines.

If the within- and between-group regression coefficients are different,

then it is often convenient to replace xij in (4.9) with the within-group
deviation score, defined as To distinguish the corresponding
parameters from those in (4.9), they are denoted by . The resulting model
is
This model is statistically equivalent to model (4.9) but has a more
convenient parametrization because the between-group regression
coefficient is now

while the within-group regression coefficient is

The use of the within-group deviation score is called within-group

centering; some computer programs for multilevel analysis have special
facilities for this.
Example 4.3 Within- and between-group regressions for IQ.
We continue Example 4.2 by allowing differences between the within-group and between-group
regressions of the language score on IQ. The results are displayed in Table 4.4. IQ here is the variable
with overall centering but no group centering. Thus, the results refer to model (4.9).
The within-group regression coefficient is 2.454 and the between-group regression coefficient is
2.454 + 1.312 = 3.766. A pupil with a given IQ obtains, on average, a higher language test score if he
or she is in a class with a higher average IQ. In other words, the contextual effect of mean IQ in the
class gives an additional contribution over and above the effect of individual IQ. The effect of
classroom average IQ is about half that of individual IQ. A graph of these results would be
qualitatively be similar to Figure 4.3 in the sense that the within-group regression lines are less steep
than the between-group regression line.
Table 4.4 represents within each class, denoted j, a linear regression equation

Table 4.4: Estimates for random intercept model with different within- and
between-group regressions.
where U0j is a class-dependent deviation with mean 0 and variance 8.68 (standard deviation 2.95).
The within-class deviations about this regression equation, Rij, have a variance of 40.43 (standard
deviation 6.36). Within each class, the effect (regression coefficient) of IQ is 2.454, so the regression
lines are parallel. Classes differ in two ways: they may have different mean IQ values, which affects
the expected results Y through the term this is an explained difference between the classes;
and they have randomly differing values for U0j, which is an unexplained difference. These two
ingredients contribute to the class-dependent intercept, given by 41.11
The within-group and between-group regression coefficients would be equal if, in formula (4.9),
the coefficient of average IQ were 0 (i.e., γ01 = 0). This null hypothesis can be tested (see Section
6.1) by the t-ratio defined as

given here by 1.312/0.262 = 5.01, a highly significant result. In other words, we may conclude that
the within- and between-group regression coefficients are indeed different.
If the individual IQ variable had been replaced by within-group deviation scores
that is, model (4.10) had been used, then the estimates obtained would have been and
cf. formulas (4.11) and (4.12). Indeed, the regression equation given above can be
described equivalently by

which indicates explicitly that the within-group regression coefficient is 2.454, while the between-
group regression coefficient (i.e., the coefficient of the group means on the group means ) is
3.766.

When interpreting the results of a multilevel analysis, it is important to

keep in mind that the conceptual interpretation of within-group and
between-group regression coefficients usually is completely different.
These two coefficients may express quite contradictory mechanisms. This is
related to the shift of meaning and the ecological fallacy discussed in
Section 3.1. For theoretically important variables in multilevel studies, it is
the rule rather than the exception that within-group regression coefficients
differ from between-group regression coefficients (although the statistical
significance of this difference may be another matter, depending as it does
on sample sizes, etc.).

4.7 Parameter estimation

The random intercept model (4.8) is defined by its statistical parameters:
the regression parameters, γ, and the variance components, σ2 and τ20. Note
that the random effects, U0j, are not parameters in a statistical sense, but
latent (i.e., not directly observable) variables. The literature (e.g., Longford,
1993; de Leeuw and Meijer, 2008b) contains two major estimation methods
for estimating the statistical parameters, under the assumption that the U0j
as well as the Rij are normally distributed: maximum likelihood (ML) and
residual (or restricted) maximum likelihood (REML).
The two methods differ little with respect to estimating the regression
coefficients, but they do differ with respect to estimating the variance
components. A very brief indication of the difference between the two
methods is that REML estimates the variance components while taking into
account the loss of degrees of freedom resulting from the estimation of the
regression parameters, while ML does not take this into account. The result
is that the ML estimators for the variance components have a downward
bias, and the REML estimators do not. For example, the usual variance
estimator for a single-level sample, in which the sum of squared deviations
is divided by the sample size minus 1, is a REML estimator; dividing
instead by the total sample size gives the corresponding ML estimator.
The relative difference between ML and REML estimates will usually
be largest for the estimated variance parameters (and covariance parameters
for random slope models as treated in Chapter 5). In a first approximation,
for a two-level random intercept model the ratio of ML to REML estimates
for the random intercept variance may be expected to be something like N/(
), where q is the number of level-two explanatory variables (to
which should be added a fractional part of level-one explanatory variables
that have a strong intraclass correlation). The difference will be important
when is small. For a large number (as a rule of thumb, ‘large’
here means the difference between the ML and the
REML estimates will be immaterial. The literature (e.g., McCulloch and
Searle, 2001, Section 6.10) suggests that the REML method is preferable
with respect to the estimation of the variance parameters (and the
covariance parameters for the more general models treated in Chapter 5).
When one wishes to carry out deviance tests (see Section 6.2), however, the
use of ML rather than REML estimates is sometimes required.6 For this
reason, the estimates for the examples in this chapter and the next have
been calculated by the ML method. On the other hand, we shall see in
Section 6.1 that for tests of fixed effects it is better, for small values of N, to
use standard errors produced by the REML method.
Example 4.4 ML and REML estimates.
Example 4.3 presents ML estimates. The REML estimates differ from these in the second or third
decimal place, with the exception of the random intercept variance. The ML estimate is 8.68
(standard error 1.10), the REML estimate 8.78 (standard error 1.11). There are 211 classes in the data
set and there is q = 1 level-one explanatory variable, average IQ. This leads to the expectation that
the REML estimate will be roughly 209/211 = 0.99 times the ML estimate. Indeed, 8.68/8.78 = 0.99
(the difference is in the third decimal place). The difference of 8.78 − 8.68 = 0.10 between the
estimates is less than one tenth of the standard error, and therefore negligible in practice.

Various algorithms are available to determine these estimates. They

have names such as expectation-maximization (EM), Fisher scoring,
iterative generalized least squares (IGLS), and residual or restricted IGLS
(RIGLS). They are iterative, which means that a number of steps are taken
in which a provisional estimate comes closer and closer to the final
estimate. When all goes well, the steps converge to the ML or REML
estimate. Technical details can be found, for example, in Raudenbush and
Bryk (2002), Goldstein (2011), Longford (1993, 1995), and de Leeuw and
Meijer (2008b). In principle, the algorithms all yield the same estimates for
a given estimation method (ML or REML). The differences are that for
some complicated models, the algorithms may vary in the extent of
computational problems (sometimes one algorithm may converge and the
other not), and the amount of computing time required. For the practical
user, the differences between the algorithms are hardly worth thinking
about.
 An aspect of the estimation of hierarchical linear model parameters that
surprises some users of this model is the fact that it is possible that the
variance parameters, in model (4.8) notably parameter τ20, can be estimated
to be exactly 0. The value of 0 is then also reported for the standard error by
many computer programs! Even when the estimate is this does not
mean that the data imply absolute certainty that the population value of τ02
is equal to 0. Such an estimate can be understood as follows. For simplicity,
consider the empty model, (4.6). The level-one residual variance σ2 is
estimated by the pooled within-group variance. The parameter τ20 is
estimated by comparing this within-group variability to the between-group
variability. The latter is determined not only by τ20 but also by σ2, since

Note that τ20, being a variance, cannot be negative. This implies that, even
if τ20 = 0, a positive between-group variability is expected. If observed
between-group variability is equal to or smaller than what is expected from
(4.13) for τ20 = 0, then the estimate is reported (cf. the discussion
following (3.11)).
If the group sizes nj are variable, the larger groups will, naturally, have a
larger influence on the estimates than the smaller groups. The influence of
group size on the estimates is, however, mediated by the intraclass
correlation coefficient. Consider, for example, the estimation of the mean
intercept, γ00. If the residual intraclass correlation is 0, the groups have an
influence on the estimated value of γ00 that is proportional to their size. In
the extreme case where the residual intraclass correlation is 1, each group
has an equally large influence, independent of its size. In practice, where
the residual intraclass correlation is between 0 and 1, larger groups will
have a larger influence, but less than proportionately.

4.8 ‘Estimating’ random group effects: posterior

means
The random group effects U0j are latent variables rather than statistical
parameters, and therefore are not estimated as an integral part of the
statistical parameter estimation. However, there may be many reasons why
it is nonetheless desirable to ‘estimate’ them.7 This can be done by a
method known as empirical Bayes estimation which produces so-called
posterior means; see, for example, Efron and Morris (1975), Gelman et al.
(2004, Section 5.4). The basic idea of this method is that U0j is ‘estimated’
by combining two kinds of information: the data from group j; and the fact
(or, rather, the model assumption) that the unobserved U0j is a random
variable just like all other random group effects, and therefore has a normal
distribution with mean 0 and variance τ20 In other words, data information
is combined with population information.
The formula is given here only for the empty model, that is, the model
without explanatory variables. The idea for more complicated models is
analogous; formulas can be found in the literature, for example, Longford
(1993, Section 2.10) and Snijders and Berkhof (2008, Section 3.3.3).
The empty model was formulated in (4.6) as

Since γ00 is already an estimated parameter, an estimate for β0j will be the
same as an estimate for U0j +γ00. Therefore, estimating β0j and estimating
U0j are equivalent problems given that an estimate for γ00 is available.
If we only used group j, β0j would be estimated by the group mean,
which is also the OLS estimate,

If we looked only at the population, we would estimate β0j by its population

mean, γ00. This parameter is estimated by the overall mean,
where M = Σj nj denotes the total sample size. Another possibility is to
combine the information from group j with the population information. The
optimal combined ‘estimate’ for β0j is a weighted average of the two
previous estimates,

where EB stands for ‘empirical Bayes’ and the weight λj is defined as the
reliability of the mean of group j (see (3.21)),

The ratio of the two weights, λj/(1– λj), is just the ratio of the true variance
τ02 to the error variance σ2/nj. In practice we do not know the true values of
the parameters σ2 and T0, and we substitute estimated values to calculate
(4.15).
Formula (4.15) is called the posterior mean or empirical Bayes estimate,
for β0j. This term comes from Bayesian statistics. It refers to the distinction
between the prior knowledge about the group effects, which is based only
on the population from which they are drawn, and the posterior knowledge
which is based also on the observations made about this group. There is an
important parallel between random coefficient models and Bayesian
statistical models, because the random coefficients used in the hierarchical
linear model are analogous to the random parameters that are essential in
the Bayesian statistical paradigm. See Gelman et al. (2004, Chapter 1 and
Section 5.4).
Formula (4.15) can be regarded as follows. The OLS estimate (4.14) for
group j is pushed slightly toward the general mean This is an example of
shrinkage to the mean just as is used (e.g., in psychometrics) for the
estimation of true scores. The corresponding estimator is sometimes called
the Kelley estimator; see, for example, Kelley (1927), Lord and Novick
(1968), or other textbooks on classical psychological test theory. From the
definition of the weight λj it is apparent that the influence of the data of
group j itself becomes larger as group size nj becomes larger. For large
groups, the posterior mean is practically equal to the OLS estimate the
intercept that would be estimated from data on group j alone.
In principle, the OLS estimate (4.14) and the empirical Bayes estimate
(4.15) are both sensible procedures for estimating the mean of group j. The
former is an unbiased8 estimate, and does not require the assumption that
group j is a random element from the population of groups. The latter is
biased toward the population mean, but for a randomly drawn group it has a
smaller mean squared error. The squared error averaged over all groups will
be smaller for the empirical Bayes estimate, but the price is a conservative
(drawn to the average) appraisal of the groups with truly very high or very
low values of β0j.The estimation variance of the empirical Bayes estimate is

if the uncertainty due to the estimation of γ00 and λj is neglected, which is a

good approximation for medium and large level-two sample sizes N. This
formula also is well known from classical psychological test theory (e.g.,
Lord and Novick, 1968).
The same principle can be applied (but with more complicated
formulas) to the ‘estimation’ of the group-dependent intercept β0j = γ00 +
U0j in random intercept models that do include explanatory variables, such
as (4.8). This intercept can be ‘estimated’ again by γ00 plus the posterior
mean of U0 j and is then also referred to as the posterior intercept.
Instead of being primarily interested in the intercept as defined by γ00 +
U0j, which is the value of the regression equation for all explanatory
variables having value 0, one may also be interested in the value of the
regression line for group j for the case where only the level-one variables 0
while the level-two variables have the values proper to this group. To
‘estimate’ this version of the intercept of group j, we use

where the values indicate the (ML or REML) estimates of the regression
coefficients. The values (4.17) also are sometimes called posterior
intercepts.
 The posterior means (4.15) can be used, for example, to see which
groups have unexpectedly high or low values on the outcome variable,
given their values on the explanatory variables. They can also be used in a
residual analysis, for checking the assumption of normality for the random
group effects, and for detecting outliers (see Chapter 10). The posterior
intercepts (4.17) indicate the total main effect of group j, controlling for the
level-one variables X1, . . . , Xp, but including the effects of the level-two
variables Z1, . . . , Zq. For example, in a study of students in schools where
the dependent variable is a relevant indicator of scholastic performance,
these posterior intercepts could be valuable information for the parents
indicating the contribution of the various schools to the performance of
their beloved children.
Example 4.5 Posterior means for random data.
We can illustrate the ‘estimation’ procedure by returning to the random digits table (Chapter 3, Table
3.1). Macro-unit 04 in that table has an average of Ȳj = 31.5 over its 10 random digits. The grand
mean of the total 100 random digits is Ȳ = 47.2. The average of macro-unit 04 thus seems to be far
below the grand mean. But the reliability of this mean is only λj = 26.7/{26.7+(789.7/10)} = 0.25.
Applying (4.15), the posterior mean is calculated as

In words, the posterior mean for macro-unit 04 is 75% (i.e., 1 — λj) determined by the grand mean of
47.2 and only 25% (i.e.,λj) by its OLS mean of 31.5. The shrinkage to the grand mean is evident.
Because of the low estimated intraclass correlation of = 0.03 and the low number of observations
per macro-unit, nj = 10, the empirical Bayes estimate of the average of macro-unit 04 is closer to the
grand mean than to the group mean. In this case this is a clear improvement: there is no between-
group variance in the population, and the posterior mean is much closer to the true value of γ00 + U0j
= 49.5 + 0 = 49.5 than the group average.

4.8.1 Posterior confidence intervals

Now suppose that parents have to choose a school for their children, and
that they wish to do so on the basis of the value a school adds to abilities
that students already have when entering the school (as indicated by an IQ
test). Let us focus on the language scores. ‘Good’ schools are schools where
students on average are ‘over-achievers’, that is to say, they achieve more
than expected on the basis of their IQ. ‘Poor’ schools are schools where
students on average have language scores that are lower than one would
expect given their IQ scores.
In this case the level-two residuals U0j from a two-level model with
language as the dependent and IQ as the predictor variable convey the
relevant information. But remember that each U0j has to be estimated from
the data, and that there is sampling error associated with each residual,
since we work with a sample of students from each school. Of course we
might argue that within each school the entire population of students is
studied, but in general we should handle each parameter estimate with its
associated uncertainty since we are now considering the performance of the
school for a hypothetical new pupil at this school.
Therefore, instead of simply comparing schools on the basis of the
level-two residuals it is better to compare these residuals taking account of
the associated confidence intervals.
These must be constructed using the standard errors of the empirical
Bayes estimates. Here we must distinguish between two kinds of standard
error. The comparative standard error expresses the deviation from the
unobserved random variable U0j, and is the standard deviation of

The diagnostic standard error expresses the deviation from the value 0,
which is the overall mean of the variables U0j, and is the standard deviation
of

The comparative standard error, also called the posterior standard deviation
of U0j, is used to assess how well the unobserved level-two contributions
U0j can be ‘estimated’ from the data, for example, to compare groups (more
generally, level-two units) with each other, as is further discussed later in
this section. The diagnostic standard error is used when the posterior means
must be standardized for model checking. Using division by the diagnostic
standard errors (and if we may ignore the fact that these are themselves
subject to error due to estimation), the standardized empirical Bayes
estimators can be regarded as standardized residuals having a standard
normal distribution if the model is correct, and this can be used for model
checking. This is discussed in Chapter 10.
A theoretically interesting property is that the sum of the diagnostic and
comparative variances is equal to the random coefficient variance:

This is shown more generally for hierarchical linear models by Snijders and
Berkhof (2008, Section 3.3.3). The interpretation is that, to the extent that
we are better able to estimate the random coefficients (smaller comparative
standard errors) – for example, because of larger group sizes nj – the
variability of the estimated values will increase (larger diagnostic
standard errors) because they capture more of the true variation between the
coefficients U0j.
The comparative standard error of the empirical Bayes estimate is
smaller than the root mean squared error of the OLS estimate based on the
data only for the given macro-unit (the given school, in our example). This
is just the point of using the empirical Bayes estimate. For the empty model
the comparative standard error is the square root of (4.16), which can also
be expressed as

This formula was also given by Longford (1993, Section 1.7). Thus, the
standard error depends on the within-group as well as the between-group
variance and on the number of sampled students for the school. For models
with explanatory variables, the standard error can be obtained from
computer output of multilevel software. Denoting the standard error for
school j shortly by S.E.j, the corresponding 90% confidence intervals can be
calculated as the intervals

Two cautionary remarks are in order, however.

 First, the shrinkage construction of the empirical Bayes estimates
implies a bias: ‘good’ schools (with a high U0j) will tend to be represented
too negatively, ‘poor’ schools (with low U0j) will tend to be represented too
positively (especially if the sample sizes are small). The smaller standard
error is bought at the expense of this bias. These confidence intervals have
the property that, on average, the random group effects U0j will be included
in the confidence interval for 90% of the groups. But for close to average
groups the coverage probability is higher, while for the groups with very
low or very high group effects the coverage probability will be lower than
90%.
Second, users of such information generally wish to compare a series of
groups. This problem was addressed by Goldstein and Healy (1995). The
parent in our example will make her or his own selection of schools, and (if
the parent is a trained statistician) will compare the schools on the basis of
whether the confidence intervals overlap. In that case, the parent is
implicitly performing a series of statistical tests on the differences between
the group effects U0j. Goldstein and Healy (1995, p. 175) write: ‘It is a
common statistical misconception to suppose that two quantities whose
95% confidence intervals just fail to overlap are significantly different at
the 5% significance level’. The reader is referred to their article on how to
adjust the width of the confidence intervals in order to perform such
significance testing. For example, testing the equality of a series of level-
two residuals at the 5% significance level requires confidence intervals that
are constructed by multiplying the standard error given above by 1.39 rather
than the well-known 5% value of 1.96. For 10% significance level, the
factor is 1.24 rather than 1.64. So the ‘comparative confidence intervals’ are
allowed to be narrower than the confidence intervals used for assessing
single groups.
Example 4.6 Comparing added value of schools.
Table 4.4.2 presents the multilevel model where language scores are controlled for IQ. The posterior
means which can be interpreted as the estimated value added, are presented graphically in
Figure 4.4. The figure also presents the confidence intervals for testing the equality of any pair of
residuals at a significance level of 5%. The end points of these intervals are given by the posterior
mean plus or minus 1.39 times the comparative standard error. For convenience the schools are
ordered by the posterior mean.
There are 18 schools having a confidence interval overlapping with the confidence interval of the
best school in this sample, implying that their value-added scores do not differ significantly. At the
lower extreme, 34 schools do not differ significantly from the worst school in this sample. We can
also deduce from the graph that about half of this sample of schools have approximately average
scores that cannot be distinguished statistically from the mean value of 0.

Example 4.7 Posterior confidence intervals for random data.

Let us also look, once again, at the random digits example. In Figure 4.5 we have graphed for each of
the 10 macro-units (now in the original order) their OLS means and their posterior means with
confidence intervals.

Figure 4.4: The added value scores for 211 schools with comparative
posterior confidence intervals.
Once again we clearly observe the shrinkage since the OLS means (×) are further apart than the
posterior means (•). Furthermore, as we would expect from a random digits example, none of the
pairwise comparisons results in any significant differences between the macro-units since all 10
confidence intervals overlap.

4.9 Three-level random intercept models

The three-level random intercept model is a straightforward extension of
the two-level model. In previous examples, data were used where students
were nested within schools. The actual hierarchical structure of educational
data is, however, students nested within classes nested within schools.
Other examples are: siblings within families within neighborhoods, and
people within regions within states. Less obvious examples are: students
within cohorts within schools, and longitudinal measurements within
persons within groups. These latter cases will be illustrated in Chapter 15
on longitudinal data. For the time being we concentrate on ‘simple’ three-
level hierarchical data structures. The dependent variable now is denoted by
Yijk, referring to, for example, pupil i in class j in school k. More generally,
one can talk about level-one unit i in level-two unit j in level-three unit k.
The three-level model for such data with one explanatory variable may be
formulated as a regression model

where β0jk is the intercept in level-two unit j within level-three unit k. For
the intercept we have the level-two model,

Figure 4.5: OLS means (×) and posterior means with comparative
posterior confidence intervals.

where δ00k is the average intercept in level-three unit k. For this average
intercept we have the level-three model,

This shows that there are now three residuals, as there is variability on three
levels. Their variances are denoted by
The total variance between all level-one units now equals σ2 + τ2 + φ2, and
the population variance between the level-two units is τ2 + φ2. Substituting
(4.22) and (4.21) into the level-one model (4.20) and using (in view of the
next chapter) the triple indexing notation γ100 for the regression coefficient
β1 yields

Example 4.8 A three-level model: students in classes in schools.

For this example we use a data set on 3,792 students in 280 classes in 57 secondary schools with
complete data (see Opdenakker and Van Damme, 1997). On entering the school students were
administered tests on IQ, mathematical ability, and achievement motivation, and data were also
collected on the educational level of the father and the students’ gender.
The response variable is the score on a mathematics test administered at the end of the second
grade of secondary school (when the students were approximately 14 years old). Table 4.5 contains
the results of the analysis of the empty three-level model (Model 1) and a model with a fixed effect
of student intelligence.
The total variance is 11.686, the sum of the three variance components. Since this is a three-level
model there are several kinds of intraclass correlation coefficient. Of the total variance, 2.124/11.686
= 18% is situated at the school level while (2.124 + 1.746) /11.686 = 33% is situated at the class and
school level. The level-three intraclass correlation expressing the likeness of

Table 4.5: Estimates for three-level model.

students in the same schools thus is estimated to be 0.18, while the intraclass correlation expressing
the likeness of students in the same classes and the same schools thus is estimated to be 0.33. In
addition, one can estimate the intraclass correlation that expresses the likeness of classes in the same
schools. This level-two intraclass correlation is estimated to be 2.124/(2.124 + 1.746)= 0.55. This is
more than 0.5: the school level contributes slightly more to variability than the class level. The
interpretation is that if one randomly takes two classes within one school and calculates the average
mathematics achievement level in one of the two, one can predict reasonably accurately the average
achievement level in the other class. Of course we could have estimated a two-level model as well,
ignoring the class level, but that would have led to a redistribution of the class-level variance to the
two other levels, and it would affect the validity of hypothesis tests for added fixed effects.
Model 2 shows that the fixed effect of IQ is very strong, with a t-ratio (see Section 6.1) of
0.121/0.005 = 24.2. (The intercept changes drastically because the IQ score does not have a zero
mean; the conventional IQ scale, with a population mean of 100, was used.) Adding the effect of IQ
leads to a stronger decrease in the class- and school-level variances than in the student-level variance.
This suggests that schools and classes are rather internally homogeneous with respect to IQ and/or
that intelligence may play its role partly at the school and class levels.

Another example in which the various different types of intraclass

correlation coefficient in a three-level model are discussed is Siddiqui et al.
(1996).
As in the two-level model, predictor variables at any of the three levels
can be added. All features of the two-level model can be generalized quite
straightforwardly to the three-level model – significance testing, model
building, testing the model fit, centering of variables, etc. – although the
researcher should now be more careful because of the more complicated
formulation.
For example, for a level-one explanatory variable there may be three
kinds of regressions. In the school example, these are the within-class
regression, the within-school/between-class regression, and the between-
school regression. Coefficients for these distinct regressions can be obtained
by using the class means as well as the school means as explanatory
variables with fixed effects.
Example 4.9 Within-class, between-class, and between-school regressions.
Continuing the previous example, we now investigate whether indeed the effect of IQ is in part a
class-level or school-level effect: in other words, whether the within-class, between-class/within-
school, and between-school regressions are different. Table 4.6 presents the results.
In Model 3, the effects of the class mean, , as well as the school mean , have been
added. The class mean has a clearly significant effect (t = 0.106/0.013 = 8.15), which indicates that
between-class regressions are different from within-class regressions. The school mean does not have
a significant effect (t = 0.039/0.028 = 1.39), so there is no evidence that the between-school
regressions are different from the between-class regressions. It can be concluded that for the
explanation of mathematics achievement, the composition with respect to intelligence plays a role at
the class level, but not at the school level.

Table 4.6: Estimates for three-level model with distinct within-class, within-
school, and between-school regressions.
 As in Section 4.6, replacing the variables by the deviation scores leads to an equivalent model
formulation in which, however, the within-class, between-class, and between-school regression
coefficients are given directly by the fixed parameters. In the three-level case, this means that we
must use the following three variables:
the within-class deviation score of the student from the class mean;
the within-school deviation score of the class mean from the school mean; and
the school mean itself.
The results are shown as Model 4.
We see here that the within-class regression coefficient is 0.107, equal to the coefficient of
student-level IQ in Model 3; the between-class/within-school regression coefficient is 0.212, equal
(up to rounding errors) to the sum of the student-level and the class-level coefficients in Model 3;
while the between-school regression coefficient is 0.252, equal to the sum of all three coefficients in
Model 3. From Model 3 we know that the difference between the last two coefficients is not
significant.

4.10 Glommary
Regression model. A statistical model for investigating how a dependent
variable can be predicted, or explained, from one or more explanatory
variables, also called independent variables or predictor variables.
The usual form for this dependence is a linear model.
Intercept. The constant term in a linear function. For the linear function Y
= a + bX, the intercept is a; this is the value of Y corresponding to X =
0.

Fixed effects. Coefficients in a regression model considered as fixed (i.e.,

nonrandom) parameters.

Random effects. Coefficients in a regression model considered as random

parameters, that is, parameters distributed according to some
probability distribution. Normal distributions are often used here, but
other distributional shapes can also be used.

Residuals. Another term used for random variables in a regression model,

sometimes called ‘errors’, and representing unexplained variability.
Random coefficients are a special case. The term Rij in (4.8) is a
residual which is not usually called a random coefficient (because
coefficients are numbers multiplying something else).

Homoscedasticity. The assumption that residuals for different cases have

the same variance. The converse is called heteroscedasticity.

Fixed effects model. A statistical model with fixed effects only, where the
only random term is the residual term at the lowest level.

Analysis of covariance model. A linear model with fixed effects, having

continuous (interval level) as well as categorical predictor variables.

Ordinary least squares (OLS). The usual way to estimate parameters

(fixed effects) in a linear fixed effects model.

Mixed model. A statistical model with fixed as well as random effects.

Intercept. In a linear function of x, the function value at x = 0.

Random intercept model. The model to which this chapter is devoted,

which is a mixed model with (for two-level structures) two kinds of
random residuals: the residual at level one and the random intercept at
 level two. A general formula is given by (4.8). The three-level random
intercept model has three kinds of random residuals, etc.

Hierarchical linear model. A more general model than the random

intercept model, and treated in Chapter 5.

Random intercept. The intercept in a random intercept model (also used

more generally in hierarchical linear models), symbolized in formula
(4.8) by γ00 + γ01z1j + . . . + γ0q Zqj + U0j. Sometimes also the values z
are assumed equal to zero for this definition, leading to γ00 + U0j. The
random intercept is associated with the intercept variance, the variance
of U0j.

Empty model. The random intercept model without any explanatory

variables, defined by (4.6).

Intraclass correlation coefficient. The correlation between two randomly

drawn individuals in the same randomly drawn group; an equivalent
definition is the fraction of total variability that is due to the group
level. This is defined in the empty model by (4.7). See also Section
3.3.

Within- and between-group regression coefficients. To avoid ecological

fallacies, it is important to distinguish within-group from between-
group regressions. The within-group regression coefficient of Y on X is
the expected difference in Y between two cases in the same group, for
a one-unit difference in X. The between-group regression coefficient of
Y on X is the expected difference in the group means on Y between two
groups differing by one unit in their mean values on X. In Section 4.6 it
was explained how these two coefficients can be obtained from the
random intercept model in which the individual variable X as well as
the group means of X are included as two separate predictor variables.

Cross-level interaction effect. An interaction between two variables

defined at different levels.
Parameter estimation. A brief intuitive overview of parameter estimation
in the random intercept model was given in Section 4.7.

Maximum likelihood estimation (ML). The best-known general statistical

method for parameter estimation, due to R.A. Fisher (1890–1962).

Restricted or residual maximum likelihood estimation (REML). A

method of parameter estimation in linear and generalized linear models
that often works better than ML, especially for smaller samples, but
gives roughly the same results for large sample sizes.

Empirical Bayes estimation. Estimation of the higher-level residuals, such

as random intercepts for the groups, by combining group-level and
population-level information. This is done using Bayesian techniques,
by so-called posterior means. The use of population-level information
implies a ‘shrinkage’ of the group-level coefficients toward the
population mean, the shrinkage being stronger accordingly as the
groups are smaller or, more generally, convey less information.

Posterior confidence intervals. Confidence intervals for higher-level

residuals, likewise combining group-level information with
population-level information.

Comparative standard error, or posterior standard deviation. The

standard error of the empirical Bayes estimate for a group-level
residual minus the true (unobserved) residual. This is used for
comparing groups.

Diagnostic standard error. The standard error of the empirical Bayes

estimate for a group-level residual. This is used for standardizing
empirical Bayes estimates for diagnostic model checking.

Higher-level random intercept models. The random intercept model can

also be defined for three- and higher-level data structures, and for the
three-level case this was discussed in Section 4.9.
1
Maas and Hox (2005) conclude that for N = 30 groups even of small sizes nj = 5, the coverage rates
of confidence intervals of fixed parameters of the hierarchical linear model are still satisfactory. The
fact that for N = 30 they obtain unacceptable coverage rates for the asymptotic confidence intervals
for variance parameters indicates that, for such a small number of groups, these asymptotic
confidence intervals should not be used for variance parameters. This is, however, not an argument
against multilevel modeling in general for N ≥ 30 groups.
2
In the first edition of this book we used a subset of this data set, from which students and schools
with missing variables on other variables were also dropped. This was not necessary and may have
induced bias; therefore we now use the larger data set. In Chapter 9 the effect of dropping the 258
students with some missing variables is investigated, and turns out to be unimportant. To avoid
complications at an early stage, we use here a data set which has no variables with missing data.
3
Mind the word valid at the beginning of this paragraph. You may have fitted model (4.5) and
obtained an estimated value for ρI (Y) that is quite low or even 0. This does not exclude the
possibility that there are group differences of the kind treated in Chapter 5: random slopes. If there is
an important random slope variance, even without an intercept variance, ordinary regression analysis
may yield incorrect results and less insight than the hierarchical linear model.
4
The subscripts on the regression coefficients γ may seem somewhat baroque. The reason is to
obtain consistency with the notation in the next chapter; see (5.12).
5
If cases with missing data are deleted in order to carry out the multilevel analysis, then one should
calculate the group mean as the average over all level-one units for which this particular variable is
available, and before deleting cases because they have missing values on other variables. But instead
of deleting cases it is better to use the methods of Chapter 9.
6
When models are compared with different fixed parts, deviance tests should be based on ML
estimation. Deviance tests with REML estimates may be used for comparing models with different
random parts and the same fixed part. Different random parts will be treated in the next chapter.
7
The word ‘estimate’ is in quotation marks because the proper statistical term for finding likely
values of the U0j, being random variables, is prediction. The term ‘estimation’ is reserved for finding
likely values for statistical parameters. Since prediction is associated in everyday speech, however,
with determining something about the future, we prefer to speak here about ‘estimation’ between
parentheses.
8
Unbiasedness means that the average of many - hypothetical - independent replications of this
estimate for this particular group j would be very close to the true value .
 5
The Hierarchical Linear Model

In the previous chapter the simpler case of the hierarchical linear model was
treated, in which only intercepts are assumed to be random. In the more
general case, slopes may also be random. In a study of students within
schools, for example, the effect of the pupil’s intelligence or socio-
economic status on scholastic performance could differ between schools.
This chapter presents the general hierarchical linear model, which allows
intercepts as well as slopes to vary randomly. The chapter follows the
approach of the previous one: most attention is paid to the case of a two-
level nesting structure, and the level-one units are called – for convenience
only – ‘individuals’, while the level-two units are called ‘groups’. The
notation is also the same.

OVERVIEW OF THE CHAPTER

First the concept of random slopes is defined, and a discussion is given of
their interpretation and the interpretation of the associated statistical
parameters. This involves the notion that random slopes, as well as random
intercepts, are group-level latent variables that can be explained by
observed group-level variables: ‘intercepts and slopes as outcomes’. Next
we discuss how to specify random slope models, in particular, the issue of
centering variables by subtracting the overall mean or the group mean. The
chapter concludes with brief treatments of parameter estimation and random
slopes in three-level models.
 5.1 Random slopes
In the random intercept model of Chapter 4, the groups differ with respect
to the average value of the dependent variable: the only random group
effect is the random intercept. But the relation between explanatory and
dependent variables can differ between groups in more ways. For example,
in the field of education (nesting structure: students within classrooms), it is
possible that the effect of socio-economic status of students on their
scholastic achievement is stronger in some classrooms than in others. As an
example in developmental psychology (repeated measurements within
individual subjects), it is possible that some subjects progress faster than
others. In the analysis of covariance, this phenomenon is known as
heterogeneity of regressions across groups, or as group-by-covariate
interaction. In the hierarchical linear model, it is modeled by random
slopes.
Let us go back to a model with group-specific regressions of Y on one
level-one variable X only, like model (4.3) but without the effect of Z,

The intercepts β0j as well as the regression coefficients, or slopes, β1j are
group-dependent. These group-dependent coefficients can be split into an
average coefficient and the group-dependent deviation:

Substitution leads to the model

It is assumed here that the level-two residuals U0j and U1j as well as the
level-one residuals Rij have mean 0, given the values of the explanatory
variable X. Thus, γ10 is the average regression coefficient just as γ00 is the
average intercept. The first part of (5.3), γ00+γ10xij, is called the fixed part of
the model. The second part, U0j+U1jxij+Rij, is called the random part.
The term U1j xij can be regarded as a random interaction between group
and X. This model implies that the groups are characterized by two random
effects: their intercept and their slope. These are called latent variables,
meaning that they are not directly observed but play a role ‘behind the
scenes’ in producing the observed variables. We say that X has a random
slope, or a random effect, or a random coefficient. These two group effects
will usually not be independent, but correlated. It is assumed that, for
different groups, the pairs of random effects (U0j, U1j) are independent and
identically distributed, that they are independent of the level-one residuals
Rij, and that all Rij are independent and identically distributed. The variance
of the level-one residuals Rij is again denoted σ2; the variances and
covariance of the level-two residuals (U0j, U1j) are denoted as follows:

Just as in the preceding chapter, one can say that the unexplained group
effects are assumed to be exchangeable.

5.1.1 Heteroscedasticity
Model (5.3) implies not only that individuals within the same group have
correlated Y-values (recall the residual intraclass correlation coefficient of
Chapter 4), but also that this correlation as well as the variance of Y are
dependent on the value of X. For example, suppose that, in a study of the
effect of socio-economic status (SES) on scholastic performance (Y), we
have schools which do not differ in their effect on high-SES children, but
do differ in the effect of SES on Y (e.g., because of teacher expectancy
effects). Then for children from a high-SES background it does not matter
which school they go to, but for children from a low-SES background it
does. The school then adds a component of variance for the low-SES
children, but not for the high-SES children: as a consequence, the variance
of Y (for a random child at a random school) will be larger for the former
than for the latter children. Further, the intraclass correlation will be nil for
high-SES children, whereas for low-SES children it will be positive.
This example shows that model (5.3) implies that the variance of Y,
given the value x on X, depends on x. This is called heteroscedasticity in the
statistical literature. An expression for the variance of (5.3) is obtained as
the sum of the variances of the random variables involved plus a term
depending on the covariance between U0j and U1j (the other random
variables are uncorrelated). Here we also use the independence between the
level-one residual Rij and the level-two residuals (U0j, U1j). From (5.3) and
(5.4), we obtain the result

Similarly, for two different individuals (i′ and with i ≠ i′) in the same group,

Formula (5.5) implies that the residual variance of Y is minimal for xij = −
τ01/τ11. (This is deduced by differentiation with respect to xij.) When this
value is within the range of possible X-values, the residual variance first
decreases and then increases again; if this value is smaller than all X-values,
then the residual variance is an increasing function of x; if it is larger than
all X-values, then the residual variance is decreasing.

5.1.2 Do not force τ01 to be 0!

All of the preceding discussion implies that the group effects depend on x:
according to (5.3), this effect is given by U0j + U1j x. This is illustrated by
Figure 5.1, a hypothetical graph of the regression of school achievement (Y)
on intelligence (X).
 Figure 5.1: Different vertical axes.

It is clear that there are slope differences between the three schools.
Looking at the Y(1)-axis, there are almost no intercept differences between
the schools. But if we add a value 10 to each intelligence score x, then the
Y-axis is shifted to the left by 10 units: the Y(2)-axis. Now school 3 is the
best, school 1 the worst: there are strong intercept differences. If we had
subtracted 10 from the x-scores, we would have obtained the Y(3)-axis,
again with intercept differences but now in reverse order. This implies that
the intercept variance τ00, as well as the intercept-by-slope covariance τ01,
depend on the origin (0-value) for the X-variable. From this we can learn
two things:

1. Since the origin of most variables in the social sciences is arbitrary, in

random slope models the intercept-by-slope covariance should be a
free parameter estimated from the data, and not a priori constrained to
the value 0 (i.e., left out of the model).

2. In random slope models we should be careful with the interpretation of

the intercept variance and the intercept-by-slope covariance, since the
intercept refers to an individual with x = 0. For the interpretation of
these parameters it is helpful to define the scale for X so that x = 0 has
an interpretable meaning, preferably as a reference situation. For
example, in repeated measurements when X refers to time, or
measurement number, it can be convenient to let x = 0 correspond to
the start, or the end, of the measurements. In nesting structures of
individuals within groups, it is often convenient to let x = 0 correspond
to the overall mean of the population or the sample – for example, if X
is IQ at the conventional scale with mean 100, it is advisable to
subtract 100 to obtain a population mean of 0.

5.1.3 Interpretation of random slope variances

For the interpretation of the variance of the random slopes, τ12, it is also
illuminating to take the average slope, γ10, into consideration. Model (5.3)
implies that the regression coefficient, or slope, for group j is γ10 + U1j. This
is a normally distributed random variable with mean γ10 and standard
deviation Since about 95% of the probability of a normal
distribution is within two standard deviations from the mean, it follows that
approximately 95% of the groups have slopes between γ10 – 2τ1 and γ10 +
2τ1. Conversely, about 2.5% of the groups have a slope less than γ10 – 2τ1
and 2.5% have a slope steeper than γ10 + 2τ1.

Example 5.1 A random slope for IQ.

We continue our study of the effect of IQ on a language test score, begun in
Chapter 4. Recall that IQ is here on a scale with mean 0 and its standard
deviation in this data set is 2.04. A random slope of IQ is added to the
model, that is, the effect of IQ is allowed to differ between classes. The
model is an extension of model (5.3): a fixed effect for the class average on
IQ is added. The model reads

The results can be read from Table 5.1. Note that the ‘Level-two random
part’ heading refers to the random intercept and random slope which are
random effects associated with the level-two units (the class), but that the
variable that has the random slope, IQ, is itself a level-one variable.

Table 5.1: Estimates for random slope model.

 Figure 5.2 presents a sample of 15 regression lines, randomly chosen
according to the model of Table 5.1. (The values of the group mean IQ were
chosen randomly from a normal distribution with mean −0.072 and standard
deviation 0.955,which are the mean and standard deviation of the group
mean of IQ in this data set.)This figure thus demonstrates the population of
regression lines that characterizes, according to this model, the population
of schools.

Should the value of 0.195 for the random slope variance be considered to be
high? The slope standard deviation is and the average slope
is γ10 = 2.48. The values of average slope

Figure 5.2: Fifteen random regression lines according to the model of Table
5.1 (with randomly chosen intercepts and slopes).

± two standard deviations range from 1.60 to 3.36. This implies that the
effect of IQ is clearly positive in all classes, but high effects of IQ are more
than twice as large as low effects. This may indeed be considered an
important difference. (As indicated above, ‘high’ and ‘low’ are respectively
understood here as those values occurring in classes with the top 2.5% and
the bottom 2.5% of the class-dependent effects.)

Interpretation of intercept–slope covariance

The correlation between random slope and random intercept is
Recall that all variables are centered (have
zero mean), so that the intercept corresponds to the language test score for a
pupil with average intelligence in a class with average mean intelligence.
The negative correlation between slope and intercept means that classes
with a higher performance for a pupil of average intelligence have a lower
within-class effect of intelligence. Thus, the higher average performance
tends to be achieved more by higher language scores of the less intelligent,
than by higher scores of the more intelligent students.
In a random slope model, the within-group coherence cannot be simply
expressed by the intraclass correlation coefficient or its residual version.
The reason is that, in terms of the present example, the correlation between
students in the same class depends on their intelligence. Thus, the extent to
which a given classroom deserves to be called ‘good’ varies across
students.
To investigate how the contribution of classrooms to students’
performance depends on IQ, consider the equation implied by the parameter
estimates:

Recall from Example 4.2 that the standard deviation of the IQ score is about
2, and the mean is 0. Hence students with an intelligence among the bottom
few percent or the top few percent have IQ scores of about ±4. Substituting
these values in the contribution of the random effects gives U0j ± 4U1j. It
follows from equations (5.5) and (5.6) that for students with IQ = ±4, we
have

and therefore

Hence, the language test scores of the most intelligent and the least
intelligent students in the same class are positively correlated over the
population of classes: classes that have relatively good results for the less
able tend also to have relatively good results for the more able students.
This positive correlation corresponds to the result that the value of IQ
for which the variance given by (5.5) is minimal, is outside the range from
−4 to +4. For the estimates in Table 5.1, this variance is (with some
rounding)

Taking the derivative of this function of x equating it to 0 yields that the

variance is minimal for x = 1.67/0.39 = 4.3, just outside the IQ range from
−4 to +4. This again implies that classes tend mostly to perform either
higher, or lower, over the entire range of IQ. This is illustrated also by
Figure 5.2 (which, however, also contains some regression lines that cross
each other within the range of IQ, illustrating that the random nature of
these regression lines will lead to exceptions to this pattern).

5.2 Explanation of random intercepts and slopes

The aim of regression analysis is to explain variability in the outcome (i.e.,
dependent) variable. Explanation is understood here in a quite limited way
– as being able to predict the value of the dependent variable from
knowledge of the values of the explanatory variables. The unexplained
variability in single-level multiple regression analysis is just the variance of
the residual term. Variability in multilevel data, however, has a more
complicated structure. This is related to the fact, mentioned in the preceding
chapter, that several populations are involved in multilevel modeling: one
population for each level. Explaining variability in a multilevel structure
can be achieved by explaining variability between individuals but also by
explaining variability between groups; if there are random slopes as well as
random intercepts, at the group level one could try to explain the variability
of slopes as well as intercepts.
In the model defined by (5.1)–(5.3), some variability in Y is explained
by the regression on X, that is, by the term γ10 xij; the random coefficients
U0j, U1j, and Rij each express different parts of the unexplained variability.
In order to try to explain more of the unexplained variability, all three of
these can be the point of attack. In the first place, one can try to find
explanations in the population of individuals (at level one). The part of
residual variance that is expressed by σ2 = var(Rij) can be diminished by
including other level-one variables. Since group compositions with respect
to level-one variables can differ from group to group, inclusion of such
variables may also diminish residual variance at the group level. A second
possibility is to try to find explanations in the population of groups (at level
two). Note that here the variables that we try to explain, U0j and U1j, are not
directly observed, but latent variables. If we wish to reduce the unexplained
variability associated with U0j and U1j, we can also say that we wish to
expand equations (5.2) by predicting the group-dependent regression
coefficients β0j and β1j from level-two variables Z. Supposing for the
moment that we have one such variable, this leads to regression formulas
for β0j and β1j on the variable Z given by

and

In words, the βs are treated as dependent variables in regression models for

the population of groups; however, these are ‘latent regressions’, because
the βs cannot be observed without error. Equation (5.7) is called an
intercepts as outcomes model, and (5.8) a slopes as outcomes model.1

5.2.1 Cross-level interaction effects

This chapter began with the basic model (5.1), reading

Substituting (5.7) and (5.8) into this equation leads to the model
which we have rearranged so that the fixed part comes first and the random
part next. Comparing this with model (5.3) shows that this explanation of
the random intercept and slope leads to a different fixed part of the model,
but does not change the formula for the random part, which remains U0j +
U1jxij + Rij. However, it is to be expected that the residual random intercept
and slope variances, τ02 and τ12, will be less than their counterparts in model
(5.3) because part of the variability of intercept and slopes now is explained
by Z. In Chapter 7 we will see, however, that this is not necessarily so for
the estimated values of these parameters.
Equation (5.9) shows that explaining the intercept β0j by a level-two
variable Z leads to a main effect of Z, while explaining the coefficient β1j of
X by the level-two variable Z leads to a product interaction effect of X and
Z. Such an interaction between a level-one and a level-two variable is called
a cross-level interaction.
For the definition of interaction variables such as the product zj xij in
(5.9), it is advisable to use component variables Z and X for which the
values Z = 0 and X = 0, respectively, have some interpretable meaning. For
example, the variables Z and X could be centered around their means, so
that Z = 0 means that Z has its average value, and analogously for X.
Another possibility is that the zero values correspond to some kind of
reference value. The reason is that, in the presence of the interaction term
γ11 zj xij, the main effect coefficient γ10 of X is to be interpreted as the effect
of X for cases with Z = 0, while the main effect coefficient γ01 of Z is to be
interpreted as the effect of Z for cases with X = 0.

Example 5.2 Cross-level interaction between IQ and group-mean IQ.

A comparison of Examples 4.2 and 4.3 shows that part of the intercept
variability in the language scores is explained by the class mean of IQ. This
confirms that this aggregated variable is a meaningful classroom
characteristic. We attempt to use it also to explain part of the random slope
variability, which means that we introduce the interaction between
individual IQ and the classroom mean of IQ. When this variable is added to
the model of Example 5.1, the parameter estimates presented in Table 5.2
are obtained.
The cross-level interaction is significant (tested by a t-ratio as explained
in Section 6.1), with t = −0.187/0.064 = −2.92, p < 0.01, and the random
slope variance has decreased from 0.195 to 0.163. (It follows from Chapter
7, however, that the expectation that a random slope variance should
decrease when a cross-level interaction effect is included, is not always
confirmed.) Thus, the classroom average of IQ is indeed successful in
explaining part of the between-group variability in the effect of IQ.
How can we assess the magnitude of this cross-level effect? The class-
dependent regression coefficient of IQ (cf. (5.8)) is

Table 5.2: Estimates for model with random slope and cross-level
interaction.

estimated as
The classroom mean of the centered IQ variable has a rather skewed
distribution, ranging from –4.8 to +2.5. For ranging between –4.8 and
+2.5, the fixed part of this expression ranges (in reverse order) between
about 2.00 and 3.36. This implies sizeable differences in the effect of
intelligence on language score between classes with students having low
average IQ and those with students of high average IQ, classrooms with
lower average IQ being associated with higher effects of intelligence on
language scores. (We must note, however, that further analysis of these data
will show that the final conclusions will be different.)

Cross-level interactions can be considered on the basis of two different

kinds of argument. The above presentation is in line with an inductive
argument: a researcher who finds a significant random slope variance may
be led to think of level-two variables that could explain the random slope.
An alternative approach is to base the cross-level interaction on substantive
(theoretical) arguments formulated before looking at the data. The
researcher is then led to estimate and test the cross-level interaction effect
irrespective of whether a random slope variance was found. If a cross-level
interaction effect exists, the power of the statistical test of this fixed effect is
considerably higher than the power of the test for the corresponding random
slope (assuming that the same model serves as the null hypothesis).
Therefore there is nothing wrong with looking for a specific cross-level
interaction even if no significant random slope was found. This is further
discussed in the last part of Section 6.4.1.

More variables
The preceding models can be extended by including more variables that
have random effects, and more variables explaining these random effects.
Suppose that there are p level-one explanatory variables, X1, . . . ,Xp, and q
level-two explanatory variables, Z1, . . . ,Zq. Then, if the researcher is not
afraid of a model with too many parameters, he can consider the model
where all X-variables have varying slopes, and where the random intercept
as well as all these slopes are explained by all Z-variables. At the within-
group level (i.e, for the individuals) the model is then a regression model
with p variables,
The explanation of the regression coefficients β0j, . . . ,β1j, . . . ,βpj is based
on the between-group model, which is a q-variable regression model for the
group-dependent coefficient βhj,

Substituting (5.11) into (5.10) and rearranging terms then yields the model

This shows that we obtain main effects of each X and Z variable as well as
all cross-level product interactions. Further, we see the reason why, in
formula (4.8), the fixed coefficients were called γh0 for the level-one
variable Xh, and γ0k for the level-two variable Zk.

The groups are now characterized by p + 1 random coefficients U0j, U1j,

. . . ,Upj. These random coefficients are independent between groups, but
may be correlated within groups. It is assumed that the vector (U0j, . . . ,
Upj) is independent of the level-one residuals Rij and that all residuals
have population mean 0, given the values of all explanatory variables. It
is also assumed that the level-one residual Rij has a normal distribution
with constant variance σ2 and that (U0j, . . . , Upj) has a multivariate
normal distribution with a constant covariance matrix. Analogously to
(5.4), the variances and covariances of the level-two random effects are
denoted by

Interactions and ecological fallacies

In Section 3.1 we saw that aggregations of level-one variables to level two
(or higher levels) can be substantively meaningful and can be important to
avoid ecological fallacies. This was discussed in Section 4.6, showing how
the group-level average of an individual-level variable can be included in
the model to represent the difference between within-group and between-
group regressions. Similar considerations apply to interaction effects. For
example, when the interaction effect of two level-one variables X1 and X2 is
being investigated, we should be aware of the possibility that this
interaction has effects at level two rather than at level one, for one or both
of the variables. Thus, the representation of the interaction by a product
variable x1ijx2ij can be extended to four possibilities,

each with their own theoretical interpretation. An example was given by

van Yperen and Snijders (2000) in a study of the interaction between job
demands and job control on the psychological health of employees in
departments of firms.

Example 5.3 A model with many fixed effects.

Next to intelligence, another important explanation of school performance
is the home environment. This is reflected by the socio-economic status of
the pupil’s family, represented here by a variable called SES (a numerical
variable with mean 0.0 and standard deviation 10.9 in this data set). We
include SES at the individual level as well as the class mean; we are also
interested in the interaction effect between IQ and SES. To try and avoid
ecological fallacies, in line with the discussion above, the four
combinations of interactions between IQ and SES are included as in (5.14).
The covariance of the random slopes of IQ and SES is excluded,
because including this covariance in the model led to failure of the
estimating algorithm to converge. We may assume that this covariance is
not different from 0 to any important extent.
Estimating model (5.12) (with p = q = 2), to which are added the
interaction between the two individual variables as well as the interaction
between the two classroom means, leads to the results presented in Table
5.3. A discussion of this table is deferred to the next example.
 Table 5.3: Estimates for model with random slopes and many effects.

Unless p and q are quite small, model (5.12) entails a number of

statistical parameters that is usually too large for comfort. Therefore, two
simplifications are often used.

1. Not all X-variables are considered to have random slopes. Note that the
explanation of the variable slopes by the Z-variables may have led to
some random residual coefficients that are not significantly different
from 0 (testing the τ parameters is discussed in Chapter 6) or even
estimated to be 0.

2. Given that the coefficients βhj of a certain variable Xh, are variable
across groups, it is not necessary to use all variables Zk in explaining
their variability. The number of cross-level interactions can be
 restricted by explaining each βhj by only a well-chosen subset of the
Zk.

Which variables to give random slopes, and which cross-level and other
interaction variables to use, will depend on subject-matter as well as
empirical considerations. The statistical aspects of testing and model fitting
are treated in later chapters. For further help in the interpretation of
interactions, in particular, methods to determine the values of zj where the
regression on xij is different from zero (or vice versa), see Curran and Bauer
(2005) and Preacher et al. (2006).

Example 5.4 A parsimonious model with several variables.

In Table 5.3 the random slope variance of SES is estimated by 0 (this
happens occasionally; cf. Section 5.4). Therefore, this random slope is
excluded from the model.
We shall see in Chapter 6 that the significance of fixed effects can be
tested by applying a t-test to the ratio of estimate to standard error. This
consideration led us to exclude all cross-level interactions. The resulting
estimates are displayed in Table 5.4. The estimates of the remaining effects
do not change much compared to Table 5.3.
All explanatory variables have mean 0. Therefore, the intercept γ00 is
the mean of students with average IQ and SES in a class with average
composition with respect to these variables. The regression coefficient of
IQ γ10 is the average effect of IQ for students of average SES, and similarly
for the regression coefficient γ20 for SES.
Compared to Example 5.2, it turns out that there is no significant cross-
level interaction effect between IQ and its classroom mean, but there are
two interaction effects between IQ and SES: the interaction between the
two individual variables and the interaction between the two classroom
means.
The model found can be expressed as a model with variable intercepts
and slopes by the formula
where X1 is IQ, X2 is SES, and X3 is their product interaction. The intercept
is

where Z1 is average IQ, Z2 is average SES, and Z3 is their product

interaction. The coefficient of X1 is

while the coefficient of X2 is not variable,

Table 5.4: Estimates for a more parsimonious model.

5.2.2 A general formulation of fixed and random parts

Formally, and in many computer programs, these simplifications lead to a
representation of the hierarchical linear model that is slightly different from
(5.12). (For example, the HLM program uses the formulations (5.10)–
(5.12), whereas R and MLwiN use formulation (5.15).) Whether a level-one
variable was obtained as a cross-level interaction or not is immaterial to the
computer program. Even the difference between level-one variables and
level-two variables, although possibly relevant for the way the data are
stored, is of no importance for the parameter estimation. Therefore, all
variables – level-one and level-two variables, including product interactions
– can be represented mathematically simply as xhij. When there are r
explanatory variables, ordered so that the first p have fixed and random
coefficients, while the last r – p have only fixed coefficients,2 the
hierarchical linear model can be represented as

The two terms,

are the fixed and the random parts of the model, respectively.
In cases where the explanation of the random effects works extremely
well, one may end up with models with no random effects at level two. In
other words, the random intercept U0j and all random slopes Uhj in (5.15)
have zero variance, and may just as well be omitted from the formula. In
this case, the resulting model may be analyzed just as well by OLS
regression analysis, because the residuals are independent and have
constant variance. Of course, this is known only after the multilevel
analysis has been carried out. In such a case, the within-group dependence
between measurements has been fully explained by the available
explanatory variables (and their interactions). This underlines the fact that
whether the hierarchical linear model is a more adequate model for analysis
than OLS regression depends not on the mutual dependence of the
measurements, but on the mutual dependence of the residuals.
 5.3 Specification of random slope models
Given that random slope models are available, the researcher has many
options to model his data. Each predictor may be assigned a random slope,
and each random slope may covary with any other random slope.
Parsimonious models, however, should be preferred, if only for the simple
reason that a strong scientific theory is general rather than specific. A good
reason for choosing between a fixed and a random slope for a given
predictor variable should preferably be found in the theory that is being
investigated. If the theory (whether this is a general scientific theory or a
practical policy theory) does not give any clue with respect to a random
slope for a certain predictor variable, then one may be tempted to refrain
from using random slopes. However, this implies a risk of invalid statistical
tests, because if some variable does have a random slope, then omitting this
feature from the model could affect the estimated standard errors of the
other variables. The specification of the hierarchical linear model, including
the random part, is discussed more fully in Section 6.4 and Chapter 10.
In data exploration, one can try various specifications. Often it appears
that the chance of detecting slope variation is high for variables with strong
fixed effects. This, however, is an empirical rather than a theoretical
assertion. Actually, it may well be that when a fixed effect is – almost –
zero, there does exist slope variation. Consider, for instance, the case where
male teachers treat boys advantageously over girls, while female teachers
do the reverse. If half of the sample consists of male and the other half of
female teachers, then, all other things being equal, the main gender effect
on achievement will be absent, since in half of the classes the gender effect
will be positive and in the other half negative. The fixed effect of students’
gender is then zero but varies across classes (depending on the teachers’
gender). In this example, of course, the random effect would disappear if
one specified the cross-level interaction effect of teachers’ gender with
students’ gender.

5.3.1 Centering variables with random slopes?

Recall from Figure 5.1 that the intercept variance and the meaning of the
intercept in random slope models depend on the location of the X-variable.
Also the covariance between the intercepts and the slopes is dependent on
this location. In the examples presented so far we have used an IQ score for
which the grand mean was zero (the original score was transformed by
subtracting the grand mean IQ). This facilitated interpretation since the
intercept could be interpreted as the expected score for a student with
average IQ. Making the IQ slope random did not have consequences for
these meanings.
In Section 4.6 a model was introduced by which we could distinguish
within- from between-group regression. Two models were discussed:

and

It was shown that , so that the two models are

equivalent.
Are the models also equivalent when the effect of Xij or is
random across groups? This was discussed by Kreft et al. (1995). Let us
first consider the extension of (4.9). Define the level-one and level-two
models

substituting the level-two model into the level-one model leads to

Next we consider the extension of (4.10):

substitution and rearrangement of terms now yields

This shows that the two models differ in the term U1j ·j which is included
in the group-mean-centered random slope model but not in the other model.
Therefore in general there is no one-to-one relation between the γ and the
parameters, so the models are not statistically equivalent except for the
extraordinary case where variable X has no between-group variability.
This implies that in constant slope models one can either use Xij and .j
or (Xij – .j) and .j as predictors, since this results in statistically equivalent
models, but in random slope models one should carefully choose one or the
other specification depending on substantive considerations and/or model
fit.
On which consideration should this choice be based? Generally one
should be reluctant to use group-mean-centered random slopes models
unless there is a clear theory (or an empirical clue) that not the absolute
score Xij but rather the relative score (Xij – .j) is related to Yij. Now (Xij –
.j) indicates the relative position of an individual in his or her group, and
examples of instances where one may be particularly interested in this
variable are:

research on normative or comparative social reference processes (e.g.,

*
Guldemond, 1994);
* research on relative deprivation;
* research on teachers’ rating of student performance.

A more detailed discussion is given in Enders and Tofighi (2007). However,

that paper pays no attention to the fact that interactions between level-one
variables can also play a role for their group means (see our discussion on
p. 83). This implies further decisions about centering, which likewise
should be based on a combination of substantive theoretical and empirical
considerations.
 5.4 Estimation
The discussion in Section 4.7 can also be applied, with the necessary
extensions, to the estimation of parameters in the more complicated model
(5.15). A number of iterative estimation algorithms have been proposed by,
for example, Laird and Ware (1982), Goldstein (1986), and Longford
(1987), and are now implemented in multilevel software.
The following may give some intuitive understanding of estimation
methods. If the parameters of the random part, that is, the parameters in
(5.13) together with σ2, were known, then the regression coefficients could
be estimated straightforwardly with the so-called generalized least squares
(GLS) method. Conversely, if all regression coefficients γhk were known,
the ‘total residuals’ (which seems an apt name for the second line of
equation (5.12)) could be computed, and their covariance matrix could be
used to estimate the parameters of the random part. These two partial
estimation processes can be alternated: use provisional values for the
random part parameters to estimate regression coefficients, use the latter
estimates to estimate the random part parameters again (and now better),
then go on to estimate the regression coefficients again, and so on ad
libitum – or, rather, until convergence of this iterative process. This loose
description is close to the iterated generalized least squares (IGLS) method
that is one of the algorithms for calculating the ML estimates; see Goldstein
(2011).
There exist other methods for calculating the same estimates – Fisher
scoring (Longford, 1993) and the EM algorithm (see Raudenbush and Bryk,
2002) – each with its own advantages. An extensive mathematical
explanation is given in de Leeuw and Meijer (2008b).
Parameters can again be estimated with the ML or REML method; the
REML method is preferable in the sense that in the case of small sample
sizes (N) at the higher level it produces less biased estimates for the random
part parameters and more reliable standard errors (Section 6.1). The ML
method, however, is more convenient if one wishes to use deviance tests
(see the next chapter). The IGLS algorithm produces ML estimates,
whereas the so-called ‘restricted IGLS’ (RIGLS) algorithm yields the
REML estimates. For the random slopes model it is also possible that
estimates for the variance parameters τh2 are exactly 0. The explanation is
analogous to the explanation given for the intercept variances in Section
4.7.
The random group effects Uhj can again be ‘estimated’ by the empirical
Bayes method, and the resulting ‘estimates’ are called posterior slopes
(sometimes posterior means). This is analogous to the treatment of posterior
means in Section 4.8.
Usually the estimation algorithms do not allow an unlimited number of
random slopes to be included. Depending on the data set and the model
specification, it is not uncommon for the algorithm to refuse to converge for
more than two or three variables with random slopes. Sometimes the
convergence can be improved by linearly transforming the variables with
random slopes so that they have (approximately) zero means, or by
transforming them to have (approximately) zero correlations.
For some data sets, the estimation method can produce estimated
variance and covariance parameters that correspond to impossible
covariance matrices for the random effects at level two. For example, τ01
sometimes is estimated to be larger than τ0 × τ1, which would imply an
intercept–slope correlation larger than 1. This is not an error of the
estimation procedure, and it can be understood as follows. The estimation
procedure is directed at the mean vector and covariance matrix of the vector
of all observations. Some combinations of parameter values correspond to
permissible structures of the latter covariance matrix that, nevertheless,
cannot be formulated as a random effects model such as (5.3). Even if the
estimated values for the τhk parameters do not combine into a positive
definite matrix τ, the σ2 parameter will still make the covariance matrix of
the original observations (cf. equations (5.5) and (5.6)) positive definite.
Therefore, such a strange result is in contradiction to the random effects
formulation (5.3), but not to the more general formulation of a patterned
covariance matrix for the observations.
Most computer programs give the standard errors of the variances of the
random intercept and slopes; some give the standard errors of the estimated
standard deviations instead. This is related to the fact that the estimated
variances may have very skewed distributions, especially if these
distributions are close to the point 0. The estimated standard deviation (i.e.,
the square root of the estimated variance) will often have a more nearly
normal distribution than the estimated variance itself. Approximate
relations between the standard errors of estimated variances and estimated
standard deviations are given in (6.2) in the next chapter.

5.5 Three or more levels

When the data have a three-level hierarchy, as discussed also in Section 4.9,
slopes of level-one variables can be made random at level two and also at
level three. In this case there will be at least two level-two and two level-
three equations: one for the random intercept and one for the random slope.
So, in the case of one explanatory variable, the model might be formulated
as follows:

(level-one model)
(level-two model for intercept)
(level-two model for slope)
(level-three model for intercept)
(level-three model for slope)

In the specification of such a model, for each level-one variable with

random slope it has to be decided whether its slope must be random at level
two, random at level three, or both. Generally one should have either strong
a priori knowledge or a good theory to formulate models as complex as this
one or even more complex models (i.e., with more random slopes). Further,
for each level-two variable it must be decided whether its slope is random at
level three.

Example 5.5 A three-level model with a random slope.

We continue with the example of Section 4.9, where we illustrated the
three-level model using a data set about a mathematics test administered to
students in classes in schools. Now we include the available covariates
(which are all centered around their grand means), and the regression
coefficient for the mathematics pretest is allowed to be random at levels two
and three. The results are shown in Table 5.5.

Table 5.5: Estimates for three-level model with random slopes.

 The interpretation of the fixed part is straightforward as in conventional
single-level regression models. The random part is more complicated. Since
all predictor variables were grand mean centered, the intercept variances
φ02 (level three) and τ02 (level two) have a clear meaning: they represent the
amount of variation in mathematics achievement across schools and across
classes within schools, respectively, for the average student whilst
controlling for differences in IQ, mathematics ability, achievement
motivation, educational level of the father, and gender. Comparing this table
with Table 4.5 shows that much of the initial level-three and (especially)
level-two variation has now been accounted for. Once there is control for
initial differences, schools and classes within schools differ considerably
less in the average mathematics achievement of their students at the end of
grade 2.
Now we turn to the slope variance. The fixed slope coefficient for the
mathematics pretest is estimated to be 0.146. The variance at level three for
this slope is 0.0024, and at level two 0.0019. So the variability between
schools of the effect of the pretest is somewhat larger than the variability of
this effect between classes. At one end of the distribution there are a few
percent of the schools that have an effect of the pretest that is only
whereas in the most selective schools this effect
is
There is also variation in this coefficient across classes within schools.
In the population of classes, when we draw a random class in a random
school, the total slope variance is 0.0024+0.0019. Therefore, the classes
with the 2.5% lowest and highest effect of pretest have regression
coefficients . Hence the gap between initial low and
initial high achievers (4 standard deviations apart; this standard deviation
for the pretest is 8.21) within the same class can become as large as points,
or on the other hand it can
remain as small as point in
the least selective classes. Given the standard deviation of 3.4 for the
dependent variable, a difference of 1 is very small, whereas 9 is quite a
large difference.

5.6 Glommary
Hierarchical linear model. The main model for multilevel analysis, treated
in this chapter.

Random slope. The random residual at level two in the hierarchical linear
model indicating group-level deviations in the effect of an explanatory
variable X on the dependent variable, symbolized in formula (5.3) by
U1j and in the more general formula (5.15) by Uhj. It is associated with
the slope variance, the variance of Uhj.
To interpret the random slope variance, it is often helpful to consider
the distribution of the group-dependent slopes. This distribution
follows from

in a model such as (5.2) without cross-level interactions. This follows

a normal distribution with mean γh0 and variance τh2, the slope
variance. If the random slope βhj, a latent variable, is explained by
 group-level variables Z1, . . . , Zq, which leads to cross-level
interactions in the model for Yij, the model for the random slope is
given in (5.11),

Heteroscedasticity. The dependent variable in a random slope model is

heteroscedastic, that is, the residual standard deviation depends on the
variables that have random slopes.

Cross-level interaction effect. An interaction between two variables

defined at different levels, which can enter the model when a random
slope is explained by a level-two variable.

Model specification. This received ample attention, including the

following points:

Covariances between random intercepts and random slopes should

1. (almost always) be estimated as free parameters, and not be
equated a priori to zero.
Random slopes can often be explained by level-two variables,
2.
leading to models with cross-level interactions.
Decisions have to be made about whether or not to apply group
3.
centering to variables that have random slopes.
To avoid ecological fallacies and have the best possible theoretical
4. interpretations, it is important also to consider interactions
involving group means of level-one variables.

Parameter estimation. A brief intuitive overview of parameter estimation

in the hierarchical linear model was given in Section 5.4.

Empirical Bayes estimation. The random slopes can be estimated by

empirical Bayes methods just like the random intercepts. Such
estimates are called posterior slopes.
Higher-level hierarchical linear models. The hierarchical linear model
can also be defined for three- and higher-level data structures, and for
the three-level case this was discussed in Section 5.5.

1
In the older literature, these equations were applied to the estimated groupwise regression
coefficients rather than the latent coefficients. The statistical estimation then was carried out in two
stages: first ordinary least squares (OLS) estimation within each group, then OLS estimation with the
estimated coefficients as outcomes. This is statistically inefficient unless the group sizes are very
large, and does not distinguish the ‘true score’ variability of the latent coefficients from the sampling
variability of the estimated groupwise regression coefficients. A two-step approach which does make
this distinction is briefly treated in Section 3.7, following equation (3.40).
2
It is mathematically possible that some variables have a random but not a fixed effect. This makes
sense only in special cases.
 6
Testing and Model Specification

OVERVIEW OF THE CHAPTER

The first part of this chapter discusses hypothesis tests for parameters in the
hierarchical linear model, with separate treatments of tests for the
parameters in the fixed part and those in the random part. The second part is
concerned with model specification issues.
It is assumed that the reader has a basic knowledge of statistical
hypothesis testing: null and alternative hypotheses, type I and type II errors,
significance level, and statistical power.

6.1 Tests for fixed parameters

Suppose we are working with a model represented as (5.15). The null
hypothesis that a certain regression parameter is 0,

can be tested by a t-test. The statistical estimation leads to an estimate

with associated standard error S.E.( ). For a small number of groups N ≤
50 the standard errors based on REML estimation should be used (the
difference between REML and ML estimation was explained in Sections
4.7 and 5.4; the need to use REML standard errors is argued by Manor and
Zucker, 2004), while for a larger number of groups this is not important.
Their ratio is a t-value:
One-sided as well as two-sided tests can be carried out on the basis of this
test statistic.1 Under the null hypothesis, T(γh) has approximately a t
distribution, but the number of degrees of freedom (df) is somewhat more
complicated than in multiple linear regression,because of the presence of
the two levels. The approximation by the t distribution is not exact even if
the normality assumption for the random coefficients holds. Suppose first
that we are testing the coefficient of a level-one variable. If the total number
of level-one units is M and the total number of explanatory variables is r,
then we can take df = M − r − 1. To test the coefficient of a level-two
variable when there are N level-two units and q explanatory variables at
level two, we take df = N − q − 1. To test the coefficient of the cross-level
interaction between level-one variable X and level-two variable Z, when the
model contains a total of q other level-two variables also interacting with
this variable X, we also use df = N − q − 1.
If the number of degrees of freedom is large enough (say, larger than
40), the t distribution can be replaced by a standard normal distribution.
This rule for the degrees of freedom in the t-test is simple and has good
properties. The literature contains proposals for various other rules. Manor
and Zucker (2004) give a review and a simulation study. Their conclusion
is, first, that these tests should use the standard errors obtained from REML
estimation, not from ML estimation. Furthermore, they found that the so-
called Satterthwaite approximation as well as the simple approximation
mentioned here give a good control of the type I error rate, and most other
methods give inflated type I error rates for fewer than 30 groups. Maas and
Hox (2004) confirmed that for 30 or more groups, the Wald tests of fixed
effects using REML standard errors have reliable type I error rates.
Example 6.1 Testing within- and between-group regressions.
Continuing the examples on students in schools of Chapters 4 and 5, we now wish to test whether
between-group and within-group regressions of language test score on IQ are different from one
another, when controlling for socio-economic status (SES). A model with a random slope for IQ is
used. Two models are estimated and presented in Table 6.1. The first contains the raw (i.e., grand-
mean-centered) IQ variable along with the group mean, the second contains the within-group
deviation variable , defined as
also together with the group mean. The variable with the random slope is in both models the grand-
mean-centered variable IQ. (It may be noted that although this variable is not listed among the
variables with a fixed effect in Model 2, it is present implicitly because it is the sum of two included
variables that do have a fixed effect.)
To test whether within- and between-group regression coefficients are different, the significance
of the group mean is tested, while controlling for the effect of the original variable, IQ. The
difference of within- and between-group regressions is discussed further in Section 4.6.
Table 6.1 shows that only the estimate for differs between the two models. This is in
accordance with Section 4.6: if IQ is variable 1 and is variable 2, so that their regression
coefficients are γ10 and γ20, respectively, then the within-group regression coefficient is γ10 in Model
1 and γ20 in Model 2, while the between-group regression coefficient is γ10+γ01 in Model 1 and γ01 in
Model 2.The models are equivalent representations of the data and differ only in the parametrization.
The deviances (explained in Section 6.2) are exactly the same.
The within-group regression and between-group regressions are the same if γ01 = 0 in Model 1,
that is, there is no effect of the group mean given that the model controls for the raw variable (i.e., the
variable without group centering). The t-test statistic for testing H0: γ01 = 0 in Model 1 is equal to
0.647/0.264 = 2.45. This is significant (two-sided p < 0.02). It may be concluded that within-group
and between-group regressions are significantly different.

Table 6.1: Estimates for two models with different between- and within-
group regressions.

The results for Model 2 can be used to test whether the within-group or between-group
regressions are 0. The t-statistic for testing the within-group regression is 2.265/0.065 = 34.9, the
statistic for testing the between-group regression is 2.912/0.262 = 11.1. Both are extremely
significant. In conclusion, there are positive within-group as well as between-group regressions, and
these are different from one another.

6.1.1 Multiparameter tests for fixed effects

Sometimes we wish to test several regression parameters simultaneously.
For example, consider testing the effect of a categorical variable with more
than three categories. The effect of such a variable can be represented in the
fixed part of the hierarchical linear model by c − 1 dummy variables, where
c is the number of categories, and this effect is nil if and only if all the
corresponding c − 1 regression coefficients are 0. Two types of test are
much used for this purpose: the multivariate Wald test and the likelihood
ratio test, also known as the deviance test. The latter test is explained in the
next section.
For the multivariate Wald test, we need not only the standard errors of
the estimates but also the covariances among them. Suppose that we
consider a certain vector γ of q regression parameters, for which we wish to
test the null hypothesis

The statistical estimation leads to an estimate and an associated estimated

covariance matrix . Here again the covariance matrix produced by REML
estimation, not by ML estimation, needs to be used, unless the number of
groups is large enough (N ≥ 50). From the estimate and its covariance
matrix, we can let a computer program calculate the test statistic
represented in matrix form by

Under the null hypothesis, the distribution of this statistic divided by q can
be approximated by an F distribution, with q degrees of freedom in the
numerator, while the degrees of freedom in the denominator are determined
as for the t-test explained above. Here, for a large number of degrees of
freedom (say, more than 40), the F distribution is approximated by a chi-
squared distribution with q degrees of freedom.
 The way of obtaining tests presented in this section is not applicable to
tests of whether parameters (variances or covariances) in the random part of
the model are 0. The reason is that, if a population variance parameter is 0,
its estimate divided by the estimated standard error does not approximately
have a t distribution. Tests for such hypotheses are discussed in the next
section.

6.2 Deviance tests

The deviance test, or likelihood ratio test, is a quite general principle for
statistical testing. In applications of the hierarchical linear model this test is
mainly used for multiparameter tests and for tests about the random part of
the model. The general principle is as follows.
When parameters of a statistical model are estimated by the maximum
likelihood method, the estimation also provides the likelihood, which can
be transformed into the deviance defined as minus twice the natural
logarithm of the likelihood. This deviance can be regarded as a measure of
lack of fit between model and data, but (in most statistical models) one
cannot interpret the values of the deviance directly, but only differences in
deviance values for several models fitted to the same data set.
Suppose that two models are fitted to one data set, model M0 with m0
parameters and a larger model M1 with m1 parameters. Thus M1 can be
regarded as an extension of M0, with m1 − m0 parameters added. Suppose
that M0 is tested as the null hypothesis and M1 is the alternative hypothesis.
Indicating the deviances by D0 and D1, respectively, their difference D0 −
D1 can be used as a test statistic having a chi-squared distribution with m1 −
m0 degrees of freedom. This type of test can be applied to parameters of the
fixed as well as of the random part.
The deviance produced by the residual maximum likelihood method can
be used in deviance tests only if the two models compared (M0 and M1)
have the same fixed parts and differ only in their random parts. Otherwise
the ML estimation method must be used.
Example 6.2 Test of random intercept.
In Example 4.2, the random intercept model yields a deviance of D1 = 24,912.2, while the OLS
regression model has a deviance of D0 = 25,351.0. There is m1 − m0 = 1 parameter added, the random
intercept variance. The deviance difference is 438.8, immensely significant in a chi-squared
distribution with df = 1. This implies that, even when controlling for the effect of IQ, the differences
between groups are strongly significant.
In Section 6.2.1 we shall see, however, that the test of the random intercept can be improved.

For example, suppose one is testing the fixed effect of a categorical

explanatory variable with c categories. This categorical variable can be
represented by c − 1 dummy variables. Model M0 is the hierarchical linear
model with the effects of the other variables in the fixed part and with the
given random part. Model M1 also includes all these effects; in addition, the
c − 1 regression coefficients of the dummy variables have been added.
Hence the difference in the number of parameters is m1 − m0 = c − 1. This
implies that the deviance difference D0 − D1 can be tested against a chi-
squared distribution with df = c − 1. This test is an alternative for the
multiparameter Wald test treated in the preceding section. These tests will
be very close to each other for intermediate and large sample sizes. For a
small number of groups (N < 40) the precision of type I error rates can be
improved by the so-called Bartlett correction (see Zucker et al., 2000;
Manor and Zucker, 2004).
Example 6.3 Effect of a categorical variable.
In the data set used in Chapters 4 and 5, schools differ according to their denomination: public,
Catholic, Protestant, nondenominational private, and a remainder category. To represent these five
categories, four dummy variables are used, contrasting the last four categories with the first. This
means that all dummy variables are 0 for public schools, the first is 1 for Catholic schools, etc. When
the fixed effects of these c − 1 = 4 variables are added to the model presented in Table 5.2, which in
this example has the role of M0 with deviance D0 = 24,856.8, the deviance decreases to D1 =
24,840.3. The chi-squared value is D0 − D1 = 16.5 with df = 4, p < 0.005. It can be concluded that,
while controlling for IQ, the group average of IQ, and the interaction of these two variables (see the
specification of model M0 in Table 5.2), there are differences between the five types of school.
The estimated fixed effects (with standard errors) of the dummy variables are 1.65 (0.55) for the
Catholic and −0.51 (0.59) for the Protestant schools, and 0.60 (1.05) and 0.23 (0.94) for the other two
categories. These effects are relative to the public schools. Calculating the t-ratios shows that the
Catholic schools show higher achievement, controlling for these three IQ variables, than the public
schools, while the other three categories do not differ significantly from the public schools.

6.2.1 More powerful tests for variance parameters

Variances are by definition nonnegative. When testing the null hypothesis
that the variance of the random intercept or of a random slope is zero, the
alternative hypothesis is therefore one-sided. This observation can be used
to arrive at a sharpened version of the deviance test for variance parameters.
This principle was derived by Miller (1977), Self and Liang (1987), and
Stram and Lee (1994). Molenberghs and Verbeke (2007) give an overview
of this and other tests for parameters in the random part. The good
properties of this procedure were confirmed in a simulation study by
LaHuis and Ferguson (2009).
First, consider the case where a random intercept is tested. The null
model M0 is the model without a random part at level two, that is, all
observations Yij are independent, conditional on the values of the
explanatory variables. This is an ordinary linear regression model. The
alternative model M1 is the random intercept model with the same
explanatory variables. There is m1 − m0 = 1 additional parameter, the
random intercept variance τ02. For the observed deviances D0 of model M0
(this model can be estimated by ordinary least squares) and D1 of the
random intercept model, the difference D0 − D1 is calculated. If D0 − D1 =
0, the random intercept variance is definitely not significant (it is estimated
as being 0 ...). If D0 − D1 > 0, the tail probability of the difference D0 − D1
is looked up in a table of the chi-squared distribution with df = 1. The p-
value for testing the significance of the random intercept variance is half
this tail value.
Second, consider the case of testing a random slope. Specifically,
suppose that model (5.15) holds, and the null hypothesis is that the last
random slope variance is zero: τp2 = 0. Under this null hypothesis, the p
covariances, τhp for h = 0, ... , p − 1, are also 0. The alternative hypothesis is
the model defined by (5.15), and has m 1 − m0 = p + 1 parameters more than
the null model (one variance and p covariances). For example, if there are
no other random slopes in the model (p = 1), m1 − m0 = 2. A similar
procedure is followed as for testing the random intercept, but now it is more
complicated.2 Both models are estimated, yielding the deviance difference
D0 − D1. If D0 − D1 = 0, the random slope variance is not significant. If D0
− D1 > 0, the tail probability of the difference D0 − D1 can be tested against
a so-called mixture distribution of the chi-squared distribution with df = p +
1 and the chi-squared distribution with df = p, with mixture proportions .
This is also called the chi-bar distribution. It is not a standard distribution.
The p-value is just the average of the p-value from the distribution and
that from the distribution. This can be easily calculated by any
program that can calculate p-values (or cumulative distribution functions)
for chi-squared distributions.
Table 6.2 gives critical values for testing a random slope in a model that
has a random intercept and no, one, or two additional random slopes. These
critical values are in between the critical values of the and
distributions, where p is the total number of random slopes in the model
including the tested one but excluding the random intercept. Therefore, a
conservative test is obtained by testing against the distribution.

Table 6.2: Critical values for 50–50 mixture of and distribution.

Example 6.4 Test of random slope.

When comparing Tables 4.4 and 5.1, it can be concluded that m1 − m0 = 2 parameters are added and
the deviance diminishes by D0 − D1 = 15,227.5 − 15,213.5 = 14.0. Testing the value of 14.0 in Table
6.2 for p = 1 yields p < 0.001. Thus, the significance probability of the random slope for IQ in the
model of Table 5.1 is p < 0.001.
As another example, suppose that one wishes to test the significance of the random slope for IQ
in the model of Table 5.4. Then the model must be fitted in which this effect is omitted and all other
effects remain. Thus, the omitted parameters are τ12 and τ01 so that df = 2. Fitting this reduced model
leads to a deviance of D0 = 24,655.1, which is 28.3 more than the deviance in Table 5.4. Table 6.2 for
p = 1 shows that again p < 0.001. Thus there is strong support for the random slope.

6.3 Other tests for parameters in the random part

Deviance tests are very convenient for testing parameters in the random
part, but other tests for random intercepts and slopes also exist. Some of
these are reviewed by Molenberghs and Verbeke (2007). We mentioned in
Section 3.3.2 the ANOVA F-test for the intraclass correlation. This is
effectively a test for randomness of the intercept. If it is desired to test the
random intercept while controlling for explanatory variables, one may use
the F-test from the ANCOVA model, using the explanatory variables as
covariates.
Raudenbush and Bryk (2002) present chi-squared tests for random
intercepts and slopes, that is, for variance parameters in the random part.
These are based on calculating OLS estimates for the values of the random
effects within each group and testing these values for equality. For the
random intercept, these are the large-sample chi-squared approximations to
the F-tests in the ANOVA or ANCOVA model mentioned in Section 3.3.2.
Another test for variance parameters in the random part was proposed
by Berkhof and Snijders (2001) and Verbeke and Molenberghs (2003), and
is explained in Section 10.2.2.

6.3.1 Confidence intervals for parameters in the random part

When constructing confidence intervals for variance parameters it is
tempting to use the symmetric confidence interval, which for a confidence
level of 95% is defined as the estimate plus or minus 1.96 times the
standard error. This confidence interval, however, is based on the
assumption that the parameter estimate is nearly normally distributed,
which is doubtful for estimated variances; for example, because these are
necessarily nonnegative. These confidence intervals are applicable here
only when the standard error is quite small compared to the estimate, so that
they do not come near the value of 0. Generally, they are more safely
applied to the standard deviation (the square root of the variance). A
symmetric confidence interval for the standard deviation then can be
transformed back to a nonsymmetric confidence interval for the variance, if
this is desired. Sometimes also the logarithm of the variance can be used.
(Note that the properties of the logarithm imply that ln(τ2) = 2 ln(τ).)
The standard errors of these transformations are related through the
approximations3
where ln denotes the natural logarithm. If this value is less than 0.1 the
distribution of will usually be close to a normal distribution; if it is
between 0.1 and 0.3 then the symmetric confidence interval still will give a
reasonable rough approximation; if it is larger than 0.3 (which means that
the number of higher-level units is probably rather low, or the higher-level
variance is quite small) the normal approximation itself will break down
and the symmetric confidence interval should not be used.
A better way to construct the confidence interval is based on the so-
called profile likelihood.4 This confidence interval consists of those values
of the parameter for which twice the logarithm of the profile likelihood is
not smaller than twice the logarithm of the maximized likelihood in the ML
estimate, minus a value c; this c is the critical value in a chi-squared
distribution on 1 degree of freedom (c = 3.84 for a confidence level of
95%). This confidence interval will be implemented in R package lme45
(see Bates, 2010). Another option is to follow a Bayesian approach (Section
12.1) and construct a confidence interval based on the posterior distribution
of τ2 or τ without relying on normal approximations.
Example 6.5 Confidence interval for intercept and slope variances
In the random intercept model of Example 4.3, the parameter estimates for the two variance
components are 02 = 8.68 (S.E. = 1.10) for the level-two variance and 2 = 40.43 (S.E. = 0.96) for
the level-one variance. The 95% confidence intervals based on the profile likelihood, obtained from
R package lme4a, are as follows. The nature of the likelihood-based method implies that the
endpoints of the confidence interval for the variances are the squares of those for the standard
deviations: for the standard deviations, the interval is

and for the variances,

The symmetric 95% confidence intervals based on the standard errors are defined as the parameter
estimates plus or minus 1.96 times the standard errors. These confidence intervals are less reliable. In
this example we use for the variances the standard errors in Table 4.4, and for the standard deviations
the standard errors from equation (6.2). The results for the standard deviations are

and for the variances,

It can be concluded that in this case, where the ratios of the variance estimates to their standard errors
are reasonably large for both parameters, the symmetric confidence intervals are very good
approximations when based on the standard deviations, especially for σ, and less good for the
variances but for σ2 still totally acceptable.
To construct a confidence interval for the intercept variance τ02 using only the information in
Table 4.4, if one does not have access to the profile likelihood method, the best way (as mentioned
above) is to construct a symmetric confidence interval for the intercept standard deviation τ0 and
transform this to a confidence interval for the variance. This operates as follows (we work to three
decimal places so as to have precision to two decimal places in the result):

2
* From 0 = 8.680, calculate 0 = 2.946.
2
* From S.E.( 0)
= 1.096 and formula (6.2), calculate S.E.( 0) = 1.096/(2 × 2.946) =
0.1860.
* The confidence interval for τ0 now extends from 2.946 − 1.960 × 0.1860 = 2.581 to 2.946
+ 1.960 × 0.1860 = 3.311.
* Hence the confidence interval for τ02 extends from 2.5812 = 6.66 to 3.3112 = 10.96. This is
still not as good as the interval based on the profile likelihood, but considerably better than
the symmetric interval directly based on the estimated intercept variance.

6.4 Model specification

Model specification is the choice of a satisfactory model. For the
hierarchical linear model this amounts to the selection of relevant
explanatory variables (and interactions) in the fixed part, and relevant
random slopes (with their covariance pattern) in the random part. Model
specification is one of the most difficult parts of statistical inference,
because there are two steering wheels: substantive (subject-matter related)
and statistical considerations. These steering wheels must be handled
jointly. The purpose of model specification is to arrive at a model that
describes the observed data to a satisfactory extent but without unnecessary
complications. A parallel purpose is to obtain a model that is of substantive
interest without wringing from the data drops that are really based on
chance but interpreted as substance. In linear regression analysis, model
specification is already complicated – as is discussed in many textbooks on
regression (e.g., Ryan, 1997) – and in multilevel analysis the number of
complications is multiplied because of the complicated nature of the
random part.
The complicated nature of the hierarchical linear model, combined with
the two steering wheels for model specification, implies that there are no
clear and fixed rules to follow. Model specification is a process guided by
the following principles.

1. Considerations relating to the subject matter. These follow from

field knowledge, existing theory, detailed problem formulation,
and common sense.
2. The distinction between effects that are indicated a priori as effects
to be tested, that is, effects on which the research is focused, and
effects that are necessary to obtain a good model fit. Often the
effects tested are a subset of the fixed effects, and the random part
is to be fitted adequately but of secondary interest. When there is
no strong prior knowledge about which variables to include in the
random part, one may follow a data-driven approach to select the
variables for the random part.
3. A preference for ‘hierarchical’ models in the general sense (not the
‘hierarchical linear model’ sense) that if a model contains an
interaction effect, then the corresponding main effects should
usually also be included (even if these are not significant); and if a
variable has a random slope, its fixed effect should normally also
be included in the model. The reason is that omitting such effects
may lead to erroneous interpretations.
4. Doing justice to the multilevel nature of the problem. This is done
as follows:
(a) When a given level-one variable is present, one should be
aware that the within-group regression coefficient may differ
from the between-group regression coefficient, as described in
Section 4.6. This can be investigated by calculating a new
variable, defined as the group mean of the level-one variable,
and testing the effect of the new variable.
(b) When there is an important random intercept variance, there
are important unexplained differences between group means.
 One may look for level-two variables (original level-two
variables as well as aggregates of level-one variables) that
explain part of these between-group differences.
(c) When there is an important random slope variance of some
level-one variable (say, X1), there are important unexplained
differences between the within-group effects of X1 on Y. Here
also one may look for level-two variables that explain part of
these differences. This leads to cross-level interactions as
explained in Section 5.2. Recall from this section, however,
that cross-level interactions can also be expected from
theoretical considerations, even if no significant random slope
variance is found.
5. Awareness of the necessity of including certain covariances of
random effects. Including such covariances means that they are
free parameters in the model, not constrained to 0 but estimated
from the data.
In Section 5.1.2, attention was given to the necessity to include in
the model all covariances τ0h between random slopes and random
intercept. Another case in point arises when a categorical variable
with c ≥ 3 categories has a random effect. This is implemented by
giving random slopes to the c − 1 dummy variables that are used to
represent the categorical variable. The covariances between these
random slopes should then also be included in the model.
Formulated generally, suppose that two variables Xh and Xh′ have
random effects, and that the meaning of these variables is such that
they could be replaced by two linear combinations, aXh + a′Xh′
and bXh + b′Xh′. (For the random intercept and random slope
discussed in Section 5.1.2, the relevant type of linear combination
would correspond to a change of origin of the variable with the
random slope.) Then the covariance τhh′ between the two random
slopes should be included in the model.
6. Reluctance to include nonsignificant effects in the model – one
could also say, a reluctance to overfit. Each of points 1–5 above,
however, could override this reluctance.
 An obvious example of this overriding is the case where one
wishes to test for the effect of X2, while controlling for the effect
of X1. The purpose of the analysis is a subject-matter
consideration, and even if the effect of X1 is nonsignificant, one
still should include this effect in the model.
7. The desire to obtain a good fit, and include all effects in the model
that contribute to a good fit. In practice, this leads to the inclusion
of all significant effects unless the data set is so large that certain
effects, although significant, are deemed unimportant nevertheless.
8. Awareness of the following two basic statistical facts:
(a) Every test of a given statistical parameter controls for all other
effects in the model used as a null hypothesis (M0 in Section
6.2). Since the latter set of effects has an influence on the
interpretation as well as on the statistical power, test results
may depend on the set of other effects included in the model.
(b) We are constantly making type I and type II errors. Especially
the latter, since statistical power often is rather low. This
implies that an effect being nonsignificant does not mean that
the effect is absent in the population. It also implies that a
significant effect may be so by chance (but the probability of
this is no larger than the level of significance – most often set
at 0.05).
Multilevel research often is based on data with a limited
number of groups. Since power for detecting effects of level-
two variables depends strongly on the number of groups in the
data, warnings about low power are especially important for
level-two variables.
9. Providing tested fixed effects with an appropriate error term in the
model (whether or not it is significant). For level-two variables,
this is the random intercept term (the residual term in (5.7)). For
cross-level interactions, it is the random slope of the level-one
variable involved in the interaction (the residual term in (5.8)). For
level-one variables, it is the regular level-one residual that one
 would not dream of omitting. This guideline is supported by
Berkhof and Kampen (2004).

These considerations are nice – but how should one proceed in practice?
To get an insight into the data, it is usually advisable to start with a
descriptive analysis of the variables: an investigation of their means,
standard deviations, correlations, and distributional forms. It is also helpful
to make a preliminary (‘quick and dirty’) analysis with a simpler method
such as OLS regression.
When starting with the multilevel analysis as such, in most situations
(longitudinal data may provide an exception), it is advisable to start with
fitting the empty model (4.6). This gives the raw within-group and between-
group variances, from which the estimated intraclass correlation can be
calculated. These parameters are useful as a general description and a
starting point for further model fitting. The process of further model
specification will include forward steps – select additional effects (fixed or
random), test their significance, and decide whether or not to include them
in the model – and backward steps – exclude effects from the model
because they are not important from a statistical or substantive point of
view. We mention two possible approaches in the following subsections.

6.4.1 Working upward from level one

In the spirit of Section 5.2, one may start with the construction of a model
for level one, that is, first explaining within-group variability, and then
explaining between-group variability. This two-phase approach is
advocated by Raudenbush and Bryk (2002) and is followed in the HLM
program (see Chapter 18).

Modeling within-group variability

Subject-matter knowledge and availability of data lead to a number of level-
one variables X1, ... , Xp which are deemed important, or hypothetically
important, to predict or explain the value of Y. These variables lead to
equation (5.10) as a starting point:
This equation represents the within-group effects of Xh on Y. The between-
group variability is first modeled as random variability. This is represented
by splitting the group-dependent regression coefficients βhj into a mean
coefficient γh0 and a group-dependent deviation Uhj:

Substitution yields

This model has a number of level-one variables with fixed and random
effects, but it will usually not be necessary to include all random effects.
For the precise specification of the level-one model, the following steps
are useful.

1. Select in any case the variables on which the research is focused.

In addition, select relevant available level-one variables on the
basis of subject-matter knowledge. Also include plausible
interactions between level-one variables.
2. Select, among these variables, those for which, on the basis of
subject-matter knowledge, a group-dependent effect (random
slope!) is plausible. If unsure, one could select the variables that
are expected to have the strongest fixed effects.
3. Estimate the model with the fixed effects of step 1 and the random
effects of step 2.
4. Test the significance of the random slopes, and exclude the
nonsignificant slopes from the model.
5. Test the significance of the regression coefficients, and exclude the
nonsignificant coefficients from the model. This may also be the
moment to consider the inclusion of interaction effects between
level-one variables.
6. As a check, one could test whether the variables for which a group-
 dependent effect (i.e., a random slope) was not thought plausible in
step 2 do indeed have a nonsignificant random slope. (Keep in
mind that including a random slope normally implies inclusion of
the fixed effect!) Be reluctant to include random slopes for
interactions; these are often hard to interpret.
7. Use the methods of Chapter 10 for checking the assumptions (such
as homoscedasticity and normal distributions of the residuals at all
levels).

With respect to the random slopes, one may be restricted by the fact that
data usually contain less information about random effects than about fixed
effects. Including many random slopes can therefore lead to long iteration
processes of the estimation algorithm. The algorithm may even fail to
converge. For this reason it may be necessary to specify only a small
number of random slopes.
After this process, one has arrived at a model with a number of level-
one variables, some of which have a random effect in addition to their fixed
effect. It is possible that the random intercept is the only remaining random
effect. This model is an interesting intermediate product, as it indicates the
within-group regressions and their variability.

Modeling between-group variability

The next step is to try and explain these random effects by level-two
variables. The random intercept variance can be explained by level-two
variables, the random slopes by interactions of level-one with level-two
variables, as was discussed in Section 5.2. It should be kept in mind that
aggregates of level-one variables can be important level-two variables.
For deciding which main effects of level-two variables and which cross-
level interactions to include, it is again advisable first to select those effects
that are plausible on the basis of substantive knowledge, then to test these
and include or omit them depending on their importance (statistical and
substantive), and finally to check whether other (less plausible) effects also
are significant.
This procedure has a built-in filter for cross-level interactions: an
interaction between level-one variable X and level-two variable Z is
considered only if X has a significant random slope. However, this ‘filter’
should not be employed as a strict rule. If there are theoretical reasons to
consider the X × Z interaction, this interaction can be tested even if X does
not have a significant random slope. The background to this is the fact that
if there is X × Z interaction, the test for this interaction has a higher power
to detect this than the test for a random slope.
It is possible that, if one carries out both tests, the test of the random
slope is nonsignificant while the test of the X × Z interaction is significant.
This implies that either a type I error has been made by the test on X × Z
interaction (this is the case if there is no interaction), or a type II error has
been made by the test of the random slope (this is the case if there is
interaction). Assuming that the significance level is 0.05 and one focuses on
the test of this interaction effect, the probability of a type I error is less than
0.05 whereas the probability of a type II error can be quite high, especially
since the test of the random slope does not always have high power for
testing the specific alternative hypothesis of an X × Z interaction effect.
Therefore, provided that the X × Z interaction effect was hypothesized
before looking at the data, the significant result of the test of this effect is
what counts, and not the lack of significance for the random slope.
By the same reasoning, if there is theoretical justification for testing the
effect of a level-two variable it is not necessary to require the precondition
that there is positive and significant random intercept variance.

6.4.2 Joint consideration of level-one and level-two variables

The procedure of first building a level-one model and then extending it with
level-two variables is neat but not always the most efficient or most
relevant. If there are level-two variables or cross-level interactions that are
known to be important, why not include them in the model right from the
start? For example, it could be expected that a certain level-one variable has
a within-group regression differing from its between-group regression. In
such a case, one may wish to include the group mean of this variable at the
outset.
In this approach, the same steps are followed as in the preceding
section, but without the distinction between level-one and level-two
variables. This leads to the following steps:
 1. Select relevant available level-one and level-two variables on the
basis of subject-matter knowledge. Also include plausible
interactions. Do not forget group means of level-one variables to
account for the possibility of difference between within-group and
between-group regressions. Also do not forget cross-level
interactions.
2. Select among the level-one variables those for which, on the basis
of subject-matter knowledge, a group-dependent effect (random
slope!) is plausible. (A possibility would be, again, to select those
variables that are expected to have the strongest fixed effects.)
3. Estimate the model with the fixed effects of step 1 and the random
effects of step 2.
4. Test the significance of the random slope variance, and exclude
from the model the random slopes of those variables for which this
variance is not significant.
5. Test the significance of the regression coefficients, and exclude the
nonsignificant coefficients from the model. This can also be a
moment to consider the inclusion of more interaction effects.
6. Check whether other effects, thought less plausible at the start of
model building, are indeed not significant. If they are significant,
include them in the model.
7. Use the methods of Chapter 10 for checking the assumptions (such
as homoscedasticity and normal distributions of the residuals at all
levels).

In an extreme instance of step 1, one may wish to include all available

variables and a large number of interactions in the fixed part. Similarly, one
might wish to give all level-one variables random effects in step 2. Whether
this is possible in practice will depend, inter alia, on the number of level-
one variables. Such an implementation of these steps leads to a backward
model-fitting process, where one starts with a large model and reduces it by
stepwise excluding nonsignificant effects. The advantage is that masking
effects (where a variable is excluded early in the model building process
because of nonsignificance, whereas it would have reached significance if
one had controlled for another variable) do not occur. The disadvantage is
that it may be a very time-consuming procedure.

6.4.3 Concluding remarks on model specification

This section has suggested a general approach to specification of multilevel
models rather than laying out a step-by-step procedure. This is in
accordance with our view of model specification as a process with two
steering wheels and without foolproof procedures. This implies that, given
one data set, a (team of) researcher(s) may come up with more than one
model, each in itself an apparently satisfactory result of a model
specification process. In our view, this reflects the basic indeterminacy that
is inherent in model fitting on the basis of empirical data. It is quite possible
for several different models to correspond to a given data set, and that there
are no compelling arguments to choose between them. In such cases, it is
better to accept this indeterminacy and leave it to be resolved by further
research than to make an unwarranted choice between the different models.
This treatment of model specification may seem rather inductive, or
data-driven. If one is in the fortunate situation of having a priori hypotheses
to test (usually about regression coefficients), it is useful to distinguish
between, on the one hand, the parameters on which hypothesis tests are
focused, and, on the other hand, the parts of the model that are required to
make it fit the data well. The latter part is important to get right in order to
obtain valid tests of the hypotheses. These two parts can be treated
distinctly: the first part must evidently be included in the model, and for the
second part an inductive approach is adequate.
Another aspect of model specification is the checking of assumptions.
Independence assumptions should be checked in the course of specifying
the random part of the model. Distributional assumptions – specifically, the
assumption of normal distributions for the various random effects – should
be checked by residual analysis. Checks of assumptions are treated in
Chapter 10.

6.5 Glommary
t-test. The easiest test for a fixed coefficient is the so-called Wald test. Its
test statistic is the t-ratio, that is, the ratio of estimate to standard error.
This ratio is tested against a t distribution. For small N the number of
degrees of freedom is important and rules were given for this purpose.

F-test. Multiparameter tests of fixed coefficients according to the principle

of theWald test lead to F-tests, which for large N can be tested against
a chi-squared distribution.

Deviance test. A general procedure for hypothesis testing is the deviance

test or likelihood ratio test. This can be easily used for multiparameter
tests, for example, tests of a categorical variable in the fixed part of the
model.
Random intercept and random slope variances can also be tested by
deviance tests. Then some power can be gained by using the fact that
variances cannot be negative, which leads to a modified null
distribution.

ML and REML estimation. For Wald tests of parameters in the fixed part
when the number of higher-level units is rather small (N ≤ 50), REML
standard errors must be used. For deviance tests of random part
parameters, ML estimates must be used.

Confidence intervals for random part parameters. These are tricky,

because the estimated variances mostly do not have nearly normal
distribution so that the usual symmetric confidence interval is
inapplicable. Accordingly, the interpretation of standard errors of
parameters in the random part is dubious unless the standard error is
quite small compared to the parameter estimate. The symmetric 95%
confidence interval, defined as parameter estimate ± 1.96 times the
standard error, is more precise when applied to the standard deviation
(τ) than to the variance (τ2); this approach requires that the resulting
interval is far enough away from 0. The profile likelihood method
provides better confidence intervals, without assuming normality for
the parameter estimates. Another possibility is to follow a Bayesian
approach (Section 12.1) and construct a posterior distribution.
Model specification. This depends on a combination of subject-matter and
statistical considerations, and it is impossible to give simple rules. A
number of considerations and possible approaches were discussed, and
more will follow in Chapter 10.

1
This is one of the common principles for construction of a t-test. This type of test is called the Wald
test, after the statistician Abraham Wald (1902–1950).
2
In the first edition, we erroneously stated that exactly the same procedure could be followed.
3
These approximations are based on linear approximations to the logarithmic and square root
functions (the so-called delta method), and are invalid if these standard errors are too large.
4
The profile likelihood function for a given parameter is the value of the likelihood function for this
parameter, maximized over all other parameters.
5
At the time of writing, it is implemented in the beta version lme4a, available from R-Forge.
 7
How Much Does the Model Explain?

OVERVIEW OF THE CHAPTER

This chapter starts by noting that, counter to what might intuitively be
expected, estimated variance parameters may go up when explanatory
variables are added to the model. This leads to issues in the definition of the
concept of explained variance, commonly referred to as R2. A definition of
R2, defined at level one, is given which does not have this problem in its
population values, although it still may sometimes have the problem in
values calculated from data.
Next an exposition is given of how different components of the
hierarchical linear model contribute to the total observed variance of the
dependent variable. This is a rather theoretical section that may be skipped
by the reader.

7.1 Explained variance

The concept of ‘explained variance’ is well known in multiple regression
analysis: it gives an answer to the question how much of the variability of
the dependent variable is accounted for by the linear regression on the
explanatory variables. The usual measure for the explained proportion of
variance is the squared multiple correlation coefficient, R2. For the
hierarchical linear model, however, the concept of ‘explained proportion of
variance’ is somewhat problematic. In this section, we follow the approach
of Snijders and Bosker (1994) to explain the difficulties and give a suitable
multilevel version of R2.
One way to approach this concept is to transfer its customary treatment,
well-known from multiple linear regression, straightforwardly to the
hierarchical random effects model: treat proportional reductions in the
estimated variance components, σ2 and in the random-intercept model for
two levels, as analogs of R2-values. Since there are several variance
components in the hierarchical linear model, this approach leads to several
R2-values, one for each variance component. However, this definition of R2
now and then leads to unpleasant surprises: it sometimes happens that
adding explanatory variables increases rather than decreases some of the
variance components. Even negative values of R2 are possible. Negative
values of R2 are clearly undesirable and are not in accordance with its
intuitive interpretation.
In the discussion of R2-type measures, it should be kept in mind that
these measures depend on the distribution of the explanatory variables. This
implies that these variables, denoted in this section by X, are supposed to be
drawn at random from the population at level one and the population at
level two, and not determined by the experimental design or the
researcher’s whim. In order to stress the random nature of the X-variable,
the values of X are denoted Xij, instead of by xij as in earlier chapters.

7.1.1 Negative values of R2?

As an example, we consider data from a study by Vermeulen and Bosker
(1992) on the effects of part-time teaching in primary schools. The
dependent variable Y is an arithmetic test score; the sample consists of 718
grade 3 pupils in 42 schools. An intelligence test score X is used as
predictor variable. Group sizes range from 1 to 33 with an average of 20. In
the following, it sometimes is desirable to present an example for balanced
data (i.e., with equal group sizes). The balanced data presented below are
the data artificially restricted to 33 schools with 10 pupils in each school, by
deleting schools with fewer than 10 pupils from the sample and randomly
sampling 10 pupils from each school from the remaining schools. For
demonstration purposes, three models were fitted: the empty Model A;
Model B, with a group mean, , as predictor variable; and Model C, with a
within-group deviation score (Xij − ) as predictor variable. Table 7.1
presents the results of the analyses both for the balanced and for the entire
data set. The residual variance at level one is denoted σ2; the residual
variance at level two is denoted

Table 7.1: Estimated residual variance parameters and for models

with within-group and between-group predictor variables.

From Table 7.1 we see that in the balanced as well as in the unbalanced
case, increases as a within-group deviation variable is added as an
explanatory variable to the model. Furthermore, for the balanced case, is
not affected by adding a group-level variable to the model. In the
unbalanced case, increases slightly when adding the group variable.
2
When R is defined as the proportional reduction in the residual variance
parameters, as discussed above, then R2 on the group level is negative for
Model C, while for the entire data set R2 on the pupil level is negative for
Model B. Estimating σ2 and using the REML method results in slightly
different parameter estimates. The pattern, however,remains the same. It is
argued below that defining R2 as the proportional reduction in residual
variance parameters σ2 and , respectively, is not the best way to define a
measure analogous to R2 in the linear regression model; and that the
problems mentioned can be solved by using other definitions, leading to the
measure denoted below by .
7.1.2 Definitions of the proportion of explained variance in two-
level models
In multiple linear regression, the customary R2 parameter can be introduced
in several ways: for example, as the maximal squared correlation coefficient
between the dependent variable and some linear combination of the
predictor variables, or as the proportional reduction in the residual variance
parameter due to the joint predictor variables. A very appealing principle to
define measures of modeled (or explained) variation is the principle of
proportional reduction of prediction error. This is one of the definitions of
R2 in multiple linear regression, and can be described as follows. A
population of values is given for the explanatory and the dependent
variables (X1i, …, Xqi, Yi), with a known joint probability distribution; β is
the value for which the expected squared error

is minimal. (This is the definition of the ordinary least squares (OLS)

estimation criterion. In this equation, β0 is defined as the intercept and X0i =
1 for all i.) If, for a certain case i, the values of X1i, …, Xqi are unknown,
then the best predictor for Yi is its expectation ε(Y), with mean squared
prediction error var(Yi); if the values X1 i, …, Xqi are given, then the linear
predictor of Yi with minimum squared error is the regression value ΣhβhXhi.
The difference between the observed value Yi and the predicted value Σh
βhXhi is the prediction error. Accordingly, the mean squared prediction error
is defined as the value of (7.1) for the optimal (i.e., estimated) value of β.
The proportional reduction of the mean squared error of prediction is
the same as the proportional reduction in the unexplained variance, due to
the use of the variables X1,…, Xq. Mathematically, it can be expressed as

This formula expresses one of the equivalent ways to define R2.

 The same principle can be used to define ‘explained proportion of
variance’ in the hierarchical linear model. For this model, however, there
are several options with respect to what one wishes to predict. Let us
consider a two-level model with dependent variable Y. In such a model, one
can choose between predicting an individual value Yij at the lowest level, or
a group mean On the basis of this distinction, two concepts of explained
proportion of variance in a two-level model can be defined. The first, and
most important, is the proportional reduction of error for predicting an
individual outcome. The second is the proportional reduction of error for
predicting a group mean. We treat the first concept here; the second is of
less practical importance and is discussed in Snijders and Bosker (1994).
First consider a two-level random effects model with a random intercept
and some predictor variables with fixed effects but no other random effects:

Since we wish to discuss the definition of ‘explained proportion of

variance’ as a population parameter, we assume temporarily that the vector
γ of regression coefficients is known.

Explained variance at level one

We define the explained proportion of variance at level one. This is based
on how well we can predict the outcome of Yij for a randomly drawn level-
one unit i within a randomly drawn level-two unit j. If the values of the
predictors Xij are unknown, then the best predictor for Yij is its expectation;
the associated mean squared prediction error is var(Yij). If the value of the
predictor vector Xij for the given unit is known, then the best linear
predictor for Yij is the regression value (where Xh0j is defined as 1
for all h, j.) The associated mean squared prediction error is

The level-one explained proportion of variance is defined as the

proportional reduction in mean squared prediction error:
 Now let us proceed from the population to the data. The most
straightforward way to estimate is to consider + for the empty
model,

as well as for the fitted model (7.2), and compute 1 minus the ratio of these
values. In other words, is just the proportional reduction in the value of
+ due to including the X-variables in the model. For a sequence of
nested models, the contributions to the estimated value of (7.3) due to
adding new predictors can be considered to be the contribution of these
predictors to the explained variance at level one.
To illustrate this, we once again use the data from the first (balanced)
example, and estimate the proportional reduction of prediction error for a
model where within-group and between-groups regression coefficients may
be different.

From Table 7.2 we see that + for model A amounts to 10.965, and for
model D to 7.964. is thus estimated to be 1−(7.964/10.965)= 0.274.

Population values of are nonnegative

What happens to when predictor variables are added to the multilevel
model? Is it possible that adding predictor variables leads to smaller values
of ? Can we even be sure at all that these quantities are positive?

Table 7.2: Estimating the level-one explained variance (balanced data).

It turns out that a distinction must be made between the population

parameter and its estimates from data. Population values of in
correctly specified models, with a constant group size n, become smaller
when predictor variables are deleted, provided that the variables U0j and Eij
on the one hand are uncorrelated with all the Xij variables on the other hand
(the usual model assumption).
For estimates of , however, the situation is different: these estimates
sometimes do increase when predictor variables are deleted. When it is
observed that an estimated value for becomes smaller by the addition of a
predictor variable, or larger by the deletion of a predictor variable, there are
two possibilities: either this is a chance fluctuation, or the larger model is
misspecified. Whether the first or the second possibility is more likely will
depend on the size of the change in , and on the subject-matter insight of
the researcher. In this sense changes in in the ‘wrong’ direction serve as a
diagnostic for possible misspecification. This possibility of misspecification
refers to the fixed part of the model, that is, the specification of the
explanatory variables having fixed regression coefficients, and not to the
random part of the model. We return to this in Section 10.2.3.

7.1.3 Explained variance in three-level models

In three-level random intercept models (Section 4.9), the residual variance,
or mean squared prediction variance, is the sum of the variance components
at the three levels, . Accordingly, the level-one explained
proportion of variance can be defined here as the proportional reduction in
the sum of these three variance parameters.
Example 7.1 Variance in maths performance explained by IQ.
In Example 4.8, Table 4.5 exhibits the results of the empty model (Model 1) and a model in which IQ
has a fixed effect (Model 2). The total variance in the empty model is 7.816 + 1.746 + 2.124 = 11.686
while the total unexplained variance in Model 2 is 6.910 + 0.701 + 1.109 = 8.720. Hence the level-
one explained proportion of variance is 1−(8.720/11.686)= 0.25.

7.1.4 Explained variance in models with random slopes

The idea of using the proportional reduction in the prediction error for
respectively, as the definitions of explained variance at either level,
can be extended to two-level models with one or more random regression
coefficients. The formulas to calculate can be found in Snijders and
Bosker (1994). However, the estimated values for usually change only
very little when random regression coefficients are included in the model.
The reason is that this definition of explained variance is based on
prediction of the dependent variable from observed variables, and the
knowledge of the precise specification of the random part only gives minor
changes to the quality of this prediction.
The formulas for estimating in models with random intercepts only
are very easy. Estimating R21 in models with random slopes is more tedious.
The simplest possibility for estimating R21 in random slope models is to re-
estimate the models as random intercept models with the same fixed parts
(omitting the random slopes), and use the resulting parameter estimates to
calculate R21 in the usual (simple) way for random intercept models. This
will normally yield values that are very close to the values for the random
slopes model.
Example 7.2 Explained variance for language scores.
In Table 5.4, a model was presented for the data set on language scores in elementary schools used
throughout Chapters 4 and 5. When a random intercept model is fitted with the same fixed part, the
estimated variance parameters are = 8.10 for level two and = 38.01 for level one. For the
empty model, Table 4.1 shows that the estimates are = 18.12 and = 62.85. This implies that
explained variance at level one is 1−(38.01 + 8.10)/(62.85 + 18.12)= 0.43. This is a quite high
explained variance. It can be deduced from the variances in Table 4.4.2 that the main part of this is
due to IQ.

7.2 Components of variance1

The preceding section focused on the total amount of variance that can be
explained by the explanatory variables. In these measures of explained
variance, only the fixed effects contribute. It can also be theoretically
illuminating to decompose the observed variance of Y into parts that
correspond to the various constituents of the model. This is discussed in this
section for a two-level model.
For the dependent variable Y, the level-one and level-two variances in
the empty model (4.6) are denoted by respectively. The total
variance of Y is therefore , and the components of variance are the
parts into which this quantity is split. The first split, obviously, is the split of
into and was extensively discussed in our treatment of the
intraclass correlation coefficient.
 To obtain formulas for a further decomposition, it is necessary to be
more specific about the distribution of the explanatory variables. It is usual
in single-level as well as in multilevel regression analysis to condition on
the values of the explanatory variables, that is, to consider those as given
values. In this section, however, all explanatory variables are regarded as
random variables with a given distribution.

7.2.1 Random intercept models

For the random intercept model, we divide the explanatory variables into
level-one variables X and level-two variables Z. Deviating from the notation
in other parts of this book, matrix notation is used, and X and Z denote
vectors.
The explanatory variables X1,…, XP at level one are collected in the
vector X with value Xij for unit i in group j. It is assumed more specifically
that Xij can be decomposed into independent level-one and level-two parts,

(i.e., a kind of multivariate hierarchical linear model without a fixed part).

The expectation is denoted by the level-one covariance matrix is

and the level-two covariance matrix is

This implies that the overall covariance matrix of X is the sum of these,

Further, the covariance matrix of the group average for a group of size n is
 It may be noted that this notation deviates slightly from the common
split of Xij into

The split (7.5) is a population-based split, whereas the more usual split (7.6)
is sample based. In the notation used here, the covariance matrix of the
within-group deviation variable is

while the covariance matrix of the group means is

For the discussion in this section, the present notation is more convenient.
The split (7.5) is not a completely innocuous assumption. The
independence between and implies that the covariance matrix of the
group means is at least as large2 as 1/(n − 1) times the within-group
covariance matrix of X.
The vector of explanatory variables Z = (Z1,..., Zq) at level two has value
Zj for group j. The vector of expectations of Z is denoted

and the covariance matrix is

In the random intercept model (4.8), denote the vector of regression

coefficients of the Xs by

and the vector of regression coefficients of the Zs by

Taking into account the stochastic nature of the explanatory variables then
leads to the following expression for the variance of Y:

This equation only holds if Z and are uncorrelated. For the special case
where all explanatory variables are uncorrelated, this expression is equal to

(This holds, for example, if there is only one level-one and only one level-
two explanatory variable.) This formula shows that, in this special case, the
contribution of each explanatory variable to the variance of the dependent
variable is given by the product of the regression coefficient and the
variance of the explanatory variable.
The decomposition of X into independent level-one and level-two parts
allows us to indicate precisely which parts of (7.7) correspond to the
unconditional level-one variance of Y, and which parts to the
unconditional level-two variance

This shows how the within-group variation of the level-one variables eats
up some part of the unconditional level-one variance; parts of the level-two
variance are eaten up by the variation of the level-two variables, and also by
the between-group (composition) variation of the level-one variables.
Recall, however, the definition of which implies that the between-group
variation of X is taken net of the ‘random’ variation of the group mean,
which may be expected given the within-group variation of Xij.

7.2.2 Random slope models

For the hierarchical linear model in its general specification given by (5.12),
a decomposition of the variance is very complicated because of the
presence of the cross-level interactions. Therefore the decomposition of the
variance is discussed for random slope models in the formulation (5.15),
repeated here as

without bothering about whether some of the xhij are level-one or level-two
variables, or products of a level-one and a level-two variable.
Recall that in this section the explanatory variables X are stochastic. The
vector X = (X1,…, Xq) of all explanatory variables has mean and
covariance matrix . The subvector (X1,…, XP) of variables that have
random slopes has mean µX(p) and covariance matrix ΣX(p). These
covariance matrices could be split into within-group and between- group
parts, but that is left to the reader.
The covariance matrix of the random slopes (U1j,..., Upj) is denoted by
T11 and the p × 1 vector of the intercept–slope covariances is denoted by
T10.
With these specifications, the variance of the dependent variable can be
shown to be given by

(A similar expression, but without taking the fixed effects into account, is
given by Snijders and Bosker (1993) as formula (21).) A brief discussion of
all terms in this expression is as follows.

1. The first term, gives the contribution of the fixed effects and may
be regarded as the ‘explained part’ of the variance. This term could be split
into a level-one and a level-two part as in the preceding subsection.

2. The next few terms,

should be seen as one piece. One could rescale all variables with random
slopes to have a zero mean (cf. the discussion in Section 5.1.2); this would
lead to and leave of this piece only the intercept variance . In
other words, (7.10) is just the intercept variance after subtracting the mean
from all variables with random slopes.

3. The penultimate term, trace (T11 Σ x(p)), is the contribution of the random
slopes to the variance of Y. In the extreme case where all variables X1,…, Xq
would be uncorrelated and have unit variances, this expression reduces to
the sum of squared random slope variances. This term also could be split
into a level-one and a level-two part.

4. Finally, is the residual level-one variability that can neither be

explained on the basis of the fixed effects, nor on the basis of the latent
group characteristics that are represented by the random intercept and
slopes.

7.3 Glommary
Estimated variance parameters. These may go up when variables are
added to the hierarchical linear model. This seems strange but it is a
known property of the hierarchical linear model, indicating that one
should be careful with the interpretation of the fine details of how the
variance in the dependent variable is partitioned across the levels of
the nesting structure.

Explained variance. Denoted by , this was defined as the proportional

reduction in mean squared error for predicting the dependent variable,
due to the knowledge of the values of the explanatory variables. This
was elaborated for two-level and three-level random intercept models.

Explained variance in random slope models. In this approach, this is so

little different from in random intercept models that it is better for
 this definition of explained variance to pay no attention to the
distinction between random slope and random intercept models.

Components of the variance of the dependent variable. These show how

not only the fixed effects of explanatory variables, reflected in , but
also the random effects contribute to the variance of the dependent
variable.

1
This is a more advanced section which may be skipped by the reader.
2
The word ‘large’ is meant here in the sense of the ordering of positive definite symmetric matrices.
 8
Heteroscedasticity

The hierarchical linear model is quite a flexible model, and it has some
other features in addition to the possibility of representing a nested data
structure. One of these features is the possibility of representing multilevel
as well as single-level regression models where the residual variance is not
constant.
In ordinary least squares regression analysis, a standard assumption is
homoscedasticity: residual variance is constant, that is, it does not depend
on the explanatory variables. This assumption was made in the preceding
chapters, for example, for the residual variance at level one and for the
intercept variance at level two. The techniques used in the hierarchical
linear model allow us to relax this assumption and replace it with the
weaker assumption that variances depend linearly or quadratically on
explanatory variables. This opens up an important special case of
heteroscedastic models, that is, models with heterogeneous variances:
heteroscedasticity where the variance depends on given explanatory
variables. More and more programs implementing the hierarchical linear
model also allow this feature. This chapter treats a two-level model, but the
techniques treated (and the software mentioned) can be used also for
heteroscedastic single-level regression models.

OVERVIEW OF THE CHAPTER

The chapter mainly treats heteroscedasticity at level one. First, the case is
treated of the variance depending linearly on an explanatory variable. Next,
the case of a quadratic dependence is treated, which also explains the
representation of heteroscedasticity (linear as well as quadratic) as a direct
extension of the hierarchical linear model of Chapter 5. Finally, a brief
treatment is given of heteroscedasticity at level two; this follows the same
principles as level-one heteroscedasticity.

8.1 Heteroscedasticity at level one

8.1.1 Linear variance functions
In a hierarchical linear model it sometimes makes sense to consider the
possibility that the residual variance at level one depends on one of the
predictor variables. An example is a situation where two measurement
instruments have been used, each with a different precision, resulting in two
different values for the measurement error variance which is a component
of the level-one variance. A linear dependence of the level-one residual
variance on some variable X1 can be expressed by

where the value of X1 for a given unit is denoted by x1ij while the random
part at level one now has two parameters, and . The reason for
incorporating the factor 2 and calling the parameter σ01 will become clear
later, when also quadratic variance functions are considered. For example,
when X1 is a dummy variable with values 0 and 1, the residual variance is
for the units with X1 = 0 and for the units with X1 = 1. When
the level-one variance depends on more than variable, their effects can be
added to the variance function (8.1) by adding terms etc.
Example 8.1 Residual variance depending on gender.
In the example used in Chapters 4 and 5, the residual variance might depend on the pupil’s gender. To
investigate this in a model that is not overly complicated, we take the model of Table 5.4 and add the
effect of gender (a dummy variable which is 0 for boys and 1 for girls). Table 8.1 presents estimates
for two models: one with constant residual variances, and one with residual variances depending on
gender.
Thus, Model 1 is a homoscedastic model and Model 2 a gender-dependent heteroscedastic model.
The girls do much better on the language test: for the fixed effect of gender, Model 1 has t =
2.407/0.201 = 12.0, p < 0.0001. According to formula (8.1), the residual variance in Model 2 is 37.85
for boys and 37.85 − 2 × 1.89 = 34.07 for girls. The residual variance estimated in the homoscedastic
Model 1 is very close to the average of these two figures. This is natural, since about half of the
pupils are girls and half are boys. The difference between the two variances is significant: the
deviance test yields χ2 = 24,486.8 − 24,482.2 = 4.6, df = 1, p < 0.05. However, the difference is not
large. The conclusion is that, controlling for IQ, SES, the school means of these variables, and their
interactions, girls score on average 2.4 higher than boys, and the results for girls are slightly less
variable than for boys.

Analogous to the dependence due to the multilevel nesting structure as

discussed in Chapter 2, heteroscedasticity has two faces: it can be a
nuisance and it can be interesting. It can be a nuisance because the failure to
take it into account may lead to a misspecified model and, hence, incorrect
parameter estimates and standard errors. On the other hand, it can also be an
interesting phenomenon in itself. When high values on some variable X1 are
associated with a higher residual variance, this means that for the units who
score high on X1 there is, within the context of the model being considered,
more uncertainty about their value on the dependent variable Y. Thus, it
may be interesting to look for explanatory variables that differentiate
especially between units who score high on X1. Sometimes a nonlinear
function of X1, or an interaction involving X1, could play such a role.
Example 8.2 Heteroscedasticity related to IQ.
Continuing the previous example, it is now investigated whether residual variance depends on IQ.
The corresponding parameter estimates are presented as Model 3 in Table 8.2.
Comparing the deviance to Model 1 shows that there is a quite significant heteroscedasticity
associated with IQ: χ2 = 24,486.8 − 24,430.2 = 56.6, df = 1, p < 0.0001. The level-one variance
function is (cf. (8.1))

Table 8.1: Estimates for homoscedastic and heteroscedastic models.

This shows that language scores of the less intelligent pupils are more
variable than language scores of the more intelligent. The standard
deviation of IQ is 2.04 and the mean in this data set is 0.04. Thus, the range
of the level-one variance, when IQ is in the range of the mean ± twice the
standard deviation, is between 36.38 − 3.38 × (0.04 + 2 × 2.04) = 22.45 and
36.38 − 3.38 × (0.04 − 2 × 2.04) = 50.04. This is an appreciable variation
around the average value of 36.00 estimated in the homoscedastic Model 1.
Prompted by the IQ-dependent heteroscedasticity, the data were
explored for effects that might differentiate between the pupils with lower
IQ scores. Nonlinear effects of IQ and some interactions involving IQ were
tried. It appeared that a nonlinear effect of IQ is discernible, represented
better by a so-called spline function1 than by a polynomial function of IQ.
Specifically, the coefficient of the square of IQ turned out to be different for
negative than for positive IQ values (recall that IQ was standardized to have
an average of 0, which turned out to change to 0.04 after dropping a few
cases with missing values). This is represented in Model 4 of Table 8.2 by
introducing two nonlinear transformations of which is the square but
only for negative values and 0 elsewhere; and , being the square only for
positive values and 0 elsewhere. The cutoff point of 0 was tried because it is
a natural value, corresponding closely to the average; some other values
also were tried

Table 8.2: Heteroscedastic models depending on IQ.

and did not yield clearly better results. Thus, we use the definitions
Adding these two variables to the fixed part gives a quite significant
decrease in the deviance (24,430.2 − 24,369.0 = 61.2 with two degrees of
freedom) and strongly reduces the random slope of IQ. Comparing the
deviance with the model in which the random slope at level two of IQ is left
out shows, however, that this random slope still is significant. The total
fixed effect of IQ is now given by

The graph of this effect is shown in Figure 8.1. It is an increasing function

which flattens out for low and especially for high values of IQ in a way that
cannot be well represented by a quadratic or cubic function, but for which a
quadratic spline is a good representation. Perhaps this can be interpreted as
a bottom and ceiling effect of the test used.

Figure 8.1: Effect of IQ on language test.

This is an interesting turn of this modeling exercise, and a nuisance only

because it indicates once again that explanation of school performance is a
quite complicated subject. Our interpretation of this data set toward the end
of Chapter 5 was that there is a random slope of IQ (i.e. schools differ in the
effect of IQ), with a variance estimated at 0.164 (a standard deviation of
0.40). Now it turns out that the data are clearly better described by a model
in which IQ has a nonlinear effect, stronger in the middle range than toward
its extreme values; in which the effect of IQ varies across schools, but with
a variance of 0.044, corresponding to a standard deviation of 0.21, much
smaller than the earlier value; and in which the combined effects of IQ,
SES, their school averages and interactions, and gender predict the
language scores for high-IQ pupils much better than for low-IQ pupils.
Thus, a considerable portion of what seemed to be between-school
differences in the effect of IQ on the language test now turns out to be
partly the consequence of nonlinear effects of IQ and heteroscedasticity.

8.1.2 Quadratic variance functions

The formal representation of level-one heteroscedasticity is based on
including random effects at level one, as spelled out in Goldstein (2011) –
see the remarks in Section 3.1 of Goldstein’s book on a complex random
part at level one. In Section 5.1.1 it was remarked that random slopes (i.e.,
random effects at level two) lead to heteroscedasticity. This also holds for
random effects at level one. Consider a two-level model and suppose that
the level-one random part is

Denote the variances of R0ij and R1ij by and , respectively, and their
covariance by . The rules for calculating with variances and covariances
imply that

Formula (8.4) is just a formal representation, useful for specifying this

heteroscedastic model. For the interpretation one should look not at (8.4)
with its would-be random effect, but rather at only the right-hand side of
(8.5). This formula can be used without the interpretation that and
are variances and σ01 a covariance; these parameters might be any numbers.
The formula only implies that the residual variance is a quadratic function
of x1ij. In the previous section, the case was used where producing
the linear function (8.1). If a quadratic function is desired, all three
parameters are estimated from the data.
Example 8.3 Educational level attained by pupils.
This example is about a cohort of pupils entering secondary school in 1989, studied by Dekkers et al.
(2000). The question is how well the educational level attained in 1995 can be predicted from
individual characteristics and school achievement at the end of primary school. The data is about
13,007 pupils in 369 secondary schools. The dependent variable is an educational attainment
variable, defined as 12 minus the minimum number of additional years of schooling it would take
theoretically for this pupil in 1995 to gain a certificate giving access to university. The range is from
4 to 13 (a value of 13 means that the pupil is already a first-year university student). Explanatory
variables are teacher’s rating at the end of primary school (indicating the most suitable type of
secondary school, range from 1 to 4), achievement on three standardized tests (so-called CITO tests,
on language, arithmetic, and information processing – all with a mean value of about 11 and a
standard deviation between 4 and 5), socio-economic status (a discrete ordered scale with values
from 1 to 6), gender (0 for boys, 1 for girls), and minority status (based on the parents’ country of
birth, 0 for the Netherlands and other industrialized countries, 1 for other countries).
Table 8.3 presents the results of a random intercept model (Model 1). Standard errors are quite
small due to the large number of pupils in this data set. The explained proportion of variance at level
one is R12 = 0.39. Model 2 shows the results for a model where residual variance depends
quadratically on SES. It follows from (8.5) that residual variance here is given by

The difference in deviance between Models 1 and 2 (χ2 = 97.3, df = 2, p < 0.0001) indicates that the
dependence of residual variance on SES is quite significant. The variance function decreases
curvilinearly from the value 2.91 for SES = 1 to a minimum value of 1.75 for SES = 6. This implies
that when educational attainment is predicted by the variables in this model, the uncertainty in the
prediction is highest for low-SES pupils. It is reassuring that the estimates and standard errors of
other effects are not appreciably different between Models 1 and 2.
The specification of this random part was checked in the following way. First, the models with
only a linear or only a quadratic variance term were estimated separately. This showed that both
variance terms are significant. Further, it might be possible that the dependence of the residual
variance on SES is a random slope in disguise. Therefore a model with a random slope (a random
effect at level two) for SES also was fitted. This showed that the random slope was barely significant
and not very large, and did not take away the heteroscedasticity effect.
The SES-dependent heteroscedasticity led to the consideration of nonlinear effects of SES and
interaction effects involving SES. Since the SES variable assumes values 1 to 6, five dummy
variables were used contrasting the respective SES values to the reference category SES = 3. In order
to limit the number of variables, the interactions of SES were defined as interactions with the
numerical SES variable rather than with the categorical variable represented by these dummies. For
the same reason, for the interaction of SES with the CITO tests only the average of the three CITO
tests (range from 1 to 20, mean 11.7) was considered. Product interactions of SES with gender, with
the average CITO score, and with minority status were considered. As factors in the product, SES
and CITO tests were centered approximately by using (SES − 3) and (CITO average − 12). Gender
and minority status, being 0–1 variables, did not need to be centered.
This implies that although SES is represented by dummies (i.e., as a categorical variable) in the
main effect, it is used as a numerical variable in the interaction effects. (The main effect of SES as

Table 8.3: Heteroscedastic model depending quadratically on SES.

a numerical variable is implicitly also included, because it can be represented also as an effect of the
dummy variables. Therefore the numerical SES variable does not need to be added to the model.)
Minority status does not have a significant main effect when added to Model 2, but the main
effect was added to facilitate interpretation of the interaction effect. The results are in Table 8.4.
Model 3 as a whole is a strong improvement over Model 2: the difference in deviance is χ2 = 44.8
for df = 7 (p < 0.0001). For the evaluation of the nonlinear effect of SES, note that since SES = 3 is
the reference category, the parameter value for SES = 3 should be taken as 0.0. This demonstrates
that the effect of SES is nonlinear but it is indeed an increasing function of SES. The differences
between the SES values 1, 2, and 3 are larger than those between the values 3, 4, 5, and 6. The
interaction effects of SES with gender and with minority status are significant. The main effect of
minority status corresponds with its effect for SES = 3, since the product variable was defined using
SES − 3, and is practically nil. Thus, it turns out that, when the other included variables (including
the main effect of SES) are controlled for, pupils with parents from a nonindustrialized country attain
a higher educational level than those with parents from the Netherlands or another industrialized
country when they come from a low-SES family, but a lower level if they come from a high-SES
family.
In Model 3, residual variance is

This decreases from 2.87 for SES = 1 to 1.71 for SES = 6. Thus, with the inclusion of interactions
and a nonlinear SES effect, residual variance has hardly decreased.

Table 8.4: Heteroscedastic model with interaction effects.

 Model 3 was obtained, in the search for heteroscedasticity, in a data-driven rather than theory-
driven way. Therefore one may question the validity of the tests for the newly included effects: are
these not the result of capitalization on chance? The interaction effect of SES with minority status has
such a high t-value, 4.4, that its significance is beyond doubt, even when the data-driven selection is
taken into account. For the interaction of SES with gender this is less clear. For convincing
hypothesis tests, it would have been preferable to use cross-validation: split the data into two subsets
and use one subset for the model selection and the other for the test of the effects. Since the two
subsets should be independent, it would be best to select half the schools at random and use all pupils
in these schools for one subset, and the other schools and their pupils for the other.

More generally, the residual variance may depend on more than one
variable; in terms of representation (8.4), several variables may have
random effects at level one. These can be level-two as well as level-one
variables. If the random part at level one is given by
while the variances and covariances of the Rhij are denoted by and ,
then the variance function is

This complex level-one variance function can be used for any values for
the parameters and , provided that the residual variance is positive.
The simplest case is to include only and the ‘covariance’ parameters
leading to the linear variance function

Correlates of diversity
It may be important to investigate the factors that are associated with
outcome variability. For example, Raudenbush and Bryk (1987) (see also
Raudenbush and Bryk, 2002, pp. 131–134) investigated the effects of
school policy and organization on mathematics achievement of pupils. They
did this by considering within-school dispersion as a dependent variable.
The preceding section offers an alternative approach which remains closer
to the hierarchical linear model. In this approach, relevant level-two
variables are considered as potentially being associated with level-one
heteroscedasticity.
Example 8.4 School composition and outcome variability.
Continuing the preceding example on educational attainment predicted from data available at the end
of primary school, it is now investigated whether composition of the school with respect to socio-
economic status is associated with diversity in later educational attainment. It turns out that socio-
economic status has an intraclass correlation of 0.25, which is quite high. Therefore the average
socio-economic status of schools could be an important factor in the within-school processes
associated with average outcomes but also with outcome diversity.
To investigate this, the school average of SES was added to Model 3 of Table 8.4 both as a fixed
effect and as a linear effect on the level-one variance. The nonsignificant SES-by-CITO interaction
was deleted from the model. The school average of SES ranges from 1.4 to 5.5, with mean 3.7 and
standard deviation 0.59. This variable is denoted by SA-SES. Its fixed effect is 0.417 (S.E. 0.109, t =
3.8). We further present only the random part of the resulting model in Table 8.5.
To test the effect of SA-SES on the level-one variance, the model was estimated also without this
effect. This yielded a deviance of 46,029.6, so the test statistic is χ2 = 46,029.6 – 45,893.0 = 136.6
with df = 1, which is very significant. The quadratic effect of SA-SES was estimated both as a main
effect and for level-one heteroscedasticity, but neither was significant.
 How important is the effect of SA-SES on the level-one variance? The standard deviation of SA-
SES is 0.59, so four times the standard deviation (the difference between the few percent highest-SA-
SES and the few percent lowest-SA-SES schools) leads to a difference in the residual variance of 4 ×
0.59 × 0.265 = 0.63. For an average residual level-one variance of 2.0 (see Model 1 in Table 8.3), this
is an appreciable difference.
This ‘random effect at level one’ of SA-SES might be explained by interactions between SA-SES
and pupil-level variables. The interactions of SA-SES with gender and with minority status were
considered. Adding these to Model 4 yielded interaction effects of −0.219 (S.E. 0.096, t = −2.28) for
SA-SES by minority status and −0.225 (S.E. 0.050, t = 4.50) for SA-SES by gender. This implies
that, although a high school average for SES leads to higher educational attainment on average (the
main effect of 0.417 reported above), this effect is weaker for minority pupils and for girls. These
interactions did, however, not lead to a noticeably lower effect of SA-SES on the residual level-one
variability.

Table 8.5: Heteroscedastic model depending on average SES.

8.2 Heteroscedasticity at level two

For the intercept variance and the random slope variance in the hierarchical
linear model it was assumed in preceding chapters that they are constant
across groups. This is a homoscedasticity assumption at level two. If there
are theoretical or empirical reasons to drop this assumption, it could be
replaced by the weaker assumption that these variances depend on some
level-two variable Z. For example, if Z is a dummy variable with values 0
and 1, distinguishing two types of groups, the assumption would be that the
intercept and slope variances depend on the group. In this section we only
discuss the case of level-two variances depending on a single variable Z;
this discussion can be extended to variances depending on more than one
level-two variable.
 Consider a random intercept model in which one assumes that the
intercept variance depends linearly or quadratically on a variable Z. The
intercept variance can then be expressed by

for parameters , , and . For example, in organizational research, when

the level-two units are organizations, it is possible that small-sized
organizations are (because of greater specialization or other factors) more
different from one another than large-sized organizations. Then Z could be
some measure for the size of the organization, and (8.7) would indicate that
it depends on Z whether differences between organizations tend to be small
or large; where ‘differences’ refer to the intercepts in the multilevel model.
Expression (8.7) is a quadratic function of zj, so that it can represent a
curvilinear dependence of the intercept variance on Z. If a linear function is
used the intercept variance is either increasing or decreasing over
the whole range of Z.
Analogous to (8.4), this variance function can be obtained by using the
‘random part’

Thus, strange as it may sound, the level-two variable Z formally gets a

random slope at level two. (Recall that in Section 5.1.1 it was observed that
random slopes for level-one variables also create some kind of
heteroscedasticity, namely, heteroscedasticity of the observations Y.)
The parameters , , and are, as in the preceding section, not to be
interpreted themselves as variances and a corresponding covariance. The
interpretation is by means of the variance function (8.8). Therefore it is not
required that To put it another way, ‘correlations’ defined
formally by may be larger than 1 or smaller than − 1, even infinite,
because the idea of a correlation does not make sense here. An example of
this is provided by the linear variance function for which and only
the parameters and are used.
In a similar way, the random slope variance of a level-one variable X1
can be made to depend on a level-two variable Z by giving both X1 and the
product X1 × Z a random slope at level two. To avoid computational
difficulties it is advisable to use a variable Z which has been centered, even
if only very approximately.

8.3 Glommary
Heteroscedasticity. Residual variance depending on explanatory variables.
The hierarchical linear model can be extended so that residual
variances depend, linearly or quadratically, on explanatory variables.
This can be a nuisance but also interesting.

Heteroscedasticity at level one. This is achieved by formally giving a

variable a random effect at level one. This should not be regarded as a
‘real’ random effect but merely as a device to achieve a linear or
quadratic variance function.

Correlates of diversity. Variables that are predictive of residual variances.

This term is used in studies where the interest is in explaining
variances rather than explaining means. This is a nice interpretation or
theoretical framing of heteroscedastic models. When heteroscedasticity
is found, this can lead to searching for variables that, when included as
fixed effects in the model, might explain this heteroscedasticity – for
example, nonlinear transformations of explanatory variables, or
interactions.

Random slopes or heteroscedasticity? As a by-product of our example of

heteroscedasticity at level one, we found that what in Chapter 5
appeared to be a random slope for IQ, could be better described in part
as a consequence of a nonlinear effect of IQ in the fixed part (modeled
by a spline function).

Heteroscedasticity at level two. This can be achieved by formally giving a

variable a random effect at level two. Again, this is not a ‘real’ random
effect but only a device to obtain the heteroscedasticity.

1
Spline functions (introduced more extensively in Section 15.2.2 and treated more fully, for
example, in Fox, 2008, Chapter 17) are a more flexible class of functions than polynomials. They are
polynomials of which the coefficients may be different on several intervals.
 9
Missing Data

In scientific practice, available data often are incomplete – because research

subjects did not participate fully, registers were not complete, an apparatus
failed, or for whatever reason. For a long time, it was not uncommon for
researchers to ignore the associated difficulties, and one usual approach was
working with complete cases only and leaving out incomplete cases: a
practice called listwise deletion or complete case analysis. Another usual
approach was available case analysis, where each parameter is estimated
using the data points directly relevant to this particular parameter; for
example, the expected value of some variable is estimated by the mean of
this variable over the cases where the variable was observed. It is now
recognized that these approaches can lead to incorrect conclusions, and
therefore it is no longer acceptable to use them unthinkingly as default
methods. Appropriate ways have been developed to deal with missing data,
and this chapter serves as an introduction to some of these methods and
their application to multilevel analysis.
There are two basic reasons for attempting to deal with incomplete data
in a suitable way. Most important, in many cases the absence of data is
related to meaningful information, and ignoring this can bias the results
obtained. Second, deleting incomplete cases is wasteful of information in
the sense that standard errors of estimators are unnecessarily large and
hypothesis tests have unnecessarily low power.
Classic texts on missing data analysis in general are Little and Rubin
(2002) and Schafer (1997). Relatively brief and accessible overviews are
Carpenter and Kenward (2008) and Graham (2009). A textbook on missing
data in longitudinal data sets analyzed by hierarchical linear models is
Molenberghs and Kenward (2007).
 OVERVIEW OF THE CHAPTER
The chapter starts by outlining the basis of the modern theory of inference
from incomplete data. This rests on the concept of ‘missingness at random’,
which expresses that the fact itself that some data points are missing does
not potentially reveal anything about the unobserved values that is not
already contained in the observed data. Advice is given about potentially
collecting auxiliary data that could make this assumption more reasonable.
How to deal with missing data in situations where this assumption is not
satisfied depends strongly on the particular case under study and the reasons
why data points are missing; this is outside the scope of this book.
Of the various techniques for analysis of incomplete data, we focus on
two that are suitable for multilevel analysis: maximum likelihood and
multiple imputation. We give most attention to the approach of multiple
imputation because it has great potential for general multilevel data
structures, but requires some explanation if one is to understand it properly.
Since this is a topic that for the special case of multilevel data is only
beginning to be implemented in statistical software, we go somewhat
deeper into the technical details of how to carry out this type of data
analysis than we do in most other chapters. In particular, we treat in some
detail the very flexible technique of multiple imputation by chained
equations. We treat incompleteness for general multilevel data structures,
without paying special attention to longitudinal data.

9.1 General issues for missing data

Suppose that we are working with an incomplete data set. To model such a
data set, we need to think about a model for the hypothetical complete data
set of which this is the currently observed part. Let us consider each
explanatory and each dependent variable for each case in the data as a data
point; in the terminology of Chapter 4, this is each variable yij, xpij, and zqj.
Then any data point could be either observed or missing. (We suppose that
we do have complete information about the nesting structure, that is, in a
two-level design, for all level-one units we know which level-two unit they
belong to. Cases where this kind of information is missing were discussed
by Hill and Goldstein, 1998). The hypothetical complete data set has
observed values for all data points, whereas the actual incomplete data set
lacks some of these data points. In the model for the incomplete data, this
can be indicated by missingness indicators defined for all data points,
which convey the information as to whether a given data point is missing or
observed, for example, by values 1 and 0, or true and false, reflecting
missing versus observed.
Here we must note a distinction that is usually glossed over. In
regression-type models, of which the hierarchical linear model is one
example, where there is a distinction between dependent and explanatory
(or independent) variables, the model is conditional on the explanatory
variables. Mostly the latter variables are treated as fixed – they are whatever
they are, and they might as well have been determined by the researcher, as
they are in purely experimental designs. If, however, we wish to make sense
of an incomplete data set in which some of the data points for explanatory
variables xpij and zqj are missing, then we must entertain a model for the
complete data set in which the explanatory variables are random, rather
than fixed, variables, and in which they have some simultaneous probability
distribution. This probability distribution, when it has been estimated, will
allow us to say that some hypothetical values for the unobserved variables
are more likely than others, and this kind of knowledge will be used in the
analysis of the incomplete data set, for example, by attempting to fill in
(‘impute’) the unobserved data points. To summarize, in the usual approach
the explanatory variables are treated as if they have fixed, deterministic
values; but in the analysis of an incomplete data set the variables that have
some missing values are treated as random variables. If there are some
variables that have no missing cases at all, these may be regarded as
deterministic values.
The extra step taken in modern treatments of missing data is to regard
the missingness indicators themselves also as random variables that are part
of the data. This alerts us to the possibility that they may contain
meaningful information; for example, if some pupils do not show up for an
exam because they expect to fail, the missingness is predictive of what
would be a poor exam mark, if it had been observed. It is sometimes
meaningful to regard the missingness variables as part of the substantively
interesting social process that is being studied.
Rubin (1976) defined three types of missingness:
 missingness completely at random (MCAR), where the missingness
indicators are independent of the complete data. An example is
randomly failing measurement apparatus.

missingness at random (MAR), where, conditionally given the

observed data, the missingness indicators are independent of the
unobserved data.1 This means that the missingness itself contains no
information about the unobserved values that is not also contained in
the observed data.2 An example is repeated measures of ill patients
where initially all patients are being observed and observation stops at
recovery, while recovery is recorded in the data.

missingness not at random (MNAR), where, conditionally given the

observed data, the missingness indicators are dependent on the
unobserved data. An example is repeated measures of ill patients
where some patients do not show up for data collection if they feel
very well or very ill, and the reason for absence is unknown.

The first situation, MCAR, is the pleasant situation in which we can almost
forget about the incompleteness of the data, and throwing away the
incomplete cases would entail only a loss of statistical efficiency (and
therefore, loss of power) but no bias. The second, MAR, is the potentially
difficult situation where naive approaches such as complete case analysis
can be biased and hence misleading, but where it is possible to make
judicious use of the observations and avoid bias. For the MCAR as well as
the MAR case it may be necessary to use special statistical methods to
avoid efficiency loss. By the way, the terminology is potentially misleading
because many people might interpret the term ‘missingness at random’ as
meaning what was defined above as ‘missingness completely at random’,
but these are the generally employed terms so we have to use them.
The third situation, MNAR, leads to serious difficulties for the
researcher because the missingness pattern contains information about the
unobserved values, but in practice we do not know which information. We
wish to say something about the world but we have observed too little of it.
To draw conclusions from the data it is necessary in this case to make
additional assumptions about how missingness is related to the values of the
data points that would have been observed if they had not been missing.
While the researcher must attempt to make plausible assumptions of this
kind, they will not be completely testable. In other words, to analyze data in
the MNAR case we need to make assumptions going beyond the available
data and beyond the model that we would use for analyzing complete data.
To still say something in this case about the uncertainty of the conclusions
drawn from the data, it will be helpful to carry out sensitivity analyses
studying the sensitivity of the conclusions to the assumptions made. In the
MNAR situation, any data analysis will leave bigger questions than in the
MAR situation.

9.1.1 Implications for design

Even though this may seem like forcing an open door, we must emphasize
that it is important in the phase of design and data collection to attempt to
collect complete data, and if this is impossible, to minimize the amount of
missing data. Second, if missing data are unavoidable, it will be helpful to
record information about what led to the missingness (e.g., in the case of
data about employees in an organization, if data for an employee are totally
absent, whether it was because of illness, or refusal to cooperate, or
negligence, or loss of the questionnaire, etc.) and about auxiliary variables
that may be predictive of missingness and that can be also collected in case
of missingness (continuing the preceding example, some of the reasons for
nonresponse will be related with the general cooperative attitude of the
employee, and the research design could include a question to supervisors
to rate the cooperative attitude for each of their subordinates). It is also
sometimes possible to observe an auxiliary variable that may serve as a
rough proxy for the missing variable (which then should be regarded as a
separate variable and collected, if possible, for all cases). In most social
research, the process leading to incompleteness of collected data is itself a
social process, and the research will gain by trying to understand it.
The analysis of an incomplete data set should start with a description of
the number of missing data points, and how these are distributed across
variables and across higher-level units. To understand the amount and
patterns of missingness, it may be helpful to employ methods such as pilot
studies, debriefing of respondents and those carrying out the data collection,
and reflection on the data collection process. An explicit model of the
missingness pattern of a given variable may be constructed by a multilevel
logistic regression analysis of the binary missingness indicators as
dependent variables (see Chapter 17). Thinking about such a model can
elucidate the difference between MAR and MNAR: ifMNAR holds then
hypothetically knowing the unobserved data would help in predicting the
missingness indicators, whereas it would not help under the MAR
condition.
Most methods for handling incomplete data are based on the MAR
assumption. This is because, on the one hand, MCAR is rather rare in the
practice of social research. On the other hand, although MNAR may often
be more realistic than MAR, it is much more difficult to handle than MAR;
and, as suggested above, by collecting auxiliary information that is
predictive of missingness the researcher may be able to push the research
design from MNAR in the direction of MAR. Also, even in the MNAR
case, a principled analysis employing the MAR assumption will mostly lead
to more credible results than a naive way of handling the missing data.

9.2 Missing values of the dependent variable

For missingness of the dependent variable, MAR means in practice (under
an additional assumption of mutual independence of missingness
indicators) that the missingness of the dependent variable may depend on
observed explanatory variables for this particular level-one unit, but not on
unobserved ones that are correlated with the dependent variable, nor on the
unobserved value of the dependent variable itself. For example, if
particularly high or low values of the dependent variable were more likely
to be missing but this is not totally predictable based on the observed
explanatory variables, then MNAR would hold. In this section it further is
assumed that MAR holds.
Suppose that there are only missing data for the dependent variable, all
other variables are fully observed, and the explanatory variables are
considered as deterministic variables (so that we are not interested in
estimating the distribution of the explanatory variables). If all observed
variables are potentially included among the explanatory variables – this
will depend on the research question – the situation is relatively
uncomplicated. Since the hierarchical linear model, like other regression-
type techniques, is conditional on the values for the explanatory variables,
such units can safely be dropped from the data set for the basic analysis
according to the hierarchical linear model (cf. Little and Rubin, 2002, p.
237).
However, the research question may be such that some observed
variables Xextra are excluded from the model that is currently being
estimated. This may be because of causality arguments, or because of the
particular focus of the current research question, as will be discussed on
p.136. The variables Xextra might also include the auxiliary variables
collected to explain the missingness as suggested in Section 9.1.1. Then it is
possible that MAR holds when Xextra is included, but MNAR would hold if
it were excluded from consideration. It also is possible that Xextra has a
strong predictive value for the dependent variable. In such cases the
variable Xextra should be included in the procedures for handling the missing
data, even if it is not to be used further among the explanatory variables for
the model that is currently being estimated.
For multivariate hierarchical linear models, that is, models with multiple
dependent variables (Chapter 16), the situation is more complicated,
because the patterns of missingness may well differ between the various
dependent variables. The treatment of incomplete data for such models,
with missing values on the dependent variables, was discussed by Schafer
and Yucel (2002).

9.3 Full maximum likelihood

An ideal way to proceed under MAR is to formulate the likelihood of the
observed data and find the corresponding maximum likelihood (ML)
estimates. This is sometimes referred to as ‘full information maximum
likelihood’ (FIML).
When there are missing values only on the dependent variable and
missingness depends only on the explanatory variables, then the
missingness mechanism is MAR. The preceding section argued that then
one may simply compute the ML estimates of the regular hierarchical linear
model in which the level-one units with a missing dependent variable are
dropped. These are then the ML estimates for the observed data. This was
also proposed for incomplete multivariate data, or incomplete longitudinal
data with a fixed occasion design, on p. 257. One of the advantages of the
algorithms available for the hierarchical linear model is that they easily
allow FIML estimation for incomplete longitudinal or multivariate data.
If there are missing values also in explanatory variables and the missing
data pattern is monotone, then the full maximum likelihood approach is
feasible. ‘Monotone’ means that the variables can be ordered such that the
cases with missing values on a given variable are a subset of the cases with
missing values on the next variable. For nonmonotone missingness patterns
the full maximum likelihood approach is more cumbersome. This is also the
case when the missingness in the dependent variable depends on auxiliary
variables Xextra (as described above) that are observed but that the
researcher does not wish to use among the explanatory variables. There are
methods for parameter estimation in such situations by FIML (see
Molenberghs and Kenward, 2008; Ibrahim and Molenberghs, 2009), but the
multiple imputation method of Section 9.4 is more flexible and we restrict
ourselves to presenting the latter method. Further discussion on the choice
between FIML and multiple imputation is given by Graham (2009, p. 560).

9.4 Imputation
Imputation means filling in something (hopefully reasonable) for the
missing data points, which then leads to a completed data set so that a
regular complete-data analysis is possible. The general literature on missing
data (e.g., Little and Rubin, 2002; Schafer and Graham, 2002) provides
arguments and examples explaining why simple methods such as
imputation by an overall mean, and also slightly more complicated methods
such as imputation by a groupwise mean, will yield biased results and
underestimation of standard errors. Rubin (1987) had the insight that
multiple stochastic imputation is an approach which leads to good
(approximately unbiased) results under MAR. This approach now is one of
the major methods employed in the analysis of incomplete data, and an
extensive discussion can be found, for example, in Graham (2009). Multiple
stochastic imputation works in three steps:

1. Construct several completed data sets, in each of which the missing

 values are imputed by a random draw from the distribution of missing
data, given the observed data. These random draws must be
independent across the multiple data sets.
2. Use a regular complete-data analysis method for each of these
multiple data sets, producing parameter estimates and standard errors
for the hypothetical situation where these are real observed data sets.
3. Combine the parameter estimates and standard errors across these
multiple analyses into a single set of results. The resulting overall
standard errors will combine the within-data-set standard errors
obtained in step 2 with the between-data-set variability of the
estimates.

Note that this procedure does not require any type of knowledge of the
model for the missingness variables beyond the validity of the MAR
assumption. The virtue of this approach is that step 2 is normally
straightforward, consisting, in our case, of the estimation of the hierarchical
linear model or one of its variants. The combination formulas of step 3 are
simple, as we shall see below. The catch is in step 1. This requires
knowledge of the distribution of the missing data given the observed data,
but this will depend on the parameters that we are trying to estimate in the
whole process3 – the dog seems to bite its own tail. The approach is still
feasible because it is possible to let the dog bite its own tail ad infinitum,
and cycle through the steps repeatedly until stability of the results is
achieved. Especially in the Bayesian paradigm this can be formulated in a
very elegant way as a kind of ‘data augmentation’ (Tanner and Wong,
1987), as we shall see below.
The model used for the imputations in step 1 is the assumed model for
the joint distribution of all observations. Let us assume that the model in
which we are interested, often referred to appropriately as the model of
interest, is a model for a dependent variable Y (e.g., a hierarchical linear
model) with an array X of explanatory variables. If the combination of X
and Y is the totality of available data, then the model of step 2 is the
conditional distribution of Y given the rest of the data according to the joint
distribution used in step 1. Then the imputation model and the model of
interest are in correspondence. This is, however, not necessary.
 Often the research question is such that some observed variables Xextra
are excluded from the model that is currently being estimated. This may be
because of causality arguments – for example, because the variable Xextra
was measured at a later or earlier moment than the other variables, or
because Xextra occupies the theoretical place of a mediating variable which
is excluded from the currently estimated model. Or a large data set may
have been collected, and the dependence of Y on Xextra is thought to be
weak or theoretically irrelevant, and is therefore not investigated. Also the
considerations of Section 9.1.1 can lead to observing variables that are
relevant for the missingness mechanism but not for the model of interest.
Then missingness might be at random (MAR) in the joint distribution of all
the variables (X, Xextra, Y), whereas missingness would not be at random
(MNAR) if only the observations of (X, Y) would be considered. In all these
cases the imputation of missing values should depend not only on the
variables in the model of interest, but also on the variables in Xextra. Note
that this makes the imputation models potentially more complex than the
model of interest. To summarize, step 1 will impute by drawing from the
conditional distribution of the missing variables given the observations in
all of (X, Xextra, Y), while step 2 analyzes the dependence of Y only on X.
Of the three steps, we shall now elaborate the first and third. The second
step consists of the estimation of parameters of the hierarchical linear model
for all randomly imputed data sets, and how to do this should be clear to
anyone who has read the relevant chapters of this book. Note that each of
the imputed data sets itself is a complete data set so that it can be analyzed
by the multilevel methods for complete data.
Several computer packages for multilevel analysis have recently added
options for multiple stochastic imputation, or are in the process of doing so.
For example, MLwiN has a macro for relatively simple imputations, and an
associated program REALCOM-impute for imputation of missing data in
more complex patterns; and the mice package in R can impute missing
values in two-level models. See Chapter 18 for more information about
these packages.

Plausible values
Multiple data sets constructed from multiple stochastic imputation are
sometimes called plausible values. These are used particularly when the
missing values are missing by design. This means that the data collection
was such that some variables were missing on purpose for some parts of the
sample, or for some waves in the case of panel surveys. For example, if
there are four related questions that are too much of a burden on the
respondent, the sample could be divided randomly into six groups, each of
which gets the same two out of these four questions. The missingness for
such variables is MCAR, which facilitates the construction of the
imputation model.

9.4.1 The imputation method

Let us now consider step 1, the imputation of the missing data, that is, the
process of replacing missing values by likely values. If there are missing
values in several variables, this will require the conditional distribution of
all these variables. The best way to obtain random draws from the
conditional distribution of the missing values given the observed data will
depend on the model under consideration. There is one flexible approach,
however, which can often be followed. This can be described as iteratively
drawing from all the conditional distributions of each variable given the
rest, and continuing this process until a stochastic kind of convergence is
obtained. It is a Bayesian approach (see Section 12.1), which means that the
parameters are themselves also regarded as random variables, and their
distribution reflects the uncertainty about their true values.
Denote all variables jointly by X1,. . ., Xp, Y. This array should also
include the variables that were referred to above as Xextra. Denote by φ the
vector of all parameters of the joint distribution of the complete data (X1,. .
., Xp, Y).
1. Initialize the process by a random draw from an approximation, perhaps
quite rough, from the conditional distribution of the missing values given
the observed values. In many cases a reasonable approach is to
approximate the distribution of (X1,. . ., Xp, Y) by a multivariate normal
distribution, ignoring the multilevel structure; and draw the missing
values from this estimated normal distribution. This is straightforward
 because many statistical software programs nowadays contain good
procedures for estimating the parameters for the multivariate normal
distribution from incomplete data, for example, using the EM algorithm
(Little and Rubin, 2002); likewise, for normal distributions many
software packages contain methods for drawing from the conditional
distributions, which again are normal. If some of the Xk or Y variables
have clearly nonnormal distributions, then a transformation may be
applied to reduce skewness, and variables may be truncated or rounded to
the set of permitted values; since this is the initialization only, a rough
procedure is acceptable. This initialization yields a completed data set.
2. Let be a random draw from the conditional distribution of φ (the so-
called posterior distribution) given the completed data set.
3. Impute the missing values in Y by random draws from the conditional
distribution of the missing cases in Y given X1,. . ., Xp and given the
observed cases in Y, using for (X1,. . ., Xp) the completed data set and
parameter value .
4. Successively for k = 1,. . ., p, impute the missing values in Xk by random
draws from the conditional distribution of the missing cases in Xk given
X1,. . ., Xk−1, Xk+1,. . ., Xp, Y and given the observed cases in Xk, using for
(X1,. . ., Xk−1, Xk+1,. . ., Xp, Y) the completed data set and parameter value
.
5. Go to step 2 and repeat the process, until it is deemed that it has lasted
long enough.

Of course steps 3 and 4 need to be carried out only for the variables that
have missing values.
This procedure defines a stochastic iterative process. The general
principle was proposed by Tanner and Wong (1987) who called it data
augmentation, because the data set is augmented by the imputed missing
data. Cycling through the conditional distributions in this way is called
Gibbs sampling (e.g., Gelman et al., 2004), which again is a particular kind
of Markov chain Monte Carlo (MCMC) procedure (Section 12.1).
Mathematical results ensure that the joint probability distribution of (X1,. . .,
Xp, Y, ) converges to the joint distribution of the complete data and the
parameters given the observed data; recall that this procedure is defined in a
Bayesian framework in which the parameters are also random variables.
This has the advantage that the uncertainty about the parameter values will
be reflected in corresponding greater dispersion in the imputed values than
if one followed a frequentist procedure and fixed the parameter at an
estimated value.
The cycle consisting of steps 2–4 will be repeated a large number of
times. First a ‘burn-in’ of a number of cycles is needed to give confidence
that the process has converged and the distribution of generated values is
more or less stable. After this point, the generated imputations can be used.
Successively generated imputations will be highly correlated. Therefore a
sampling frequency K has to be chosen, and one in every K generated
imputations will be retained for further use. K will have to be large enough
that the imputed values separated by K cycles have a negligible correlation.
It may be difficult to assess when convergence has taken place, however
– in step 5 this was hidden behind the word ‘deemed’. The literature on
MCMC procedures contains material about convergence checks, and an
important sign is that the consecutive values generated for the components
of the parameter have achieved some kind of stability, as can be seen from
plotting them. More detailed diagnostics for convergence may be used in
more complicated models. Abayomi et al. (2008) propose to study the
convergence of the imputation procedure by considering the marginal
distributions and bivariate scatterplots of the imputed data sets, looking in
particular at the deviations between the observed and the imputed data
points. Evidently, it is quite acceptable that there are differences; when
missingness depends on observed variables, then differences between
distributions of observed and imputed data are indeed to be expected. But
one should check that the imputed data seem plausible and are in line with
presumed mechanisms of missingness.
The procedure of steps 1–5 was proposed and developed by Yucel
(2008) for normally distributed variables in two-level and three-level
models; and by Goldstein et al. (2009) for quite general two-level models,
including models for binary and categorical variables, where level-one as
well as level-two variables may have missing values. Such procedures are
implemented in REALCOM-impute (Goldstein, 2011) and the R package
mlmmm.
Example 9.1 Simulated example: missingness dependent on a third variable.
As a simple artificial example, consider the model

where X1ij and Rij are independent standard normal variables, and the random intercept U0j has
variance τ02 = 0.16. Let the group sizes be constant at n = 20 and let there be N = 200 groups; the
number of groups is sufficiently large that one simulated data set should give a good indication of
how parameters are being recovered. Suppose that there is a third variable X2ij given by

where the Eij likewise are independent standard normal variables; and let missingness be determined
according to the rule that Yij is observed only for X2ij ≤ 2. We simulated such a data set, and 834 out
of the 4,000 values of Yij turned out to be missing. Suppose the model of interest is

Since the data are generated according to (9.1), the true parameter values are γ00 = 0.5, γ10 = 0.2, γ01
= 0, τ02 = 0.16, σ2 = 1.
The variables for imputation are Y, X1, X2. Thus, in terms of the discussion on p. 136, X = X1
while Xextra = X2. The missingness mechanism for the three variables Y, X1, X2 jointly here is MAR.
Since missingness of Y depends on X2, the model for only Y and X1 is MNAR.
Model 1 of Table 9.1 gives the result of an analysis of the model of interest after dropping all
level-one units with a missing value. The estimated parameter values are far off; for example, γ10 is
estimated as 0.028 with a standard error of 0.018, whereas the true value is 0.2. This illustrates the
pernicious effect of listwise deletion of missing data on parameter estimates.

Table 9.1: Estimates for random intercept model with listwise deletion of
missing values (Model 1) and with randomly imputed values according to a
multivariate normal distribution (Model 2).
 Model 2, on the other hand, was obtained after imputing the missing
values based on an assumed model that the random vectors (Yij, X1ij, X2ij)
are independent for different values of (i, j), and have a three-variate normal
distribution. This model is correct for the marginal distribution of (Yij, X1ij,
X2ij) but ignores the multilevel structure. It is the model suggested as step 1,
the initialization, of the procedure above. This model recovers the fixed
parameters quite well, with, for example, The recovery of the
variance parameters is less good, which may be expected since the
dependence structure is not well represented by the imputation model. The
results are not trustworthy, in addition, because the estimation is done as if
the imputed data were observed in a bona fide way, whereas they were
made up. This has the consequence that the standard errors will be too low,
suggesting undue precision. This issue will be dealt with in the next
subsection, where the example is continued.
Model 1 is based on considerably fewer cases than Model 2. The
deviance in Model 2 is based on an inflated data set, and therefore is totally
incomparable with that of Model 1.

9.4.2 Putting together the multiple results

We now continue with steps 2 and 3 of page 135, the combination of the
multiple imputations so as to produce estimates for the model of interest.
 We suppose that we have a number of data sets in which the missing
values have been randomly imputed according to the methods of the
preceding section. It is also assumed that, given the observed parts of the
data, the imputed parts are independent across these data sets. In the
MCMC method of the preceding section, this means that the sampling
frequency K must be high enough, ensuring negligible correlation between
consecutively generated imputed data sets. Denote the number of imputed
data sets by M and the data sets themselves by D(1), D(2),. . ., D(M). Thus,
each data set contains the variables X and Y, and all have the same values
for the observed data points but different imputed values for the missing
data points. We first discuss how to combine parameter estimates across the
imputed data sets, and then how to combine deviance tests across the
imputations.
In step 2, for each data set D(m) the parameters are estimated by the
method appropriate to the model of interest, for example, the hierarchical
linear model. Let us focus on a single parameter, for example, one of the
regression coefficients (fixed effects) or variance parameters. This
parameter is arbitrarily denoted θ. Denote the estimate for this parameter
from data set D(m) by and the associated standard error, calculated under
the erroneous assumption of complete data, by S.E.(m). If there is very little
missing data, or the other variables permit quite accurate imputations (e.g.,
because for the variable that has missing values there exists another, highly
correlated variable, which was always observed), then the estimates will
be scarcely different from each other. On the other hand, if the missingness
entails a considerable loss of information then the estimates will be very
different from one another, depending on the random imputations. Thus,
there are two sources of imprecision in the parameter estimates: the within-
data-set uncertainty which is reflected by the standard errors S.E.(m), and
the imputation uncertainty reflected by the variation between the estimates
. These two sources are combined by the following formulas from Little
and Rubin (2002).
First, from the multiple imputations we compute the average estimate
the average within-data-set variance

and the between-imputation variance

Then the combined estimate is the mean parameter estimate (9.3) over all
imputed data sets, and its standard error is

If the different imputations lead to almost the same estimates , then the
between-imputation variance B will be very small, and the standard error
(9.6) is practically the same as the individual standard errors S.E.(m).
However, often the situation is different, and the overall standard error (9.6)
will be larger.
To test the null hypothesis that the parameter θ is equal to some value
θ0, the usual t ratio,

can be tested against a t distribution with degrees of freedom given by

This is based on the assumption that the estimates are approximately

normal, discussed below (p. 144). Similar multi-dimensional formulas are
possible for vector-valued parameters θ; see Little and Rubin (2002).
A useful measure is the fraction of missing information for estimating a
given parameter. Intuitively, it is evident that if W is larger, as compared to
B, there is a larger difference between the ‘complete data sets that could
have been’, so that a larger value of W compared to B points to a more
important amount of missing information. The formula for the estimated
fraction of missing information is

This expresses how much of the information about the value of a given
parameter, potentially available in the complete data set, has been lost
because of the missingness. This fraction will depend on the parameter
under consideration, as we shall see in the example below. The estimated
value may be rather unstable across independent repetitions of the
imputation process, but will indicate the rough order of magnitude.
Wald tests as well as deviance (likelihood-ratio) tests can be combined
by methods discussed by Little and Rubin (2002). Suppose that a null
hypothesis is being tested by a Wald test or deviance test and the multiple
imputations yield, for the completed data sets obtained, test statistics C1,
C2,. . ., CM which in the complete data case have under the null hypothesis
an asymptotic chi-squared distribution with q degrees of freedom. Test
statistic Cm will be either the Wald test, or the difference between the
deviances for the estimated models corresponding to the null and alternative
hypotheses, each for the same data set D(m). Define by the average test
statistic,

and by V the sample variance of the square roots of Cm,

where
Then the combined test statistic is

This test statistic must be tested against an F distribution with degrees of

freedom q and b, where

These rather intimidating equations work out in such a way that when the
deviance statistics Cm are practically the same, so that V is very small, then
q × is very close to each of these deviances and the result of the combined
tests is practically the same as the result based on the individual deviance
tests. For deviance tests, Little and Rubin (2002) also present another
combined test, but this requires more information than the individual test
statistics.
The original work in which Rubin (e.g., 1987) developed the multiple
imputation method suggested that a very low number of imputations such as
K = 5 is sufficient. However, this may lead to instability of the estimated
between-imputation variance B. Further, a larger fraction of missing data
will require a larger number of imputations. Graham et al. (2007) developed
recommendations for the number of imputations; for example, their advice
is to use at least 20 imputations if 10–30% of the information is missing,
and at least 40 imputations if half of the information is missing. These
proportions of missing information refer to the quantity estimated by (9.8).
If multiparameter tests are required, for example, for testing the effect
of a categorical variable with three or more categories, then the required
number of imputations is also larger. Carpenter and Kenward (2008)
recommend multiplying the number of imputations by q(q + 1) /2, where q
is the dimension of the multiparameter test.
Example 9.2 Simulated example continued: multiple imputations.
We continue the preceding example. Recall that the model of interest is
(see (9.2)), there are missing values only in Y, and the missingness mechanism is known to be MAR,
given the additional variable X2.
The first set of results in Table 9.2 repeats Model 2 of Table 9.1, based on a single random
imputation under a multivariate normal model for (Y, X1, X2) without a multilevel structure, and
treating this as an observed data set. The second set of results present the result of multiple
imputations, using the imputation model

which is in accordance with the conditional distribution of the missing values given the observed
data, although some of the coefficients are 0. We used M = 50 imputed data sets.

Table 9.2: Estimates for random intercept model with a single imputation
and with M = 50 multiple imputations.

Comparing the two sets of results shows that in this case the parameter
estimates from the single imputation were quite close already, and the more
realistic standard errors for the multiple imputations all are somewhat larger
than those for the single imputation, in accordance with what is expected
from (9.6), but also for the standard errors the differences are not large. For
the level-two variance τ02 a clear improvement is obtained, the true value
being 0.16. The imputations using an imputation model with a multilevel
structure led to good recovery of the intraclass correlation, which was
underestimated by imputations using a multivariate normal model without
multilevel structure.
 The fractions of missing information range from 0.05 for the regression
coefficient of to 0.32 for the level-one variance. The proportion of cases
with a missing value was 834/4,000= 0.21. The fractions of missing
information for the parameters pertaining to the level-two model are smaller
than this value, which means that the multiple imputation succeeds in
recovering a greater part of the information for these parameters; this is
possible because there the other level-one units in the same level-two unit
also contribute to the information. That the fraction of missing information
for the level-one parameters are larger than 0.21 might be a random
deviation.

This example is quite extreme, because the missingness is determined

deterministically by the observed variable X2 which is strongly correlated
with the dependent variable. Still, this example teaches us a number of
things:

1. An analysis using only the complete data and ignoring the existence
of missing data can be hopelessly biased.
2. If missingness is at random (MAR) and a correct imputation model is
used, then multiple stochastic imputation can yield approximately
unbiased results.
3. It is not necessary to know the missingness mechanism, it is sufficient
to know that MAR holds for the given set of variables, and to have an
imputation model that correctly represents the dependencies between
these variables.
4. If there are missing values only in the dependent variable, but other
variables are available that are predictive of the dependent variable
and not included in the model of interest, cases with a missing
dependent variable should not be dropped and multiple imputation
can lead to more credible results.
5. Imputation models that do not represent the dependencies in the data
well may yield results that in a similar way underestimate these
dependencies.

Normality assumptions for multiple imputation

This method of combining the multiple imputations assumes that the
parameter estimates in the complete data set have a normal sampling
distribution. In multilevel models, this will always require that the number
of units at the highest level is large enough, say, 30 or more (with values as
low as 20 still possible but giving poorer approximations); and the total
number of level-one units is high enough, say, at least 100. For the
estimated regression coefficients as well as for the level-one residual
variance, normality will then be a reasonable assumption. For the estimated
variance components at level two and higher, however, the approximate
normality will break down if the true variance is relatively small. It is safer
to apply equations (9.3)–(9.6) to the estimated standard deviation, and
transform back to the variance, if this is desired. Note that in Section 6.3 we
commented on the normal approximation for the distribution of variance
parameters, and presented the approximate equation (6.2) for transforming
standard errors between variances and standard deviations, with conditions
under which this is a reasonable approximation. If these conditions are not
met, a Bayesian approach (Section 12.1) will still be possible.

9.5 Multiple imputations by chained equations

If there are missing values in more than one variable, then step 4 in the
multiple imputation procedure as described on p. 137 can be difficult in
practice. It requires the conditional distributions to be specified for all
variables that have any missing values, conditional on all the other
variables; and these conditional distributions must be coherent in the sense
that they belong to the simultaneous distribution of all variables together.
The multilevel dependencies between the variables may make this difficult.
Goldstein et al. (2009) developed a quite general model for this, applicable
to combinations of continuous and discrete variables, which is implemented
in REALCOM-impute (Goldstein, 2011). However, difficulties still may
remain for various model specifications, for example, if interactions are
involved or there are differences between within-group and between-group
regressions.
A very flexible approximate procedure for imputing missing values is
imputation by chained equations. This has been proposed by various
authors and is also known under various other names, such as fully
conditional specification and partially incompatible MCMC. Van Buuren
(2007) gives an overview of this literature. Although the theoretical
grounding of the method is incomplete, it is intuitively very appealing and
has yielded good results in simulations, as shown by Raghunathan et al.
(2001), van Buuren et al. (2006), van Buuren (2007), and for clustered data
by Zhao and Yucel (2009). This method was further developed for
multilevel data structures by van Buuren (2011).
It proceeds as follows. We consider a statistical model with basic
variables which for the moment we shall call Y1, Y2,. . ., YK, all of which
may contain some missing data points. We do not specify the simultaneous
distribution of all variables jointly, but only model the conditional
distribution of each variable on all the others. Each of these distributions is
a regression-type model with a response variable Yk and a set of explanatory
variables Y1,. . ., Yk−1,Yk+1,. . ., YK, which may be augmented by
interactions, group means, and other variables constructed from these.
Depending on the type of variable, this might be a linear regression model
if Yk is continuous, a logistic regression model if Yk is binary, etc. For the
multilevel case, these models would be the multilevel variants: a
hierarchical linear model, a multilevel logistic regression model, etc.
Denote by the vector of parameters for this model. The principle of
imputation by chained equations is to treat in turn the imputation for each of
the variables Yk while considering the others as given, using these
regression-type models; and cycle repeatedly through the variables. This
can be described as follows:

1.Initialize the process in a reasonable way. This can be done, for

example, by doing the following for each k. Estimate φk from the
available data, and impute by drawing from the conditional
distribution of the missing values in Yk given the observed data. This
initialization yields a completed data set.
2.For each k = 1,. . ., K in turn, do the following.
Make a random draw from the conditional (‘posterior’)
(a) distribution of φ given the completed data set.
k

(b)Impute the missing values in Yk by random draws from the model

 for Yk given the other variables Y1,. . ., Yk−1, Yk+1,. . ., YK, using the
completed data set for these variables, conditioning on the
observed cases for Yk, and using the estimated parameter value .
If Yk is a variable in a multilevel structure, and group means of Yk
for groups defining a higher level are among the explanatory
variables for some of the other variables Yh, then update these
(c)
group means from the completed data for Yk. (The same will be
done if interactions or other kinds of aggregates than group means
play a role.)
3.Repeat the process of step 2 until it is deemed that it has lasted long
enough.
If for some variables there are no missing data, they can evidently be
skipped in the imputation process.
The secret here is that we do not worry whether the models for Yk given
Y1,. . ., Yk−1, Yk+1,. . ., YK hang together in one multivariate distribution. This
is a theoretical defect but a practical benefit. For example, for a two-level
model of interest, the variables Y1,. . ., YK would be the dependent variables
and all the explanatory variables with any missing values, the level-one as
well as the level-two variables. Each of these can be modeled separately in
the most appropriate way. The level-one variables can be modeled by the
hierarchical linear model if they are continuous, by a multilevel logistic
regression model if they are binary, etc.; level-two variables can be modeled
by single-level linear or logistic or other generalized linear regression
models, with as predictors the other level-two variables and the group
means of the level-one variables; interactions can be accommodated
without any problems.
Multilevel software is now beginning to implement the procedures of
this and the preceding sections. Since their implementation is not yet
widespread, we give some more details on how these steps can be carried
out.
Drawing from the posterior distribution of the parameters as mentioned
in step 2(a) is a basic element in Bayesian estimation procedures treated in
Section 12.1 (see p. 196).
This Bayesian step can also be replaced by the following.
 Estimate the parameter from the completed data set, together with
a′. its estimated covariance matrix Σk. Draw from the multivariate
normal distribution with mean and covariance matrix Σk.

For regression coefficients especially this will be hardly any different from
the Bayesian procedure (step 2(a)). An even simpler method mentioned by
van Buuren (2007) is the plug-in method:
a″. Estimate the parameter from the completed data set, and use =
.
If sample sizes are reasonably large, this will usually yield quite similar
results to the full procedure.
The procedure for imputing as in step 2(b) depends on the type of
variable that is being considered. Let us assume that we are dealing with a
two-level data structure where the group sizes are rather homogeneous.
(Strongly heterogeneous group sizes may lead to variance heterogeneity
between the groups, which leads to further complications.) First consider
the case where Yk is a level-two variable for which a linear regression
model is reasonable. The parameter value comprises the regression
coefficients and the residual variance, to be denoted . Then for each
level-two unity j, using the regression coefficients and the explanatory
variables (which are completely known in the imputed data set), calculate
the predicted value Generate independent normally distributed random
variables with expected value 0 and variance 52. Then the imputed value
is + Rkj.
Now consider the case where Yk is a level-one variable for which a
random intercept model is used; denote this by

Here the parameter value comprises the regression coefficients, the

intercept variance , and the residual variance . A complication arises
from the fact that the observed values for for i′ = 1,. . ., nj offer
information about U0j which should be used in the imputation. We proceed
by first calculating the predicted value from the regression coefficients
and the explanatory variables. Then we calculate the empirical Bayes
estimate (posterior mean; see Section 4.8) denoted by and the
associated posterior standard deviation (or comparative standard deviation,
see p. 65), denoted S.E.j. These values depend both on the parameters and
on the observed (nonmissing) data for variable Yk. Next we generate a
normally distributed random variable with mean 0 and variance S.E.j2.
Then the imputed value for U0j is the sum4 Finally, for all
missing values of Ykij we generate independent normally distributed random
variables Rkij with expected value 0 and variance . The missing values are
then imputed by

Example 9.3 Comparison of results without and with imputations.

Here we again use the data set for which many examples were given in Chapters 4 and 5. The model
of interest is a random intercept model with effects of verbal IQ and SES at the individual level as
well the group level,

As auxiliary variables Xextra in the imputation model we used three variables that are closely related
to the variables in the model of interest: an earlier language test, a measure of ‘performal IQ’, and a
dummy variable indicating belonging to an ethnic minority. In these variables we have the following
numbers of missing values:

Language test Y 114

IQ 17
SES 133
Performal IQ 8
Earlier language test 318
Minority 0

For the three variables with most missing cases, there is very little overlap in the individuals for
whom there are missing values, which leads to good possibilities for imputation.
The imputation was done by the method of chained equations, where for each variable with
missing values the imputation model was defined as a random intercept model using the other five
variables, as well as their group means, as predictors. From a preliminary estimation it appeared that
several interactions seemed to have significant effects: the product interactions between IQ and SES,
between the group means of IQ and of SES, and between IQ and the classroom proportion (i.e.,
group mean) of minority students. These interactions were added to the imputation models for the
variables that were not involved in these interactions.
 Table 9.3 presents two sets of results for the language test, as a function of SES and IQ, in a
random intercept model. The left-hand column gives the parameter estimates of the naive analysis,
where all students with a missing value on at least one of these variables have been dropped. The
right-hand column gives the estimates from the analysis based on imputation by chained equations,
with 50 imputations. The results for the two analyses are very similar, which suggests that in this case
the missingness is not related to the ways in which intelligence, socio-economic status, and schools
influence this language test. This permitted us to use casewise deletion in the examples in Chapters
4-5.

Table 9.3: Estimates for random intercept model with complete cases only,
and using multiple imputations.

9.6 Choice of the imputation model

In any imputation procedure, the choice of the imputation model is a crucial
decision. The discussion here applies to imputation based on the
simultaneous distribution as well as to imputation by chained equations.
In the first place, an ample set of variables should be used for the
imputations. If reasonable predictors are available for variables with
missing cases, or for missingness itself, then it is wise to include those in
the imputation model; if they are not part of the model of interest they will
have the role of the Xextra variable discussed on p. 136. In the case of
complex sample surveys, variables that determine inclusion probabilities
should also be included in the imputation model (cf. Chapter 14). All
variables that make the MAR assumption more plausible as compared to
MNAR should be included in this way, even if they have some missing
values themselves.
Second, a general rule is that it does not hurt to use relatively complex
models for the imputations, as long as they are feasible in practice. Issues of
overfitting as are known for parameter estimation in most cases do not
matter for imputation. One exception is the phenomenon of ‘complete
separation’ in logistic regression: sometimes for a logistic regression model
there is a threshold value such that for values of the linear predictor higher
or lower than the threshold, the observed values of the dependent variable
are all zeros or all ones. When this is encountered in a model used for
imputation, extra care is required to check if this would be realistic, and
avoid it if it is not.
Third, while simplification certainly is allowed and even unavoidable, it
should be realized that dependencies in the total data set that are not
represented in the imputation model will also be underrepresented in the
imputed values, which may lead to bias in the estimates for the model of
interest. For example, as was illustrated by Example 9.2, using a single-
level imputation model for a two-level data set is likely to lead to
misrepresentation of the intraclass correlations. As another example, not
using group means of level-one variables as explanatory variables for the
imputation of other level-one variables may lead to biased estimation of
differences between within-group and between-group regression
coefficients. This implies that especially those aspects in which the
researcher is mainly interested should be represented well in the imputation
model.
When constructing the imputation model from separate models for each
variable that has any missing cases (e.g., if one uses chained equations), a
useful procedure may be the following. Start by fitting preliminary models
for the distribution of all variables given the others, and for the distribution
of the missingness indicators given the others. These preliminary models
could be based on one simple and rough imputed data set (e.g., obtained
using the assumption of multivariate normality and without taking the
multilevel structure into account). For each pair of variables Yk and Yh, if
either variable is important for predicting the other variable, and/or variable
Yh is important for predicting missingness in Yk, then include Yh in the
imputation model for Yk.
Van Buuren (2011) suggests using imputation models that are multilevel
models with heterogeneous level-one variances. This may be relevant
mainly when group sizes are strongly heterogeneous, because then a regular
hierarchical linear model of Y on a variable X which itself has a
homogeneous within-group variance may lead to a conditional distribution
of X given Y with heterogeneous level-one variances. This is unlikely to be
much of a concern when group sizes do not vary strongly.
Finally, researchers should not be afraid of having to make some
approximations and reasonable guesses. Always they must realize that
imputation is guesswork unless the missing data mechanism is known a
priori, for example, because of having been built into the data collection
design. Results of statistical analyses based on imputation of missing
values, in cases where the missingness mechanism is a priori unknown, will
be surrounded by even more uncertainty than are statistical results obtained
from complete data sets.

9.7 Glommary
Missingness indicators. Binary random variables indicating that a data
point is missing or observed.

Complete data. The hypothetical data structure that is regarded as the data
set that would be observed if there were no missing data points.

Missingness completely at random (MCAR). The missingness indicators

are independent of the complete data.

Missingness at random (MAR). Conditionally given the observed data,

the missingness indicators are independent of the unobserved data.
Under MAR, naive approaches such as analysing the complete cases
only have the risk of serious bias and loss of information, but it is
possible to follow approaches that do not have these defects.
Missingness not at random (MNAR). This is the negation of MAR. The
analysis of incomplete data where missingness is not at random will
always depend on untestable assumptions, and will be more
complicated and leave open more questions than the MAR case. This
book does not treat methods of analysis for the MNAR case. For
further literature the reader is referred to Molenberghs and Kenward
(2007), Carpenter and Kenward (2008), Graham (2009), Hogan et al.
(2004), Yuan and Little (2007), Roy and Lin (2005), and Mason et al.
(2010).

Implications for design. It is recommended to collect auxiliary data that

are predictive of missingness indicators and of the values of
unobserved data points. Including such auxiliary data can push the
design in the direction of MAR.

Model of interest. The model on which the research is substantively

focused.

Missing values on the dependent variable. If there are missing data points
only for the dependent variable, while there are auxiliary variables that
are predictive of missingness or of the unobserved values, and these
auxiliary variables are not included in the model of interest, then bias
may be reduced by methods of analysis that use these auxiliary
variables and retain the data points with a missing dependent variable.

Full information maximum likelihood. When possible, an analysis based

on the likelihood of the observed data is preferable. This is called full
information maximum likelihood. This approach is sometimes
possible, but cumbersome, in cases with missing data points for
several variables in a nonmonotone pattern.

Imputation. Filling in the missing data.

Multiple stochastic imputation. Constructing multiple imputed data sets

by drawing the imputations for the missing data points from the
conditional distribution of their values, given the observed data.
Imputation model. The probability model used for the stochastic
imputation. It is recommended, if possible, to use auxiliary variables
that are not included in the model of interest but that are predictive of
missingness or of the unobserved values, and to use these auxiliary
variables in the imputation model.

Combination of results. Each imputed data set is a complete data set for
which parameter estimates can be obtained for the model of interest.
These estimates can be combined across the multiple imputed data sets
to obtain, in the MAR case, and if the imputation model is correct,
approximately unbiased parameter estimates and standard errors.

Multiple imputation by chained equations. An imputation method that

needs, for each variable that has missing data, a model for this variable
conditional on the other variables; but does not require these models to
be the conditional distributions for the simultaneous distribution of all
variables jointly. This is a flexible approach that can be utilized
fruitfully for the complexities of multilevel data structures.

Choice of the imputation model. It is advisable to take care that the

important dependencies in the data are represented in the imputation
model; for example, the multilevel data structure, interactions, and
differences between within-group and between-group regressions.

1
The phrase ‘conditionally given the observed data’ refers to the assumption that the observed values
are known, but without knowing that these variables are all that what is observed; thus, the
conditioning is not on the missingness indicators.
2
Formally, for some of what is asserted below about MAR, an extra condition is needed. This is that
the parameters for the missingness process and those for the model explaining the primary dependent
variable are distinct. This condition is called separability. It is necessary to ensure, together with
MAR, that the missingness variables themselves are not informative for the primary research
question. This will be tacitly assumed below when the MAR condition is considered. MAR and
separability together define ignorability of the missingness mechanism.
3
At least it will depend on the parameters of the imputation model; not on parameters of the model
of interest that are not included in the imputation model.
4
Incidentally, it may be noted that and ∆U0j are uncorrelated (cf. (4.18)), so that var(U0j) =
var(Û0jEB) +S.E.j2.
 10
Assumptions of the Hierarchical
Linear Model

Like all statistical models, the hierarchical linear model is based on a

number of assumptions. If these assumptions are not satisfied, the
procedures for estimating and testing coefficients may be invalid. The
assumptions are about the linear dependence of the dependent variable, Y,
on the explanatory variables and the random effects; the independence of
the residuals, at level one as well as at the higher level or levels; the
specification of the variables having random slopes (which implies a certain
variance and correlation structure for the observations); and the
homoscedastic normal distributions for the residuals. It is advisable, when
analyzing multilevel data, to devote some energy to checks of the
assumptions. This chapter is concerned with such checks.
Why are model checks important at all? One danger of model
misspecification is the general misrepresentation of the relations in the data
(e.g., if the effect of X on Y is curvilinear increasing for X below average
and decreasing again for X above average, but only the linear effect of X is
tested, then one might obtain a nonsignificant effect and conclude
mistakenly that X has no effect on Y). Another is the invalidity of
hypothesis tests (e.g., if the random part of the model is grossly
misspecified, the standard errors and therefore the hypothesis tests for fixed
effects may be completely off the mark). To construct an adequate model
for a given data set, good insight into the phenomenon under study is
necessary, combined with an inspired application of the relevant social
theories; in addition, statistical tools such as those described in this chapter
can be helpful.
 OVERVIEW OF THE CHAPTER
First, the basic assumptions of the hierarchical linear model are
summarized. This model in itself is already much more flexible than the
general linear model that is the basis of standard regression analysis. The
chapter continues by discussing several approaches that can be
straightforwardly applied within the regular hierarchical linear model to
check model fit and perhaps improve it by adding additional hierarchical
linear model components. This extends what has already been said about
this topic in Section 6.4. Specification of the fixed part is treated, with
attention to transforming explanatory variables. Then the specification of
the random part is discussed, with special attention given to testing
heteroscedasticity.
The second half of the chapter is about residuals, which can be defined
for each of the levels. Inspection of the residuals may be important in
diagnosing units that do not conform well to the model. Measures are
explained, analogous to Cook’s distance in linear regression, that indicate
the influence of higher-level units on the results of the multilevel fit.
Finally, a short section is devoted to methods developed to handle data for
which the residuals have nonnormal distributions.
One of the threads running through this chapter is, in line with Section
6.4, that model specification in most cases should be done upward: first try
to get the model at the lower levels well specified, then continue with the
higher levels.
The topics of this chapter are treated with more of the mathematical
background in Snijders and Berkhof (2008). With respect to software
implementation, there is some variation among multilevel analysis
programs in methods for model checking. MLwiN macros to compute the
diagnostics mentioned in Sections 10.4–10.7, and R scripts for doing the
same, can be downloaded from the website for this book,
http://www.stats.ox.ac.uk/~snijders/mlbook.htm.

10.1 Assumptions of the hierarchical linear model

The hierarchical linear model was introduced in Chapter 5. In this chapter
we consider only the two-level model. The basic definition of the model for
observation Yij on level-one unit i within level-two unit j, for i = 1,. . ., nj, is
given in formula (5.15) as

As before, we shall refer to level-two units also as ‘groups’. The

assumptions were formulated as follows after equation (5.12). The vectors
(U0j, U1j,. . ., Upj) of level-two random coefficients, or level-two residuals,
are independent between groups. These level-two random coefficients are
independent of the level-one residuals Rij, which also are mutually
independent, and all residuals have population mean 0, given the values of
all explanatory variables. Further, the level-one residuals Rij have a normal
distribution with constant variance σ2. The level-two random effects
(U0j,U1j,. . ., Upj) have a multivariate normal distribution with a constant
covariance matrix. The property of constant variance is also called
homoscedasticity, nonconstant variance being referred to as
heteroscedasticity.
To check these assumptions, the following questions may be posed:

1. Does the fixed part contain the right variables (now X1,…, Xr)?
2. Does the random part contain the right variables (now X1,. . ., Xp)?
3. Are the level-one residuals normally distributed?
4 Do the level-one residuals have constant variance?
5. Are the level-two random coefficients normally distributed?
Do the level-two random coefficients have a constant covariance
6.
matrix?

These questions are answered in this chapter in various ways. The answers
are necessarily incomplete, because it is impossible to give a completely
convincing argument that a given specification is correct. For complicated
models it may be sensible, if there are enough data, to employ cross-
validation (e.g., Mosteller and Tukey, 1977). Cross-validation means that
the data are split into two independent halves, one half being used for the
search for a satisfactory model specification and the other half for testing of
effects. This has the advantage that testing and model specification are
separated, so that tests do not lose their validity because of capitalization on
chance. For a two-level model, two independent halves are obtained by
randomly distributing the level-two units into two subsets.

10.2 Following the logic of the hierarchical linear

model
The hierarchical linear model itself is an extension of the linear regression
model and relaxes one of the crucial assumptions of that model, the
independence of the residuals. The following model checks follow
immediately from the logic of the hierarchical linear model, and were
mentioned accordingly in Section 6.4.

10.2.1 Include contextual effects

In the spirit of Chapters 4 and 5, for every level-one explanatory variable
one should consider the possibility that the within-group regression is
different from the between-group regression (see Section 4.6), and the
possibility that X has a random slope. The difference between the within-
and between-group regression coefficients of a variable X is modeled by
also including the group mean of X in the fixed part of the model. This
group mean is a meaningful contextual variable, as follows from Chapters 2
and 3.
For interaction variables it can also be checked whether there are
contextual effects. This is done by applying the principles of Section 4.6 to,
for example, the product variable Xij Zj rather than the single variable Xij.
The cross-level interaction variable Xij Zj can be split into
and it is possible that the two variables and have different
regression coefficients. This is tested by checking whether the level-two
interaction variable has an additional fixed effect when Xij Zj already
belongs to the fixed part of the model. If there is such an additional effect,
one has the choice between including in the model either and ,
or Xij Zj and ; these two options will yield the same fit (cf. Section 4.6).
The assumption that the population mean of Uhj is 0, conditionally
given all the explanatory variables, implies that these random intercepts and
slopes are uncorrelated with all explanatory variables. If this assumption is
incorrect, the problem can be remedied by adding relevant explanatory
variables to the model. A nonzero correlation between Uhj and a level-two
variable Zj is remedied by including the product variable Xhij Zj, as
discussed in the preceding paragraph (cf. Snijders and Berkhof, 2008,
Section 3.2). A nonzero correlation between Uhj and a level-one variable
Xkij is remedied by including the product variable . Since for h = 0 we
have Xhij = 1, applying this procedure to the random intercept U0j leads to
including the main effect variables Zj and respectively.
Sometimes it makes sense also to include other contextual variables, for
example, the standard deviation within group j of a relevant level-one
variable. For example, heterogeneity of a school class, as measured by the
standard deviation of the pupils’ intelligence or of their prior achievements,
may be a complicating factor in the teaching process and thus have a
negative effect on the pupils’ achievements.

10.2.2 Check whether variables have random effects

Definition of levels
A first issue is the definition of the ‘levels’ in the model. Formulated more
technically, these are the systems of categories where the categories have
random coefficients: the random intercepts and possibly also random
slopes. Since this is at the heart of multilevel modeling it has already been
given much attention (see Chapter 2 and Section 4.3; Section 6.4 is also
relevant here). Several papers have studied the errors that may be made
when a level is erroneously left out of the model; see Tranmer and Steel
(2001), Moerbeek (2004), Berkhof and Kampen (2004), van den Noortgate
et al. (2005), and Dorman (2008). A general conclusion is that if a level is
left out of the model, that is, a random intercept is omitted for some system
of categories, then the variance associated with that level will be
redistributed mainly to the next lower and (if it exists) next higher levels.
Furthermore, erroneous standard errors may be obtained for coefficients of
variables that are defined on this level (i.e., are functions of this system of
categories), and hence tests of such variables will be unreliable. This will
also hold for variables with strong intraclass correlations for the omitted
level. Similarly, if a random slope is omitted, then the standard errors of the
corresponding cross-level interaction effects may be incorrectly estimated.
This leads to the general rule that if a researcher is interested in a fixed
effect for a variable defined at a certain level (where a level is understood as
a system of categories, such as schools or neighborhoods), then it is
advisable to include this level with a randomly varying intercept in the
multilevel model – unless there is evidence that the associated random
intercept variance is negligible. Similarly, if there is interest in a fixed effect
for a cross-level interaction X × Z where Z is defined at a certain level, then
X should have a random effect at this level – again, unless the random slope
variance is negligible. Furthermore, if there is interest in the amount of
variability (random intercept variance) associated with a given level, and
there exist next higher and next lower levels of nesting in the phenomenon
under study, then these levels should be included in the model with random
intercepts.

Specification of random slopes

It is possible that the effect of explanatory variables differs from group to
group, that is, some variables have random slopes. In the analysis of
covariance (ANCOVA), this is known as heterogeneity of regression. It is
usual in ANCOVA to check whether regressions are homogeneous.
Similarly, in the hierarchical linear model it is advisable to check for each
level-one variable in the fixed part whether it has a random slope.
Depending on the amount of data and the complexity of the model,
estimating random slopes for each explanatory variable may be a time-
consuming affair. A faster method for testing random slopes was proposed
by Berkhof and Snijders (2001) and also by Verbeke and Molenberghs
(2003). These tests do not require the complete ML or REML estimates for
the random slope models. Suppose that some model, denoted by M0, has
been estimated, and it is to be tested whether a random slope for some
variable should be added to the model. Thus, M0 is to be tested as the null
hypothesis. The method proceeds as follows.
Recall from Section 4.7 that the estimation algorithms for the
hierarchical linear model are iterative procedures. If the IGLS, RIGLS, or
Fisher scoring algorithm is used, it is already possible to base a test of the
random slope on the results of the first step of the algorithm, making it
unnecessary to carry out the full estimation process with all iteration steps.
To test whether some variable has a random slope, first estimate the
parameters for model M0 and then add this random slope to M0 (it is
sufficient to add only the variance parameter, without the corresponding
covariance parameters). Now carry out just one step of the (R)IGLS or
Fisher scoring algorithm, starting from the parameter estimates obtained
when fitting M0. Denote the provisional slope variance estimate after this
single step by , with associated standard error (also after the single step)
. Then the t-ratio,

can be tested against the standard normal distribution. (Note that, as is

explained in Berkhof and Snijders (2001), this test of a t-ratio for a variance
parameter can be applied only to the result of the first step of the estimation
algorithm and not to the final ML estimate. The main reason is that the
standard error of the usual estimator, obtained on convergence of the
algorithm, is to a certain extent proportional to the estimated variance
component, in contrast to the standard error of the single-step estimate.)
Another question that is relevant when a random slope is specified for a
variable X of which the between-group regression coefficient differs from
the within-group coefficient, is whether the original variable Xij or the
within-group deviation variable should get the random slope. This
was discussed in Section 5.3.1.
 10.2.3 Explained variance
In the discussion in Section 7.1 of the concept of explained variance for
multilevel models, it was mentioned that the fraction of explained variance
at level one, , can be used as a misspecification diagnostic. If this fraction
of explained variance decreases when a fixed effect is added to the model,
this may be a sign of misspecification. A small decrease, however, may be a
result of chance fluctuations. For reasonably large data sets, a decrease by a
magnitude of 0.05 or more should be taken as a warning of possible
misspecification.

10.3 Specification of the fixed part

Whether the fixed part contains the right variables is an equivocal question
in almost all research in the social and behavioral sciences. For many
dependent variables there is no single correct explanation, but different
points of view may give complementary and valid representations. The
performance of a pupil at the final moment of compulsory education can be
studied as a function of the characteristics of the child’s family; as a
function of the child’s achievements in primary school; as a function of the
child’s personality; or as a function of the mental processes during the final
examination. All these points of view are valid, taken alone and also
combined. Therefore, what is the right set of variables in the fixed part
depends in the first place on the domain in which the explanation is being
considered.
In the second place, once the data collection is over one can only use
variables based on the available data. Both of these considerations imply
that some variable, say Xq, may be omitted from the fixed part of the model.
Part or all of this variable will then be represented by the residuals in the
random part. If the fixed part includes some variables that are correlated
with Xq, then the effects of the variables included will take up some of the
effect of the variable omitted. This is well known in regression analysis. It
implies that (unless the variables are experimentally manipulated) one
should be careful with interpretation and very reluctant to make causal
interpretations. For the hierarchical linear model the consequences of
omitting variables from the fixed part of the model are discussed in
Raudenbush and Bryk (2002, Chapter 9).
In a multilevel design, one should be aware in any case of the
possibility that supposed level-one effects are in reality, completely or
partially, higher-level effects of aggregated variables. This was discussed in
Section 10.2.1.

Transformation of explanatory or dependent variables

One important option for improving the specification of the fixed part of the
hierarchical linear model is the transformation of explanatory variables.
Aggregation to group means or to group standard deviations is one kind of
transformation. Calculating products or other interaction variables is
another kind. The importance given in multilevel modeling to cross-level
interactions should not diminish the attention paid to the possibility of
within-level interactions.
A third kind of transformation is the nonlinear transformation of
variables. Such transformations may be applied to the dependent as well as
the explanatory variables. Our treatment focuses on the latter, but the
former possibility should not be forgotten. As examples of nonlinear effects
of an explanatory variable X on Y, one may think of an effect which is
always positive but levels off toward low or high values of X; or a U-shaped
effect, expressed by a function that first decreases and then increases again.
The simplest way to investigate the possible nonlinearity of the effect of
a variable X is to conjecture some nonlinear transformation of X, such as X2,
ln(X) (for positive variables), or 1/X, add this effect to the fixed part of the
model – retaining the original linear effect of X – and test the significance
of the effect of this nonlinear transformed variable. Instead of adding just
one nonlinear transformed variable, flexibility is increased by adding
several transformations of the same variable X, because the linear
combinations of these transformations will be able to represent quite a
variety of shapes of the effect of X on Y.
Rather than conjecturing a specific good transformation, one may also
choose the best one from within a family of transformations. A well-known
family of transformations is the Box–Cox transformation (Box and Cox,
1964) defined by
where ln denotes the natural logarithm and the parameter λ must be
determined so as to give the best fit. An interesting feature of this family is
that it shows how the logarithm is embedded in a family of power
transformations. It is a family of transformations often applied to the
dependent variable, in cases where there is skewness in the distribution of
the residuals. This transformation is treated in Cook and Weisberg (1982,
Chapter 5) and Atkinson (1985), and for transforming dependent variables
in multilevel models by Hodges (1998, p. 506) and Goldstein et al. (2009).
Another convenient family of transformations for explanatory variables
is constituted by spline functions. Spline functions are discussed in Section
15.2.2. More extensive discussions are given in Seber and Wild (1989,
Section 9.5) and Fox (2008, Chapter 17); and for multilevel models in
Goldstein (2011, Chapter 15) and Snijders and Berkhof (2008, Sections
3.8–3.11). We find quadratic splines simple to work with and quite
effective. An example is given in Figure 8.1. A quadratic spline
transformation of some variable X can be chosen as follows. Choose a value
x0 in the middle of the range of X (e.g., the grand mean of X) as a reference
point for the square of X and choose a few other values x1,. . ., xK within the
range of X. Usually, K is a low number, such as 1, 2, or 3. Higher values of
K yield more flexibility to approximate many shapes of functions of X.
Calculate, for each k = 1,. . ., K, the ‘half-squares’ fk(X), defined as 0 left of
xk and (X - xk)2 right of xk:

In the fixed part use the linear effect X, the quadratic effect (X - x0)2, and the
effects of the ‘half squares’ f1(X),. . ., fK(X). Together these functions can
represent a wide variety of smooth functions of X, as is evident from Figure
8.1. If some of the fk(X) have nonsignificant effects they can be left out of
the model, and by trial and error the choice of the so-called nodes xk may be
improved. Such an explorative procedure was used to obtain the functions
displayed in this figure.

10.4 Specification of the random part

The specification of the random part has been discussed above and in
Section 6.4. If certain variables are mistakenly omitted from the random
part, the tests of their fixed coefficients may be unreliable. Therefore it is
advisable to check the randomness of slopes of all variables of main
interest, and not only those for which a random slope is theoretically
expected.
The random part specification is directly linked to the structure of the
covariance matrix of the observations. In Section 5.1.1 we saw that a
random slope implies a heteroscedastic specification of the variances of the
observations and of the covariance between level-one units in the same
group (level-two unit). When different specifications of the random part
yield a similar structure for the covariance matrix, it will be empirically
difficult or impossible to distinguish between them. But also a
misspecification of the fixed part of the model can lead to a
misspecification of the random part; sometimes an incorrect fixed part
shows up in an unnecessarily complicated random part. For example, if an
explanatory variable X with a reasonably high intraclass correlation has in
reality a curvilinear (e.g., quadratic) effect without a random component,
whereas it is specified as having a linear fixed and random effect, the
excluded curvilinear fixed effect may show up in the shape of a significant
random effect. The latter effect will then disappear when the correctly
specified curvilinear effect is added to the fixed part of the model. This was
observed in Example 8.2. The random slope variance of IQ was attenuated
in this example when a curvilinear effect of IQ was considered.

10.4.1 Testing for heteroscedasticity

In the hierarchical linear model with random slopes, the observations are
heteroscedastic because their variances depend on the explanatory
variables, as expressed by equation (5.5). However, the residuals Rij and Uhj
are assumed to be homoscedastic, that is, to have constant variances.
In Chapter 8 it was explained that the hierarchical linear model can also
represent models in which the level-one residuals have variances depending
linearly or quadratically on an explanatory variable, say, X. Such a model
can be specified by the technical device of giving this variable X a random
slope at level one. Similarly, giving a level-two variable Z a random slope at
level two leads to models for which the level-two random intercept variance
depends on Z. Also random slope variances can be made to depend on some
variable Z. Neglecting such types of heteroscedasticity may lead to
incorrect hypotheses tests for variables which are associated with the
variables responsible for this heteroscedasticity (X and Z in this paragraph).
Checking for this type of heteroscedasticity is straightforward using the
methods of Chapter 8. However, this requires variables to be available that
are thought to be possibly associated with residual variances.
A different method, described in Raudenbush and Bryk (2002, Chapter
9), can be used to detect heteroscedasticity in the form of between-group
differences in the level-one residual variance, without a specific connection
to some explanatory variable. It is based on the estimated least squares
residuals within each group, called the ordinary least squares (OLS)
residuals. This method is applicable only if many (or all) groups are
considerably larger than the number of level-one explanatory variables. The
level-two explanatory variables are, for the moment, neglected. What
follows applies to the groups for which nj–r–1 is not too small (say, 10 or
more), where r is the number of level-one explanatory variables. For each
of these groups separately an OLS regression is carried out with the level-
one variables as explanatory variables. Denote by the resulting estimated
residual variance for group j and by dfj = nj – r – 1 the corresponding
number of degrees of freedom. The weighted average of the logarithms,

must be calculated. If the hierarchical linear model is well specified this

weighted average will be close to the logarithm of the maximum likelihood
estimate of σ2.
 From the group-dependent residual variance a standardized residual
dispersion measure can be calculated using the formula

If the level-one model is well specified and the population level-one

residual variance is the same in all groups, then the distribution of the
values dj is close to the standard normal distribution. The sum of squares,

can be used to test the constancy of the level-one residual variances. Its null
distribution is chi-squared with N′ – 1 degrees of freedom, where N′ is the
number of groups included in the summation.
If the within-groups degrees of freedom dfj are less than 10 for many or
all groups, the null distribution of H is not chi-squared. Since the null
distribution depends only on the values of dfj and not on any of the
unknown parameters, it is feasible to obtain this null distribution by straight
forward computer simulation. This can be carried out as follows: generate
independent random variables Vj according to chi-squared distributions with
dfj degrees of freedom, calculate = Vj / dfj, and apply equations (10.3),
(10.4) and (10.5). The resulting value H is one random draw from the
correct null distribution. Repeating this, say, 1,000 times gives a random
sample from the null distribution with which one can compare the observed
value from the real data set.
If this test yields a significant result, one can inspect the individual dj
values to investigate the pattern of heteroscedasticity For example, it is
possible that the heteroscedasticity is due to a few unusual level-two units
for which dj has a large absolute value. For other approaches for dealing
with heteroscedasticity, see Section 10.4.2.
Example 10.1 Level-one heteroscedasticity.
The example of students’ language performance used in Chapters 4, 5, and 8 is considered again. We
investigate whether there is evidence of level-one heteroscedasticity where the explanatory variables
at level one are IQ, SES, and gender, specifying the model as Model 1 in Table 8.1. Note that the
nature of this heteroscedasticity test is such that the level-two variables included in the model do not
matter. Those groups were used for which the residual degrees of freedom are at least 10. There were
133 such groups. The sum of squared standardized residual dispersions defined in (10.5) is H =
155.5, a chi-squared value with df = 132, yielding p = 0.08. Hence this test does not give evidence of
heteroscedasticity.
Although nonsignificant, the result is close enough to significance that it is still worthwhile to try
and look into this small deviation from homoscedasticity. The values dj can be regarded as standard
normal deviates in the case of homoscedasticity. The two schools with the largest absolute values had
dj = –3.2 and –3.8, expressing low within-school variability; while the other values were all less than
2.5 in absolute value, which is of no concern. The two homogeneous schools were also those with the
highest averages for the dependent variable, the language test. They were not particularly large or
small, and had average compositions with respect to socio-economic status and IQ. Thus, they were
outliers in two associated ways: high averages and low internal residual variability. The fact that the
homoscedasticity test gave p = 0.08, not significant at the conventional level of 0.05, suggests,
however, that these outliers are not serious.
An advantage of this test is that it is based only on the specification of the within-groups
regression model. The level-two variables and the level-two random effects play no role at all, so
what is checked here is purely the level-one specification. However, the null distributions of the dj
and of H mentioned above do depend on the normality of the level-one residuals. A heavier-tailed
distribution for these residuals in itself will also lead to higher values of H, even if the residuals do
have constant variance. Therefore, if H leads to a significant result, one should investigate the
possible pattern of heteroscedasticity by inspecting the values of dj, but one should also inspect the
distribution of the OLS within-group residuals for normality.

10.4.2 What to do in case of heteroscedasticity

If there is evidence for heteroscedasticity, it may be possible to find
variables accounting for the different values of the level-one residual
variance. These could be level-one as well as level-two variables.
Sometimes such variables can be identified on the basis of theoretical
considerations. In addition, plots of dj versus relevant level-two variables or
plots of squared unstandardized residuals (see Section 10.5) may be
informative for suggesting such variables. When there is a conjecture that
the nonconstant residual variance is associated with a certain variable, one
can apply the methods of Chapter 8 to test whether this is indeed the case,
and fit a heteroscedastic model.
In some cases, a better approach to deal with heteroscedasticity is to
apply a nonlinear transformation to the dependent variable, such as a square
root or logarithmic transformation, as discussed in Section 10.3. How to
choose transformations of the dependent variable in single-level models is
discussed in Atkinson (1985), Cook and Weisberg (1982), and Goldstein et
al. (2009). Transforming the dependent variable can be useful, for example,
when the distribution of the dependent variable (or rather, of the residuals)
is highly skewed.
When there is heteroscedasticity and the dependent variable has a small
number of categories, another option is to use the multilevel ordered logit
model for multiple ordered categories (Section 17.4) or to dichotomize the
variable and apply multilevel logistic regression (Section 17.2).

10.5 Inspection of level-one residuals

A plethora of methods have been developed for the inspection of residuals
in ordinary least squares regression; see, for example, Atkinson (1985) and
Cook and Weisberg (1982, 1999). Inspection of residuals can be used, for
example, to find outlying cases that have an undue high influence on the
results of the statistical analysis, to check the specification of the fixed part
of the model, to suggest transformations of the dependent or explanatory
variables, or to point to heteroscedasticity.
These methods may likewise be applied to the hierarchical linear model,
but some changes are necessary because of the more complex nature of the
hierarchical linear model and the fact that there are several types of
residuals. For example, in a random intercept model there is a level-one and
also a level-two residual. Various methods of residual inspection for
multilevel models were proposed by Hilden-Minton (1995). He noted that a
problem in residual analysis for multilevel models is that the observations
depend on the level-one and level-two residual jointly, whereas for model
checking it is desirable to consider these residuals separately. It turns out
that level-one residuals can be estimated so that they are unconfounded by
the level-two residuals, but the other way around is impossible.
Analysis of level-one residuals can be based on the OLS regressions
calculated within each group separately. This was proposed by Hilden-
Minton (1995). We made use of the within-group OLS regression in Section
10.4.1, where we remarked that an advantage of this approach is its being
based only on the specification of the level-one model. Thus, it is possible
to inspect estimated level-one residuals without confounding with level-two
residuals or with level-two misspecification.
The estimated residuals are obtained by carrying out OLS regressions in
each group separately, using only the level-one variables. It is advisable to
omit the residuals for groups with low within-group residual degrees of
freedom, because those residuals are quite unstable.
The variance of the OLS residuals depends on the explanatory variables.
The OLS residuals can be standardized by dividing them by their standard
deviation. This yields the standardized OLS residuals; see Atkinson (1985)
or other texts on regression diagnostics. This provides us with two sets of
level-one residuals: the raw OLS residuals and the standardized OLS
residuals. The advantage of the standardized residuals is their constant
variance; the disadvantage is that the standardization may distort the
(possibly nonlinear) relations with explanatory variables.
Various possibilities for employing residuals to inspect model fit were
discussed by Hilden-Minton (1995). For basic model checking we propose
to use the level-one within-group OLS residuals in the following ways:
1. Plot the unstandardized OLS residuals against level-one explanatory
variables to check for possible nonlinear fixed effects.
If there are in total a few hundred or more level-one units a raw plot may
seem to exhibit just random scatter, in which case it is advisable to make
a smoothed plot. The residual plot can be smoothed as follows.
Often the explanatory variable has a limited number of categories; not
only when it is composed of a few ordered categories, but also, for
example, when it is based on an underlying scale with integer values,
possibly transformed, as in Example 10.2 below. Then the mean residual
can be calculated for each category (‘binning’) and plotted together with
vertical bars extending to twice the standard error of the mean within this
category. (This standard error can be calculated in the usual way, without
taking account of the nonconstant variance and the mutual dependence of
the residuals.)
For continuous explanatory variables many statistical packages offer
some method of smoothing, for example, a locally weighted scatterplot
smoother abbreviated as loess or lowess (Cleveland, 1979). If such a
smoother is not available, a simple and usually effective way of
smoothing is the following construction of a simple moving average.
Denote by x1,. . ., xM the ordered values of the explanatory variable under
consideration and by r1,. . ., rM the corresponding values of the residuals.
This means that xi and ri are values of the explanatory variable and the
within-group residual for the same level-one unit, reordered so that x1 ≤
 x2 ≤ . . . ≤ xM. (Units with identical x-values are ordered arbitrarily.)
Since groups with a small number of units may have been omitted, the
number M may be less than the original total sample size. The smoothed
residuals now are defined as averages of 2K consecutive values,

for i = K,. . ., M – K. The value of K will depend on data and sample size;
for example, if the total number of residuals M is at least 1,500, one
could take moving averages of 2K =100 values.
One may add to the plot horizontal lines plotted at r equal to plus or
minus twice the standard error of the mean of 2K values, that is,
is the variance of the OLS residuals. This indicates
roughly that values outside this band that is, may be
considered to be relatively large.
Make a normal probability plot of the standardized OLS residuals to
check the assumption of a normal distribution. This is done by plotting
the values where is the standardized OLS residual and zi, the
corresponding normal score (i.e., the expected value from the standard
normal distribution according to the rank of . Especially when this
2. shows that the residual distribution has longer tails than the normal
distribution, there is a danger of parameter estimates being unduly
influenced by outlying level-one units.
The squared standardized residuals can also be plotted against
explanatory variables to assess the assumption of level-one
homoscedasticity. Here also, averaging within categories, or smoothing,
can be very helpful in showing the patterns that may exist.

When a data exploration is carried out along these lines, this may
suggest model improvements which can then be tested. One should realize
that these tests are suggested by the data; if the tests use the same data that
were used to suggest them, which is usual, this will lead to capitalization on
chance and inflated probabilities of type I errors. The resulting
improvements are convincing only if they are ‘very significant’ (e.g., p <
0.01 or p < 0.001). If the data set is large enough, it is preferable to employ
cross-validation (see p. 126).
Example 10.2 Level-one residual inspection.
We continue Example 10.1, in which the data set also used in Chapters 4 and 5 is considered again,
the explanatory variables at level one being IQ, SES, and gender, with an interaction between IQ and
SES. Within-group OLS residuals were calculated for all groups with at least 10 within-group
residual degrees of freedom.
The variables IQ and SES are both based on integer scales from which the mean was subtracted,
so they have a limited number of categories. For each category, the mean residual was calculated and
the standard error of this mean was calculated in the usual way. The mean residuals are plotted in
Figure 10.1 for IQ categories containing 12 or more students; for SES categories were formed as
pairs of adjacent values, because otherwise they would be too small.
The vertical lines indicate the intervals bounded by the mean residual plus or minus twice the
standard error of the mean. For IQ the mean residuals exhibit a clear pattern, while for SES they do
not. The left-hand figure suggests a nonlinear function of IQ having a local minimum for IQ between
–2 and 0 and a local maximum for IQ somewhere about 2. There are few students with IQ values less

Figure 10.1: Mean level-one OLS residuals (with bars extending to twice
the standard error of the mean) as function of IQ (left) and SES (right).

Figure 10.2: Mean level-one OLS residuals (with bars extending to twice
the standard error of the mean) as function of IQ (left) and SES (right),for
model with nonlinear effect of IQ.
than –4 or greater than +4, so in this IQ range the error bars are very wide and not very informative.
Thus the figures point toward a nonlinear effect for IQ, in which the deviation from linearity has a
local minimum for a negative IQ value and a local maximum for a positive IQ value. Examples of
such functions are third-degree polynomials and quadratic spline functions with two nodes (cf.
Section 15.2.2). The first option was explored by adding IQ2 and IQ3 to the fixed part. The second
option was explored by adding IQ−2 and IQ+2, as defined in (8.2), to the fixed part. The second
option gave a much better model improvement and therefore was selected. When IQ_2 and IQ+2 are
added to Model 1 of Table 8.1, which yields the same fixed part as Model 4 of Table 8.2, the
deviance goes down by 54.9 (df = 2, p < 0.00001). This is strongly significant, so this nonlinear
effect of IQ is convincing even though it was not hypothesized beforehand but suggested by the data.
The mean residuals for the resulting model are graphed as functions of IQ and SES in Figure 10.2.
These plots do not exhibit a remaining nonlinear effect for IQ.
A normal probability plot of the standardized residuals for the model that includes the nonlinear
effect of IQ is given in Figure 10.3. The distribution looks quite normal except for the very low
values, where the residuals are somewhat more strongly negative (i.e., larger in absolute value) than
expected. However, this deviation from normality is rather small.

Figure 10.3: Normal probability plot of standardized level-one OLS

residuals.

10.6 Residuals at level two

Estimated level-two residuals always are confounded with the estimated
level-one residuals. Therefore one should check the specification of the
level-one model before moving on to checking the specification at level
two. The level-two specification can be checked by using the level-two
residuals treated in this section and the influence measures treated in the
next.
The empirical Bayes estimates (also called posterior means) of the
level-two random effects, treated in Section 4.8, are the ‘estimates’ of the
level-two random variables Uhj (h = 0, 1,. . ., p) in the model specification
(10.1). These can be used as estimated level-two residuals. They can be
standardized by dividing by their diagnostic standard error, as defined in
Section 4.8.1.
Similarly to the level-one residuals, one may plot the unstandardized
level-two residuals as a function of relevant level-two variables to check for
possible nonlinearity; normal probability plots may be made of the
standardized level-two residuals; and the squared standardized level-two
residuals may be plotted as a function of level-two variables to check
homoscedasticity. Smoothing the plots will be less often necessary because
of the usually much smaller number of level-two units. If the plots contain
outliers, one may inspect the corresponding level-two units to check
whether anything unusual can be found.
Normal probability plots of standardized residuals are sensitive to both
nonnormal residual distributions and misspecification of the fixed part of
the model (Verbeke and Molenberghs, 2000, Section 7.8; Eberly and
Thackeray, 2005). Therefore one should realize that what seems to be a
deviation from normality might also be caused by an incorrect specification
of the fixed effects.
Checking empirical Bayes residuals is discussed more extensively in
Langford and Lewis (1998, Section 2.4). If the diagnostic variances differ
considerably between level-two units, greater precision can be obtained by
using the diagnostic variances also to weight the contributions of the
standardized residuals to their cumulative distribution function; this was
proposed by Lange and Ryan (1989).
Example 10.3 Level-two residual inspection.
The example above is continued now with the inspection of the level-two residuals. The model is
specified with fixed effects of IQ and SES and their interaction; the school means of IQ and SES and
their interaction; the two nonlinear functions of IQ; and gender. The random part at level two
includes a random intercept and a random slope for IQ. (This is like Model 4 of Table 8.2, except that
the level-one heteroscedasticity is left out.) Level-two empirical Bayes residuals are calculated for
the intercept and for the slope of IQ; these may be called ‘posterior intercepts’ and ‘posterior slopes’,
respectively. Figure 10.4 shows the unstandardized posterior intercepts as a function of the school
averages of IQ and SES. Locally weighted scatterplot (‘lowess’) smoothers (Cleveland, 1979), as
mentioned on p. 162, are added as nonlinear approximations to the averages of the residuals as a
function of these school variables to provide a better view of how the average residuals depend on
the school means. Figure 10.5 shows the unstandardized posterior slopes for IQ, as a function of the
school averages of IQ and SES. The four plots show some weak patterns, but not enough to justify
further model extensions.

Figure 10.4: Posterior intercepts as function of (left) average IQ and (right)

average SES for each school. Smooth lowess approximations are indicated
by ο.

Figure 10.5: Posterior IQ slopes as function of (left) average IQ and (right)

average SES for each school. Smooth lowess approximations are indicated
by ο.
In Figure 10.6 the normal probability plots of the standardized residuals are shown. These normal
plots support the approximate normality of the level-two residuals.
 Figure 10.6: Normal probability plot of standardized level-two residuals:
(left) intercepts, (right) slopes.

10.7 Influence of level-two units

The diagnostic value of residuals can be supplemented by so-called
influence diagnostics, which give an indication of the effect of certain parts
of the data on the parameter estimates obtained. In this section we present
diagnostics to investigate the influence of level-two units. These diagnostics
show how strongly the parameter estimates are affected if unit j is deleted
from the data set. The residuals treated here are based on the deletion
principle (Atkinson, 1985), that is, they are obtained by comparing the data
for a given unit with the values expected under the model, estimated from a
data set from which this unit has been deleted. This treatment is based on
Snijders and Berkhof (2008, Section 3), in which a modification of Lesaffre
and Verbeke (1998) and of the corresponding section in the first edition of
the current textbook was developed. Other influence measures for
multilevel models may be found in Langford and Lewis (1998, Section 2.3).
For the definition of the diagnostics we shall need matrix notation, but
readers unfamiliar with this notation can base their understanding on the
verbal explanations.
Denote by γ = (γ0, γ1,…,γr) the vector of all parameters of the fixed part,
consisting of the general intercept and all regression coefficients. Further,
denote by the vector of estimates produced when using all the data and by
the covariance matrix of this vector of estimates (so the standard errors
of the elements of are the square roots of the diagonal elements of ).
Denote the parameter estimate obtained if level-two unit j is deleted from
the data by (-j). Then unit j has a large influence if (-j) differs much from .
This difference can be measured on the basis of the covariance matrix ,
because this matrix indicates the uncertainty that exists anyway in the
elements of . For increased diagnostic precision, a modified estimate (–j)
is used, based on the data from which group j has been deleted.
Unit j has a large impact on the parameter estimates if, for one or more
of the individual regression coefficients h, the difference between h and
(-j)h is not much smaller, or even larger, than the standard error of h. Given
that the vector γ has a total of r + 1 elements, a standardized measure of the
difference between the estimated fixed effects for the entire data set and the
estimates for the data set excluding unit j is given by

This can be interpreted as the average squared deviation between the

estimates with and those without unit j, where the deviations are measured
proportional to the standard errors (and account is taken of the correlations
between the parameter estimates). For example, if there is only one fixed
parameter and the value of for some unit is 0.5, then leaving out this unit
changes the parameter estimate by a deviation equal to times its
standard error, which is quite appreciable.
This influence diagnostic is a direct analog of Cook’s distance (e.g.
Cook and Weisberg, 1982; Atkinson, 1985) for linear regression analysis. If
the random part at level two is empty, then is equal to Cook’s distance.
It may be quite time-consuming to calculate the estimates (–j) for each
group j. Since the focus here is on the diagnostic value of the influence
statistic rather than on the precise estimation for the data set without group
j, an approximation can be used instead. Pregibon (1981) proposed, in a
different context, substituting the one-step estimator for (–j) This is the
estimator which starts from the estimate obtained for the full data set and
then carries out a single step of the IGLS, RIGLS, or Fisher scoring
algorithm. Denoting this one-step estimator by and the corresponding
one-step estimator for the covariance matrix by , we obtain the
influence diagnostic

A similar influence diagnostic can be defined for the influence of group

j on the estimates of the random part. Denote by φ the vector of all
parameters of the random part and by q the number of elements of this
vector. If the covariance matrix of the random effects is unrestricted, then φ
contains the variances and covariances of the random slopes as well as the
level-one residual variance. Given that there are p random slopes, this is a
total of q =(p + 1) (p + 2) /2 + 1 parameters. Now define as the estimate
based on the complete data set and as the covariance matrix of this
estimate (having on its diagonal the squared standard errors of the elements
of ); and (–j) and as the one-step estimates when deleting level-two
unit j. Then the diagnostic for the influence of unit j on the parameters of
the random part is defined by

One can also consider the combined influence of group j on the

parameters of the fixed and those of the random part of the model. Since the
estimates and are approximately uncorrelated (Longford, 1987), such a
combined influence diagnostic can be defined simply as the weighted
average of the two previously defined diagnostics,

Lesaffre and Verbeke (1998) propose local influence diagnostics which are
closely related to those proposed here.1

Whether a group has a large influence on the parameter estimates for the
whole data set depends on two things. The first is the leverage of the group,
that is, the potential of this group to influence the parameter estimates
because of its size nj and the values of the explanatory variables in this
group. Groups with a large size and with strongly dispersed values of the
explanatory variables have a high leverage. The second is the extent to
which this group fits the model as defined by (or estimated from) the other
groups, which is closely related to the values of the residuals in this group.
A poorly fitting group that has low leverage, for example, because of its
small size, will not strongly affect the parameter estimates. Likewise, a
group with high leverage will not strongly affect the parameter estimates if
it has deletion residuals very close to 0.
If the model fits well and the explanatory variables are approximately
randomly distributed across the groups, then the expected value of
diagnostics (10.6), (10.7), and (10.8) is roughly proportional to the group
size nj. A plot of these diagnostics as a function of nj will reveal whether
some of the groups influence the fixed parameter estimates more strongly
than should be expected on the basis of group size.
The fit of the level-two units to the model can be measured by the
deletion standardized multivariate residual for each unit. This measure was
also proposed by Snijders and Berkhof (2008) as a variant of the ‘squared
length of the residual’ of Lesaffre and Verbeke (1998), modified by
following the principle of deletion residuals. It is defined as follows. The
predicted value for observation Yij on the basis of the fixed part, minimizing
the influence of the data in unit j on this prediction, is given by

The deletion multivariate residual for group j is the vector of deviations

between the observations and these predicted values,

This residual can be standardized on the basis of the covariance matrix of

the vector of all observations in group j. For the hierarchical linear model
with one random slope, for example, the elements of this covariance matrix
are given by (5.5) and (5.6). Denote, for a general model specification, the
covariance matrix of the observations in group j by

This covariance matrix is a function of the parameters of the random part of

the model. Substituting estimated parameters, again obtained by the
deletion principle to minimize the influence of the outcomes observed for
group j itself, yields the estimated covariance matrix, . Now the
standardized multivariate residual is defined by
This approach can also be used to define standardized deletion level-one
residuals as

where is the ith diagonal element of the covariance matrix .

Unlike the level-one OLS residuals of Section 10.5, these definitions
can also be applied to models with heteroscedasticity at level one (see
Chapter 8).
The standardized residual can be interpreted as the sum of squares of
the multivariate residual for group j after transforming this multivariate
residual to a vector of uncorrelated elements each with unit variance. If the
model is correctly specified and the number of groups is large enough for
the parameter estimates to be quite precise, then Sj2 has a chi-squared
distribution with nj degrees of freedom. This distribution can be used to test
the value of Sj2 and thus investigate the fit of this group to the hierarchical
linear model as defined by the total data set. (This proposal was also made
by Waternaux et al., 1989.) Some caution should be exercised with this test,
however, because among all the groups some ‘significant’ results can be
expected to occur by chance even for a well-specified model. For example,
if there are 200 groups and the model is correctly specified, it may be
expected that Sj2 has a significantly large value at the 0.05 significance level
for ten groups, and at the 0.01 level for two groups. One can apply here the
Bonferroni correction for multiple testing, which implies that the model
specification is doubtful if the smallest p-value of the standardized
multivariate residuals is less than 0.05/N, where N is the total number of
groups.
Values of Cj are assessed only in comparison with the other groups,
while values of Sj2 can be assessed on the basis of the chi-squared
distribution. When checking the model specification, one should worry
mainly about groups j for which the influence diagnostic Cj is large
compared to the other groups, and simultaneously there is a poor fit as
measured by the standardized multivariate residual Sj2. For such groups one
should investigate whether there is anything particular. For example, there
may be data errors or the group may not really belong to the population
investigated. This investigation may, of course, also lead to a model
improvement.
Example 10.4 Influence of level-two units.
We continue checking the model for the language test score which was also investigated in Example
10.3. Recall that this is Model 4 of Table 8.2, but without the level-one heteroscedasticity. The 20
largest influence diagnostics (10.8) are presented in Table 10.1 together with the p-values of the
standardized multivariate residuals (10.10). When assessing the p-values one should realize that,
since these are the most influential schools, this set will contain some schools that differ relatively
strongly from the others and therefore have relatively low p-values.
The school with the largest influence has a good fit (p = 0.293). School 15 fits very poorly, while
schools 18, 52, 107, 122, and 187 have a moderately poor fit. The Bonferroni-corrected bound for the
smallest p-value is 0.05/211 = 0.00024. The poorest fitting school has a p-value of 0.00018, smaller
than 0.00024, which suggests indeed that the fit is unsatisfactory, although not to a very strong
extent.
As a further model improvement, the heteroscedasticity as a function of IQ was considered,
leading to exactly Model 4 of Table 8.2. The 20 largest influence diagnostics for this model are
presented in Table 10.2.
The influence diagnostic of school 213 has considerably increased; but this school does not have
a very small number of students in the data set (nj = 19) and it does not have a very poor fit to the

Table 10.1: The 20 largest influence diagnostics.

rest of the data: p = 0.010 is not alarming giving that it is observed for the school selected as the one
with the strongest influence. The fit of school 15 has improved, with p going from 0.00018 to
0.00049, and now exceeds the Bonferroni-corrected critical value of 0.00024. Therefore, the results
shown in Table 10.2 are quite reassuring.
There is one school with a smaller p-value for the standardized multivariate residual. This is
school 108, with nj = 9 students and pj = 0.00008. This is less than 0.00024 and therefore does lead
to significance, for the poorest fitting school, with the Bonferroni correction. Inspection shows that
the poor fit is caused by the fact that two out of the nine pupils in this school have very low values on
the language test (14 and 17) while not having unusual values on IQ or SES. Since this school does
not have a large influence (Cj = 0.008), however, the fact that it does not fit well is of no real
concern.
A look at the data for school 108 reveals that the predicted values Ŷij for this school are quite
homogeneous and close to the average, but two of the nine outcome values Yij are very low.
Correspondingly, the standardized deletion level-one residuals (10.5) are quite small, −3.0 and −4.3.
When these two students are deleted from the data set, the parameter estimates hardly change, but the
lowest p-value for the standardized multivariate residuals is no longer significant when the
Bonferroni correction is applied. Analyzing this data set while excluding the two schools with the
poorest fit, 15 and 108, yields practically the same estimates as the full data set.
 To conclude, the investigation of the influence statistics led to an improved fit in line with the
heteroscedasticity models of Chapter 8, and to the detection of two deviant cases in the data set.
These two cases out of a total of 3,758, however, did not have any noticeable influence on the
parameter estimates.

Table 10.2: The 20 largest influence diagnostics for the extended model.

10.8 More general distributional assumptions

If it is suspected that the normality and homoscedasticity assumptions of the
hierarchical linear model are not satisfied, one could employ methods based
on less restrictive model assumptions. This is an area of active research.
 One approach is to derive estimators that are robust in the sense that the
influence of individual cases or higher-level units is bounded. Such
estimators were proposed by Richardson (1997), Yau and Kuk (2002),
Dueck and Lohr (2005), and Copt and Victoria-Feser (2006). See also
Demidenko (2004, Section 4.4) and Wu (2010, Chapter 9).
Another approach is to postulate distributions for residuals with heavier
tails than the normal distribution; this will reduce influence for outliers.
Models with t-distributed residuals were developed by Seltzer (1993),
Seltzer et al. (1996), Pinheiro et al. (2001), Seltzer and Choi (2003), and
Staudenmayer et al. (2009); and with more general residual distributions by
Chen et al. (2002) and Burr and Doss (2005). Verbeke and Lesaffre (1996)
propose mixtures of normal distributions for the residuals; see also Verbeke
and Molenberghs (2000). Alonso et al. (2010) discuss the fact that greater
difficulties may arise in the interpretation of estimated random effect
variances for such methods.
Kasim and Raudenbush (1998) develop methods for models with
heterogeneous level-one variances that are not (as in Chapter 8) a function
of observed level-two variables. For these models as well as for those
assuming heavy-tailed distributions for higher-level residuals, computer-
intensive methods such as Gibbs sampling are very useful; see Gelman et
al. (2004, Chapter 15) and, more specifically, Congdon (2010, Sections 5.4,
6.5 and 10.4).
A different but very practical approach is provided by the latent class
models of Section 12.3. These models approximate arbitrary distributions of
random effects by discrete distributions and therefore can be used in
principle for arbitrary random effect distributions.
‘Sandwich’ estimators for standard errors, also called cluster-robust
standard errors, make no assumptions about the distributional shape of the
random coefficients. Verbeke and Lesaffre (1997) and Yuan and Bentler
(2002) proposed sandwich estimators to provide standard errors applicable
for nonnormally distributed random effects. They are treated in Section
12.2. However, sandwich estimators do require large enough numbers of
units at the highest level; see Verbeke and Lesaffre (1997), Maas and Hox
(2004), and the further discussion in Section 12.2.

10.9 Glommary
Specification of the hierarchical linear model. This is embodied in the
choice of the dependent variable, the choice of levels, of variables with
fixed effects, and of variables with random effects with possible
heteroscedasticity.

Assumptions of the hierarchical linear model. These are the linearity of

the expected value of the dependent variable as a function of the
explanatory variables, expressed by (10.1); the residuals at all levels
having expected values 0 (independently of the predictor variables),
constant variances (unless a heteroscedastic model is employed), and a
normal distribution; and independence of the residuals across levels.

Model misspecification. This is a deviation between the process or

‘mechanism’ that generated the data and the specification of the model
used to analyze the data. Misspecification can be serious because
interesting conclusions may be masked, and erroneous inferences are
possible – the ‘interesting’ and ‘nuisance’ aspects discussed in Chapter
2.

The logic of the hierarchical linear model. Following this can be useful in
specifying the model: specify the levels correctly; consider whether
variables with fixed effects may have different within-group and
between-group effects (a choice with even more possibilities if there
are three or more levels); ask similar questions for interaction effects;
consider whether variables should have random slopes, and whether
there might be heteroscedasticity. A considerable decrease of the
explained variance (Section 7.1) when adding a fixed effect to the
model is also a sign of potential misspecification of the model.

Transformations. Transforming the dependent as well as the explanatory

variables can be helpful, for example, when variables have skewed
distributions.

Box–Cox transformation. This is a family of transformations including

powers (such as squares and square roots) and the logarithm. All these
transformations are embedded in a one-parameter family so that the
choice of a good power or logarithmic transformation is implemented
 by choosing this parameter. This is a useful family of transformations
for dependent as well as explanatory variables.

Spline functions. Polynomial functions (e.g., quadratic or cubic

polynomials) between certain points (the ‘knots’ of the spline) linked
smoothly at these knots. These are useful transformations for the
explanatory variables.

Working upward. This is often a good way to check a multilevel model:

first check the level-one model, then the level-two model (and so on, if
there are higher levels). This is because the level-one model
specification can be checked without confounding by the level-two
model specification, but not the other way around.

Level-one residuals. The difference between predicted and observed values

for the basic outcomes; to define these, a particular choice for
‘predicted’ must be made. A simple choice is predictions by OLS
regressions for each level-two unit (‘group’) separately. This leads to
the OLS within-group residuals. These can be used to test the
specification of the within-group model: choice of fixed effects and
homoscedasticity. To check the linearity of effects, smoothed plots of
the level-one residuals can be used.

Normal probability plot. This is a plot of the residuals as a function of

expected values from a normal distribution (also called a normal
quantile–quantile or normal QQ plot). The plot will be approximately
linear in the case of normal distributions. It can be employed to check
the normality of level-one and level-two residuals.

Level-two residuals. These can be defined as empirical Bayes estimates, or

posterior means, of the level-two random effects; since these have
expected value 0, these estimates themselves are also the residuals.
Empirical Bayes estimates were defined in Section 4.8. Here also plots
can be made of the smoothed residuals as a function of relevant level-
two explanatory variables to check linearity, as well as normal
probability plots to check normality.
Leverage. The potential of a part of the data, for example, a level-two unit,
to influence the parameter estimates because of its size and the values
of the explanatory variables in this part of the data set.

Goodness of fit. This is how well the data, or a part of the data,
corresponds to the model. The residuals are an important expression of
this: residuals close to 0 point to a good fit.

Outliers. Data points that do not correspond well to the model, as will be
indicated by large residuals.

Influence diagnostics. Statistics expressing influence of parts of the data

set on the parameter estimates. In this chapter we considered influence
diagnostics for two-level units. Higher-level units potentially have
larger influence than level-one units, because a higher-level unit
contains more information. Large influence is the combined result of
large leverage and poor or mediocre fit. If it is found that some level-
two units have a high influence then it will be important to assess to
what extent this is due to large leverage and poor fit: we mostly do not
mind if a large unit with a reasonable to medium fit has a strong
influence.

Deletion diagnostics. Diagnostics based on deleting part of the data, re-

estimating the parameters, and assessing how strongly these parameter
estimates differ from those obtained for the full data. This is a general
principle for constructing residuals; in this chapter we followed an
approximation to this principle (following it completely may be time-
consuming).

Cook’s distance. A well-known influence diagnostic for linear regression

models. The level-two influence diagnostics treated in this chapter are
similar in spirit to Cook’s distance, but now either applied separately
to the fixed coefficients and the parameters of the random part, or to all
parameters of the hierarchical linear model together.

Multivariate residual. This is the vector of deviations between the

observations and predicted values for a given level-two unit, and can
 be standardized by its covariance matrix, estimated according to the
deletion principle.

Standardized deletion level-one residuals. These can be defined in a

corresponding way, but now for the individual level-one units. In
contrast to the OLS within-group residuals, these residuals can also be
defined for models with heteroscedasticity at level one, and are
meaningful even for small groups.

Bonferroni correction. This is a simple, generally applicable way of

dealing with multiple testing. If several hypotheses are tested
simultaneously at significance level α0, the probability that at least one
of the null hypotheses is erroneously rejected will be larger than α0. If
the number of tested hypotheses is K, the simultaneous probability of
this last event cannot be larger than K × α0. Therefore, the Bonferroni
correction consists of testing each hypothesis at level α/K, so that the
probability of erroneously rejecting at least one null hypothesis is no
more than α. This is justified if the K tests should indeed be regarded
as one simultaneous test; otherwise it will lead to an unnecessarily
conservative test. In this chapter, the Bonferroni correction was applied
to testing whether a given level-two unit fits the model; this should be
done for all level-two units jointly, which makes this correction
appropriate.

Robust estimators. Estimators that retain good performance for data that
do not satisfy the default model assumptions very well. Different types
of robust estimators have been developed, for example, incorporating
robustness against outliers or against misspecification of the
distribution of the random effects.

1
Lesaffre and Verbeke’s diagnostics are very close to diagnostics (10.6), (10.7), and (10.8),
multiplied by twice the number of parameters, but using different approximations. We have opted for
the diagnostics proposed here because their values are on the same scale as the well-known Cook’s
distance for the linear regression model, and because the one-step estimates can be conveniently
calculated when one uses software based on the Fisher scoring or (R)IGLS algorithms.
 11
Designing Multilevel Studies

Up to now it has been assumed that the researcher wishes to test interesting
theories on hierarchically structured systems (or phenomena that can be
thought of as having a hierarchical structure, such as repeated data) on
available data. Or that multilevel data exist, and that one wishes to explore
the structure of the data. This, of course, is the other way around. Normally
a theory (or a practical problem that has to be investigated) will direct the
design of the study and the data to be collected. This chapter focuses on one
aspect of this research design, namely, the sample sizes. Sample size
questions in multilevel studies have also been treated by Snijders and
Bosker (1993), Mok (1995), Cohen (1998), Donner and Klar (2000),
Rotondi and Donner (2009), and others. Snijders (2001, 2005) and Cohen
(2005) provide primers for this topic.
Another aspect, the allocation of treatments to subjects within groups, is
discussed in Raudenbush (1997) and by Moerbeek and Teerenstra (2011).
Raudenbush et al. (2007) compare matching and covariance adjustment in
the case of group-randomized experiments. Hedges and Hedberg (2007)
treat the case where sample size is a function of the intraclass correlation
and covariate effects, while Rotondi and Donner (2009) provide an
approach to sample size estimation in which the distribution of the
intraclass correlation varies. Hedeker et al. (1999) as well as Raudenbush
and Liu (2001) present methods for sample size determination for
longitudinal data analyzed by multilevel methods.
This chapter shows how to choose sample sizes that will yield high
power for testing, or (equivalently) small standard errors for estimating,
certain parameters in two-level designs, given financial and practical
constraints. A problem in the practical application of these methods is that
sample sizes that are optimal, for example, for testing some cross-level
interaction effect are not necessarily optimal, for example, for estimating
the intraclass correlation. The fact that optimality depends on one’s
objectives, however, is a general problem of life that cannot be solved by
this textbook. If one wishes to design a good multilevel study it is advisable
to determine first the primary objective of the study, express this objective
in a parameter to be tested or estimated, and then choose sample sizes for
which this parameter can be estimated with a small standard error, given
financial, statistical, and other practical constraints. Sometimes it is possible
to check, in addition, whether also for some other parameters
(corresponding to secondary objectives), these sample sizes yield
acceptably low standard errors.
A relevant general remark is that the sample size at the highest level is
usually the most restrictive element in the design. For example, a two-level
design with 10 groups, that is, a macro-level sample size of 10, is at least as
uncomfortable as a single-level design with a sample size of 10.
Requirements on the sample size at the highest level, for a hierarchical
linear model with q explanatory variables at this level, are at least as
stringent as requirements on the sample size in a single-level design with q
explanatory variables.

OVERVIEW OF THE CHAPTER

Before discussing the issue of power in multilevel designs in more detail,
we make some general introductory remarks on the power of statistical tests
and the factors that determine power in single-level models: sample size,
effect size, and significance level. Then two not too complicated situations
are considered: the design of a study to estimate a population mean whilst
using a two-stage sampling design, and a design to determine the number of
measurements per level-two unit to get a reliable estimate for those level-
two units. Next, we introduce the problem of budget constraints when
determining optimal sample sizes at both levels of the hierarchy, and
present a series of examples of sample size calculations when one wishes to
estimate associations between variables in a multilevel situation. The issue
whether one should randomize groups or individuals within groups when
conducting an experiment is then discussed. Finally, we treat sample size
issues in cases where one is interested in precisely estimating the intraclass
correlation coefficient or one of the variance components.
Level-two units will be referred to as ‘groups’ or ‘clusters’.

11.1 Some introductory notes on power

When a researcher is designing a multistage sampling scheme, for example,
to assess the effects of schools on the achievement of students, or to test the
hypothesis that citizens in impoverished neighborhoods are more often
victims of crime than other citizens, important decisions must be made with
respect to the sample sizes at the various levels. For the two-level design in
the first example the question may be phrased as follows: should one
investigate many schools with a few students per school or a few schools
with many students per school? Or, for the second example: should we
sample many neighborhoods with only a few citizens per neighborhood or
many citizens per neighborhood and only a few neighborhoods? In both
cases we assume, of course, that there are budgetary constraints on the
research being conducted. To phrase this question more generally and more
precisely: how should researchers choose sample sizes at the macro and
micro level in order to ensure a desired level of power given a relevant
(hypothesized) effect size and a chosen significance level? The average
micro-level sample size per macro-level unit will be denoted by n and the
macro-level sample size by N. In practice the sizes of the macro-level units
will usually be variable (even if only because of unintentionally missing
data), but for calculations of desired sample sizes it normally is adequate to
use approximations based on the assumption of constant ‘group’ sizes.
A general introduction to power analysis can be found in the standard
work by Cohen (1988), or, for a quick introduction, Cohen’s (1992) power
primer. The basic idea is that we would like to find support for a research
hypothesis (H1) stating that a certain effect exists, and therefore we test a
null hypothesis about the absence of this effect (H0) using a sample from
the population of interest. The significance level α represents the risk of
mistakenly rejecting H0. This mistake is known as a type I error. In
addition, β denotes the risk of disappointingly not rejecting H0, when the
effect does exist in the population. This mistake is known as a type II error.
The statistical power of a significance test is the probability of rejecting H0
given the effect size in the population, the significance level α, and the
sample size and study design. Power is therefore given by 1 – β. As a rule
of thumb, Cohen suggests that power is moderate when it is 0.50 and high
when it is at least 0.80. Power increases as α increases, and also as the
sample size and/or the effect size increase. The effect size can be conceived
as the researcher’s idea about ‘the degree to which the null hypothesis is
believed to be false’ (Cohen, 1992, p. 156).
We suppose that the effect size is expressed by some parameter γ that
can be estimated with a certain standard error, denoted by S.E. . Bear in
mind that the size of the standard error is a monotone decreasing function of
the sample size: the larger the sample size the smaller the standard error! In
most single-level designs, the standard error of estimation is inversely
proportional (or roughly so) to the square root of sample size.
The relation between effect size, power, significance level, and sample
size can be presented in one formula. This formula is an approximation that
is valid for practical use when the test in question is a one-sided t-test for γ
with a reasonably large number of degrees of freedom (say, df ≥ 10). Recall
that the test statistic for the t-test can be expressed by the ratio t = .
The formula is

where z1– α, z1– β, and zβ are the z-scores (values from the standard normal
distribution) associated with the cumulative probability values indicated. If,
for instance, α is chosen at 0.05 and 1 – β at 0.80 (so that β = 0.20), and an
effect size of 0.50 is what we expect, then we can derive that we are
searching for a minimum sample size that satisfies

Formula (11.1) contains four ‘unknowns’. This means that if three of

these are given, then we can compute the fourth. In most applications that
we have in mind, the significance level α is given and the effect size is
hypothetically considered (or guessed) to have a given value; either the
standard error is also known and the power 1 – β is calculated, or the
intended power is known and the standard error calculated. Given this
standard error, we can then try to calculate the required sample size.
For many types of design one can choose the sample size necessary to
achieve a certain level of power on the basis of Cohen’s work. For nested
designs, however, there are two kinds of sample size: the sample size of the
micro-units within each macro-unit n and the sample size of the macro-units
N, with N × n being the total sample size for the micro-units.
Thus, more units are being observed at the micro level than at the macro
level, therefore the tradeoff between significance level and power can also
differ between the levels. When one is interested in assessing the effect of a
macro-level variable on a micro-level outcome whilst controlling for a set
of micro-level covariates, one could employ different significance levels at
the different levels. In this case one may, for example, set α for testing
effects at the micro level to be as small as 0.01, but for effects of the macro-
level variable at 0.05, in order to obtain sufficient power.

11.2 Estimating a population mean

The simplest case of a multilevel study occurs when one wishes to estimate
a population mean for a certain variable of interest (e.g., income, age,
literacy), and one is willing to use the fact that the respondents are
regionally clustered. This makes sense, since if one is interviewing people it
is a lot cheaper to sample, say, 100 regions and then interview 10 persons
per region, than to randomly sample 1,000 persons that may live scattered
all over a country. In educational assessment studies it is of course more
cost-efficient to sample a number of schools and then to take a subsample
of students within each school, than to sample students completely at
random (ignoring their being clustered in schools).
Cochran (1977, Chapter 9) provides formulas to calculate desired
sample sizes for such two-stage sampling. On p. 242, he defines the design
effect for a two-stage sample, which is the factor by which the variance of
an estimate (which is the square of the standard error of this estimate) is
increased because of using a two-stage sample rather than a simple random
sample with the same total sample size. Since estimation variance for a
simple random sample is inversely proportional to total sample size, the
design effect is also the factor by which total sample size needs to be
increased to achieve the same estimation variance (or, equivalently, the
same standard error) as a simple random sample of the given size. In
Section 3.4 we saw that the design effect is given by

where n is the average sample size in the second stage of the sample and ρI
is the intraclass correlation. Now we can use (3.17) to calculate required
sample sizes for two-stage samples.
Example 11.1 Assessing mathematics achievement.
Consider an international assessment study on mathematics achievement in secondary schools. The
mathematics achievement variable has unit variance, and within each country the mean should be
estimated with a standard error of 0.02. If a simple random sample of size n is considered, it can
readily be deduced from the well-known formula

that the sample size should be n = 1 /(0.022) = 2,500.

What will happen to the standard error if a two-stage sampling scheme is employed (first schools
then students), if the intraclass correlation is 0.20, and assuming that there are no direct extra
budgetary consequences of sampling schools (this might be the case where one is estimating costs,
when the standard errors are imposed by some international board)?
The design effect, when one opts for sampling 30 students per sampled school, is now 1 +(30 – 1)
× 0.20 = 6.8. So instead of having a simple random sample size of 2,500 one should use a two-stage
sample size of 2,500 x 6.8 = 17,000 in order to achieve the same precision.
Suppose further that one also wishes to test whether the average maths achievement in the
country exceeds some predetermined value. The power in this design, given this sample size,
depends on the effect size and α. The effect size here is the difference between the actual average
achievement and the tested predetermined value. Let us assume α is set at 0.01 and that one wishes to
know the effect size for which the power is as large as β = 0.80. According to (11.1), this will be the
case when the effect size is at least 0.02 × (2.33 + 0.84) = 0.063.

11.3 Measurement of subjects

In Section 3.5 the situation was discussed where the macro-unit may be, for
example, an individual subject, a school, or a firm, and the mean μ + Uj is
regarded as the ‘true score’ of unit j which is to be measured. The level-one
deviation Rij is then regarded as measurement error. This means that,
whereas in the previous section the question was how to measure the
overall population mean, now we are interested in measuring the mean of a
level-two unit, which we shall refer to as a subject. The observations Yij can
then be regarded as parallel test items for measuring subject j. Suppose we
wish to use an unbiased estimate, so that the posterior mean of Section 4.8
is excluded (because it is biased toward the population mean; see p. 63).
The true score for subject j therefore will be measured by the observed
mean, .j.
In equation (3.21) the reliability of an estimate defined as the mean of nj
measurements was given by

Now suppose that it is desired to have a reliability of at least a given value

λ0. The equation for the reliability implies that this requires at least

measurements for each subject.

Example 11.2 Hormone measurement.
Consider individual persons as subjects, for whom the concentration of some hormone is to be
measured from saliva samples. Suppose it is known that the intraclass correlation of these
measurements is 0.40, so that the reliability of an individual measurement is 0.40 (cf. equation
(3.19)). When an aggregate measurement based on several saliva samples is to be made with a
reliability of λ0 = 0.80, this requires at least nmin = (0.80 × 0.60) /(0.20 × 0.40) = 6 measurements per
subject.

11.4 Estimating association between variables

When one is interested in the effect of an explanatory variable on some
dependent variable in a two-level hierarchical linear model, two situations
are possible: both variables are micro-level variables, or the dependent
variable is a micro-level variable whereas the explanatory variable is
defined at the macro level. Generally, in both cases, standard errors are
inversely proportional to , the square root of the sampled number of
macro-units, which implies that taking as many macro-units as possible
(while n, the sample size within each sampled macro-unit, is constant)
reduces the standard errors in the usual way, which in turn has a positive
effect on the power of the statistical tests. Now let us suppose that the
researcher is restricted to a given total sample size M = N × n, but that he or
she is free in choosing either N large with n small or N small with n large.
Snijders and Bosker (1993) developed formulas for standard errors of
fixed effects in hierarchical linear models. These formulas can be of help in
accurately making this decision. Their software program PinT (see Chapter
18) can help researchers to settle these problems. This program calculates
standard errors for estimating fixed effects in two-level designs, as a
function of sample sizes. The greatest practical difficulty in using this
program is that the user has to specify the means, variances, and
covariances of all explanatory variables as well as the variances and
covariances of all random effects. This specification has to be based on
prior knowledge and/or reasonable guesswork. If one is uncertain about
these input values, one may try several combinations of them to see how
sensitive the resulting optimal sample sizes are for the input values within
the likely range.
We refrain from presenting the exact formulas, but give two examples
that can help in gaining some notion of how to decide in such a situation.
More information can be found in the PinT manual and in Snijders and
Bosker (1993).
Example 11.3 A design to study level-one associations.
Suppose one wishes to assess the association between income (INCOME) and total number of years
spent in school (YEARS) as part of a larger national survey using face-to-face interviews. Since
interviewers have to travel to respondents it seems worthwhile to reduce traveling costs and to take a
two-stage sample: randomly select neighborhoods and within each neighborhood select a number of
respondents.
First we have to make an educated guess as to the possible structure of the data. The dependent
variable is INCOME, while YEARS is the explanatory variable. For convenience let us assume that
both variables are standard normal scores (z-scores, with mean 0 and variance 1). (Perhaps monetary
income will need to be logarithmically transformed to obtain an approximately normal distribution.)
Suppose that for INCOME the intraclass correlation is 0.50 and for YEARS it is 0.30. Now let us
guess that the correlation between both variables is 0.447 (which implies that YEARS accounts for
20% of the variation in INCOME). The residual intraclass correlation for INCOME may be 0.40
(since the between-neighborhood regression of INCOME on YEARS generally will be larger than
the within-neighborhood regression: dwellings of educated, rich people tend to cluster together). The
standard errors of the effect of YEARS on INCOME for various combinations of the sample sizes n
and N, restricted so that the total sample size M = n × N is at most 3,000, are given in Table 11.1.
 Since N and n are integer numbers and M may not exceed 3,000, the total sample size fluctuates
slightly. Nevertheless, it is quite clear that in this example it is optimal to take n = 1 respondent per
neighborhood while maximizing the number of neighborhoods in the sample. If we test some target
value and assume an effect size of 0.05 (the degree to which the results may be ‘off target’) and set α
at 0.01, for n = 1 the power 1 – β would be 0.92, as follows from applying (11.1):

Using n = 1 implies a single-level design, in which it is impossible to estimate the intraclass

correlation. If one wishes to estimate the within-neighborhood variance and the intraclass correlation,
one needs at least two respondents per neighborhood.

Table 11.1: Standard errors as a function of n and N.

In the example given the researchers clearly had an almost unlimited

budget. Of course it is far more expensive to travel to 3,000 neighborhoods
and interview one person per neighborhood, than to go to 100
neighborhoods and interview 30 persons per neighborhood. It is therefore
common to impose budgetary constraints on the sampling design. One can
express the costs of sampling an extra macro-unit (neighborhood, school,
hospital) in terms of micro-unit costs.
 To sketch the idea in the context of educational research, assume that
observation costs are composed of salary, travel costs, and the material
required. Assume that the costs of contacting one school and the travel for
the visit are $150. Further assume that the salary and material costs for
testing one student are $5. For example, investigating 25 students at one
school costs a total of $150 + 25 x $5 = $275. More generally, this means
that the cost function can be taken as $150N + $5Nn = $5N(n + 30).
In many practical cases, for some value of c (in the example, c = 30) the
cost function is proportional to N(n + c). The number c is the ratio
indicating how much more costly it is to sample one extra macro-unit
(without changing the overall total number of micro-units sampled), than it
is to sample one extra micro-unit within an already sampled macro-unit.
Usually for mail or telephone surveys using a two-stage sampling
design there is no efficiency gain in using a two-stage sample as compared
to using a simple random sample, so that c = 0. But for research studies in
which face-to-face interviewing or supervised testing is required, efficiency
gains can be made by using a two-stage sampling design.
So we are searching for sample sizes at micro and macro level that
satisfy the inequality

in which K is the available budget, expressed in monetary units equal to the

cost of sampling one additional micro-level unit. In the example given, K
would be the budget in dollars divided by 5.
Example 11.4 Setting up a class size experiment.
Let us now turn to an example in which a macro-level variable is supposed to have an effect on a
micro-level variable, and apply the idea just described on cost restrictions.
We might wish to set up an experiment with class size, in which a class size reduction of 6 pupils
per class is compared to a usual class size of 26 in its effect on achievement of young pupils. Let us
assume that achievement is measured in z-scores, that is, their variance is 1. If there are additional
covariates in the design to control for differences between the experimental and control group we
assume that the residual variance, given these covariates, is equal to 1.
Let us suppose that the experiment is set up to detect an effect of 0.20 (or larger) in the
population of interest. Transforming the effect size d into a correlation coefficient r using the formula
(Rosenthal, 1991, p. 20)
results in r = 0.10. Stated otherwise, the experimental factor will account for r2 = 1 percent of the
variation in the achievement score. The astonishing thing about this example is that most researchers
would be inclined to test all pupils within a class and would take the class size of 23 (being the
average of 20 and 26) as given. There is, however, no need at all to do so. If the total sample size M =
N × n were predetermined, it would be optimal (from a statistical point of view) to take as many
classes as possible with only one pupil per class. It is only because of budgetary constraints that a
two-stage sample may be preferred. Let us take c = 23 (taking an extra school is 23 times the price of
taking one extra student within an already sampled school). Running the software program PinT with
a budget constraint N(n + 23) ≤ 1,000 leads to the results in Table 11.2.
The optimum for the standard error associated with the effect of the treatment appears to be when
we take a sample of 9 pupils per school and 31 schools. In that case the standard error of interest
reaches its minimum of 0.09485. It will be clear from the outset that this standard error is too large to
reach a satisfactory power for this effect size of 0.20. Even with α as high as 0.10, the power will be
only 0.42.
The first solution to this problem would be to increase the budget. Applying formula (11.1) shows
that to reach a power of 0.80 with α = 0.10, the standard error should be made twice as small. This
implies a multiplication of the number of schools by 22 = 4, which amounts to 4 × 31 = 124 schools
in each of which 9 pupils are sampled. To make matters even worse, politicians do not like to spend
money on interventions that may not work, and for that reason it may be required to have α as low as
0.01. Applying formula (11.1) shows that the original standard error of 0.09485 should be brought
down to one third of its value, implying a required sample size of 32 × 31 = 279 schools.
If, instead of the optimum value n = 9, one used the earlier mentioned value of n = 23, the
standard error would become 0.10346/0.09485 = 1.09 times as large, which could be offset by
sampling 1.092 = 1.19 as many schools. To reach a power of 0.80 with α = 0.01 and n = 23, one
would need to sample 32 × 1.19 × 31 = 332 schools.

Table 11.2: Standard errors in case of budget constraints.

 This example illustrates that when one is interested in effects of macro-
level variables on micro-level outcomes it usually is sensible (in view of
reducing the standard error of interest) to sample many macro-units with
relatively few micro-units. The example given may be regarded as a cluster
randomized trial: whole classes of students are randomly assigned to either
the experimental (a class size of 20) or the control (a class size of 26)
condition. Raudenbush (1997) shows how much the standard error for the
experimental effect can be decreased (and thus the power increased) if a
covariate is added that is strongly related to the outcome variable. In this
case this would imply, in practice, to administer a pretest or intelligence
test.
The software program Optimal Design (Spybrook et al., 2009) provides
power calculations for a macro-level effect on a micro-level outcome with
varying intraclass correlations, significance levels, sizes of clusters,
numbers of clusters, and varying degrees of associations between the
covariate and the dependent variable. If one is interested in a macro-level
variable effect on a micro-level outcome, and the macro-units differ
considerably with respect to another macro-level variable, one has the
choice between matching and covariance adjustment. Raudenbush et al.
(2007) discuss this situation, pointing to the fact that matching in general
leads to a larger loss of degrees of freedom, which in general implies that
covariance adjustment results in higher power.

11.4.1 Cross-level interaction effects

Of particular interest in social research are cross-level interaction effects. If
one is interested in such an effect, how can one improve the power of the
test being used? Should one increase the sample size of the macro-units or
increase the sample size of the micro-units? If only the total sample size is
fixed (i.e., c = 0 in (11.3)), then it usually is optimal to take the sample size
for the macro-units as large as possible. This is illustrated in the following
example, taken from Snijders and Bosker (1993).
Example 11.5 Design for the estimation of a cross-level effect.
Suppose we wish to assess the effect of some school policy measure to enhance the achievement of
students from low socio-economic status families, while taking into account aptitude differences
between students. More specifically, we are interested in whether this policy measure reduces
differences between students from low and high socio-economic status families. In this case we are
interested in the effect of the cross-product variable POLICY × SES on achievement, POLICY being
a school-level variable and SES being a student-level variable. Assume that we control for the
students’ IQ and that SES has a random slope. The postulated model is

where Yij is the achievement of student i in school j.

The primary objective of the study is to test whether γ11 is 0. It is assumed that all observed
variables are scaled to have zero mean and unit variance. With respect to the explanatory variables it
is assumed that SES is measured as a within-school deviation variable (i.e., all school means are 0),
the intraclass correlation of IQ is 0.20, the correlation between IQ and SES is 0.30, and the
covariance between the policy variable and the school mean for IQ is –0.13 (corresponding to a
correlation of –0.3). With respect to the random effects it is assumed that the random intercept
variance is 0.09, the slope variance for SES is 0.0075, and the intercept–slope covariance is –0.01.
Residual variance is assumed to be 0.5. Some further motivation for these values is given in Snijders
and Bosker (1993).
Given these parameter values, the PinT program can calculate the standard errors for the fixed
coefficients γ. Figure 11.1 presents the standard errors of the regression coefficient associated with
the cross-level effect as a function of n (the micro-level sample size), subject to the constraint that
total sample size M is not more than 1,000. N is taken as M/n rounded to the nearest lower integer. As
n increases, N will decrease (or, sometimes, remain constant).
 The graph clearly illustrates that it is optimal to take n as small as possible, and (by implication)
N as large as possible, since the clustering of the data will then affect the standard errors the least.
This may appear to be somewhat counter-intuitive: in the two-step formulation of the model, at the
school level we regress the SES slope on the POLICY variable, which suggests that we need a
reliable estimate for the SES slope for each school, and that ‘thus’ n should be large. In this case,
however, one should realize that the cross-product variable POLICY × SES is simply a student-level
variable, and that as long as we have variation in this variable, the effect can be estimated. This is
even when n is as small as 1. Although to estimate intercept and slope variances we need n to be at
least 3.
This conclusion can be generalized to studies into the development over time of a certain
characteristic (cf. Chapter 15). A series of cross-sectional measurements for people of different ages,
for example, can be used to make assessments on growth differences between the sexes.

Figure 11.1: Standard error of a cross-level effect (for N × n ≤ 1,000).

We again assumed in the previous example that, given the total sample size, it is not extra costly
to sample many schools. This assumption is not very realistic, and therefore we will look at the
behavior of the relevant standard error once again but now assuming that the price of sampling an
extra school is c = 5 times the price of sampling a pupil, and the total budget in (11.3) still is K =
1,000. Running the software program PinT once again, we obtain the results presented in Figure 11.2.
Because of the budget restriction the total sample size M decreases and the number of groups N
increases as the group size n decreases. Balancing between N large (with n small) and N small (with
n large), we see that in this case it is optimal to take n somewhere between 15 and 20 (implying that
N is between 50 and 40, and M between 750 and 800). More precisely, the minimum standard error is
0.0286, obtained for n equal to 17 and 18. If one is willing to accept a standard error that is 5%
higher than this minimum, then n can be as low as 9.
 Figure 11.2: Standard error of a cross-level effect (for a budget restriction
with c = 5, K = 1,000).

11.5 Allocating treatment to groups or individuals

Up to now we have assumed that the design of a multilevel study involves
the choice of the number of macro-units and the number of micro-units
within those macro-units. In observational designs this is obvious. In many
experimental designs, such as the previous example where one is interested
in the manipulable macro-variable class size, this is also the case. In some
experimental designs, however, one has the choice of manipulating the
predictor variable either at the macro level defined by the clusters, or within
those clusters. This is known as the difference between cluster randomized
trials and multisite trials (Raudenbush and Liu, 2000; Moerbeek, 2005;
Moerbeek and Teerenstra, 2011). An example of such a situation is a
clinical trial in hospitals, which may be conducted at the level of hospitals
(some hospitals giving the experimental treatment to their patients and other
hospitals not) or at the level of groups of patients (some receiving the
experimental treatment and others not) within hospitals. Or to give an
educational example, an experiment where pupils learn words interactively
whilst working with a personal computer, which can be delivered at the
level of either whole classes or pupils within classes.
Suppose that the research investigates the difference between two
experimental treatments. If one determines the treatment at the macro level,
the power is determined by the residual intraclass correlation given the
treatment administered, ρI(Y∣X), and the number of micro-level units nj to
be sampled within each macro-level unit j (since these show up in the
design effect formula 1 + (nj – 1) ρI, given in (3.17)). In this case smaller
values for nj and ρI lead to improved power, if the total sample size N is
kept constant. On the other hand, if one randomizes within the macro-level
clusters, then the residual intraclass correlation coefficient has two parts:
the similarity of micro-units within the treatment groups within the clusters
(ρa) and the similarity of micro-units across the treatment groups within the
clusters (ρb). The latter can be regarded as the consequence of random
treatment-by-cluster interaction. The total residual intraclass correlation is
ρI = ρa + ρb. If for simplicity we assume that the two treatment conditions
both are allocated to half of the total sample, the power now is a function of
1 – ρa + (nj – 1) ρb (Moerbeek, 2005). This shows that a multisite trial is to
be preferred over a cluster randomized trial when considering power,
because 1 – ρa ≤ 1 and ρb ≤ ρI.
From a practical point of view, however, multisite trials may be more
difficult to implement. For example, one has to make sure that there will be
no contamination between the two treatment conditions within the same
site. On the other hand one can argue that a cluster randomized control may
lead to selection bias: for example, certain parents might subsequently
avoid sending their children to the control schools. A compromise between
these two situations is the so called pseudo-cluster randomized trial, where
randomization occurs first at the macro level and thereafter at the micro
level within the clusters. In this case relatively more micro-units are in the
treatment condition within those clusters that were randomly assigned to the
treatment condition, whereas the reverse is true for micro-units in the
control condition within those clusters that were randomly assigned to the
control condition (see Moerbeek and Teerenstra, 2011, for a further
discussion on this issue).
 11.6 Exploring the variance structure
In all the examples given so far, we have presented cases in which we
wished to assess either the mean value of a variable or the association
between two variables. Another series of interesting questions may have to
do with unspecified effects: does it matter which school a child attends? Do
neighborhoods differ in their religious composition? Do children differ in
their growth patterns? Is the effect of social status on income different in
different regions? In all these instances we are interested in the magnitude
of the variance components, and the power of the tests on the variance
components is affected by n and N in a different manner from the power of
the tests on the fixed effects.

11.6.1 The intraclass correlation

These example questions about unspecified effects can be answered in the
first instance by estimating or testing the corresponding intraclass
correlation coefficient. Note that in Sections 11.2 and 11.3, the minimum
required sample sizes were expressed in terms of the intraclass correlation
coefficient, so that also to determine those required sample sizes it is
necessary to estimate this coefficient.
In equation (3.13) the standard error for estimating the intraclass
correlation was given as

To investigate whether it is better to take many small groups or few large

ones, a graph of this formula can be drawn as a function of n, the group
size, assuming either a constant total sample size or the budget constraint
(11.3). For simplicity, let us forget that N must be an integer number, and
employ the sample size constraint M = N × n and the budget constraint N(n
+ c) = K as equations for computing N as a function of n.
For a fixed total sample size M, the substitution N = M/n implies that
the factor with the square root sign in (3.13) must be replaced by
The optimal group size is the value of n for which the resulting expression
is minimal. This optimum for a fixed total sample size depends on ρI. If ρI
is precisely equal to 0, the standard error decreases with n and the optimum
is achieved for N = 2 groups each of size n = M/2. This implies that to test
the null hypothesis ρI = 0, it is best to use larger group sizes. For ρI close to
1, on the other hand, the standard error increases with n and it is best to
have N = M/2 groups each of size 2. It is unpleasant that the optimum
values can be so totally different, depending on the value of ρI which is
unknown to begin with! However, in most cases the researcher will have
some prior knowledge about the likely values of ρI. In most social science
research, the intraclass correlation ranges between 0.0 and 0.4, and often
narrower bounds can be indicated. For such a range of values, the
researcher can determine a group size that gives a relatively small standard
error for all ρI values within this range.

Figure 11.3: Standard error for estimating the intraclass correlation

coefficient for M = 400.
 For a fixed total sample size M = 400, the standard error is graphed in
Figure 11.3 for several values of ρI. This figure shows that the optimum
group size does indeed depend strongly on the value of ρI. For example, if
ρI = 0.4, group sizes over 20 are unattractive whereas for ρI = 0.0 this is the
case for group sizes of fewer than 5. This suggests that if prior knowledge
implies that the intraclass correlation might be as small as 0.0 or as large as
0.4, and if the focus of the research is on estimating ρI while the total
sample size is fixed at 400, one should preferably use group sizes between 5
and 20.
Example 11.6 A good compromise for a fixed total sample size.
Suppose that, as in Figure 11.3, the total sample size must be close to M = 400, while it is believed to
be likely that the intraclass correlation ρI is between 0.05 and 0.2. Calculating the standard errors for
these two values of ρI shows that for ρI = 0.05 the smallest standard error is equal to 0.0308, achieved
for n = 19; for ρI = 0.2 the minimum is 0.0510, achieved for n = 6. This implies that it is best to use
an intermediate group size, not too high because that would lead to a loss in standard error for low
values of ρI, and not too low because that would entail a disadvantage for high ρI.
Inspecting the standard errors for intermediate group sizes (this can be done visually from Figure
11.3) shows that group sizes between 9 and 12 are a good compromise in the sense that, for each
value 0.05 < ρI < 0.2, the standard error for 9 ≤ n ≤ 12 is not more than 10% higher than the
minimum possible standard error for this value of ρI.

If the optimum group size is to be found for the budget constraint

(11.3), the value N = K/(n + c) must be substituted in expression (3.13) for
the standard error, and the standard error calculated or graphed for a
suitable range of n and some representative values of ρI.
Example 11.7 A good compromise for a budget constraint.
Suppose that the budget constraint must be satisfied with c = 10, so each additional group costs ten
times as much as an additional observation in an already sampled group, and with a total available
budget of K = 600. The extreme possible situations are then N = 50 groups of size n = 2, and N = 2
groups of size n = 290. Suppose that practical constraints imply that the group size cannot be more
than 20. The results obtained when substituting N = 600/(n + 10) in (3.13) and calculating the
resulting expression are shown in Figure 11.4.
 Figure 11.4: Standard error for estimating the intraclass correlation
coefficient for a budget constraint with c = 10, K = 600.
The figure shows that for larger values of ρI, the optimum group size has an intermediate value
(e.g., n = 8 for ρI = 0.4), but the standard error does not depend very strongly on the group size as
long as the group size is at least 4. For the smaller values of ρI, it is best to take the group size as
large as possible within the allowed range. Group sizes n = 17 and 18 give a good compromise
between what would be best for ρI as small as 0.0 and what would be best for ρI = 0.4, but smaller
group sizes may also be reasonable.
Now suppose, to modify the example, that from the point of view of estimating fixed effects it is
desirable to use smaller group sizes, and that it is likely that ρI is at least 0.05. On the basis of this
figure, one could then decide to sacrifice some of the standard error for low values of ρI, and use a
group size as low as 8 or 10.

11.6.2 Variance parameters

For a two-level random intercept model, approximations to the estimation
variances of the level-one variance σ2 and the intercept variance are
given by Longford (1993, p. 58). For the standard errors, these formulas are
and

If one prefers standard errors for the estimated standard deviations,

according to formula (6.2) these formulas can be transformed by

and similarly for S.E. .

The standard error of is minimized by using large level-one sample
sizes n, but in practice the standard error is, for a given reasonable budget
constraint, so little sensitive to the value of n that one does not really care
about obtaining precisely the minimal possible standard error for this
parameter (this is also the conclusion reached by Cohen, 1998). To obtain a
good design for estimating the variance parameters, the intercept variance is
more important.
If optimal sample sizes are to be determined to estimate the intercept
variance, the same approach can be applied as in the preceding section. One
can plot these standard errors against n, with M = N × n fixed to a given
value, for various values of the intraclass correlation coefficient ρI = / (σ2
+ ). If there is a budget constraint, one can substitute N = K /(n + c) into
Longford’s approximation expressions and plot the standard errors while
keeping N × (n + c) fixed. An example is given by Cohen (1998, Section
5.1).
Standard errors of random slope variances are quite complicated
functions of the sample sizes, the variation of the explanatory variables
within and between the groups, and the variances themselves. Further
research is needed for elaborating guidelines on how to choose sample sizes
appropriately for the estimation of variance parameters in random slope
models. If precise estimation of such parameters is required, it seems
advisable to use large sample sizes (30 or higher) at either level and to make
sure that the explanatory variable of which one wishes to estimate the
random slope variance has enough dispersion within the level-two units.
This may be achieved, for example, by some kind of stratification.
For complicated models for which no formulas are available for the
standard errors, Monte Carlo simulation is an alternative way to form an
impression about standard errors. Simulation analysis of multilevel models
is possible in various programs, such as MLwiN, MLPowSim (Browne et
al., 2009), and the R package pamm (see Chapter 18).

11.7 Glommary
Hypothesis testing. For the general theory of statistical hypothesis testing
we refer to the general statistical literature, but in this glommary we
mention a few of the basic concepts in order to refresh the reader’s
memory.

Null hypothesis. A set of parameter values such that the researcher would
like to determine whether these are tenable in the light of the currently
analyzed data, or whether there is empirical evidence against them; the
latter conclusion is expressed as ‘rejecting the null hypothesis’.

Alternative hypothesis. A set of parameter values different from the null

hypothesis, such that if these values were obtained, the researcher
would like the statistical test to conclude that there is evidence against
the null hypothesis. In practice, the alternative hypothesis often
embodies the theory proposed or investigated by the researcher, and
the evidence against the null hypothesis will also be interpreted as
support for the alternative hypothesis.

Effect size. Numerical expression for the degree to which a parameter value
is believed to deviate from the null hypothesis. This is an ambiguous
concept, because the importance of deviations from a null value may
be expressed in so many ways. Cohen (1992) gives a list of definitions
of effect sizes.
To express differences between groups, the difference between the
 means divided by the within-group standard deviation is a popular
effect size, called Cohen’s d. Rosenthal (1991) proposes to express
effect sizes generally as correlation coefficients. For multilevel
analysis this can be exemplified by noting that variances on their
original scales have arbitrary units and therefore are not useful as
effect sizes, but the intraclass correlation coefficient transforms the
within- and between-group variances to the scale of a correlation
coefficient, and thereby becomes comparable across data sets and
useful as an effect size index.

Type I error. Mistakenly rejecting the null hypothesis based on the

observed data whereas in actual fact the hypothesis is true.

Significance level. Also called type I error rate, and usually denoted α, this
is the probability of making a type I error, given parameter values that
indeed conform to the null hypothesis. In classical statistical
hypothesis testing, the significance level is predetermined by the
researcher (e.g., α = 0.05; the popularity of this value according to the
statistician R.A. Fisher is caused by the fact that humans have 20
fingers and toes). In practice, significance levels are set at lower values
accordingly as sample sizes are higher, in view of the tradeoff
mentioned below.

Type II error. Failing to reject the null hypothesis when in fact the null
hypothesis is false, and the alternative hypothesis is true.

Type II error rate. Often denoted β, this is the probability of making a type
II error, given parameter values that indeed conform to the alternative
hypothesis. There is a tradeoff between the probabilities of making
type I and type II errors: having a higher threshold before rejecting the
null hypothesis leads to lower type I and higher type II error rates.

Power. This is the probability 1 – β of correctly rejecting the null

hypothesis given parameters that conform to the alternative hypothesis.
Power increases as the sample size increases and as effect size
increases. But power is also dependent on the type I error rate: a
smaller level of significance (smaller type I error rate) will lead to less
power (higher type II error rate).
Design effect of a two-stage sample. The ratio of the variance of
estimation obtained with the given sampling design, to the variance of
estimation obtained for a simple random sample from the same
population, and with the same total sample size.
Required sample sizes for two-stage samples, needed to obtain a given
standard error of the estimated population mean, and thereby a given power
for distinguishing the population mean from a hypothesized value, can be
calculated using the design effect.

Power analysis. A study in which one investigates the power of a statistical

test to be employed, as a function of the type I error rate α, the sample
size, and the effect size. Typically in multilevel studies there are two
other considerations: the sample size of the macro-level units of
clusters versus the size of the sample of the micro-level units within
those clusters, and the residual intraclass correlation, being the degree
of resemblance of the micro-level units on the variable of interest, after
conditioning on a set of covariates.

Cost function. The cost of the study design in terms of, for example, the
sample size. In multilevel situations one weighs the costs of sampling
macro-level units against the costs of sampling micro-level units
within macro-level units. Such a cost function can then be used to
determine optimal sample sizes at either level given a budget
constraint.

Cluster randomized trial. An experiment in which whole clusters are

randomly assigned to either the intervention or control condition. A
typical case is a study in which one assigns schools to conditions and
studies the relevant outcomes of the intervention at the level of the
pupils. The risk is selection bias if, for instance, certain groups of
parents subsequently avoid sending their children to the control
schools.

Multisite trial. An experiment in which groups of micro-level units are

randomly assigned to the experimental conditions within the clusters.
A typical case is a study in which patients within hospitals are
assigned to either an experimental treatment or a standard treatment.
The risk is contamination: the physician may apply aspects of the
 experimental treatment to the patients in the control group, without
meaning to do so.

Pseudo-cluster randomized trial. A compromise between the cluster

randomized trial and the multisite trial to avoid the risk of both
contamination and selection bias as much as possible. In this case
clusters are randomly assigned to treatment and control conditions, and
thereafter micro-level units are randomly assigned to either the
experimental or the control condition within their cluster, with a
probability depending on whether the cluster is in either the
experimental or control condition. In this case a patient may be in the
experimental group although the hospital as such was chosen to be in
the control group, and vice versa.
 12
Other Methods and Models

Several alternative methods for estimation exist, as well as several models

closely related to the hierarchical linear model, which can be used in
conjunction with, or as alternatives to, the maximum likelihood-based
methods presented in the other chapters. Two alternative estimation
methods and one alternative model are briefly presented in this chapter.

OVERVIEW OF THE CHAPTER

We begin by introducing the Bayesian approach, with the associated
Markov chain Monte Carlo algorithms. This is an alternative estimation
method which is more time-consuming but also more flexible, and can be
used to estimate more complex models than the maximum likelihood
procedures. A global overview is given, with pointers to the literature.
Then we mention the so-called sandwich estimators for standard errors,
also called robust standard errors, because they give some degree of
robustness to deviations from the model assumptions. Of interest here are
the cluster-robust standard errors used in the approach of generalized
estimating equations, where parameters are estimated according to a
traditional single-level technique and combined with robust standard errors
that control for the clustered nature of the data.
We conclude the chapter by a brief treatment of latent class models,
which can be used as multilevel models in which the random effects do not
have normal distributions but discrete distributions. If the discrete
distribution for the random effects has a lot of values, then such a
distribution can approximate any distribution. Therefore this can be
regarded as a nonparametric or semi-parametric approach, being a
combination of a linear or generalized linear model for the mean structure
and a nonparametric model for the distribution of the random effects.

12.1 Bayesian inference

The Bayesian approach to statistics is based on a different interpretation of
probability than the more usual frequentist approach. Probability in a
Bayesian sense expresses the belief in how likely it is that an event will
occur or that a variable has a value in a certain range, a belief which can
range from generally shared to subjective. In the frequentist approach,
probability is the relative frequency of occurrence of an event in a
hypothetical large number of independent replications; this approach is the
unmentioned sea in which we are swimming in the rest of this book.
Since the 1980s the Bayesian approach has received increased attention
because of the development of computer-intensive so-called Markov chain
Monte Carlo (MCMC) methods which allow very flexible modeling in the
Bayesian paradigm. This has made Bayesian statistical procedures available
for quite complicated statistical models, and for models that are hard to
manage by frequentist procedures such as maximum likelihood (ML)
estimation. An additional advantage of the Bayesian approach is how it
handles what statisticians sometimes call error propagation, the
consequences of uncertainty in various parameters, or aspects of model
specification, for the inferential conclusions drawn. The philosophical
sharpness of the contrast between the Bayesian and frequentist inferential
paradigms has subsided because of a convergence of interpretations and
procedures. One reason is that Bayesian procedures often have good
frequentist properties for medium and large sample sizes. Another reason is
the possibility, in various models, of interpreting probabilities in a
frequentist as well as a Bayesian way; for example, the probability
distribution of random effects in multilevel models. Treatments of all these
issues can be found in recent good textbooks on Bayesian statistics, for
example, Gelman et al. (2004) and Jackman (2009).

Outline of Bayesian statistics

A basic element in Bayesian statistics is that the parameters of the statistical
model, of which the values are unknown, have a probability distribution,
expressing the beliefs and uncertainty of the researcher as to their likely
values. The probability distribution of the parameters, entertained by the
researcher before having seen the data, is called the prior distribution.
Observing the data, and applying the statistical procedures, transforms this
into the posterior distribution. This transformation is effectuated formally
through Bayes’ rule, which is a direct application of the rules of conditional
probability. A major question for Bayesian inference is the choice of the
prior distribution. If the researcher has good prior knowledge about the
parameters, then this can be reflected by a corresponding choice of the
prior. If there is no such prior knowledge, however, it is desirable that the
inferential results are not sensitive to the choice of the prior distribution
within a reasonable range. The choice of adequate priors is a matter of
ongoing debate and research.
The basic correspondence between the posterior distribution and
frequentist inferential procedures is as follows:

As a point estimate, the mean or the mode of the posterior distribution

can be used: the posterior mean or posterior mode.

As a standard error, the standard deviation of the posterior distribution

can be used: the posterior standard deviation.

A confidence interval of confidence level 1 – α can be obtained as an

interval of parameter values that has probability 1 – α under the
posterior distribution.

A hypothesis test can be obtained by rejecting the null hypothesis if the

parameter value defining the null hypothesis is not included in the
confidence interval so obtained; and the p-value is the smallest value
of α such that the null parameter value is not included in the
confidence interval of confidence level 1 – α.

An important result underlying the convergence between Bayesian and

frequentist statistics is that these transformations of Bayesian to frequentist
procedures have approximately the desired frequentist properties (e.g.,
correct type I error rates of hypothesis tests) for medium and large sample
sizes (Gelman et al., 2004, Chapter 4).
Bayesian MCMC procedures are simulation-based algorithms
producing values that can be used as a sample, that is, a sequence of
random draws, from the posterior distribution. An important issue here is
the convergence of the algorithm, and the dependence between
consecutively produced values. Convergence is an issue because the
algorithm produces a sequence of values that converges to the posterior
distribution; for the usual procedures at any given moment there is no full
certainty that the algorithm has come close enough to convergence, and
convergence diagnostics must be used to check this. Dependence is an issue
because the consecutive values are a Markov chain and will be dependent,
therefore a larger sample of draws from the posterior distribution will be
required than would be the case if the draws were independent. The checks
and requirements for coping with these issues depend on the
implementation of the algorithm. More and more computer programs for
multilevel analysis have incorporated Bayesian MCMC procedures, and
these issues are treated in the manuals.
There is a similarity between the Bayesian approach and the
hierarchical linear model, which already appeared in the use of empirical
Bayes methods in Chapter 4. This can be explained as follows. The
Bayesian model for statistics has a hierarchy where first the parameter is
drawn from the prior, and then, for this parameter value, the data are drawn
from the data-generating distribution. This can be symbolized by

where θ is the parameter, Y is the data, the vertical bar | denotes

conditioning, and Pθ is the probability distribution of the data given
parameter value θ. The hierarchical linear model has the same type of
hierarchy: in a frequentist setup, the two-level model can be symbolized as
where U is the vector of all level-two random effects, Qτ its distribution
with parameter τ, and Pθ (U) is the probability distribution of the data,
conditional on the random effects U, and given parameter θ. In a Bayesian
approach, the two hierarchies can be nicely superimposed:

Here τ contains the parameters of the level-two model, and θ those of the
level-one model.
Browne and Draper (2000) presented Bayesian MCMC methods for
linear and nonlinear multilevel models; see also further publications by
these authors, presented in Draper (2008). In extensive simulation studies,
Browne and Draper (2006) found confirmation for the good frequentist
properties of Bayesian methods for multilevel modeling. Concise
introductions to Bayesian methods for multilevel modeling are given in
Raudenbush and Bryk (2002, Chapter 13), Gelman et al. (2004, Chapter
15), Gelman and Hill (2007, Chapter 18), and Hamaker and Klugkist
(2011). More extensive treatments are Draper (2008) and the textbook by
Congdon (2010).
For the regular (linear) hierarchical linear model, Bayesian and ML
procedures yield very similar results. Jang et al. (2007) present an extensive
comparative example. An advantage of the Bayesian approach may be the
more detailed insight into the likely values of parameters of the random part
of the model, because of the deviation from normality of their posterior
distributions; in frequentist terms, this is reflected by nonnormal
distributions of the estimators and by log-likelihoods, or profile log-
likelihoods, that are not approximately quadratic. For more complicated
models (e.g., with crossed random effects or with nonnormal distributions),
Bayesian approaches may be feasible in cases where ML estimation is
difficult, and Bayesian approaches may have better properties than the
approximations to some of the likelihood-based methods proposed as
frequentist methods. Examples of more complicated models fitted by a
Bayesian approach are given in Chapter 13 and in Browne et al. (2007).
 12.2 Sandwich estimators for standard errors
For many statistical models it is feasible and sensible to use ML estimators
or their relatives (e.g., restricted maximum likelihood (REML) estimators).
Statistical theory has a standard way of assessing the precision of ML
estimators if the model is correct. This uses the so-called Fisher information
matrix, which is a measure for how strongly the probabilities of the
outcomes change when the parameters change. The key result is that the
large-sample covariance matrix of the ML estimator is the inverse of the
information matrix. Recall that standard errors are contained in the
covariance matrix of the estimator, as they are the square roots of its
diagonal elements. For the hierarchical linear model, this is discussed in de
Leeuw and Meijer (2008b, pp. 38–39).
Sometimes, however, the researcher works with a misspecified model
(i.e. one that is not a good approximation of reality), often because it is
easier to do so; sometimes also the estimation was done by a method
different from ML. In such cases, other means must be utilized to obtain
standard errors or the covariance matrix of the estimator. A quite general
method is the one affectionately called the sandwich method because the
mathematical expression for the covariance matrix features a matrix
measuring the variability of the residuals, sandwiched between two
matrices measuring the sensitivity of the parameter estimates to the
observations as far as these are being used for obtaining the estimates. The
method is explained in many places, for example, de Leeuw and Meijer
(2008b, Section 1.6.3) and Raudenbush and Bryk (2002). It was proposed
by White (1980), building on Eicker (1963) and Huber (1967), with the
purpose of obtaining standard errors for ordinary least squares (OLS)
estimates valid in the case that the linear model holds, but the assumption of
independent homoscedastic (constant-variance) residuals is not valid.
Therefore he called them robust standard errors.
One of the applications is to multilevel structures. There the sandwich
estimator yields cluster-robust standard errors (Zeger et al., 1988). The
term clusters here refers to the highest-level units in a multilevel data
structure. Thus, even for a multilevel data structure where a linear model is
postulated and one suspects within-cluster correlations, one could estimate
the parameters of the linear model by ordinary least squares, which is the
ML estimator for the case of independent homoscedastic residuals; and then
obtain the standard errors from an appropriate sandwich estimator to
‘correct’ for the multilevel structure in the data. This can also be done for
three- and higher-level structures, where the clusters are defined as the
highest-level units, because these are the independent parts of the data set.
This has been elaborated in the technique of generalized estimating
equations (GEE), proposed by Liang and Zeger (1986) and Zeger et al.
(1988); see the textbook by Diggle et al. (2002). Formulated briefly, this
method assumes a linear or generalized linear model for the expected values
of the dependent variable in a multilevel data structure, conditional on the
explanatory variables; but no assumptions are made with respect to the
variances and correlations, except that these are independent between the
clusters (the highest-level units). The parameters of the linear model are
estimated under a so-called ‘working model’, which does incorporate
assumptions about the variances and correlations; in the simplest case, the
working model assumes uncorrelated residuals with constant variance. The
standard errors then are estimated by a sandwich estimator. For large
numbers of clusters, this will be approximately correct, provided that the
linear model for the means is correct and the highest-level units are
independent. The comparison between GEE modeling and the hierarchical
linear model is discussed by Gardiner et al. (2009).
Thus, the sandwich method for obtaining standard errors is a large-
sample method for use when the model used for estimation is misspecified,
or when a different estimator than the ML or REML estimator is used, as in
the methods explained in Section 14.5 taking account of design weights in
surveys. The sandwich estimator will be less efficient than the ML or
REML estimator if the model is correctly specified, and a practical question
is for which sample sizes it may be assumed to work in practice. This will
also depend on the type of misspecification, and here we should distinguish
between misspecification of the correlation structure, such as GEE
estimation of a data set satisfying a random slope model with a working
model of independent residuals, and misspecification of the distributional
form of the residuals, for which the default assumption of normality might
be violated.
The quality of cluster-robust standard errors is better when the clusters
have equal sizes and the same distributions of explanatory variables, and
suffers when the clusters are very different. Bell and McCaffrey (2002)
proved that the cluster-robust sandwich estimator for the GEE with the OLS
working model is unbiased if all clusters have the same design matrix, and
otherwise has the tendency to underestimate variances. Corrections to the
sandwich estimator for standard errors were proposed by Mancl and
DeRouen (2001), Bell and McCaffrey (2002), and Pan and Wall (2002).
These corrections lead to clear improvements of the type I error rates of
tests based on cluster-robust standard errors, particularly for clusters with
unequal sizes or different distributions of explanatory variables. These
corrections should be implemented more widely as options, or even
defaults, in software implementations of cluster-robust standard errors. The
minimum correction one should apply in the case of a small number of
clusters is the following very simple correction mentioned by Mancl and
DeRouen (2001): multiply the estimated covariance matrix as defined by
White (1980) by N /(N – q – 1) (this is the HC1 correction of McKinnon
and White, 1985), and test single parameters against a t distribution with N
– q – 1 degrees of freedom (rather than a standard normal distribution), and
multidimensional parameters against an F distribution with N – q – 1
degrees of freedom in the denominator (rather than a chi-squared
distribution). Here N is the number of clusters and q the number of
explanatory variables at the highest level. In a simulation study for two-
level binary data, Mancl and DeRouen (2001) found that this gives
satisfactory results for N = 40, and in the case of equal cluster sizes and
identical distributions of explanatory variables per cluster, for considerably
smaller N. But it is preferable to use one of the more profound and more
effective corrections.
Verbeke and Lesaffre (1997) proved that for a hierarchical linear model
in which the random part of the model is correctly specified but has
nonnormal distributions, an appropriate sandwich-type estimator is
consistent, and found support for its performance for moderate and large
sample sizes. Yuan and Bentler (2002) also derived sandwich estimators for
the hierarchical linear model, together with rescaled likelihood ratio tests
for use under nonnormality of residuals. Maas and Hox (2004) reported a
simulation study designed to assess the robustness of the sandwich standard
error estimator for nonnormal distributions of the level-two random
coefficients. They compared the ML estimator and the sandwich estimator,
as implemented in MLwiN, for linear models where the covariance
structure is correctly specified but the distributions of the level-two random
effects are nonnormal. For the fixed effects they found good coverage rates
of the confidence intervals based on either standard error for a small
number (30) of clusters. For confidence intervals for the level-two variance
parameters they found that the sandwich estimator led to better coverage
rates than the ML estimator, but even for 100 clusters the coverage rates
based on the sandwich estimator were still far from satisfactory for some of
the nonnormal distributions. However, Verbeke and Lesaffre (1997)
reported better results for a differently defined sandwich estimator under
similar distributional assumptions. It can be concluded that the robustness
of the sandwich estimator for inference about variance parameters depends
on the number of clusters being sufficiently large, or on the appropriate use
of specific small-sample versions.
Next to the sandwich estimators, other estimators for standard errors
also are being studied in the hope of obtaining robustness against deviations
from the assumed model. Resampling methods (van der Leeden et al.,
2008) may be helpful here. Cameron et al. (2008) proposed a bootstrap
procedure that offers further promise for cluster-robust standard errors.
To decide whether to use an incompletely specified model and the
associated sandwich estimator, the following issues may be considered:

1. Can the research questions be answered with the incompletely

specified model? In practice, the model will specify the fixed
effects and not attempt a correct specification of the variance and
correlation structure. In Section 2.2 we argued that the variance
and correlation structure, with its parameters such as the intraclass
correlation and issues such as heteroscedasticity, may often be of
interest, even if they may lie outside the primary scope of the
originally conceived aims of the researcher.
An issue in the use of partially specified models is that only those
parameters can be meaningfully studied for which there is a
correspondence between the true model and the model used for
estimation (the ‘working model’ of the GEE approach). Freedman
(2006) goes further, and doubts the wisdom of working with
misspecified probability models. One could say that a minimum
requirement for the correspondence between parameters in the
working model and the true model is consistency and approximate
unbiasedness of the parameter estimates. In the example where the
 true model is a hierarchical linear model as in (5.15) and the
working model is the linear regression model with independent
homoscedastic residuals, there is such a correspondence for the
fixed effects but not for the parameters of the random part. A case
where the correspondence does extend to the parameters of the
random part is when the covariance structure of the random part is
well specified but the distributions are nonnormal, as in Verbeke
and Lesaffre (1997) and Maas and Hox (2004).
2. If one uses a completely specified model, is there enough
confidence that its assumptions are a reasonable approximation to
reality? To answer such questions, diagnostic methods and model
improvements as discussed in Chapter 10 can be used.
3. Does the sandwich estimator provide approximately correct
standard errors of an approximately unbiased estimator? Here the
finite-sample corrections discussed above are relevant. It should be
realized that if the model is approximately correct, REML
estimators of standard errors will be more precise (have smaller
variability) than the sandwich estimator. For 40 or more clusters,
the sandwich estimator may be expected usually to have good
properties if the HC1 correction and finite-sample degrees of
freedom are employed, as mentioned above, unless there is great
imbalance between the clusters. For fewer clusters, this depends on
the design and parameter values, and one may need additional
arguments or better corrections to justify the use of the sandwich
estimator.
4. Is the loss of efficiency (of parameters and tests) acceptable as
compared to the efficiency of using ML methods under an
approximately correctly specified model – assuming that it is
feasible to obtain such a model? The loss of efficiency will depend
on the study design, the imbalance between the clusters, and the
parameters of the random part.
Example 12.1 Sandwich estimator for the language test.
Here we take up the model of Example 5.4. In the GEE approach, the parameters of the linear model
are estimated by OLS, and the standard errors by the sandwich estimator. For 211 groups, the
correction methods mentioned above are small; with q = 3 level-two variables, the correction factor
of the HC1 method is N /(N – q – 1) = 1.02.
 Table 12.1 presents the ML estimates for the hierarchical linear model
alongside the OLS estimates (with REML estimate for the residual
variance) and the sandwich estimates for the standard errors. The difference
between the OLS estimates and those from the hierarchical linear model
seem small, but for the classroom average of IQ the difference of 0.986 –
0.760 = 0.226 is more than 75% of a standard error, and therefore not really
small. The estimates from the hierarchical linear model here may be
considered to be more efficient, because they take account of the
dependence between individuals in the same class. The sandwich standard
errors are very close to the standard errors from the hierarchical linear
model.

Table 12.1: Estimates for the hierarchical linear model compared with OLS
estimates and sandwich standard errors.

12.3 Latent class models

The usual assumption for random coefficients in a hierarchical linear model
is that they have normal distributions (Chapter 10); this is sometimes
weakened to heavier-tailed distributions such as t distributions (e.g., Seltzer
and Choi, 2003) or mixtures of normals (Verbeke and Lesaffre, 1996;
Verbeke and Molenberghs, 2000, Chapter 12). Another possibility is to
assume that these random coefficients have arbitrary discrete distributions.
Consider the two-level hierarchical linear model as defined in (5.15):

If the vector of random coefficients (U0j, U1j,. . ., Upj) has a discrete

distribution with C values u1,. . ., uC, then each level-two unit has one of
these values, so that the population of all level-two units can be divided into
C categories, the so-called latent classes, each with a separate value uc (c =
1,. . ., C). This can be called a latent class model or a finite mixture model,
because the distribution is a mixture of C basic distributions. Such models
were studied by, among others, Wedel and DeSarbo (2002), Vermunt
(2004), and Rabe-Hesketh et al. (2005). Good and extensive treatments are
given by Vermunt (2008) and Muthén and Asparouhov (2009). A similar
kind of modeling can be applied to generalized linear models where the
dependent variable is not normally distributed. Such models are included in
the very general framework of Skrondal and Rabe-Hesketh (2004, Chapter
4). They can be fitted by software such as Latent Gold, gllamm, Mplus, and
WinBUGS. A theoretical exposition and some examples for discrete and
continuous dependent variables, for two and three levels, are given in
Vermunt (2011).
If the number C of latent classes is allowed to be large enough, this
model yields the nonparametric ML estimator for the hierarchical linear
model, that is, the ML estimator for the model where the hierarchical linear
model is specified as in (5.15) but without any assumption as to the
distribution of the random coefficients U0j, U1j,. . .,Upj (Laird, 1978). ML
estimation has a tendency toward overfitting, however, meaning here
toward overestimating the number of latent classes. More latent classes will
always yield a better fit in the sense of a larger likelihood (i.e., a smaller
deviance). Therefore it may be preferable to choose the number of latent
classes based on an information criterion such as the Bayesian information
criterion (BIC); see, for example, Skrondal and Rabe-Hesketh (2004,
Section 8.4). The BIC is defined by

where p is the number of parameters and ln(M) is the natural logarithm of

the total sample size. It is used by selecting the model for which the BIC is
minimal; this gives a tradeoff between good fit (low deviance) and a small
number of parameters.
Lukočienė and Vermunt (2007, 2009) compared the model with
normally distributed random effects with the latent class model for two-
level logistic regression models. They found that for the random intercept
model the latent class model performs very well, even for random effects
that in reality have a normal distribution. For a model with a random slope
it appeared that the latent class model combined with the ML estimator is
less satisfactory; the model with normally distributed random effects, as
well as the latent class model where the number of classes is estimated by
the BIC, both have good properties, and which of these is preferable
depends on the true random effects distribution. It should be noted that this
is specific to the model for binary outcomes, and does not carry over to
models that have, within the latent classes, normally distributed dependent
variables.
Example 12.2 Latent class version of the model for the language test.
Again we take up the model of Example 5.4, and now estimate it with a latent class model, that is, a
model with the same linear specification but a discrete distribution for the random coefficients (U0j,
U1j). If the number of classes is C, this means that there are C possible values for the pair (U0j, U1j),
each with their own probability. The software ‘Latent Gold’ was used.
The estimation for these models has the problem of local optima: the estimation algorithm finds a
solution which is not sure to be the best (the best is the global optimum). Therefore the algorithm is
executed several times with independent starting values, and one hopes that the best solution is
indeed found.
Table 12.2 gives the deviance, the number of parameters p, and the BIC, as a function of the
number of latent classes C; this data set has a total sample size of M = 3,758. The deviance becomes
smaller and smaller as C increases, which is usual. The BIC favors a model with C = 3 latent classes,
a very low number. By way of comparison, the deviance of the model with normally distributed
random coefficients is 24,626.8 with p = 11 parameters (see Table 5.4), yielding BIC = 24,717.3,
considerably smaller than for any of the latent class models. This shows that for this data set, the
model with normally distributed random coefficients is somewhat more satisfactory than a latent
class model. The parameter estimates for the fixed effects for latent class models with C ≥ 4 are close
to the parameter estimates in Table 5.4.
 Table 12.2: Deviance and BIC values.

12.4 Glommary
Bayesian statistics. An approach to statistical inference based on the notion
that probability is a measure of how likely one thinks it is that a given
event will occur or that a variable has a value in a given range. In
particular, beliefs and uncertainties about parameters in a statistical
model are expressed by assuming that these parameters have a
probability distribution.

Frequentist statistics. The more usual approach to statistical inference

based on the notion that probability is the relative frequency of
occurrence of an event in a hypothetical large number of independent
replications. Parameters are considered just as unknown numbers or
vectors.

Prior distribution. The probability distribution of the parameters,

entertained by the researcher before having seen the data.

Posterior distribution. The probability distribution of the parameters,

entertained by the researcher after having seen and analyzed the data.

Posterior mean. The expected value of the posterior distribution, used as a

point estimate of the parameters.

Posterior standard deviation. The standard deviation of the posterior

distribution, used as a measure of posterior uncertainty about the
parameters.

Markov chain Monte Carlo (MCMC) methods. Simulation-based

algorithms that produce values that can be used as a sample, that is, a
sequence of random draws, from the posterior distribution.
Bayesian procedures for multilevel modeling. These are described in
various articles and specialized textbooks.

Sandwich estimator. A special type of estimator for standard errors, or

more generally of the covariance matrix of an estimator. This estimator
generally is more robust against deviation from assumptions
concerning the random part of the model. A special case is the cluster-
robust standard error estimator, which is designed to keep account of
the multilevel dependencies (also referred to as dependencies due to
clustering) even if parameters were estimated by a single-level
method. For large numbers of units at the highest level, this standard
error estimator performs quite well.

Generalized estimating equations (GEE). An estimation method

assuming a linear or generalized linear model for the expected values
of the dependent variable in a multilevel data structure, conditional on
the explanatory variables, but making no specific assumptions with
respect to the variances and correlations.

Multilevel latent class models. Variations of the hierarchical linear model,

where the random effects have a discrete rather than a normal
distribution. This can be regarded as a nonparametric or semi-
parametric approach, because discrete distributions with a sufficient
number of categories can approximate any distribution. Another name
for latent class models is mixture models.
 13
Imperfect Hierarchies

Up to now we have treated multilevel data structures, and statistical models

to handle these, in which a lower-level unit (e.g., a pupil) is perfectly nested
in one, and only one, higher-level unit (e.g., a school). Reality, of course, is
often far more complicated. Pupils not only belong to schools but also to
neighborhoods, and one may be interested in assessing the relative effects
on pupils’ citizenship behavior in both social contexts simultaneously.
Studying interpersonal teacher behavior implies asking pupils how they rate
their teacher on several dimensions. But classes of pupils are taught by
more than one teacher, and teachers teach more than just one class. Studies
about effects of secondary schools on pupil achievement are based on the
idea of ‘value added’, that is, the gain a pupil has made during a certain
period (say, five years). But pupils sometimes switch secondary schools
during their school career, so more than one secondary school is related to
their school success. The aforementioned situations are referred to as
imperfect hierarchies. Specific models have been developed as special
cases of the hierarchical linear model to deal with these structures. They are
called cross-classified models, multiple membership models, and multiple
membership multiple classification models. In models for cross-classified
data a lower-level unit belongs uniquely to one higher-level unit of the first
type (e.g., a school) and also uniquely to one higher-level unit of the second
type (e.g., a neighborhood), but the two types of unit are not nested either
way. In models for multiple membership data the higher-level units are all
of one type (e.g., schools), but the lower-level units (pupils in this example)
may belong to more than one higher-level unit. In multiple membership
multiple classification models we have both situations at the same time. In
this chapter we will introduce these models for imperfect hierarchies.
 OVERVIEW OF THE CHAPTER
In this chapter we first treat cross-classified data structures, implying that a
model with crossed random factors is needed to deal with it. After
introducing a two-level model with crossed random intercepts, a three-level
model with cross-classifications is treated. Multiple membership models are
then explained, and some special attention is given to how to weight the
duration of membership of a lower-level unit in a higher-level unit. Having
presented and clarified the models for cross-classifications and multiple
membership, we then introduce the straightforward combination of these:
the multiple membership multiple classification model. These models may
become quite complex, and therefore brief mention is made of the Bayesian
approach and the associated Markov chain Monte Carlo estimation
algorithm, which was treated in Section 12.1.

13.1 A two-level model with a crossed random

factor
Consider the case of cross-classified data, where individuals belong to
different social units (e.g., schools, neighborhoods, families, peer groups,
sports clubs, companies) at the same time. One can think of this situation as
crossed factors, that is, individuals are nested within companies but also
within neighborhoods, and these two factors are crossed. Each individual
uniquely belongs to one combination of both: a specific company as well as
a specific neighborhood. These two factors are not nested, as not all
employees of a given company live in the same neighborhood, nor do
employed people living in a certain neighborhood all work in the same
company. One can think of this situation as lower-level units being nested
within two factors that cross, or as a two-level data structure (that handles
one of the levels of nesting) with another crossed factor (representing the
other level of nesting).
To understand the situation a classification graph as suggested by
Rasbash and Browne (2001, 2008) may be helpful, as presented in Figure
13.1.
 Figure 13.1: An example of pupils nested within a cross-classification of
schools and neighborhoods.

In this figure, 25 pupils are nested within five schools, but also within
four different neighborhoods. The lines indicate to which school and
neighborhood a pupil belongs. The seven pupils of school 1 live in three
different neighborhoods. The six pupils who live in neighborhood 3 attend
four different schools. Another way to present the cross-classification is by
making a cross-table, which of course readers can do themselves.
Following up on this example, we now continue with an explanation of
how to incorporate a crossed random factor in a multilevel model. We
consider the case of a two-level study, for example, of pupils (indicated by
i) nested within schools (indicated by j), with a crossed random factor, for
example, the neighborhood in which the pupil lives. The neighborhoods are
indexed by the letter k, running from 1 to K (the total number of
neighborhoods in the data). The neighborhood of pupil i in school j is
indicated by k(i, j). The hierarchical linear model for an outcome variable
Yij without the neighborhood effect is given by

(cf. (5.15)). The random effect of neighborhood k, indicated by W0k, is

simply added to this model:
 There is some ambiguity concerning the second level, as is represented
by the notation Yi(j, k). There are two indices at the second level, j as well as
k, denoting the nesting in school j and simultaneously in neighborhood k.
The nesting in schools is represented by the random effects Uhj like in the
preceding chapters. For the neighborhood effects W0k, the usual assumption
is made that W01,…, W0K are mutually independent random variables, also
independent of the other random effects (U0J and Rij), W0k having the
normal distribution with mean 0 and variance τ2w. The interpretation is just
like that of other random effects: part of the variability in the dependent
variable is accounted for by neighborhoods, and these are regarded as a
random sample from a (perhaps hypothetical) population of neighborhoods.
The next subsection indicates how to solve the statistical complications
caused by the fact that the structure is not neatly nested.
The analysis by multilevel software of random effects models that
combine nested and crossed effects is technically and computationally more
demanding than the usual multilevel models. Therefore, it is advisable to
make preliminary analyses focusing on first one and then the other crossed
factor; such analyses can be done employing the usual two-level model. An
important question is to what extent the two crossed factors are associated.
In a study of schools and neighborhoods, these two factors will be
associated, and when only one of them is incorporated as a random effect in
the model it will draw to itself part of the other (temporarily neglected)
random effect. In a generalizability study, on the other hand, usually the
various crossed factors are not associated and they will have (nearly)
independent effects.
This model can be estimated in a frequentist framework as described in
Raudenbush (1993) and Rasbash and Goldstein (1994). The software
packages HLM, MLwiN, R (packages lme4 and nlme), SAS, and Stata offer
ways to do these computations. It turns out, however, that the more
complex models can often more easily be estimated within a Bayesian
framework, applying the Markov chain Monte Carlo algorithm as discussed
in Section 12.1. This is implemented in MLwiN as well (Browne, 2009).
Example 13.1 Sustained primary school effects.
This example is derived from a study on disentangling complex school effects (Timmermans et al.,
forthcoming), using data from a large-scale Dutch cohort study (Kuyper and Van der Werf, 2003).
Around 20,000 pupils in approximately 100 schools were followed after having entered secondary
education in 1999 up to the point where they did their final national examinations in a range of
subjects, in most cases six or seven. The dependent variable is the average examination score across
those subjects, which in theory can range from 1 (low) to 10 (high), but empirically ranges from 3 to
9. A subset of this cohort is taken, namely those pupils who took their final national examination in
the lower general education track. This subsample consists of 3,658 pupils in 185 secondary schools,
the number being larger than the original 100 because of pupil mobility. It is also known which
primary schools these students attended before they entered secondary education. They stem from
1,292 different primary schools. In order to find out whether the effect of the primary school has a
sustained effect on the examination score of the pupils at the end of the secondary stage, a cross-
classified multilevel model analysis is performed on the data. The nesting structure is pupils within
the secondary schools in which they did their final examination, crossed with the primary schools
they stem from. Table 13.1 contains the results of a model without (Model 1) and a model with the
inclusion of primary school effects (Model 2).1

Table 13.1: Models with and without cross-classification for examination

grades.

In Model 1 in Table 13.1, the average examination score is 6.36 and the total variance 0.466
(difference due to rounding). Of this variance, 14% is associated with the secondary schools. Model 2
presents the results in which the available information on the primary schools previously attended by
the pupils is also taken into account. Of course the average examination grade remains the same, but
now we see some changes in the variance components. The variance between secondary schools
marginally decreases to 0.066, and the within-school variance decreases somewhat as well, since now
the primary schools take up a variance component of 0.006. The decrease in deviance is 7025.7 –
6977.4 = 48.3, highly significant when compared to a chi-squared distribution with df = 1. We
observe here an instance where, for a variance component, the estimate divided by the standard error
is not very large, but yet the variance is significantly different from 0 (cf. Section 6.1).
There are now three different intraclass correlations: the correlation between examination grades
of pupils who attended the same primary school but went to a different secondary school is
the correlation between grades of pupils who attended the same secondary school but came from
different primary schools is

and the correlation between grades of pupils who attended both the same primary and the same
secondary school is

We can elaborate the models by including predictor variables. The original data set contained
many potential candidate predictors, such as IQ, gender, and the socio-economic status of the pupil’s
family. For almost all of these variables scores were missing for some pupils, and therefore we
employed the chained equations technique described in Chapter 9 to impute values, using all the
information available.
From the potential predictors we selected an entry test (grand mean centered) composed of
arithmetic, language and information processing subscales; socio-economic status (SES, grand mean
centered); the primary school teacher’s advice about the most suitable level of secondary education
(grand mean centered); and ethnicity. Table 13.2 contains the results with a model including these
four predictor variables (Model 3) as an extension to the previously fitted Model 2 (presented again
in this table). According to the principles of multiple imputation (Section 9.4) we constructed 25 data
sets with imputed values. The results in the table, as well as those in the following tables in this
chapter, are the syntheses of 25 analyses run on these imputed data sets, using equations (9.3)–(9.6).

The four predictor variables all have highly significant effects, indicating that pupils with higher
entry test scores, with higher recommendations from their primary school teachers, and from more
affluent families have higher average examination scores. Moreover, pupils from ethnic minorities
have lower examination results than pupils from the Dutch majority group. Most important, however,
are the estimates of the variance components. Comparing Models 2 and 3, all variance components
have decreased. The between-pupils within-schools variance decreases from 0.395 to 0.330. The
between-secondary-schools variance (0.034, and significant) is almost half its original estimate
(0.066), which also turns out to be the case for the between-primary-schools variance: this decreases
from 0.006 to 0.003. This indicates that secondary schools appear to have a value-added effect on
pupil achievement measured at the final examination, but that primary schools, given the
achievement levels attained by pupils at the end of primary education and given their family
background, have only a marginally lasting effect as measured four or five years later at the
secondary school examinations.

Although for the models in the example the predictor variables only had
fixed effects, it is straightforward to include these variables with random
slopes at the level of one or both of the crossed random factors as shown in
equation (13.1).
 Table 13.2: Models with and without cross-classification for examination
grades (continued).

13.2 Crossed random effects in three-level models

A crossed random effect in a three-level model can occur in two ways. An
example of the first is pupils nested in classes nested in schools, with
neighborhoods as a crossed effect. In this case the extra factor is crossed
with the level-three units: neighborhoods are crossed with schools.
The second kind of random effect in a three-level model is exemplified
by pupils nested in schools nested in towns, again with neighborhoods as
crossed effects. Here the extra factor is crossed with the level-two units and
nested in the level-three units: neighborhoods are crossed with schools and
nested in towns.
Such models are rather complex, but fortunately can be handled
straightforwardly in a Bayesian framework (Section 12.1) using the MCMC
algorithm, as implemented, for example, in MLwiN.
 13.3 Multiple membership models
When schools and pupils are of focal interest in a study following pupils
over time, pupils may have attended more than one school during the period
under investigation. Which school is now ‘responsible’ for a pupil’s school
success? The first, the last, the ones in between? In Example 13.1 we chose
the last school, but others might be more interested in the first secondary
school these pupils attended, as this school gave them the essential head
start. Multiple membership models have been developed to analyze this
complex reality, thus avoiding the kind of choices just mentioned. That
some pupils have attended more than one school can be modeled explicitly.
Once again a classification graph, given in Figure 13.2, may clarify what is
meant by multiple membership.

Figure 13.2: An example of pupils nested within multiple schools.

The first six pupils all attended only school 1, but pupil 7 attended both
school 1 and school 2. The most extreme case is pupil 11, who attended
four different schools. One can think of a situation where the pupil is being
bullied and his parents try to resolve the situation by choosing another
school. Later they move house, so the pupil once again has to change
schools, and exactly the same problem of being bullied occurs, and again a
switch to another school is seen as the solution. The main difference with
the cross-classified situation, where the individuals belong to units of two
different populations of interest at level two (schools and neighborhoods, or
primary and secondary schools), is that in the multiple membership
situation there is only one such population of interest at level two, but the
individuals belong to more than one level-two unit in that population. Other
examples are teachers working in more than one school, pupils rating their
teachers of different subjects, families who changed neighborhoods during
the period under investigation, or employees changing firms.
 The solution to the problem is to use some kind of weighting, where the
membership weights can be proportional to the time a level-one unit spent
at a level-two unit, with the weights summing to 1 (Hill and Goldstein,
1998). So a pupil who attended the same primary school from grade 1 to
grade 6 has a membership weight of 1 for that school, and 0 for all other
schools. A pupil attending the first three grades in her first school and the
last three grades in another school has a membership weight of ½ for the
first, ½ for the second and 0 for all other schools. The membership weights
are denoted by wih, for pupil i in school h, adding up to 1 for every pupil:

The multilevel model for this situation, following the notation of Rasbash
and Browne (2001) and ignoring explanatory variables, is:

The subscript {j} denotes that a level-one unit does not necessarily belong
to one unique level-two unit. Therefore the level-two residuals U0h are
weighted by wih,. For a given pupil i, the schools h that this pupil never
attended have wih = 0, so they do not contribute anything to the outcome of
(13.2). For example, for a pupil who spent a quarter of his time in school 1
and the remainder in school 2, this part of the formula gives the
contribution

If there are explanatory variables at level two, then a similar kind of

weighting is used for these variables.
Next to this contribution from multiple random effects, the division of a
pupil’s school career over multiple schools could also have a fixed effect,
for example, because of the difficulties of adjustment. To represent this as a
consequence of the fractionalization of the weights wih, one may give a
fixed effect to the variable

which is 0 if one weight is 1 and all others 0, and positive if two or more of
the weights are non-zero; for example, if for pupil i there are K weights of
equal values 1/K and all other weights are 0, then Wi = K – 1.
Example 13.2 Multiple membership in secondary schools.
Following up on the previous example and paying attention only to the secondary school period of a
pupil’s school career, thus ignoring the sustained effects of the primary schools, we now try to do
justice to the fact that many pupils attended more than one secondary school. In this sample 3,438 of
the 3,658 pupils did their final examination in the school where they originally enrolled. So 94% of
the pupils never changed schools. Of the remaining 6%, 215 pupils attended two, and 5 pupils
attended three different schools. For the pupils who attended two different schools the membership
weights for their first school vary between 0.20 and 0.80. For the five pupils who attended three
different schools the membership weights for their first school vary between 0.20 and 0.25,
indicating that they rather quickly left this school. Unlike the analyses in the previous example, we
start with a model (Model 4) in which the second level is defined by the first school enrolled, of
which there are 86 in total in this subset of the sample. Then a multiple membership model (Model 5)
is fitted to the data.
From Table 13.3 it can be seen that the results for Model 4 are very similar to those presented in
Model 1 of Table 13.1. The variance for the first school a pupil attended is close to the variance for
the last school attended. In Model 5 the results of the multiple membership modeling of the data can
be seen. There is a marginal increase in the between-school variance, from 0.062 to 0.064, indicating
that the multiple membership model gets somewhat closer to the data than the ‘simple’ multilevel
model. To see whether there is an impact of the more advanced modeling on the estimated school
effects, Figure 13.3 shows a scatterplot of the residuals (‘school effects’) for the school where the
pupils did their final examinations.
On the horizontal axis the school-level residuals derived from the ‘simple’ multilevel (ML)
model (Model 4) are plotted. On the vertical axis the school-level residuals derived from the ‘more
realistic’ multiple membership (MM) model are pictured. As can be seen, there is a high correlation
(0.86 to be precise) between the two types of residuals, but the most striking point is that the
residuals from the multiple membership model are slightly less dispersed. That is, of course, because
part of the between-schools variation is now accounted for by the other schools previously attended
by some of the pupils.

Table 13.3: Models without and with multiple membership.

Figure 13.3: Scatterplot of ML- and MM-type school level-two residuals.

13.4 Multiple membership multiple classification

models
As may be expected, cross-classified and multiple membership data
structures can both be present and of interest to the researcher (Browne et
al., 2001). She may, for instance, be interested in the effects of
neighborhood as well as of multiple school membership on pupils’ attitudes
toward ethnic minorities. We can simply combine Figures 13.1 and 13.2 to
depict such a situation, yielding the classification graph of Figure 13.4.
The formula to be used for such a multiple membership multiple
classification situation consists of a merger of formula (13.1) for a cross-
classified model with formula (13.2) for the multiple membership model,
where the weights go with the U0j and, of course, not with the W0k.

Figure 13.4: An example of pupils nested within multiple schools crossed

with neighborhoods.

Table 13.4: A model with cross-classification and multiple membership for

examination grades
Example 13.3 Multiple membership in secondary schools cross-classified with primary schools.
We continue with the multiple membership model from the previous example, but now also
simultaneously model the sustained effects of the primary schools. In Model 7 there are no predictor
variables, and in Model 8 we once again include predictor variables. The results are presented in
Table 13.4.
The table shows that the parameter estimates have not changed much as compared to the cross-
classification model results presented in Table 13.1 and the multiple membership model results
presented in Table 13.3.

The multiple membership multiple classification models can also be

extended to situations with explanatory variables having random slopes (for
further details see, for instance, Beretvas, 2011), and adding these effects
may improve the last model as presented in Table 13.4. Models have also
been developed for dependent variables with nonnormal distributions, as is
discussed in the overview given by Rasbash and Browne (2008).

13.5 Glommary
Cross-classification or multiple classification. A situation where levels
are not nested in each other. For example, lower-level units belong to
higher-level units of two or more different but intersecting populations
of interest. An example of such populations is given by pupils in
neighborhoods and in schools, where pupils from the same
 neighborhood may attend different schools that also are attended by
pupils from other neighborhoods.

Crossed random factor. A random factor that is crossed with another

random factor.

Classification graph. A figure depicting the nesting of lower-level units in

higher-level units.

Multiple membership. A situation where lower-level units belong to more

than one higher-level unit of one population of interest. An example is
given by pupils migrating from one secondary school to another.

Membership weight. A weight, for instance proportional to the length of

membership duration, by which a lower-level unit is assumed to be
affected by a higher-level unit.

Multiple membership multiple classification. A situation where lower-

level units belong to more than one higher-level unit of one population
of interest, and at the same time belong to higher-level units of one or
more different but intersecting populations of interest.

1
The cross-classified models in this chapter are estimated using the Bayesian MCMC algorithm of
Browne (2004) with 10,000 iterations, and parameter expansion at the level of the primary school.
For Bayesian procedures, see Section 12.1.
 14
Survey Weights

Some data collection methods use complex surveys, that is, surveys which
cannot be regarded as simple random samples. In a simple random sample,
which may be regarded as the standard statistical data collection design,
each subset of the population having the same size has the same probability
of being drawn as the sample.1 In Chapter 2 we saw that two-stage samples
and cluster samples are quite natural sampling designs for multilevel
analysis. Another sampling method is stratification, where the population is
divided into relatively homogeneous subsets called strata and samples are
drawn independently within each stratum, with sampling proportions which
may differ between strata. Many big surveys use a combination of
stratification and multistage design. This will imply that elements of the
population may have different probabilities of being included in the sample
– these probabilities are known as the sampling probabilities or inclusion
probabilities. This must be taken into account in the data analysis.

OVERVIEW OF THE CHAPTER

This chapter starts by discussing the choice between model-based and
design-based approaches to inference. In brief, the model-based approach
consists of incorporating aspects of the design (variables used for
stratification, etc.) as far as necessary into the model, and proceeding with a
hierarchical linear model in the usual way; design-based inference uses
special techniques incorporating weights based on the sampling
probabilities. We shall argue (Section 14.3) in favor of pursuing a model-
based approach as much as possible, because this approach is more in line
with the primary aims of analytic social science and because there are still
issues with the procedures developed for multilevel design-based inference,
in particular for small and medium sample sizes.
Section 14.1 introduces in general terms the concepts of model-based
and design-based inference, and the debate – which has been more fervent
among statisticians than among social scientists – about the choice between
these modes of inference. In Section 14.2 the use of weights is discussed,
and the reader is warned about the distinction between survey weights and
precision weights. While these two first sections are about surveys in
general, the following sections focus on multilevel analysis. Section 14.3
presents a variety of methods for assessing whether a model-based
approach is appropriate in a survey study with a multilevel data structure.
Then in Section 14.5 the main methods proposed for multilevel model-
based inference are discussed.
This chapter is about statistical methods that are still being developed
and discussed, and therefore presents a less clear-cut account than some of
the more basic chapters.

14.1 Model-based and design-based inference

14.1.1 Descriptive and analytic use of surveys

Surveys can have many purposes, and when choosing between methods of
analysis it is helpful to distinguish between the descriptive and the analytic
use of surveys (e.g., Pfeffermann, 1993). A survey is a study of a finite
population, and descriptive inference is the estimation and testing of
descriptive parameters of the population, as is done in official statistics.
Consider, for example, the population of all pupils enrolled in a given type
of school at a given date in some country. An example of such a descriptive
parameter is the proportion of pupils who fail a specific test in mathematics.
Analytic inference, on the other hand, is concerned with questions about the
way in which a variable of interest depends on, or is associated with, other
variables. Continuing the example, an analytic question would be how
failing in mathematics depends on the gender of the pupil and of the
teacher, not only in this particular finite population but more generally for
comparable pupils in comparable schools. Here the researcher wishes to
generalize to a larger population, which may be hypothetical and is usually
rather vaguely defined – as signaled by the word ‘comparable’ in the
example. The difficulty of describing the population to which one wishes to
generalize is compounded in the multilevel case by the fact that there are
several populations – in most cases, one for each level. But the difficulty in
pinning down the population is not a grave concern because in practice we
can live with a bit of vagueness.
When working with complex surveys, the probabilistic nature of
statistical inference can be based on two approaches. One is the model-
based approach, which is the basis of this book and of the majority of
statistical texts. Here the researcher makes the assumption that the data can
be regarded as the outcome of a probability model with some unknown
parameters. Note that the word ‘model’ here refers not to the theoretical
model from a social science viewpoint, but to what also might be called the
‘data generating mechanism’ – more precisely, the probabilistic mechanism
that could have generated the values of the dependent variable, conditional
on the explanatory variables and the multilevel structure. In linear
regression-type models like the hierarchical linear model, the probabilistic
element is represented by the residuals (sometimes called ‘errors’) which
are the deviations between the model – which is always idealized – and the
data.2 The probability model is defined by the assumptions about the
simultaneous probability distribution of all residuals occurring in the model.
A discussion about the various types of probability models in statistical
inference is presented by Cox (1990). Here we can take the view that
randomness represents the influence of variables not included in the model
together with the approximate nature of assumptions of linearity,
measurement error, etc.
In the case of the two-level hierarchical linear model, for instance, the
probability model is expressed as in Chapter 5 by the assumptions that the
level-one residuals are independent and identically distributed across level-
one units, that the vectors of random intercepts and slopes are independent
and identically distributed across level-two units, that these have normal
distributions, that there is a linear dependence of the dependent variable on
some array of explanatory variables, etc. These are assumptions that should
a priori be plausible approximations when carefully considering the
phenomena under study, and that can be tested (cf. Chapter 10), although
uncertainty will always remain. A model never is more than an
approximation.
The other is the design-based approach. Here the inference is based on
the probability mechanism used for the sample selection. Probabilities that
elements of the population are included in the sample may differ between
elements; for example, in a survey of the adult population one might wish
to oversample employed individuals, that is, give them a higher probability
of inclusion than nonemployed persons. In the ideal case the design-based
model is totally convincing because it only expresses the rules used in the
data collection; in practice there will be some give and take because of
nonresponse, imperfections in the sampling frame, and other so-called
nonsampling errors. The inference is from the data to the finite population.
In the boundary case where the data constitute a census, that is, the
population is observed in its entirety, there is no need for statistical
modeling because everything is perfectly known. The finite population is
less idealized and less far away than the infinite population of the
probability model, but regularities of human and social behavior belong to
the realm of the idealized model rather than the finite population. For the
finite population no assumptions about, for example, normal distributions
need to be made; indeed no such assumptions can be made because for
normal distributions the number of possible values is infinite. A crucial
element in the design-based approach is the use of weights. In a one-level
design, denote by πi the probability that population element i is included in
the sample. Then the data for this unit, if indeed included in the observed
sample, will be weighted by a factor wi = 1/πi. This weighting
counterbalances the lower frequency of observation of elements with a low
inclusion probability.
It may be noted that in many introductory textbooks there is confusion
because statistical modeling is often argued by underlining the importance
of random sampling combined with making assumptions about independent
and normally distributed residuals, thereby confounding the distinction
between design-based and model-based inference.
This book is aimed at researchers who wish to obtain knowledge about
social and behavioral processes and mechanisms, with some degree of
generalizability beyond the data set under consideration. Thus it is oriented
naturally toward model-based inference. However, a survey in which
inclusion probabilities are not constant might give a biased representation
of these social and behavioral processes, and the question then is how to
analyze such a data set to obtain unbiased, or approximately unbiased,
conclusions. This issue has led to a debate between statisticians favoring a
design-based and those favoring a model-based approach. Although the
sharpness of the debate has attenuated, there are still clear differences of
opinion. An overview written for social scientists, with an interesting
presentation of the history of this debate going back to Fisher and Neyman,
is given by Sterba (2009). Those who would like to consult some of the
original sources could have a look at DuMouchel and Duncan (1983), Kish
(1992), Pfeffermann (1993), Smith (1994), Little (2004), Gelman (2007),
and articles cited therein.
The contrasting positions in their barest forms are as follows.
Proponents of a model-based approach (Bayesians3 as well as frequentists)
say that if the model is true and the sampling mechanism is independent of
the residuals in the probability model, then the sampling design is irrelevant
and taking account of the sampling design entails a loss of efficiency (Kish,
1992). This efficiency loss will be serious if the weights wi are highly
variable across population elements: a few sample elements will then be
weighted very strongly. This also implies that the residuals associated with
these elements will be weighted strongly, and if residual variance is
constant this will lead to large standard errors. Proponents of the design-
based approach, on the other hand, say that one never can be confident that
the model is true, and therefore a model-based estimator might be seriously
biased and thus worthless. Bias here means a misrepresentation of the true
data-generating mechanism, induced by the distortions due to the survey
design. Design-based estimators yield protection against such distortions,
and are possible because the design, and therefore the potential distortions,
are exactly known as a function of the design variables (e.g., DuMouchel
and Duncan, 1983).
Some degree of reconciliation between the opposing positions has taken
place (e.g., Pfeffermann, 1993; Smith, 1994). Modelers realize that it may
be important to include design-related elements in the models. The use of
the hierarchical linear model for two-stage samples is an example (see, for
example, Little, 2004). Survey-oriented statisticians have developed
methods incorporating models. In this chapter we follow such a
compromise approach, but with an argued preference in favor of model-
based inference, as will be made clear below.
 14.2 Two kinds of weights
Weights in statistical analysis can be of several kinds, and the distinction
between these is not always clear in statistical software manuals. Avoiding
confusion between the two main kind of weights is essential, and this is the
purpose of this section. We start by treating this distinction for single-level
linear regression analysis. More extensive treatments are given by
DuMouchel and Duncan (1983) and Kish (1992).
Sampling weights or survey weights are used in surveys from finite
populations and reflect inversely the probability that any particular
population element is included in the sample. These are constant for simple
random samples, and then they may safely be ignored; for complex
samples, such as stratified and two-stage samples, they usually are
nonconstant and the issue of how to take this into account is the topic of this
chapter. For example, a sample of adults could be stratified with respect to
employment status, with nonemployed people having a higher inclusion
probability than employed.
In a one-level design where πi is the probability that population element
i is included in the sample, the (nonstandardized) weight is wi = 1/πi. Cases
that are included with a low probability are in a part of the population that is
under-represented in the sample, and must receive a larger weight to obtain
unbiased estimators of properties of the population. Large weights will be
associated with large uncertainty, because one sampled individual is used to
represent a large portion of the population. Denote the parameters by β and
the fitted value for case i by ŷi(β). The probability weighted estimate (PWE)
for β is then obtained by finding the value of β for which

is as small as possible.
For the case where only a population mean is being estimated, that is, β
is one-dimensional and yi(β) = β, this reduces to
with estimated standard error

This design-based standard error, that is, estimator for the sampling
standard deviation of under random sampling according to the design
with inclusion probabilities πi, is a special case of the formula presented, for
example, in Fuller (2009, p. 354).
Precision weights, also called analytical weights, indicate that the
residual standard deviations of some cases are larger than for others, and
estimates will be more precise if cases with higher residual standard
deviation get lower weight. Denote in a one-level design the residual
standard deviation for case i by si. One reason why residual standard
deviations might be nonconstant is that a case in the data set might be an
average of a variable number of more basic units; then the weights may be
called frequency weights. Another possible reason is that the dependent
variable might represent a count or quantity, and higher quantities mostly
are naturally associated with larger variability. The (nonstandardized)
weight associated with standard deviation si is wi = 1/s2i. Here, larger
weights reflect smaller standard errors (i.e., less uncertainty); this is
opposite to the case explained above of sampling weights. Here also, the
weighted parameter estimate is obtained by finding the minimum of

as a function of β. This is called the weighted least squares (WLS)

estimator.
The PWE estimator (14.1) and the WLS estimator (14.4) are defined by
the same expression, but they are justified differently. Since they are based
on different assumptions, the expressions usually given for their variances
are different. The appendix to this chapter (Section 14.6) contains the
matrix expressions for these variances. For the case of estimating a sample
mean, if the variance for Yi is var(Yi) = σ2/wi, then the WLS estimator (14.4)
again reduces to (14.2). The standard error can then also be estimated by
(14.3), which here is a simple (one-dimensional) case of the sandwich
estimator proposed by Liang and Zeger (1986); see Section 12.2.
This estimator (14.3) for the standard error is heteroscedasticity-
consistent, that is, for large samples it will give good estimates even if the
variables Yi have nonconstant variances. However, the estimated standard
error often given is

This is correct and efficient if indeed var(Yi) = σ2/wi, but not

heteroscedasticity-consistent: it is not a reliable standard error if the
variances are not inversely proportional to the weights.

Precision
The reason for using the WLS estimator (14.4) with precision weights is to
obtain high precision.4 In the case of survey modeling, however, the PWE
estimator (14.1) is used to avoid bias, and its precision is not necessarily
optimal. Under the usual assumption of homoscedasticity (constant residual
standard deviations), there might be some cases with large sampling
weights wi which happen to have large residuals (which means that through
random causes their fit with the model is less good than for most other
cases), and the minimization of (14.1) will give such cases a large
influence. The extent of inefficiency of the PWE estimator (14.1) in the
homoscedastic case can be expressed by the effective sample size (Pothoff
et al., 1992), defined as

It is called this because when estimating by the PWE estimator (14.2) the
mean of some variable which has variance σ2 independently of the sampling
weights wi, the variance of the estimator is σ2/neff, just as if it were the
regular mean of a sample of size neff. The effective sample size is strictly
smaller than the regular sample size, unless the weights are constant;
therefore in the homoscedastic situation the use of sampling weights leads
to a loss of precision, and this can be serious if the variability of the weights
is large, as will be expressed by a small value for the effective sample size.5
If the residual variance is nonconstant (heteroscedasticity), however, the
effect on the estimation variance of using the weighted estimation (14.1)
will be different. The assumptions underlying survey weights and those
underlying precision weights are independent, and therefore they can both
be true – in which case the use of sampling weights yields a gain, rather
than a loss, of efficiency compared to unweighted estimation.

Bias
The reason for using sampling weights is to avoid bias. A sampling design
with heterogeneous inclusion probabilities will yield a sample that is a
biased representation of the population. This does not, however, necessarily
lead to bias in the estimators for the parameters. The hierarchical linear
model is a model for the dependent variable, conditional on the explanatory
variables and the multilevel structure. If the predictor variables are
represented differently in the sample than their distribution in the
population – which is one kind of bias – it does not follow that there must
be bias in the estimators of the parameters of the hierarchical linear model.
The parameters are biased only if the distribution of the residuals is affected
by the sampling design. Then the survey design is said to be informative. If
the residuals in the model are independent of the sampling design and of the
sampling weights, that is, the model is correctly specified given all
variables including the design variables, then the use of weights is
superfluous. In multilevel analysis of survey data, if we can be confident of
working with a well-specified hierarchical linear model and the sample
design is unrelated to the residuals, it is better to take no explicit account of
the survey design when doing the analysis and proceed as usual with
estimating the hierarchical linear model. The difficulty is, of course, that we
never can really be sure that the model is well specified.
As an example of the meaning of this type of bias, consider a multilevel
study of pupils in schools with language achievement as the dependent
variable, and socio-economic status (SES) as one of the explanatory
variables. Assume that the sampling design is such that the inclusion
probability of schools is likely to be related to the ethnic composition of the
schools, but no data about ethnic composition are available. Assume also
that language achievement depends on ethnic background of the pupil, and
SES is the main explanatory variable included that is related to ethnic
background. The hierarchical linear model for language achievement is then
misspecified because ethnic background is not included. and this will
mainly affect the estimate for the effect of SES. The SES variable in this
data set and model reflects a variable where different ethnic subgroups of
any given value on SES are distributed in proportions that differ from the
true population proportions, thus leading to a biased parameter estimate for
SES. This argument hinges on the fact that there is an omitted variable,
correlated with an included predictor variable, and correlated with the
inclusion in the sample.
Clearly, the most desirable solution here would be to observe the ethnic
background of pupils and include it in the model. A second solution is to
come up with a variable that reflects the sampling design and that is closely
related to the ethnic composition of the school, and control for this variable.
A third solution is to weight schools according to their inverse inclusion
probability. In the rest of this chapter we shall discuss such solutions in
more detail.

14.3 Choosing between model-based and design-

based analysis
This section presents arguments and diagnostics for choosing between a
model-based and a design-based analysis of a given multilevel data set. It is
assumed that the purpose of the statistical inference is analytic: we wish to
find out how a dependent variable depends on relevant explanatory
variables. This is expressed in a hierarchical linear model for the dependent
variable, including the relevant explanatory variables, which will be called
the model of interest. This may be either a model given on the basis of
theory, or a model where some of the details are still to be specified
depending on the data. The design is assumed to depend on a number of
variables called the design variables. These are the variables on which the
inclusion probabilities depend. With a design-based approach to analytic
inference we refer to an approach that analyzes the model of interest while
taking the design into account. Thus, given that it is about analytic
inference, this approach is based on a combination of model-based and
design-based considerations. For single-level studies we hinted at such a
combined approach by presenting the probability weighted estimator (14.1)
based on fitted values, which presupposes a model; for multilevel studies
this will be discussed further in Section 14.5.
Our position is that it will be preferable, if possible, to carry out a
model-based analysis, possibly according to a model that is augmented by
also using some of the design variables; and that a design-based analysis
will be done only if there are clear arguments against a model-based
analysis. The arguments for this preference are as follows:

Finding a well-fitting model for the dependent variable is the main

purpose of analytic inference. In the model-based approach,
augmented by design variables, one tries to find those elements of the
1.
design variables that may contribute to understanding the dependent
variable. This serves the main purpose more directly than a design-
based approach where one tries to ‘weight the design away’.
Most researchers reading this book will be better acquainted with the
model-based than with the design-based approach, and they will be
2. able to carry out a model-based analysis with more confidence and
less risk of error. In this sense, the model-based analysis has the
virtue of simplicity.
The methods developed for the design-based analysis of multilevel
data structures are good for estimation of fixed effect parameters but
3.
less so for the estimation of variance parameters and standard errors
and for hypothesis tests unless sample sizes are fairly large.
Thus we shall lay out an approach where the design variables are
scrutinized for their effect on the dependent variable, and it is investigated
whether it is sufficient to analyze the model of interest, possibly extended
with design variables, by a regular hierarchical linear model; and where an
approach that analyzes the data in a way that also takes account of the
design, by using survey weights, is followed only if there are reasons not to
trust a model-based analysis. We could say that we wish to assess whether
the sampling design is noninformative (Pfeffermann, 1993), meaning that
the residuals in the statistical model are independent of the variables used
for the survey design. In that case we can proceed with the analysis without
taking the design into account.
It should be noted that the multilevel analysis according to the
hierarchical linear model employs a model which already accounts for one
element present in many sample designs, namely, the clustering present in
cluster samples and multi-stage samples. But different sampling
probabilities occurring, e.g., in stratified or post-stratified sampling are not
accounted for in a straightforward way, and how to do this is the issue of
this chapter.

14.3.1 Inclusion probabilities and two-level weights

The treatment focuses on two-level models. We shall sometimes use the
term clusters to refer to level-two units, and use cluster size for the sample
size nj for a given level-two unit j.
Given the two-level structure, we assume a two-stage sample design,
where clusters are chosen independently with some probability, and given
that a cluster j has been selected a sample of level-one units within this
cluster is chosen. The inclusion probabilities are defined as

It is possible that πj = 1 for all j: then all clusters in the population are
included in the sample. An example is a survey in countries as level-two
units, where all countries satisfying some criterion (defining the population)
are included in the study, and within each country a sample of the
population is taken. A different possibility is πi\j = 1 for all i and j; then the
sample design is called a cluster sample, and clusters are observed either
completely or not at all. An example is a survey of schools as level-two
units, where within each sampled school all pupils are observed. The
marginal probability of observing level-one unit i in cluster j is given by the
product,
 The weights or design weights are the inverse of the inclusion
probabilities. Thus, the weights are defined by

To use weights in two-level models, the separate sets of weights at level one
and level two are needed, corresponding to the separate inclusion
probabilities at level one and level two.
From these weights the effective sample sizes can be calculated:

The effective sample size is defined such that a weighted sample gives the
same amount of information as a simple random sample with sample size
equal to the effective sample size.

The ratios of effective sample sizes to actual sample sizes are called the
design effects at level two and level one,

The design effects give a first indication of the potential loss of statistical
efficiency incurred by following a design-based rather than model-based
analysis (but see footnote 5 on p. 221).
In single-level studies the scales of these weights are irrelevant,6 that is,
all weights could be multiplied by the same number without in any way
changing the results of the analysis. In multilevel designs scaling the level-
two weights still is irrelevant, but the scale of the level-one weights is
important, and the literature is still not clear about the best way of scaling.
Pfeffermann et al. (1998) propose two methods of scaling. These are
used in design-based estimation for multilevel designs (see below in
Section 14.5). The first is

which yields scaled weights summing to the effective sample size njeff; this
is called ‘method B’ by Asparouhov (2006). The second is

yielding weights summing to the actual sample size nj, and called ‘method
A’ by Asparouhov.

14.3.2 Exploring the informativeness of the sampling design

If there is an association between the sampling design and the residuals in
the hierarchical linear model, then the sampling design is informative
(Pfeffermann et al., 1996) and a model-based approach risks being biased
and inconsistent, that is, the parameter estimates may be a
misrepresentation of the true parameters in the hierarchical linear model
even in large samples. Here we propose a variety of methods that can be
followed to assess the association of the sampling design with the model of
interest. They do not have the primary aim of obtaining a yes/no verdict
about the informativeness of the design for the given model of interest.
Their purpose is rather to explore how the design variables may be
associated with the model of interest, to get ideas about how to extend the
model of interest with relevant variables reflecting the design, and, after this
exploratory phase in which perhaps the model of interest was extended, to
assist the researcher in deciding between a model-based and a design-based
analysis. In any single study it would probably be overkill to use all
methods presented here. Which of these methods are useful will depend on
the data structure and the aims of the analysis, to be decided by the insight
and experience of the researcher.

Look at the variability of the weights. Explore the variability of the

weights, for the level-two weights wj as well as for the level-one
weights wi|j within each sampled cluster j. This is done in the first
place by calculating effective sample sizes at level two, and at
level one for each cluster, defined above in (14.9). If actual cluster
sizes nj are strongly variable, it may be illuminating to plot the
1. design effects (14.10) as a function of actual cluster sizes.
Depending on the situation, more details of the weight distribution
may be helpful, for example, by making boxplots. High variation
of weights and low design effects are warning signals: they imply
that the design is far from a random sample, which leads to the
possibility of considerable biases of model-based estimators, and
to low efficiency of design-based estimators.
Consider adding design variables to model of interest. In the ideal
case, all variables determining inclusion probabilities – the design
variables – are known and are part of the data set. This allows a
much clearer and more satisfactory analysis than knowing only the
2. weights. However, especially for large surveys, it is not uncommon
that information about the design is only partial, for example, for
confidentiality reasons or when weights have been determined not
only because of nonconstant sampling probabilities but also to
correct for nonresponse.
If the design variables are available to the researcher, then they
may be included in the model; it can be tested whether they affect
the dependent variable and whether the other results are sensitive
to their inclusion. This makes sense especially as additions to a
model of interest that is already rich enough in explanatory and
control variables; the only concern in this exploration is whether
the design variables have effects over and above the variables in
the model of interest.
The design variables should be included not only as main fixed
effects but also in interactions with other variables and possibly in
 the random part of the model. It is advisable to test their
interactions with the main explanatory variables in the model, and
also to test whether they have random slopes or are associated with
heteroscedasticity (cf. Chapter 8). General rules cannot be given
here because everything depends on the specific properties of the
design, the data structure, and the model of interest. For example,
if the design consists of a stratification (i.e., distinct subsamples) at
level one or level two, the best approach may be to estimate the
hierarchical linear model separately for each of the subsamples and
assess the importance of the differences between the results of the
subsamples. Then in a next step this can lead, if desired, to a
common analysis where a number of main effects of the
subsamples and of interactions are included to represent these
differences. Another possibility here is to include the subsamples
(strata) as an additional level in the model and, given these,
random main effects and random slopes according to what is a
good model specification. This use of random effect models is also
discussed by Little (2004, equation (14)), and Gelman (2007).

If the researcher is certain that the design variables have been included in a
satisfactory way, then a model-based analysis can be pursued further. In
practice, of course, one never is really sure about this, especially since
complicated interactions and nonlinear effects might possibly be involved.
To explore whether the model differs in ways associated to the design
weights, the following procedures can give some insights.

Apply the hierarchical linear model to parts of the data set

differing with respect to inclusion probabilities or other design
3. factors. First we present one way to investigate the influence of the
design on the estimation of the model of interest which is easily
carried out and can give insight in the magnitude of this influence.
The procedure is to split the data set into parts according to the
weights at level one and those at level two, analyze each part
separately by the hierarchical linear model according to the model
of interest, and assess the magnitude of the differences in results
between the various parts. How to make this split depends on the
 sample sizes and the distribution of the weights; a possibility is the
following. Divide the level-two units (clusters) into two or three
groups based on the level-two weights wj; within each cluster,
divide the level-one units into two or three groups based on the
scaled level-one weights w*i|j. (The choice between scaling
methods 1 (14.11a) and 2 (14.11b) is rather arbitrary here and in
most cases will not make much difference; we suggest method 1.)
The combination of these two divisions splits the data set into four
to nine (or a different number, depending on how the split is done)
disjoint parts, and for each part the model of interest is analyzed
using a regular multilevel model. Each of the resulting sets of
parameter estimates reflects a part of the populations at levels one
and two that was sampled with different inclusion probabilities. If
these sets of results are more or less the same (given their standard
errors and given what is deemed to be an important difference),
then it is unlikely that the design has a large impact on the
estimation of the model of interest. On the other hand, if the sets of
results do differ from each other, then there is evidence for such an
impact. The type of difference may give some insight in the type of
impact, for example, whether it is a consequence of mainly level-
two weights, mainly level-one weights, or both. In some cases this
may even lead to finding design-related variables that can be added
to the model of interest, after which there might be no further
appreciable impact of the design on the estimation results.
Add weight variables to the model. Noninformativeness of the
design means that the design variables are independent of the
residuals. This can be tested by using the weights as predictor
variables added to the hierarchical linear model (cf. DuMouchel
4.
and Duncan, 1983). Note that we have two sets of weights, at level
one and level two. Moreover, the level-one weights can be scaled
and there are no convincing guidelines on how to scale. We
tentatively propose to use scaling method 1.
The procedure is as follows. For the level-two weights wj and the
scaled level-one weights w*i|j of formula (14.11a), we test the main
effects as well as all interactions with explanatory variables in the
 model of interest. This test is carried out by adding the fixed
effects of these variables to the hierarchical linear model. Since
this leads to a multiparameter test, a straightforward way of doing
this is by using the ML estimators and the deviance test (Section
6.2). If this test is not significant, there is no evidence for bias of
the model-based estimator.
This procedure can be extended by testing heteroscedasticity of the
level-one residuals as a function of the level-one and level-two
weights according to the methods of Section 8.1. It follows from
Section 14.2 that if residual variance increases as a function of the
weights, then the design-based estimator is particularly inefficient;
if it decreases as a function of the weights, then the design-based
estimator may be relatively efficient.
Depending on the research questions and variables at hand, these
tests could also be applied to nonlinear transformations of the
weights. If a good search for effects of weight-related effects on
the dependent variables does not lead to findings of such effects, it
may be justified to continue with a model-based analysis that does
not take the design into account.

Depending on the design, it can be helpful to compare estimates that are

purely model-based with estimates that also take design-based estimates for
meaningful parts of the data set, or for the entire data.

Compare model-based and design-based estimators separately for

each cluster. If a great deal of detail is required with respect to the
consequences of the sampling design within the clusters, and if the
clusters are reasonably large (e.g., a difference of 30 or more
5. between each cluster size and the number of level-one predictors),
the comparison between the model-based and the design-based
analysis can be done on a cluster-by-cluster basis. If the model of
interest is the regular hierarchical linear model (5.15) without
level-one heteroscedasticity,
then the model-based estimator within each cluster is the ordinary
least squares (OLS) estimator. After all, in line with (5.1), the
within-cluster model is just a single-level linear regression model.
The within-cluster OLS estimator estimates the regression parameter
including the random slope (if any), that is, the parameter γh + Uhj
for the variables h = 0,1,…, p which do have a random slope, and γh,
for variables h = p + 1,…, r which do not. Denote this estimator (a
vector with r + 1 components) by jOLS and its covariance matrix
(which has the squared standard errors on the diagonal) by ΣjOLS.
Denote the design-based estimator (14.1) by jw with covariance
matrix Σjw. (For this single-level situation referring to a single
cluster, the scaling of the weights wi\j is immaterial.)

If the model is valid and the design variables are uncorrelated with
the level-one residuals Rij, then the difference jOLS — jw is just
random noise; on the other hand if the design variables are correlated
with the level-one residuals, then this difference estimates the bias of
the OLS estimator. A theoretically interesting property is that, if this
model is valid, the covariance matrix of the difference is the
difference of the covariance matrices:7

DuMouchel and Duncan (1983) discussed this and showed that the
test for the difference jw — jOLS can also be carried out as follows
in a linear OLS model. Compute the product of all level-one
variables xhij with the weights wi\j, and test in an OLS model the
effect of the weight wi/j itself and all r products xhijwi/j, controlling
for X1,. . ., Xr. This yields an F-test with r + 1 and nj−2(r + 1)
degrees of freedom which is exactly the same as the F-test for the
difference between the design-based and the model-based estimators.
It can be regarded as the test for the main effect of the weight
variable, together with all its interactions with the explanatory
variables. A version of this test for the case of multilevel models for
discrete outcome variables was developed by Nordberg (1989).

This yields a test outcome for each of the N clusters. This

information can be used for diagnosing the noninformativeness of
the sampling design in all clusters individually. If one also wishes to
have an overall assessment, the following procedure can be used to
combine the N tests into one. It consists of two steps. Suppose that
the test for cluster yields the F-statistic Fj, and denote the degrees of
freedom by df1 for the numerator and df2j for the denominator. In the
first step, the Fj values are transformed to an approximate standard
normal value by Fisher’s Z transformation8 (see Johnson et al., 1995,
Section 27.5) defined by

where ln(Fj) is the natural logarithm of Fj. In the second step, add
the normalized values

this test statistic can be tested in a standard normal distribution.

Another procedure for combining the tests is Fisher’s combination of
p-values (3.35).

These F-tests are exact (if the level-one residuals are normally
distributed with constant variance) and therefore also are valid for
small sample sizes nj. (Fisher’s combination of p-values also is
exact, but Fisher’s Z transformation is an approximation.)

If the denominator degrees of freedom are reasonably large (not less

than 20, preferably greater than 30), the differences between the
OLS and the weighted estimates of the regression coefficients are
reliable enough to be used for exploration of the effect of the within-
 group designs. For the most important level-one variables, these
differences can be plotted as a function of the OLS estimate or other
relevant cluster characteristics, and this can give insight in the
pattern of how the level-one design is associated with the model of
interest.

If the combination of the F-tests is not significant, this is an

indication that in the analysis of the model of interest it is not
necessary to take the level-one sampling design into account, so that
only the level-two sampling design remains. The case of having only
level-two weights and no (that is to say, constant) level-one weights
is considerably simpler than the case of variable weights at both
levels, so that this will give an important simplification.

It should be realized, however, that these F-tests only look at the

estimated values and not, for example, at the variances and standard
errors, which might also be affected by the design; and that a
significant test result can be interpreted as a signal of
informativeness, but nonsignificance cannot be interpreted as a
reliable confirmation of noninformativeness. Therefore, these
procedures can be very helpful in exploring the consequences of the
design, but they cannot give clear confirmation that nothing is being
overlooked.

Test differences between model-based and design-based estimators.

The model-based and design-based estimators can both be calculated
and compared. The test of the difference between these two
estimators, explained above for the case of a single-level
homoscedastic (OLS) model, can be generalized to testing this
6. difference for any model and any design. With the currently available
methods this does not lead to a reliable procedure in the multilevel
case, however. Therefore, although it is meaningful to compare the
two estimates, a formal statistical test is not available in practice. We
explain the basic procedure and the associated difficulties, and then
give a proposal which is practical but not a formal test.
Pfeffermann (1993) noted that the test of DuMouchel and Duncan
(1983) is a special case of the Hausman (1978) principle of
comparing an efficient estimator with a nonefficient one, and that
(14.12) holds quite generally for large sample sizes. To explain this
test, denote by the ML or REML estimator for the vector of
fixed parameters γ in model (5.15) and by a design-based
estimator. Similarly, denote the estimated covariance matrices of
these estimators (with squared standard errors on the diagonal) by
and . The test statistic for comparing the two estimators is
then

and under the null hypothesis that the sampling design is

noninformative, it has an asymptotic chi-squared distribution with r
+ 1 degrees of freedom, where r + 1 is the number of elements of
parameter γ. This test can also be applied componentwise: if γh is the
fixed coefficient of variable Xh (which is the intercept if h = 0),9 then

can be used for testing the noninformativeness of the design

specifically for coefficient γh; under the null hypothesis of
noninformativeness this has an asymptotic standard normal
distribution.

A difficulty with these tests, however, is that the estimated

covariance matrix of the weighted estimator may be quite unstable,
especially for small or moderately large level-two units. For
example, it is possible that the difference under the square root in the
denominator of (14.16) is negative. This makes the tests unreliable.
This difficulty was stressed by Asparouhov (2006). He proposed
instead to use these calculations to calculate a measure for the
informativeness of the sampling design, defined by
 This can be regarded as an effect size for the deviation between the
model-based and the design-based estimators: when this difference is
small, the difference between the two estimates is not important,
whether or not it is significant.

We should note that the caveats, expressed at the end of item 4,

against relying too strongly on testing differences between model-
based and design-based estimates, apply also here.

14.4 Example: Metacognitive strategies as

measured in the PISA study, USA 2009
As an elaborate example, this section is concerned with an analytic
investigation of metacognitive strategies as measured in the international
PISA study, using the data collected in 2009 in the USA (OECD, 2010).
The dependent variable is indicated by the name METASUM in the PISA
documentation. This is a measure of a metacognitive aspect of learning,
namely the awareness of appropriate strategies to summarize information.

14.4.1 Sampling design

The survey design for the PISA 2009 study in the USA is a stratified two-
stage sample. The first step was stratification by the combined classification
of school type (private and public) and region (Midwest, Northeast, South,
and West). Then within each stratum a two-stage sample was taken, in total
of 5,233 students within 165 schools.
The numbers of sampled schools and students are listed in Table 14.1. It
can be concluded that the number of private schools sampled is so small
that meaningful inferences for the population of private schools, as distinct
from the population of public schools, are hardly possible. It is possible to
estimate a model for the public and private schools jointly, and test for
differences between these two populations; but this test will have very low
power in view of the small sample size for private schools.

Table 14.1: Numbers of sampled schools (left) and sampled pupils (right)
per stratum, for PISA data, USA, 2009.

The data set contains two weight variables, W_FSTUWT at the student
level and W_FSCHWT at the school level. The documentation (OECD,
2010, p. 143) mentions that these are to be transformed as follows. The
student-level weights are to be divided by their school average:

In other words, the sum ∑i w1ij of the student-level weights per school is
equal to the sample size nj of this school. This is scaling method 2 of
(14.11b). The school-level weights are to be normalized so that their sum
over all students in the data set is equal to the total number of sampled
students:

The technical documentation for the PISA 2009 survey was unavailable
at the time of writing (early 2011), and therefore the technical report of the
PISA 2006 survey (OECD 2009) is used for further background
information. This report indicates (Chapter 8) that within the strata, schools
were sampled proportional to total enrollment of the schools, and within
schools pupils were sampled randomly (with equal probability). The
weights given in the data set reflect not only the selection probabilities but
also adjustment for nonresponse (so-called post-stratification).
To investigate how the sampling design could or should be included in
the analysis, we shall potentially consider the eight strata separately. This is
because they were sampled independently; the public–private dimension is
substantively meaningful and the geographic dimension is potentially
meaningful; and the number (eight) of separate analyses is still just
manageable.
As a first step we wish to gain some insight into the variability of the
weights. We calculate the weights (14.18) and (14.19) and compute the
design effects (14.10). It turns out that all level-one design effects are
between 0.95 and 1. This means they are hardly variable, and unequal
weights between pupils cannot lead to important adjustments or biases. The
level-two design effects must be considered per stratum. They are listed
only for the public schools in Table 14.2; these numbers would be
meaningless for the private schools, given the low number of such schools.
This means, for example, that, for the purpose of estimating the mean of a
homoscedastic variable, the sample for public schools in the South and
West has the precision of a sample which has less than 20% of the size of a
simple random sample. The precise definition of homoscedasticity in this
case is that the absolute deviation from the mean is unrelated to the
sampling weights.

Table 14.2: Level-two design effects for PISA data (USA 2009).

Illustration: descriptive analysis

To illustrate the implications of the nonconstant sampling weights for
schools, consider first the descriptive analysis of three parameters, for the
population of public schools in the South: the proportion of schools in
villages, in towns, and in cities. Table 14.2 reports the design effect for
schools within this stratum as 0.16, which indicates rather high although not
outrageous variability of sampling weights. Since sampling of schools was
done with probabilities proportional to school enrollment, and larger
populations will be served by bigger schools, the school weights will be
rather strongly associated with schools being situated in a village
(positively) or in a city (negatively), and not or only weakly with being
situated in a town.
We compare the design-based and model-based approaches. The design-
based approach uses the probability weighted estimates (14.2) with the
corresponding standard errors (14.3). The model-based approach makes the
usual model assumption of an ‘independent and identically distributed
random sample’, implying that the sample elements have a constant
variance (homoscedasticity), for which the estimate of the population
average is the unweighted sample mean with the usual standard error of the
mean. Table 14.3 presents the results. The estimated proportions of schools
situated in villages, towns, and cities are, respectively, 0.18, 0.65, and 0.17.
Large schools are oversampled, leading to smaller average weights in cities,
whereas small schools are undersampled, leading to larger average weights
in villages. This is reflected by the sample proportions: relatively more
sampled schools are in cities, relatively fewer in villages. The sample
proportions for schools in villages and towns would be biased if considered
as estimates for the population proportions.

Table 14.3: Design-based and model-based estimates for three proportions,

for public schools in the South.

The estimated proportion in towns, 0.65, does not differ much from the
sample proportion, 0.56. The difference is less than one standard error
(0.13). This is in line with the small association between sampling weights
and being situated in a town. The total sample size here (number of sampled
public schools in the South) is 55 (see Table 14.1). If in a simple random
sample of size 55 a proportion of 0.65 were observed, the standard error of
the estimated proportion would be 0.06, against the standard error of the
weighted estimator S.E.p( p) = 0.13. This indicates the loss of efficiency in
using a weighted instead of a simple random sample, when the variable of
interest is not associated with the weights.

14.4.2 Model-based analysis of data divided into parts

Let us suppose that the research question is how metacognitive competence
depends on gender, age (also reflected by grade), socio-economic status,
and immigrant status. The following explanatory variables are used:

* Gender; female = 0, male = 1.

Grade, originally ranging from 8 to 12; centered at 10, new range –2
*
to +2.
* Age, between 15 and 17 years, centered at 16 years.
ESCS, the PISA index of economic, social and cultural status of
* students; in this data set this has mean 0.15 and standard deviation
0.92.
Immigration status, recoded to: 0, at least one parent born in the
USA; 1, second generation (born in the USA, both parents born
*
elsewhere); 2: first generation (born outside the USA, parents
likewise). This ordinal variable is treated as having a linear effect.

For ESCS and immigration status, the school average is also included in the
model, to allow differences between within-school and between-school
regression coefficients. These school means are denoted Sch-imm and Sch-
ESCS. The centering means that the intercept refers to girls of age 16 and
grade 10, with the other variables mentioned taking value 0.
The main design variables are the stratification variables, public/private
and region; and school size, determining inclusion probabilities within
strata. Of these variables, public/private is of primary interest, the other two
variables are of secondary interest only. Therefore, we do not initially
include region and school size in the model and rather follow method 3 (p.
226): divide the data set into parts dependent on the design, and apply the
hierarchical linear model to each of the parts. Since only 11 out of the 165
schools are private and this is a category of interest, the private schools are
considered as one part, and the public schools are divided into four groups
according to the school weights (14.19). The private schools have too few
level-two units for a good multilevel analysis, but this is an intermediate
step only and will give an impression of whether they differ from the public
schools. Table 14.4 presents the results. The classes denoted by ‘weight 1’
to ‘weight 4’ are the schools with the smallest to largest weights, therefore
the relatively largest to smallest schools. Pupils with missing data on one or
more variables were left out; this amounted to about 10% of the data set.
The group of private schools seems to be the most different from the
rest, but the sample size is small. The main difference in this respect are the
larger effects of the school means of ESCS and of immigrant status; the
lower intercept for the private schools may be related to the higher effect of
school mean of ESCS together with the fact that these schools on average
have a higher ESCS than the public schools. The private schools also have a
larger intercept variance, but this is an estimate based on only 11 schools
and therefore not very precise. The estimates for the four weight groups of
public schools do not differ appreciably, given their standard errors. These
differences can be assessed more formally using test (3.42). The variable
that leads to the most strongly different estimates is the school mean of
immigration status, for which this test applied to the estimates and their
standard errors in Table 14.4 leads to the test statistic C = 7.0, a chi-squared
variable with 3 degrees of freedom, with p = 0.07. Being the smallest p-
value out of a set of eight variables, this is not really alarming. However, it
turns out that also the average immigrant status in the sample is different in
the four different weight groups of public schools, ranging from 0.08 in
weight group 4 to 0.59 in weight group 1. It is likely that this reflects
urbanization rather than weights, or school size, per se. Indeed, the average
immigrant status in the sample ranges from 0.07 in villages to 0.76 in large
cities. Therefore urbanization, a variable in five categories, is added to the
data and the estimation is repeated. The results are in Table 14.5. It should
be noted that for the private schools this is hardly a meaningful model, as it
includes four school variables for 11 schools.
Controlling for urbanization hardly affects the effect of within-school
variables and makes the coefficients for the two school averages, especially
of immigration status, more similar across the four weight categories of
public schools. The parameters of the school averages for the private
schools, on the other hand, now deviate even more from those of the public
schools than without the control for urbanization.
 Table 14.4: Estimates for model for metacognitive competence for five
parts of the data set.

From this exercise we have learnt that there are differences within the
set of public schools, dependent on the design, relating to the between-
school effect of immigrant status, and that these differences can be
attenuated by controlling for urbanization. Furthermore, we saw that the
private schools differ from the public schools with respect to the
coefficients of various of the explanatory variables. These conclusions will
be utilized in the further analyses.

14.4.3 Inclusion of weights in the model

We now continue with the model as specified in Table 14.5, separately for
the data sets of the public schools and of the private schools. The level-one
weights (14.18) vary so little that it is meaningless to investigate their
effects. Including the weights (14.19) in the data set for the public schools,
as main effects, as interactions, or in the random part of the model, does not
lead to significant improvements of the model. The same holds for adding
school size, also after a square root transformation to make its distribution
less skewed. It is not interesting to report the numbers here. This suggests
that for constructing a model for the metacognitive competence, given that
the model controls for urbanization, it might be possible to get good results
without further taking the design into account. In the sequel we shall
present the model-based estimates and compare them with the design-based
estimates.

Table 14.5: Estimates for model for metacognitive competence, including

urbanization, for five parts of the data set.

For the private schools a satisfactory analysis is impossible because

there are only 11 schools in the sample. Further, the effective sample size as
calculated by (14.9a), using the weights (14.19) for only the set of private
schools, is only 3.1. This suggests that on the basis of this design it is not
possible to say anything about effects of school variables on metacognitive
competence for private schools.

14.5 How to assign weights in multilevel models

Design-based analytic inference usually is based on superpopulation
models. As usual, the survey takes a sample from a finite population, for
example, the set of all pupils of a specific school type in a certain country.
In addition, it is assumed that this finite population is a sample from a
hypothetical infinite population, representing the scientific regularities that
are being analyzed, one may say, alternative situations that also ‘could have
been the case’. Thus there are two steps of inference: from the sample to the
finite population, and from the finite population to the infinite population.
The first step uses the sampling weights to remove the biases that may have
been caused by informative sampling designs. The second step is a model-
based part of the analysis. In extreme cases the sample covers the entire
finite population, and only the second step of inference remains. (Inferential
procedures are based, however, on a combined model for the finite
population and the superpopulation, and one should not expect two steps to
be discernible in the procedures.) This shows why it is meaningful to apply
statistical inference to analytic questions even if one has observed the entire
(finite) population. It also shows that a superpopulation model leads to
larger uncertainty (larger standard errors) than a sampling model with only
a finite population: the inference from finite population to superpopulation
adds extra uncertainty.
In the multilevel case, the superpopulation has a multilevel structure and
the hierarchical linear model holds here; the finite population can be
regarded as a multi-stage sample from this superpopulation, so that the
hierarchical linear model also holds for the finite population. The observed
sample is drawn from the finite population according to a sampling design
which may be complex, involving stratification and clustering. If the
residuals are correlated with the design variables, a situation called
informative sampling, the hierarchical linear model does not hold in the
same form for the observed sample as it does for the population, and
analytic inference has to take the sample design into account to provide
unbiased estimators. The preceding sections presented ways to find out
whether it seems plausible that the sampling design is noninformative, and
proposed to proceed with a regular hierarchical linear model analysis if this
is indeed the case. The current section discusses methods of parameter
estimation when there is likely to be some association between sampling
design and residuals. These methods use the sampling weights. Like the rest
of this chapter, the presentation here focuses on two-level models. Level-
two units also are called clusters.
In the first place it should be mentioned that if there are weights only at
level two, and sampling at level one (within the sampled clusters) was done
randomly, so that the level-one sampling weights wi|j do not depend on i,
then the weighting is relatively straightforward, and the problems and
ambiguities mentioned below do not occur. Complications occur if there are
nonconstant weights at level one.
A second point is that the design-based methods may be unreliable if
the number of clusters is small, where the boundary line between small and
large enough will normally be between 20 and 40. Multilevel surveys with a
small number of units at the highest level, for example, country-
comparative studies with rather few countries, in many cases fortunately
have large sample sizes at the lower level. In such cases a two-step
approach (Section 3.7; Achen, 2005) is preferable to a multilevel design-
based estimation for the entire survey: the first step consists of separate
analyses for each cluster, the second step of combination of the results,
using the estimated clusterwise regression coefficients as dependent
variables. If the effective sample sizes per cluster are large, then the
parameters can be estimated very precisely for each cluster separately. Then
it is not necessary to borrow strength from other highest-level units, so then
this argument for conducting a multilevel analysis across clusters does not
apply. A practical advantage of this approach is that the within-cluster
survey will often have only one level of weights, so that the design-based
estimation by cluster can proceed in the more straightforward single-level
weighted way (e.g., Fuller, 2009, Chapter 6; Lohr, 2010, Chapter 11).

Procedures and scaling methods

The weighting methods for hierarchical linear models most often used in
the literature are the method developed by Pfeffermann et al. (1998), using
the sampling weights in a probability-weighted iterative generalized least
squares (PWIGLS) method; and the method developed by Rabe-Hesketh
and Skrondal (2006) and also by Asparouhov (2006), extending earlier
work by Grilli and Pratesi (2004), maximizing a pseudo-likelihood
incorporating the weights. The method of Rabe-Hesketh and Skrondal
(2006) and Asparouhov (2006) is applicable to more general multilevel
designs (e.g., three or more levels), and for nonnormally distributed
dependent variables (generalized linear mixed models) may be expected to
perform better than the method of Pfeffermann et al. (1998).
These two weighting methods also differ in the estimates used for the
standard errors (or, more generally, the covariance matrices of the
estimators). Pfeffermann et al. (1998) use a design-based estimator, while
Rabe-Hesketh and Skrondal (2006) and Asparouhov (2006) use a sandwich-
type estimator (see Section 12.2). MLwiN, HLM 6, and LISREL 8.7
implement the methods of Pfeffermann et al. (1998), with a choice between
the design-based estimator and a sandwich estimator for standard errors.
The method of Rabe-Hesketh and Skrondal (2006) and Asparouhov (2006)
is implemented in Mplus and Stata (gllamm). A warning is in order when
using weights in statistical software: it is always important to determine
whether the weighting method uses probability weights (also called design
weights) or precision weights (also called frequency weights) (cf. the
discussion in Section 14.2). For the purposes of design-based estimation the
former type of weights is required. For multilevel models, a telltale sign of
the distinction is that probability weighting methods require separate
weights (14.8) for the different levels.
In addition to the choice between estimation methods, there is a choice
of the scaling method for the level-one weights: no scaling, or the two types
of scaling 1 and 2 mentioned above in (14.11).
Thus quite a number of procedures exists, based on the combination of
estimation method and scaling method. A clear comparison from a
mathematical point of view is given by Bertolet (2008, Section 2.2; 2010).

Properties of these model-based estimators

The following results have been obtained concerning the properties of the
PWIGLS method of Pfefferman et al. (1998) and the pseudo-maximum
likelihood method of Asparouhov (2006) and Rabe-Hesketh and Skrondal
(2006).
The weighted estimators for the fixed parameters according to the
method of Asparouhov (2006) and Rabe-Hesketh and Skrondal (2006) are
approximately unbiased if the effective cluster sizes are large, for both types
of scaling, provided that the level-two random effects are independent of
the design variables (Asparouhov, 2006). The estimators for any of these
methods are consistent for the fixed effects if the number of clusters
becomes large, and for the parameters of the random part if the number of
clusters as well as all effective cluster sizes become large (Bertolet, 2008).
Simulation studies comparing various specifications for a single estimation
method were presented in the original articles by Pfeffermann et al. (1998),
Rabe-Hesketh and Skrondal (2006), and Asparouhov (2006). Bertolet
(2008, 2010) presents a careful summary of earlier simulation studies, and
presents a simulation study comparing the two estimation methods and
various scaling methods.
With respect to the choice of a scaling method, it can be concluded from
published theoretical and simulation studies that, if it is likely that the
design is informative, some kind of scaling should be used. For large cluster
sizes, the differences between the scaling methods are small. However,
evidence about the choice between scaling methods 1 and 2 for smaller
cluster sizes is equivocal. Based on theoretical arguments for simple
models, Potthoff et al. (1992) as well as Korn and Graubard (2003) prefer
scaling method 1. Based on simulations, Pfeffermann et al. (1998), Rabe-
Hesketh and Skrondal (2006), and Asparouhov (2006) conclude with a
tentative preference for scaling method 2. Stapleton (2002) reports
simulation studies of multilevel structural equation models where scaling
method 1 performs better than method 2. Grilli and Pratesi (2004), on the
other hand, found that the procedure of Pfeffermann et al. with scaling
method 2 performs well in many realistic situations. Bertolet (2008, 2010)
finds in simulations on balance a preference for scaling method 1.
The results are quite encouraging for the estimators of the fixed effects
for medium sample sizes, but suggest that coverage probabilities of
confidence intervals require large sample sizes. These coverage
probabilities, which are equivalent to type I error rates of hypothesis tests,
are essential criteria for the reliability of these methods and integrate bias,
estimation variance, and standard error estimation, but unfortunately many
published simulation studies fail to report them. In a simple linear two-level
model with 35 clusters, Pfeffermann et al. (1998) report reasonable
precision for cluster sizes of about 40, but do not report coverage
probabilities. Rabe-Hesketh and Skrondal (2006) report good coverage
probabilities for a simulation of a two-level logistic regression model of 500
clusters with cluster size 50. Asparouhov (2006) presents a simulation study
of an empty linear model with 100 clusters, and finds coverage probabilities
for the fixed effect which are sometimes good and sometimes (depending
on intraclass correlations and degree of informativeness of the sampling
design) much too low for cluster size 20, and mostly good coverage
probabilities for cluster size 100. Bertolet (2008, 2010) finds unacceptably
low coverage probabilities for linear models and normal distributions in
simulations of 35 clusters with cluster size 20. It must be concluded that
published simulations have only covered a few combinations of sample
sizes and model specifications, that tests and confidence intervals can be
unreliable for small sample sizes, and that as yet there is no support for
good coverage probabilities unless there are at least 100 clusters and cluster
sizes are at least 50.
The various methods have also been compared for empirical data sets.
This was also done in the original articles introducing the methods. Further,
Carle (2009) compares all these methods for a data set on children with
special health care needs in the USA. Zaccarin and Donati (2009) compare
different weighting and scaling methods for a subset of the 2003 PISA data,
an international comparative study of educational outcomes. Carle finds
only quite small differences between different methods for his data set, but
Zaccarin and Donati find large differences in cases where the level-one
weights are strongly variable. Neither of these studies, however, attempted
to achieve extensive control for relevant level-one predictor variables that
might have explained away the association between the outcome variable
and the sampling design.

Conclusions
All these simulation studies are based on limited designs and cannot be
regarded as conclusive. It can be concluded that for the estimation of fixed
effects in linear multilevel models, the methods mentioned all give good
results, perhaps excluding the case of a very low number of clusters. For the
estimation of random effects and nonlinear multilevel models, the method
of Rabe-Hesketh and Skrondal (2006) and Asparouhov (2006) seems better
than the method of Pfeffermann et al. (1998). It is necessary to apply
scaling to the level-one weights, but it is not clear which of the two methods
(14.11a) or (14.11b) is better, and this may depend on the design, the
parameter values, and what is the parameter of main interest. Carle (2009)
suggests using both scaling methods; and, if these produce different results,
to conduct simulation studies to determine, for the given design and
plausible parameter values, which is the preferable scaling method. This
seems to be good advice.
The estimation of standard errors, however, still is a problem. Coverage
probabilities of confidence intervals (and therefore also type I error rates of
hypothesis tests) can be too low for small sample sizes. Extant simulation
studies suggest that a survey of 35 clusters of size 20 is too small to apply
the design-based methods discussed here (Bertolet, 2008, 2010), while 100
clusters of size 100 is sufficient (Asparouhov, 2006). How generalizable
this is, and what happens between these two boundaries, is still unknown.
Perhaps bootstrap standard errors (see below) give better results, but this
still is largely unexplored.
For two-level surveys with few clusters, but large cluster sizes, a good
alternative may be to follow a two-step approach with a first step of
separate estimations for each cluster, where single-level weighting may be
used to account for sampling weights (cf. Skinner, 1989; Lohr, 2010,
Chapter 11) and a second step combining the results for the different
clusters (cf. Section 3.7).
Example 14.1 Design-based and model-based analysis of PISA data.
We continue the example of Section 14.4 of metacognitive competence in the PISA 2009 data for the
USA. In Section 14.4 we learned that level-two weights are important in this data set; schools were
sampled with inclusion probabilities dependent on school size. Level-one weights are hardly variable,
and therefore not important. We also learned that it may be good to control for urbanization.
Furthermore, private schools were perhaps different from public schools; but the number of private
schools sampled, 11, was too low to say much about this question with any confidence.
Based on this information, it was decided first to retain the entire data set, control for main effects
of urbanization as well as private/public; and compare the design-based estimates, computed by
MLwiN according to the method of Pfeffermann et al. (1998), employing scaling method 2, with the
model-based estimates according to the hierarchical linear model. The results are presented in Table
14.6. Deviances are not given, because these are not meaningful for the pseudo-likelihood method.
To compare the two kinds of estimate, Asparouhov’s (2006) informativeness measure I2 is given, as
presented in (14.17).
The results for the student-level effects correspond rather well, but those for the school-level
effects do not. I2 expresses the difference in terms of the model-based standard error. Values larger
than 2 can be considered unacceptable, and these occur for the two school averages and also for some
of the control variables for urbanization.
To diagnose the differences, first the private schools were left out, because of the possibility of
differences between private and public schools and the lack of enough information on the private
schools, and the two estimates were compared again. The differences remained (results not shown
here). Therefore a residual analysis was done (Chapter 10), with particular attention to the school
sizes, as these were the main determinants of the inclusion probabilities. It turned out that there is one
outlying school, having 6,694 pupils while the other school sizes ranged from 100 to 3,592, and with
results concerning metacognitive competence quite different from the other schools. This school was
also left out of the data set, leaving a total sample of 153 public schools with 4,316 pupils. (An
alternative would be to retain this school in the data set and represent it by a dummy variable.) In
addition, school size was used as an additional control variable; in view of the skewness of school
size, the square root of school size was used, centered at 35 (corresponding to a school size of 1,225,
close to average school size). Table 14.7 presents the results of the design-based and model-based
estimates for this smaller data set.

Table 14.6: Design-based and model-based estimates for model for

metacognitive competence, entire data set.

The results of the two approaches now are more in line with each other. The main difference now
is the coefficient of age, which is just significant for the model-based analysis, and small but still
positive for the design-based analysis, and has an informativeness measure slightly above 2. Further
explorations led to the conclusion that this difference may have to do with an interaction between age
and school size, or urbanization – these variables are too strongly associated to disentangle their
interaction effects. Table 14.8 shows the results of the two types of analysis, now including the
interaction of age with school size.
It turns out that in this model the differences between the two approaches are further reduced,
with all informativeness measures less than 2, and identical conclusions with respect to significance
of the variables in the original model of interest. The remaining differences may be regarded as
random variation. The model-based results have smaller standard errors, confirming the greater
precision of these estimators.

Table 14.7: Design-based and model-based estimates for model for

metacognitive competence, public schools without outlier.

As a final conclusion of this data analysis, it seems reasonable to present the model-based results
of Table 14.8. The main interpretations are the following:

* Girls have on average higher metacognitive competences than boys.

* Metacognitive competences increase with grade level.
Within the same grade, older students have on average lower metacognitive competence than
* younger ones. This may be because of reasons that also led to delays in the school careers of
the older pupils in the same grade.
Students from families with more economic, social and cultural resources, and those in schools
* where on average families have more of these resources, have higher metacognitive
competences.
* Unexplained variation is mainly at the individual level, and quite small at the school level.

This should be accompanied, however, by the following caveats:

 The observational and cross-sectional nature of the data precludes causal interpretations of the
*
above conclusions.

Table 14.8: Design-based and model-based estimates for model for

metacognitive competence, public schools without outlier, with more
extensive controls.

This represents the public schools only. The private schools seem to present a somewhat
* different picture, but their sampled number is too small to draw conclusions about these
differences.
There was one outlying very large public school which also seemed to present a different
*
picture, and which was left out of the data for the final results.
There may be interactions of age with other variables on metacognitive competence. Note that
the effect of age is already controlled for grade, and the age range in this data is small (between
*
15.2 and 16.4 years). These interactions could be with school size or with urbanization. These
variables are associated and their interaction effect cannot be disentangled.
A further analysis could investigate differences between the four regions used for the
*
stratification. These differences were not considered here.
 Other methods
Since this is an area in development, it may be good to mention some other
procedures that have been proposed in the literature.
Korn and Graubard (2003) proposed another weighting method,
requiring knowledge of higher-order inclusion probabilities which often is
not available; therefore it is less widely applied. However, in their
simulations their method does appear to perform well. Bertolet (2008,
Section 2.2; 2010) explains how their method compares to the two methods
mentioned above.
Estimation of standard errors by bootstrap methods was proposed and
studied by Grilli and Pratesi (2004) and by Kovacěvicć et al. (2006), and
seems to perform well. Given the difficulties for standard error estimation,
this seems a promising method.
Pfefferman et al. (2006) propose a totally different approach. They
construct a simultaneous model for the random sample (the ‘inclusion
events’) and the dependent variable, conditional on inclusion in the sample.
This approach does not use weighting: it is model-based, not design-based.
Little (2004) and Gelman (2007) suggest a Bayesian approach by giving
random effects to subsets of the population having constant (or similar)
inclusion probabilities. This approach sacrifices unbiasedness, but may
obtain smaller posterior variances or smaller mean squared errors in return.

14.6 Appendix. Matrix expressions for the single-

level estimators
For readers with a knowledge of matrix algebra it may be helpful to see the
matrix expressions for the PWE and WLS estimators. The linear model is
defined by

In both cases, (14.1) as well as (14.4), the estimator is defined by

In the first case, (14.1), the most frequently made assumption is that the
residuals are homoscedastic, cov(R)= σ2I, giving the covariance matrix

In the second case (14.4) it is assumed that the residuals are heteroscedastic,
cov(R) = σ2W−1, yielding

14.7 Glommary
Two-stage sample. A sample design where level-two units are chosen
independently with some probability, and given that a level-two unit j
has been selected, a sample of level-one units within this level-two
unit is chosen.

Cluster sample. A sample design where level-two units are chosen

independently with some probability, and given that a level-two unit j
has been selected, all level-one units within this level-two unit are
included in the sample.

Stratified sample. A sample design where the population is divided into

subsets (strata), and samples are drawn independently within each
stratum, with sampling proportions which may differ between strata.

Inclusion probabilities, sampling probabilities. At level two, the

probabilities that given level-two units are included in the sample; at
level one, the probabilities that given level-one units are included in
the sample, under the condition that the level-two unit of which they
are a part is already included.

Design variables. Variables that determine the inclusion probabilities and

other elements of the design.
Descriptive inference. The estimation and testing of descriptive parameters
of the population, as done in official statistics.

Analytic inference. Statistical inference about the way in which a variable

depends on, or is associated with, other variables.

Model-based inference. Statistical inference based on the assumption that

the data can be regarded as the outcome of a probability model (e.g., a
linear model with normal residuals) with some unknown parameters.
The inference is directed at obtaining knowledge about the likely
values of these parameters.

Design-based inference. Statistical inference based on the probability

mechanism used for the sample selection. Somewhat more generally, a
design-based approach to analytic inference refers to analyzing the
survey data according to a postulated model, in this book a hierarchical
linear model, while taking the design into account.

Sampling weights or survey weights. Weights for data points, inversely

reflecting the probability that any particular population element is
included in the sample.

Precision weights, analytical weights, frequency weights. Weights for

data points inversely reflecting the residual standard deviations. An
example, the case of frequency weights, is the case where a data point
with a larger weight is the average of a larger number of basic
population elements.

Probability weighted estimator. An estimator obtained by using weights

that are reciprocals of inclusion probabilities. If appropriately used,
this gives unbiased or nearly unbiased estimators.

Weighted least squares. An estimation method defined by minimizing the

sum of weighted squared deviations between the observed values of
the dependent variable and their fitted values.

Effective sample size. A sample, with a complex sample design and having
effective sample size neff, gives the same amount of information as a
 simple random sample of size neff; the usual formula holds, however,
only for the estimation of population means of homoscedastic random
variables.

Design effect. Effective sample size divided by actual sample size.

Model of interest. The probability model on which the research is

substantively focused; in this book, it takes the form of a hierarchical
linear model.

Informative design. A design is noninformative for a model of interest if

the residuals in the model of interest are independent of the design
variables.

Scaling of weights. Methods to scale the level-one weights within a given

level-two unit, either (method 1) to make them sum to the effective
sample size, or (method 2) to make them sum to the actual sample size.

Superpopulation model. The assumption of an (infinite) superpopulation

for which the hierarchical linear model holds; and from which the
actual data are a sample, drawn perhaps according to a complex
sample design.

Pseudo-maximum likelihood. An approach to estimation in

superpopulation models, based on a modified maximum likelihood
equation in which weights are inserted that reflect the sample design.

Sandwich estimator or robust estimator. A procedure for estimating

standard errors for which the mathematical equation has the form of a
sandwich (one matrix sandwiched between two other matrices), and
which is more robust against deviations from the model than standard
errors calculated directly from the model equations.

1
The distinction between sampling with and without replacement is not discussed here.
2
For generalized linear models (e.g., multilevel logistic regression), it may be more enlightening not
to talk about residuals at the lowest level, but about the conditional distribution of the data given the
linear predictor.
3
See Section 12.1.
4
In other words, low mean squared error.
5
It should be noted that this measure for the design effect and its interpretation of the loss of
precision applies only to the case of estimating a mean of a homoscedastic population; for the
estimation of regression coefficients the formulas are more complicated.
6
It is assumed here that finite population corrections are ignored.
7
This relation does not hold in general. It is valid here because of the independence between the
OLS estimator and the difference jw− jOLS. The matrix expression for (14.12) is σ2 ((X′WX)−1
(X′W2X) (X′WX)−1 — (X′X)−1); cf. (14.20).
8
There are two different Fisher Z transformations; the better-known of these is for the correlation
coefficient, while this one is for the F distribution. See Fisher (1924) and Fisher and Yates (1963, p.
2) (‘The quantity z…’). For this reference we are indebted to sir David Cox.
9
Note the difference in notation: what in the clusterwise method was denoted by γj is the (r + 1)-
dimensional parameter vector for the jth cluster, whereas what is here denoted by γh is the single
number which is the parameter for variable Xh.
 15
Longitudinal Data

The paradigmatic nesting structure used up to now in this book has been the
structure of individuals (level-one units) nested in groups (level-two units).
The hierarchical linear model is also very useful for analyzing repeated
measures, or longitudinal data, where multiple repeated measurements of
the same variable are available for a sample of individual subjects. Since
the appearance of the path-breaking paper by Laird and Ware (1982), this is
the main type of application of the hierarchical linear model in the
biological and medical sciences. This chapter is devoted to the specific two-
level models and modeling considerations that are relevant for data where
the level-one units are measurement occasions and the level-two units
individual subjects.
Of the advantages of the hierarchical linear model approach to repeated
measures, one deserves to be mentioned here. It is the flexibility to deal
with unbalanced data structures, for example, repeated measures data with
fixed measurement occasions where the data for some (or all) individuals
are incomplete, or longitudinal data where some or even all individuals are
measured at different sets of time points. Here it must be assumed that the
fact that a part of the data was not observed does not itself contain some
information about the unobserved values; see Chapter 9 for a treatment of
incomplete data.
The topic of repeated measures analysis is too vast to be treated in one
chapter. For a general treatment of this topic we refer the reader to, for
example, Maxwell and Delaney (2004) and Fitzmaurice et al. (2004). In
economics this type of model is discussed mainly under the heading of
panel data; see, for example, Baltagi (2008), Chow (1984), and Hsiao
(1995). Some textbooks treating repeated measures specifically from the
perspective of multilevel modeling are Verbeke and Molenberghs (2000),
Singer and Willett (2003, Part I), and Hedeker and Gibbons (2006). The
current chapter only explains the basic hierarchical linear model
formulation of models for repeated measures.
This chapter is about the two-level structure of measurements within
individuals. When the individuals, in their turn, are nested in groups, the
data have a three-level structure: longitudinal measurements nested within
individuals nested within groups. Models for such data structures can be
obtained by adding the group level as a third level to the models of this
chapter. Such three-level models are not explicitly treated in this chapter.
However, this three-level extension of the fully multivariate model of
Section 15.1.3 is the same as the multivariate multilevel model of Chapter
16.

OVERVIEW OF THE CHAPTER

The chapter is divided into two parts. The first part deals with fixed
occasion data structures, defined by a fixed set of measurement occasions
for all individual subjects – where any number of measurements may,
however, be randomly missing for any of the subjects. For these data
structures we treat several specifications of the residual covariance patterns,
representable by a random intercept and random slopes of transformed time
variables. The most encompassing specification is the fully multivariate
model, which leaves the residual covariance matrix completely free. This
model can also be used for multivariate data that do not have a longitudinal
nature, that is, for multivariate regression analysis.
The second part of the chapter is about variable occasion designs, where
the number and timing of measurement occasions is unrestricted across
subjects. The measurements here must be arranged along an underlying
continuum such as time or age, and may be regarded as measurements of a
population of curves. Polynomial, piecewise linear, and spline functions are
treated as families of curves that may be useful to represent such
populations. These curves may then be explained by individual (level-two)
variables, by interactions between individual variables and transformed
time variables, and by changing covariates. Finally, some remarks are made
about the possibility of replacing the assumption of independent level-one
residuals by weaker assumptions under which the level-one residuals have
some kind of autocorrelation structure.

15.1 Fixed occasions

In fixed occasion designs, there is a fixed set t = 1,. . ., m of measurement
occasions. For example, in a study of some educational or therapeutic
program, we might have an intake test, pretest, mid-program test, post-test,
and follow-up test (m = 5). Another example is a study of attitude change in
early adulthood, with attitudes measured shortly after each birthday from
the 18th to the 25th year of age. If data are complete, each individual has
provided information on all these occasions. It is quite common, however,
that data are incomplete. When they are, we assume that the absent data are
missing at random, and the fact that they are missing does not itself provide
relevant information about the phenomena studied (cf. Chapter 9).
The measurement occasions are denoted by t = 1,. . ., m, but each
individual may have a smaller number of measurements because of missing
data. Yti denotes the measurement for individual i at occasion t. It is allowed
that, for any individual, some measurements are missing. Even individuals
with only one measurement do not need to be deleted from the data set:
they do contribute to the estimation of the between-individual variance,
although they do not give information about within-individual variability.
Note that, differently from the other chapters, the level-one units are
now the measurement occasions indexed by t, while the level-two units are
the individuals indexed by i.
The models treated in this section differ primarily with respect to the
random part. Choosing between them amounts to specifying the random
part of the model, and there are three kinds of motive that should be taken
into account for this purpose.
The first motive is that if one works with a random part that is not an
adequate description of the dependence between the m measurements (i.e.,
it does not satisfactorily represent the covariance matrix of these
measurements), then the standard errors for estimated coefficients in the
fixed part are not reliable. Hence, the tests for these coefficients also are
unreliable. This motive points to the least restrictive model, that is, the fully
multivariate model, as the preferred one.
The second motive is the interpretation of the random part. Simpler
models often allow a nicer and easier interpretation than more complicated
models. This point of view favors the more restrictive models, for example,
the random intercept (compound symmetry) and random slope models.
The third motive is the possibility of getting results at all and the
desirability of standard errors that are not unnecessarily large (while not
having a bias at the same time). If the amount of data is relatively small, the
estimation algorithms may not converge for models that have a large
number of parameters, such as the fully multivariate model. Or even if the
algorithm does converge, having a bad ‘data-to-parameter ratio’ may lead to
large standard errors. This suggests that, if one has a relatively small data
set, one should not entertain models with too many parameters.
To conclude, one normally should use the simplest model for the
random part (i.e., the model having the smallest number of parameters and
being the most restrictive) that still yields a good fit to the data. The model
for the random part can be selected with the aid of deviance tests (see
Section 6.2).

15.1.1 The compound symmetry model

The classical model for repeated measures, called the compound symmetry
model (e.g., Maxwell and Delaney, 2004), is the same as the random
intercept model. It is sometimes referred to as the random effects model or
the mixed model, but one should realize that it is only a simple specification
of the many possibilities of the random effects model or mixed model. If
there are no explanatory variables except for the measurement occasions,
that is, the design is a pure within-subjects design, the expected value for
measurement occasion t can be denoted by μt and this model can be
expressed by

The usual assumptions are made: the U0i and Rti are independent normally
distributed random variables with expectations 0 and variances τ20 for U0i
and σ2 for Rit.
To fit this model, note that the fixed part does not contain a constant
term, but is based on m dummies for the m measurement occasions. This
can be expressed by the following formula. Let dhti be m dummy variables,
defined for h = 1,…, m by

Then the fixed part μt in (15.1) can be written as

and the compound symmetry model can be formulated as

the usual form of the fixed part of a hierarchical linear model.

Example 15.1 Life satisfaction.
The German Socio-Economic Panel (SOEP) is a longitudinal panel data set for the population in
Germany. We use the question on life satisfaction, which has answer categories ranging from 0
(totally unhappy) to 10 (totally happy), and study how it evolves during the ages of 55–60 years, and
how it is influenced by retirement. The survey was held each year. We use data collected in 1984–
2006 for a total of 1,236 individuals, each of whom participated from 1 to 6 years in this part of the
panel study. This is part of a sample of individuals representative of the population of Germany in the
year 1984. Individuals were included if they were heads of household, aged 55 years at one of the
moments of data collection, and had information about their employment status.
We use six measurement occasions: for ages 55, 56, ..., 60. Thus, the time variable t assumes the
values 1 (age 55) through 6 (age 60). The number of observations available for the six ages decreased
from 1,224 at t = 1 to 799 at t = 6. The first number is less than 1,236 because a few observations
with missing data were dropped. The nonresponse or nonavailability at various moments may be
considered to be independent of the variables being investigated here, so missingness may be
considered to be at random.
We begin by considering two models with a random intercept only, that is, with a compound
symmetry structure for the covariance matrix. The first is the empty model of Chapter 4. In this
model it is assumed that the six measurement occasions have the same population mean. The second
model is model (15.1), which allows the means to vary freely over time. In this case, writing out
(15.3) results in the model
Applying definition (15.2) implies that the expected score for an individual, for example, at time t =
2 (age 56 years), is

The results are in Table 15.1. However, it is dangerous to trust results for the compound
symmetry model before its assumptions have been tested, because standard errors of fixed effects
may be incorrect if one uses a model with a random part that has an unsatisfactory fit. Later we will
fit more complicated models and show how the assumption of compound symmetry can be tested.
The results suggest that individual (level-two) variation is a little more important than random
differences between measurement occasions (level-one variation). Further, the means change only
slightly from time t = 1 to time t = 6. The deviance test for the difference in mean between the six
time points is borderline significant: χ2 = 21,791.34–21,780.70 = 10.64, df = 5; p = 0.06. This near-
significance is not meaningful in this case, in view of the large data set.

Just as in our treatment of the random intercept model in Chapter 4, any

number of relevant explanatory variables can be included in the fixed part.
Often there are individual-dependent explanatory variables Zk, k = 1, ..., q,
for example, traits or background characteristics. If these variables are
categorical, they can be represented by dummy variables. Such individual-
dependent variables are level-two variables in the multilevel approach, and
they are called between-subjects variables in the terminology of repeated
measures analysis. In addition, there are often one or more numerical
variables describing the measurement occasion. Denote such a variable by
s(t); this may, for example, be a measure of the time elapsing between the
occasions, or the rank order of the occasion. Such a variable is called a
within-subjects variable. In addition to the between-subjects and within-
subjects variables having main effects, they may have interaction effects
with each other. These within–between interactions are a kind of cross-level
interaction in the multilevel terminology and represent, for example,
differences between individuals in their rate and pattern of change in the
dependent variable. Substantive interest in repeated measures analysis often
focuses on these interactions.

Table 15.1: Estimates for random intercept models.

 When the main effect parameter of the between-subjects variable zki is
denoted αk and the interaction effect parameter between zki and s(t) is
denoted γk, the model is given by

The fixed part is an extension of (15.1) or the equivalent form (15.3), but
the random part is still the same. Therefore this is still called a compound
symmetry model. Inclusion in the fixed part of interactions between
individual-level explanatory variables and time variables such as s(t)
suggests, however, that the random part could also contain random slopes
of these time variables. Therefore we defer giving an example of the fixed
part (15.4) until after the treatment of such a random slope.
Classical analyses of variance methods are available to estimate and test
parameters of the compound symmetry model if all data are complete (e.g.,
Maxwell and Delaney, 2004). The hierarchical linear model formulation of
this model, and the algorithms and software available, also permit the
statistical evaluation of this model for incomplete data without additional
complications.
 Covariance matrix
In the fixed occasion design one can talk about the complete data vector

Even if there were no subject at all with complete data, the complete data
vector would still make sense from a conceptual point of view.
The compound symmetry model (15.1), (15.3) or (15.4) implies that for
the complete data vector, all variances are equal and also all covariances are
equal. The expression for the covariance matrix of the complete data vector,
conditional on the explanatory variables, is the m × m matrix

cf. p. 49. This matrix is referred to as the compound symmetry covariance

matrix. In it, all residual variances are the same and all residual within-
subject correlations are equal to

the residual intraclass correlation which we encountered in Chapter 4.

The compound symmetry model is a very restrictive model, and often
an unlikely one. For example, if measurements are ordered in time, the
correlation is often larger between nearby measurements than between
measurements that are far apart. (Only for m = 2 is this condition not so
very restrictive. In this case, formula (15.5) only means that the two
measurements have the same variance and a positive correlation.)
Example 15.2 Covariance matrix for life satisfaction.
In Table 15.1 for the empty model the estimates = 1.455 and τ02 = 1.994 were obtained. This
implies that the estimated variances of the observations are 1.455 + 1.994 = 3.449 and the estimated
within-subjects correlation is This number, however, is conditional on
the validity of the compound symmetry model; we shall see below that the compound symmetry
model gives a reasonable description of the covariance matrix, although more complicated models do
give a significantly better fit.

15.1.2 Random slopes

There are various ways in which the assumption of compound symmetry
(which states that the variance of the observations is constant over time and
that the correlation between observations is independent of how far apart
they are, as is expressed by (15.5)) can be relaxed. In the hierarchical linear
model framework, the simplest way is to include one or more random
slopes in the model. This makes sense if there is some meaningful
dimension, such as time or age, underlying the measurement occasions. It
will be assumed here that the index t used to denote the measurement
occasions is a meaningful numerical variable, and that it is relevant to
consider the regression of Y on t. This variable will be referred to as ‘time’.
It is easy to modify the notation so that some other numerical function of
the measurement occasion t gets a random slope. It is assumed further that
there is some meaningful reference value for t, denoted by t0. This could
refer, for example, to one of the time points, such as the first. The choice of
t0 affects only the parameter interpretation, not the fit of the model.
Since the focus is now on the random rather than on the fixed part, the
precise formulation of the fixed part is left out of the formulas.
The model with a random intercept and a random slope for t is given by

This model means that the rates of increase have a random, individual-
dependent component U1i, in addition to the individual-dependent random
deviations U0i which affect all values Yti in the same way. The random
effect of time can also be described as a random time-by-individual
interaction.
The value t0 is subtracted from t in order to allow the intercept variance
to refer not to the (possibly meaningless) value t = 0 but to the reference
point t = t0; cf. p. 76. The variables (U0i, U1i) are assumed to have a joint
bivariate normal distribution with expectations 0, variances τ02 and τ12, and
covariance τ01.
The variances and covariances of the measurements Yti, conditional on
the explanatory variables, are now given by

where t ≠ s; we saw the same formulas in (5.5) and (5.6). These formulas
express the fact that the variances and covariances of the outcome variables
are variable over time. This is called heteroscedasticity. Differentiating with
respect to t and equating the derivative to 0 shows that the variance is
minimal at (if t were allowed to assume any value).
Furthermore, the correlation between different measurements depends on
their spacing (as well as on their position).
Extensions to more than one random slope are obvious; for example, a
second random slope could be given to the squared value (t-t0)2. In this
way, one can perform a polynomial trend analysis to improve the fit of the
random part. This means that one fits random slopes for a number of
powers of (t - t0) to obtain a model that has a good fit to the data and where
unexplained differences between individuals are represented as random
individual-dependent regressions of Y on (t - t0), (t - t0)2, (t - t0)3, etc.
Functions other than polynomials can also be used, for example, splines
(see Section 15.2.2). Polynomial trend analysis is also discussed in Singer
and Willett (2003, Section 6.3) and Maas and Snijders (2003).
Example 15.3 Random slope of time in life satisfaction.
We continue the earlier example of life satisfaction for 55–60-year-olds. Note that the observations
are spaced a year apart, and a natural variable for the occasions is s(t) = t as used above, that is, age
recoded so that t = 1 is 55 years and t = 6 is 60 years. Thus, the time dimension here corresponds to
age. A random slope of age is now added (Model 3). The reference value for the time dimension,
denoted by t0 in formula (15.7), is taken as t0= 1, corresponding to 55 years of age.
We now investigate whether part of the differences between individuals in life satisfaction and in
the rate of change can be explained by birth cohort. Years of birth here range from 1929 to 1951. This
variable is centered at the mean, which is 1940. To avoid very small regression coefficients the
variable is divided by 10, leading to a variable Z ranging from –1.1 to +1.1, used as a between-
subjects variable. A difference of one unit in Z corresponds to a difference of 10 years in birth year.
The effect of cohort on the rate of change, that is, the interaction effect of birth year with age, is
represented by the product of Z with t. The resulting model is an extension of a model of the type of
(15.4) with a random slope:

Parameter α is the main effect for birth year, while γ is the interaction effect between birth year and
age.

Table 15.2: Estimates for random slope models.

The results are given in Table 15.2. To allow deviance tests, all results were calculated using the
maximum likelihood estimation procedure. Comparing the deviances of Models 2 and 3 shows that
the random slope of age is significant: χ2 = 32.7, df = 2, p < 0.0001. This implies that there is a
significant deviation from the compound symmetry model, Model 2 of Table 15.1. However, the
differences between individuals in rate of change are still rather small, with an estimated inter-
individual standard deviation of One standard deviation difference between
individuals with respect to slope for age will add up for the age range of 5 years to a difference of
0.16 × 5 = 0.8 which is not negligible on the scale from 0 to 10, but still less than the within-
individual standard deviation of
The effect of birth year is significant, with a t-ratio of –0.394/0.078 = 5.1, p < 0.0001. Here also
the effect size is rather small, however. The difference between lowest and highest value of Z (birth
year divided by 10) is 1.1–(–1.1) = 2.2. Given the presence of an interaction, the main effect α of
birth year corresponds to the value t = t0 where the interaction parameter γ in (15.9) cancels, that is,
55 years of age. Thus the main effect of birth year translates to a difference in expected life
satisfaction, at 55 years of age, equal to 2.2 × 0.394 = 0.87.
 The birth year × age interaction is significant, t = 0.049/0.019 = 2.6, p < 0.01. The contribution of
0.049 is of the same order of magnitude as the rather irregular differences between the estimated
means μ1 to μ6. Those born later on average experience slightly higher life satisfaction during the
period when they are 55–60 years old, those born earlier slightly lower or about the same.
The fitted covariance for the complete data vector under Model 4 has elements given by (15.8).
Inserting the estimated parameters τ02, τ12, and τ01 from Table 15.2 into (15.8) yields the covariance
matrix

and the correlation matrix

These matrices show that the variance does not change a lot, and correlations attenuate slightly as age
differences increase, but they are close to the intra-subject correlation estimated above under the
compound symmetry model as
However, these values are conditional on the validity of the model with one random slope. We
return to these data below, and will investigate the adequacy of this model by testing it against the
fully multivariate model.

15.1.3 The fully multivariate model

What is the use of restrictions, such as compound symmetry, on the
covariance matrix of a vector of longitudinal measurements? There was a
time (say, before 1980) when the compound symmetry model was used for
repeated measures because, in practice, it was impossible to get results for
other models. At that time, a complete data matrix was also required. These
limitations were gradually overcome between 1970 and 1990.
A more compelling argument for restrictions on the covariance matrix is
that when the amount of data is limited, the number of statistical parameters
should be kept small to prevent overfitting and to avoid convergence
problems in the calculation of the estimates. Another, more appealing,
argument is that sometimes the parameters of the models with restricted
covariance matrices have nice interpretations. This is the case, for example,
for the random slope variance of model (15.7). But when there are enough
data, one can also fit a model without restrictions on the covariance matrix,
the fully multivariate model. This also provides a benchmark to assess the
goodness of fit of the models that do have restrictions on the covariance
matrix.
Some insight in the extent to which a given model constrains the
covariance matrix is obtained by looking at the number of parameters. The
covariance matrix of an m-dimensional vector has m(m + 1)/2 free
parameters (m variances and m(m – 1)/2 covariances). The covariance
matrix for the compound symmetry model has only 2 parameters. The
model with one random slope has 4 parameters; the model with q random
slopes has {(q + 1) (q + 2)/2} + 1 parameters, namely, (q + 1) (q + 2)/2
parameters for the random part at level two and 1 parameter for the variance
of the random residual. This shows that using some random slopes will
quickly increase the number of parameters and thus lead to a better-fitting
covariance matrix. The maximum number of random slopes in a
conventional random slope model for the fixed occasions design is q = m –
2, because with q = m – 1 there is one parameter too many.
This suggests how to achieve a perfect fit for the covariance matrix. The
fully multivariate model is formulated as a model with a random intercept
and m – 1 random slopes at level two, and without a random part at level
one. Alternatively, the random part at level two may consist of m random
slopes and no random intercept. When all variances and covariances
between m random slopes are free parameters, the number of parameters in
the random part of the model is indeed m(m + 1)/2.
The fully multivariate model is little more than a tautology,

This model is reformulated more recognizably as a hierarchical linear

model by the use of dummy variables indicating the measurement
occasions. These dummies dhti were defined in (15.2). This leads to the
formulation
The variables Uti for t = 1, ..., m are random at level two, with expectations
0 and an unconstrained covariance matrix. (This means that all variances
and all covariances must be freely estimated from the data.) This model
does not have a random part at level one. It follows immediately from
(15.11) that the covariance matrix of the complete data vector, conditional
on the explanatory variables, is identical to the covariance matrix of (U1i,...,
Upi). This model for the random part is saturated in the sense that it yields a
perfect fit for the covariance matrix.
The multivariate model of this section can also be applied to data that
do not have a longitudinal nature. This means that the m variables can be
completely different, measured on different scales, provided that they have
(approximately) a multivariate normal distribution. Thus, this model and the
corresponding multilevel software enable the multivariate analysis of
incomplete data (provided that missingness is at random as defined in
Chapter 9).
Example 15.4 Incomplete paired data.
For the comparison of the means of two variables measured for the same individuals, the paired-
samples t-test is a standard procedure. The fully multivariate model provides the possibility of
carrying out a similar test if the data are incomplete. This is an example of the simplest multilevel
data structure: all level-two units (individuals) contain either one or two level-one units
(measurements). Such a data structure may be called an incomplete pretest–post-test design if one
measurement was taken before, and the other after, some kind of intervention.
With a different sample from the SOEP, we now consider the question whether average life
satisfaction at age 60 differs from the value at age 50. We specify this for persons born in the years
1930–1935, and alive in 1984, and consider their responses at ages 50 and 60. Of the 591 respondents
for whom a measurement for at least one of these ages is available, there are 136 who have
measurements for both ages, 112 who have a measurement only for age 50 (t = 0), and 343 who have
a measurement only for age 60 (t = 1). With the multilevel approach to multivariate analysis, all these
data can be used. The REML estimation method is used (cf. Section 4.7) because this will reproduce
exactly the conventional t-test if the data are complete (i.e., all individuals have measurements for
both variables).
The model fitted is

where the dummy variable d1 equals 1 or 0, respectively, depending on whether or not t = 1. The null
hypothesis that the means are identical for the two time points can be represented by ‘γ1 = 0’.
 Table 15.3: Estimates for incomplete paired data.

The results are in Table 15.3. The estimated covariance matrix of the complete data vector is

The test of the equality of the two means, which is the test of γ1, is not significant (t = –0.053/.141 =
0.38).

Example 15.5 Fully multivariate model for life satisfaction.

Continuing the example of life satisfaction, we now fit the fully multivariate model to the data for the
six time points t = 1,. . ., 6 corresponding to ages 55,. . ., 60. First we consider the model that may be
called the two-level multivariate empty model, of which the fixed part contains only the dummies for
the effects of the measurement occasions. This is an extension of Model 3 of Table 15.2. The
estimates of the fixed effects are presented in Table 15.4.

Table 15.4: The empty multivariate model for life satisfaction.

The estimated covariance matrix of the complete data vector is

These estimates are the ML estimates (cf. Section 4.7) of the parameters of a multivariate normal
distribution with incomplete data. They differ slightly from the estimates that would be obtained by
computing means and variances from available data with pairwise deletion of missing values. The
ML estimates are more efficient in the sense of having smaller standard errors.
An eyeball comparison with the fitted covariance matrix for the model with one random slope,
given in (15.10), shows that the differences are minor. The variances (diagonal elements) in (15.10)
are a smoother function of age than in (15.13), because the model with one random slope implies that
the variance must be a quadratic variable on age; see (15.8).
The deviance difference is χ2 = 21,748.03–21,702.42 = 45.61, df = 11 (the covariance matrix of
Model 3 has four free parameters, this covariance matrix has 15), with p < 0.0001. Thus, the fully
multivariate model results in a fit that is significantly better, but the estimated covariance matrix is
not strongly different. Again, this is a consequence of the large sample size. An advantage of the
random slope model is the clearer interpretation of the random part in terms of between-subject
differences, as was discussed in Example 15.3. Such an interpretation is not directly obvious from the
results of the fully multivariate model.
As a next step the effect of employment situation on life satisfaction is studied, while also taking
into account the birth year included in Model 4 of Table 15.2. Employment situation is a changing
explanatory variable, measured at the same moments as the dependent variable. There are three
categories: working full-time (reference category), working part-time, and not working. As this is a
level-one variable, the issue of the difference between within-group and between-group regressions,
treated in Section 4.6, comes up again. In the longitudinal case, the analog of the group mean is the
person mean. Including this in the model would, however, imply an effect from events that occur in
the future because, at each moment except the last, the person mean of a changing explanatory
variable depends on measurements to be made at a future point in time. Therefore we do not include
the person mean but, rather, the employment situation at the start of the study period, that is, at 55
years of age. This is regarded as a baseline situation, and the effects of current employment situation
are interpreted as effects of changes with respect to this baseline.

Table 15.5: The multivariate model with the effects of cohort and
employment situation.
 The estimates of the fixed effects are presented in Table 15.5. The effect of birth year and its
interaction with age is similar to Model 4 of Table 15.2. As to employment situation, compared to
full-time working, not working at age 55 has a strong negative effect on life satisfaction
and working part-time at age 55
a weak negative effect Currently
working part-time (compared to full-time) also has a weak negative effect.

Concluding remarks on the fully multivariate model

There is nothing special about this formulation as a hierarchical linear
model of a fully multivariate model for repeated measures with a fixed
occasion design. It is just a mathematical formula. What is special is that
available algorithms and software for multilevel analysis accept this
formulation, even without a random part at level one, and can calculate ML
or REML parameter estimates. In this way, multilevel software can
compute ML or REML estimates for multivariate normal distributions with
incomplete data, and also for multivariate regression models with
incomplete data and with sets of explanatory variables that are different for
the different dependent variables. Methods for such models have been in
existence for a longer period (see Little and Rubin, 2002), but their practical
use has been much facilitated by the development of multilevel software.
 The use of the dummy variables (15.2) has the nice feature that the
covariance matrix obtained for the random slopes is exactly the covariance
matrix for the complete data vector. However, one could also give the
random slopes to other variables. The only requirement is that there are
random slopes for m linearly independent variables, depending only on the
measurement occasion and not on the individual. For example, one could
use powers of (t–t0), where t0 may be any meaningful reference point, such
as the average of all values for t. This means that one uses variables

still with model specification (15.12). Since the first ‘variable’ is constant,
this effectively means that one uses a random intercept and m – 1 random
slopes.
Each model for the random part with a restricted covariance matrix is a
submodel of the fully multivariate model. Therefore the fit of such a
restricted model can be tested by comparing it to the fully multivariate
model by means of a likelihood ratio (deviance) test (see Chapter 6).
For complete data, an alternative to the hierarchical linear model exists
in the form of multivariate analysis of variance (MANOVA) and
multivariate regression analysis. This is documented in many textbooks
(e.g., Maxwell and Delaney, 2004; Stevens, 2009) and implemented in
standard software such as SPSS and SAS. The advantage of these methods
is the fact that under the assumption of multivariate normal distributions the
tests are exact, whereas the tests in the hierarchical linear model
formulation are approximate. For incomplete multivariate data, however,
exact methods are not available. Maas and Snijders (2003) elaborate the
correspondence between the MANOVA approach and the hierarchical linear
model approach.

15.1.4 Multivariate regression analysis

Because of the unrestricted multivariate nature of the fully multivariate
model, it is not required that the outcome variables Yti are repeated
measurements of conceptually the same variable. This model is applicable
to any set of multivariate measurements for which a multivariate normal
distribution is an adequate model. Thus, the multilevel approach yields
estimates and tests for normally distributed multivariate data with randomly
missing observations (see also Maas and Snijders, 2003).
The fully multivariate model is also the basis for multivariate regression
analysis, but then the focus usually is on the fixed part. There are supposed
to be individual-dependent explanatory variables Z1,. . ., Zq. If the
regression coefficients of the m dependent variables on the Zh are all
allowed to be different, the regression coefficient of outcome variable Yt on
explanatory variable Zh being denoted by γht, then the multivariate
regression model with possibly incomplete data can be formulated as

This shows that the occasion dummies are fundamental, not only because
they have random slopes, but also because in the fixed part all variables are
multiplied by these dummies. If some of the Zk variables are not to be used
for all dependent variables, then the corresponding cross-product terms
dhtizhi can be dropped from (15.14).

15.1.5 Explained variance

The proportion of explained variance at level one can be defined, analogous
to explained variance for grouped data treated in Section 7.1, as the
proportional reduction in prediction error for individual measurements,
averaged over the m measurement occasions. A natural baseline model is
provided by a model which can be represented by

that is, the fixed part depends on the measurement occasion but not on other
explanatory variables, and the random part is chosen so as to provide a
good fit to the data.
The proportion of explained variance at level one, R12, is then the
proportional reduction in the average residual variance,
when going from the baseline model to the model containing the
explanatory variables.
If the random part of the compound symmetry model is adequate, the
baseline model is (15.1). In this case the definition of R12 is just as in
Section 7.1.
If the compound symmetry model is not adequate one could use, at the
other end of the spectrum, the multivariate random part. This yields the
baseline model

which is the fully multivariate model without covariates; cf. (15.11) and
(15.12).
In all models for fixed occasion designs, the calculation of the
proportions of explained variance can be related to the fitted complete data
covariance matrix, The value of R12 is the proportional reduction in
the sum of diagonal values of this matrix when going from the baseline
model to the model including the explanatory variables.
Example 15.6 Explained variance for life satisfaction.
We continue the example of life satisfaction of people aged 55–60, now computing the proportion of
variance explained by cohort (birth year), employment situation, and the interaction of the cohort
with age as included in the model of Table 15.5; to which are added the total income after taxes,
logarithmically transformed, and satisfaction with health, also measured on a scale from 0 to 10. The
last two covariates were centered to have overall means 0. Their standard deviations are 0.67 and 2.3,
respectively.
First suppose that the compound symmetry model is used. The baseline model then is Model 2 in
Table 15.1. The variance per measurement is 1.991 + 1.452 = 3.443. The results of the compound
symmetry model with the effect of cohort and employment situation are given in Table 15.6. In
comparison with Model 2 of Table 15.1, inclusion of these fixed effects has led especially to a
smaller variance at the individual level; these variables, although some of them are changing
covariates, contribute little to explaining year-to-year fluctuations in life satisfaction. The residual
variance per measurement is 1.038 + 1.351 = 2.389. Thus, the proportion of variance explained at
level one is R12 = 1–(2.389/3.443) = 0.306.
 Table 15.6: The compound symmetry model with the effects of cohort,
employment situation, income, and health satisfaction.

Next suppose that the fully multivariate model is used. The estimated fixed effects do not change
much. The estimated covariance matrix in the fully multivariate model is given in (15.13), while the
residual covariance matrix of the model with these fixed effects, as estimated by employing the fully
multivariate model, is
The sum of diagonal values is 20.32 for the first covariance matrix and 14.13 for the second. Hence,
calculated on the basis of the fully multivariate model, R12 = 1–(14.13/20.32) = 0.305. Comparing
this to the values obtained above shows that for the calculation of R12 it does not make much
difference which random part is used. The calculations using the fully multivariate model are more
reliable, but in most cases the simpler calculations using the compound symmetry (‘random
intercept’) model will lead to almost the same values for R12.

15.2 Variable occasion designs

In data collection designs with variable measurement occasions, there is no
such thing as a complete data vector. The data are ordered according to
some underlying dimension (e.g. time or age), and, for each individual, data
are recorded at some set of time points not necessarily related to the time
points at which the other individuals are observed. For example, body
heights are recorded at a number of moments during childhood and
adolescence, the moments being determined by convenience rather than
strict planning.
The same notation can be used as in the preceding section: for
individual i, the dependent variable Yti is measured at occasions s = 1,. . .,
mi. The time of measurement for Yti is t. This ‘time’ variable can refer to
age, clock time, etc., but also to some other dimension such as one or more
spatial dimensions, the concentration of a poison, etc.
The number of measurements per individual, mi, can be anything. It is
not a problem that some individuals contribute only one observation. This
number must not in itself be informative about the process studied,
however. Therefore it is not allowed that the observation times t are defined
as the moments when some event occurs (e.g., change of job); such data
collection designs should be investigated using event history models (e.g.,
Singer and Willett, 2003, Part II; Mills, 2011). Larger numbers mi give
more information about intra-individual differences, of course, and with
larger average mis one will be able to fit models with a more complicated,
and more precise, random part.
Because of the unbalanced nature of the data set, the random intercept
and slope models are easily applicable, but the other models of the
preceding section either have no direct analog, or an analog that is
considerably more complicated. This section is restricted to random slope
models, and follows the approach of Snijders (1996), where further
elaboration and background material may be found. Other introductions to
multilevel models for longitudinal data with variable occasion designs may
be found, for example, in Raudenbush (1995) and Singer and Willett (2003,
Chapter 5).

15.2.1 Populations of curves

An attractive way to view repeated measures in a variable occasion design,
ordered according to an underlying dimension t (referred to as time), is as
observations on a population of curves. (This approach can be extended to
two- or higher-dimensional ordering principles.) The variable Y for
individual i follows a development represented by the function Fi(t), and the
population of interest is the population of curves Fi. The observations yield
snapshots of this function on a finite set of time points, with superimposed
residuals Rti which represent incidental deviations and measurement error:

Statistical modeling consists of determining an adequate class of functions

Fi and investigating how these functions depend on explanatory variables.

15.2.2 Random functions

It makes sense here to consider models with functions of t as the only
explanatory variables. This can be regarded as a model for random
functions that has the role of a baseline model comparable to the role of the
empty model in Chapter 4. Modeling can proceed by first determining an
adequate random function model and subsequently incorporating the
individual-based explanatory variables.
One could start by fitting the empty model, that is, the random intercept
model with no explanatory variables, but this is such a trivial model when
one is modeling curves that it may also be skipped. The next simplest
model is a linear function,
The value t0 is subtracted for the same reasons as in model (15.7) for the
fixed occasion design. It may be some reference value within the range of
observed values for t.
If the βhi (h = 0, 1) are split into their population average1 γh0 and the
individual deviation Uhi = βhi – γh0, similar to (5.2), this gives the following
model for the observations:

The population of curves is now characterized by the bivariate distribution

of the intercepts β0i at time t = t0, and the slopes β1i. The average intercept
at t = t0 is γ00 and the average slope is γ10. The intercepts have variance τ20
= var(U0i), the slopes have variance τ12 = var(U1i), and the intercept–slope
covariance is τ01 = cov(U0i, U1i). In addition, the measurements exhibit
random deviations from the curve with variance σ2 = var(Rti). This level-
one variance represents deviations from linearity together with
measurement inaccuracy, and can be used to assess the fit of linear
functions to the data.
Even when this model fits only moderately well, it can have a
meaningful interpretation, because the estimated slope and slope variance
still will give an impression of the average increase of Y per unit of time.
Example 15.7 Retarded growth.
In several examples in this section we present results for a data set of children with retarded growth.
More information about this research can be found in Rekers-Mombarg et al. (1997). The data set
concerns children who saw a pediatric endocrinologist because of growth retardation. Height
measurements are available at irregular ages varying between 0 and 30 years. In the present example
we consider the measurements for ages from 5.0 to 10.0 years. The linear growth model represented
by (15.17) is fitted to the data.
For this age period there are data for 336 children available, establishing a total of 1,886 height
measurements. Height is measured in centimeters (cm). Age is measured in years and the reference
age is chosen as t0 = 5 years, so the intercept refers to heights at age 5. Parameter estimates for model
(15.17) are presented in Table 15.7.
 The level-one standard deviation is so low ( = 0.9 cm) that it can be
concluded that the fit of the linear growth model for the period from 5 to 10
years is quite adequate. Deviations from linearity will be usually smaller
than 2 cm (which is slightly more than two standard errors). This is
notwithstanding the fact that, given the large sample size, it is possible that
the fit can be improved significantly from a statistical point of view by
including nonlinear growth terms.
The intercept parameters show that at 5 years, these children have an
average height of 96.3 cm with a variance of and an
associated standard deviation of The growth per year is
which has an estimated mean of 5.5 cm and standard deviation of
This implies that 95% of these children have growth rates,
averaged over this five-year period, between 5.5 – 2 × 1.3 = 2.9 and 5.5 + 2
× 1.3 = 8.1 cm per year. The slope–intercept covariance is negative:
children who are relatively short at 5 years grow relatively fast between 5
and 10 years.

Table 15.7: Linear growth model for 5–10-year-old children with retarded
growth.

Polynomial functions
To obtain a good fit, however, one can try and fit more complicated random
parts. One possibility is to use a polynomial random part. This means that
one or more powers of (t - t0) are given a random slope. This corresponds to
polynomials for the function Fi,

where r is a suitable number, called the degree of the polynomial. For

example, one may test the null hypothesis that individual curves are linear
by testing the null hypothesis r = 1 against the alternative hypothesis that
the curves are quadratic, r = 2. Any function for which the value is
determined at m points can be represented exactly by a polynomial of
degree m – 1. Therefore, in the fixed occasion model with m measurement
occasions, it makes no sense to consider polynomials of degree higher than
m – 1.
Usually the data contain more information about inter-individual
differences (corresponding to the fixed part) than about intra-individual
differences (the random part). For some of the coefficients βhi in (15.18),
there will be empirical evidence that they are nonzero, but not that they are
variable across individuals. Therefore, as in Section 5.2.2, one can have a
fixed part that is more complicated than the random part and give random
slopes only to the lower powers of (t – t0). This yields the model

which has the same structure as (5.15). For h = p + 1,. . ., r, parameter γh0 is
the value of the coefficient βhi, constant over all individuals i. For h = 1,…,
p, the coefficients βhi are individual-dependent, with population average γh0
and individual deviations Uhi = βhi – γh0. All variances and covariances of
the random effects Uhi are estimated from the data. The mean curve for the
population is given by
Numerical difficulties appear less often in the estimation of models of this
kind when t0 has a value in the middle of the range of t-values in the data
set than when t0 is outside or at one of the extremes of this range.
Therefore, when convergence problems occur, it is advisable to try and
work with a value of t0 close to the average or median value of t. Changing
the value of t0 only amounts to a new parametrization, that is, a different t0
leads to different parameters γh0 for which, however, formula (15.20)
constitutes the same function and the deviance of the model (given that the
software will calculate it) is also the same.
The number of random slopes, p, is not greater than r and may be
considerably smaller. To give a rough indication, r + 1 may not be larger
than the total number of distinct time points, that is, the number of different
values of t in the entire data set for which observations exist; also it should
not be larger than a small proportion, say, 10%, of the total number of
observations ∑imi. On the other hand, p will rarely be much larger than the
maximum number of observations per individual, maxi mi.2
Example 15.8 Polynomial growth model for children with retarded growth.
We continue the preceding example of children with retarded growth for which height measurements
are considered in the period between the ages of 5 and 10 years. In fitting polynomial models,
convergence problems occurred when the reference age t0 was chosen as 5 years, but not when it was
chosen as the midpoint of the range, 7.5 years.
A cubic model, that is, a model with polynomials of the third degree, turned out to yield a much
better statistical fit than the linear model of Table 15.7. Parameter estimates are shown in Table 15.8
for the cubic model with t0 = 7.5 years. So the intercept parameters refer to children of this age.
For the level-two random slopes (U0i, U1i U2i, U3i) the estimated correlation matrix is

The fit is much better than that of the linear model (deviance difference 496.12 for 9 degrees of
freedom). The random effect of the cubic term is significant (the model with fixed effects up to the
power r = 3 and with p = 2 random slopes, not shown in the table, has deviance 6,824.63, so the
deviance difference for the random slope of (t – t0)3 is 221.12 with 4 degrees of freedom).

Table 15.8: Cubic growth model for 5–10-year-old children with retarded
growth.
 The mean curve for the population (cf. (15.20)) is given here by

This deviates hardly at all, and not significantly, from a straight line.
However, the individual growth curves do differ from straight lines, but the
pattern of variation implied by the level-two covariance matrix is quite
complex. We return to this data set in the next example.

Other functions
There is nothing sacred about polynomial functions. They are convenient,
reasonably flexible, and any reasonably smooth function can be
approximated by a polynomial if only you are prepared to use a polynomial
of sufficiently high degree. One argument for using other classes of
functions is that some function shapes are approximated more
parsimoniously by other functions than polynomials. Another argument is
that polynomials are wobbly: when the value of a polynomial function F(t)
is changed a bit at one value of t, this may require coefficient changes that
make the function change a lot at other values of t. In other words, the fitted
value at any given value of t can depend strongly on observations for quite
distant values of t. This kind of sensitivity is often undesirable.
One can use other functions instead of polynomials. If the functions
used are called f1(t),. . ., fp(t), then instead of (15.19) the random function is
modeled as

and the observations as

What is a suitable class of functions depends on the phenomenon under

study and the data at hand. Random functions that can be represented by
(15.22) are particularly convenient because this representation is a linear
function of the parameters γ and the random effects U. Therefore, (15.22)
defines an instance of the hierarchical linear model. Below we treat various
classes of functions; the choice among them can be based on the deviance
but also on the level-one residual variance σR2, which indicates the size of
the deviations of individual data points with respect to the fitted population
of functions.
Sometimes, however, theory or data point to function classes where the
statistical parameters enter in a nonlinear fashion. In this chapter we restrict
attention to linear models, that is, models that can be represented as (15.22).
Readers interested in nonlinear models are referred to more specialized
literature such as Davidian and Giltinan (1995), Hand and Crowder (1996,
Chapter 8), and Singer and Willett (2003, Chapter 6).

Piecewise linear functions

A class of functions which is flexible, easy to comprehend, and for which
the fitted function values have a very restricted sensitivity for observations
made at other values of t, is the class of continuous piecewise linear
functions. These are continuous functions whose slopes may change
discontinuously at a number of values of t called nodes, but which are linear
(and hence have constant slopes) between these nodes. A disadvantage is
their angular appearance.
The basic piecewise linear function is linear on a given interval (t1, t2)
and constant outside this interval, as defined by

Often one of the constant values a and b is chosen to be 0. The nodes are t1
and t2. Boundary cases are the functions (choosing a = t1 and b = 0 and
letting the lower node t1 tend to minus infinity)

and (choosing a = 0 and b = t2 and letting the the upper node t2 tend to plus
infinity)

and also linear functions such as f (t) = t (where both nodes are infinite).
Each piecewise linear function can be obtained as a linear combination of
these basic functions. The choice of nodes sometimes will be suggested by
the problem at hand, and in other situations has to be determined by trial
and error.
Example 15.9 Piecewise linear models for retarded growth.
Let us try to improve on the polynomial models for the retarded growth data of the previous example
by using piecewise linear functions. Recall that the height measurements were considered for ages
from 5.0 to 10.0 years. For ease of interpretation, the nodes are chosen at the children’s birthdays, at
6.0, 7.0, 8.0, and 9.0 years. This means that growth is assumed to proceed linearly during each year,
but that growth rates may be different between the years. For the comparability with the polynomial
model, the intercept again corresponds to height at the age of 7.5 years. This is achieved by using
piecewise linear functions that all are equal to 0 for t = 7.5. Accordingly, the model is based on the
following five basic piecewise linear functions:
 The results for this model are presented in Table 15.9.
The level-two random effects (U0i, U1i,. . ., U5i) have estimated correlation matrix

The average growth rate is between 5 and 6 cm/year over the whole age range from 5 to 10 years.
There is large variability in individual growth rates. All slope variances are between 3 and 4, so the
between-children standard deviations in yearly growth rate are almost 2. The correlations between
individual growth rates in different years are not very high, ranging from –0.23 (between ages 7–8
and 9–10) to 0.48 (between ages 6–7 and 9–10). This indicates that the growth rate fluctuates rather
erratically from year to year around the average of about 5.5 cm/year.
The deviance for this piecewise linear model is 121.88 less than the deviance of the polynomial
model of Table 15.8, while it has 13 parameters more. This is a large deviance difference for only 13
degrees of freedom. The residual level-one variance also is smaller. Although the polynomial model
is not a submodel of the piecewise linear model, so that we cannot test the former against the latter by
the usual deviance test, yet we may conclude that the piecewise model not only is more clearly
interpretable but also has a better fit than the polynomial model.

Table 15.9: Piecewise linear growth model for 5–10-year-old children with
retarded growth.
 Spline functions
Another flexible class of functions which is suitable for modeling
longitudinal measurements within the framework of the hierarchical linear
model is the class of spline functions. Spline functions are smooth
piecewise polynomials. A number of points called nodes are defined on the
interval where the spline function is defined; between each pair of adjacent
nodes the spline function is a polynomial of degree p, while these
polynomials are glued together so smoothly that the function itself and its
derivatives, of order up to p – 1, are also continuous at the nodes. For p = 1
this leads again to piecewise linear functions, but for p > 1 it yields
functions which are smooth and do not have the kinky appearance of
piecewise linear functions. An example of a quadratic spline (i.e., a spline
with p = 2) was given in Chapter 8 by equation (8.3) and Figure 8.1. This
equation and figure represent a function of IQ which is a quadratic for IQ <
0 and also for IQ > 0, but which has different coefficients for these two
domains. The point 0 here is the node. The coefficients are such that the
function and its derivative are also continuous at the node, as can be seen
from the graph. Therefore it is a spline. Cubic splines (p = 3) are also often
used. We present here only a sketchy introduction to the use of splines. For
a more elaborate introduction to the use of spline functions in (single-level)
regression models, see Seber and Wild (1989, Section 9.5). The use of
spline functions in longitudinal multilevel models was discussed by Pan and
Goldstein (1998).
Suppose that one is investigating the development of some
characteristic over the age of 12–17 years. Within each of the intervals 12–
15 and 15–17 years, the development curves might be approximately
quadratic (this could be checked by a polynomial trend analysis for the data
for these intervals separately), while they are smooth but not quadratic over
the entire range of 12–17. In such a case it would be worthwhile to try a
quadratic spline (p = 2) with one node, at 15 years. Defining t1 = 15, the
basic functions can be taken as

The functions f2 and f3 are quadratic functions to the left of t1 and right
of t1 respectively, and they are continuous and have continuous derivatives.
That the functions are continuous and have continuous derivatives even at
the node t = t1 can be verified by elementary calculus or by drawing a
graph.
The individual development functions are modeled as

If β2i = β3i, the curve for individual i is exactly quadratic. The freedom to
have these two coefficients differ from each other allows us to represent
functions that look very different from quadratic functions; for example, if
these coefficients have opposite signs then the function will be concave on
one side of t1 and convex on the other side. Equation (8.3) and Figure 8.1
provide an example of exactly such a function, where t is replaced by IQ
and the node is the point IQ = 0.
The treatment of this model within the hierarchical linear model
approach is completely analogous to the treatment of polynomial models.
The functions f1, f2, and f3 constitute the fixed part of the model as well as
the random part of the model at level two. If there is no evidence for
individual differences with respect to the coefficients β2i and/or β3i, then
these could be deleted from the random part.
Formula (15.25) shows that a quadratic spline with one node has one
parameter more than a quadratic function (4 instead of 3). Each node added
further will increase the number of parameters of the function by one. There
is considerable freedom of choice in defining the basic functions, subject to
the restriction that they are quadratic on each interval between adjacent
nodes, and are continuous with continuous derivatives at the nodes. For two
nodes, t1 and t2, a possible choice is the following. This representation
employs a reference value t0 that is an arbitrary (convenient or meaningful)
value. It is advisable to use a t0 within the range of observation times in the
data. The basic functions are

When these four functions are used in the representation

coefficient β2i is the quadratic coefficient in the interval between t1 and t2,
while β3i and β4i are the changes in the quadratic coefficient that occur
when time t passes the nodes t1 or t2, respectively. The quadratic coefficient
for t < t1 is β2i + β3i, and for t > t2 it is β2i + β4i.
The simplest cubic spline (p = 3) has one node. If the reference point t0
is equal to the node, then the basic functions are

For more than two nodes, and an arbitrary order p of the polynomials,
the basic spline functions may be chosen as follows, for nodes denoted by
t1,. . ., tM:

The choice of nodes is important to obtain a well-fitting approximation.

Since the spline functions are nonlinear functions of the nodes, formal
optimization of the node placement is more complicated than fitting spline
functions with given nodes, and the interested reader is referred to the
literature on spline functions for the further treatment of node placement. If
the individuals provide enough observations (i.e., mi is large enough),
plotting observations for individuals, together with some trial and error, can
lead to a good placement of the nodes.
Example 15.10 Cubic spline models for retarded growth, 12–17 years.
In this example we again consider the height measurements of children with retarded growth studied
by Rekers-Mombarg et al. (1997), but now the focus is on the age range from 12 to 17 years. After
deletion of cases with missing data, in this age range there are a total of 1,941 measurements which
were taken for a sample of 321 children.
Some preliminary model fits showed that quadratic splines provide a better fit than piecewise
linear functions, and cubic splines fit even better. A reasonable model is obtained by having one node
at the age of 15.0 years. The basic functions accordingly are defined by (15.27) with t0 = 15.0. All
these functions are 0 for t = 15 years, so the intercept refers to height at this age. The parameter
estimates are given in Table 15.10.
 Table 15.10: Cubic spline growth model for 12–17-year-old children with
retarded growth.

The correlation matrix of the level-two random effects (U0i, U1i,. . ., U4i) is estimated as

Testing the random effects (not reported here) shows that the functions f1,. . ., f4 all have
significant random effects. The level-one residual standard deviation is only
which demonstrates that this family of functions fits rather closely to the
height measurements. Notation is made more transparent by defining
Thus, we denote f3(t) = (t–15)–3, f4(t) = (t–15)+3. The mean height curve can be obtained by filling in
the estimated fixed coefficients, which yields

This function can be differentiated to yield the mean growth rate. Note that df3(t)/dt =–3 (t–15)–2 and
df4(t)/dt = 3 (t–15)+2. This implies that the mean growth rate is estimated as

For example, for the minimum of the age range considered,t = 12 years, this is 6.43 – (0.25 × 3) +
(0.114 × 9) + 0 = 6.71 cm/year. This is slightly larger than the mean growth rate found in the
preceding examples for ages from 5 to 10 years. For the maximum age in this range, t = 17 years, on
the other hand, the average growth rate is 6.43 + (0.25 × 2) + 0 – (1.587 × 4) = 0.58 cm/year,
indicating that growth has almost stopped at this age.
The results of this model are illustrated more clearly by a graph. Figure 15.1 presents the average
growth curve and a sample of 15 random curves from the population defined by Table 15.10 and the
given correlation matrix. The average growth curve does not deviate noticeably from a linear curve
for ages below 16 years. It levels off after 16 years. Some of the randomly drawn growth curves are
decreasing in the upper part of the range. This is an obvious impossibility for the real growth curves,
and indicates that the model is not completely satisfactory for the upper part of this age range. This
may be related to the fact that the number of measurements is rather low at ages over 16.5 years.
 Figure 15.1: Average growth curve (*) and 15 random growth curves for
12–17-year-olds for cubic spline model.

15.2.3 Explaining the functions

Whether one has been using polynomial, spline, or other functions, the
random function model can be represented by (15.21), repeated here:

The same can now be done as in Chapter 5 (see p. 83): the individual-
dependent coefficients βhi can be explained with individual-level (i.e., level-
two) variables. Suppose there are q individual-level variables, and denote
these variables by Z1, . ., Zq. The inter-individual model to explain the
coefficients β0i, β1i,. . ., βri is then
Substitution of (15.29) into (15.21) yields

We see that cross-level interactions here are interactions between

individual-dependent variables and functions of time. The same approach
can be followed as in Chapter 5. The reader may note that this model
selection approach, where first a random function model is constructed and
then the individual-based coefficients are approached as dependent
variables in regression models at level two, is just what was described in
Section 6.4.1 as ‘working upward from level one’.
Example 15.11 Explaining growth by gender and height of parents.
We continue the analysis of the retarded growth data of the 12–17-year-olds, and now include the
child’s gender and the mean height of the child’s parents. Omitting children with missing value for
mother’s or father’s height left 321 children with a total of 1,941 measurements. Gender is coded as
+1 for girls and –1 for boys, so that the other parameters give values which are averages over the
sexes. Parents’ height is defined as the average height of father and mother minus 165 cm (this value
is approximately the mean of the heights of the parents). For parents’ height, the main effect and the
interaction effect with f1 (age minus 15) was included. For gender, the main effect was included as
well as the interactions with f1 and f2. These choices were made on the basis of preliminary model
fits. The resulting parameter estimates are in Table 15.11.
The correlation matrix of the level-two random effects (U0i, U1i,. . ., U4i) is estimated as

The fixed effect of gender (γ01) is not significant, which means that at 15 years there is not a
significant difference in height of boys and girls in this population with retarded growth. However,
the interaction effects of gender with age, γ11 and γ21, show that girls and boys do grow in different
patterns during adolescence. The coding of gender implies that the average height difference between
girls and boys is given by
The girl–boy difference in average growth rate, which is the derivative of this function, is equal to –
2.532 – 1.448 (t – 15). This shows that from about the age of 13 years (more precisely, for t > 13.25),
the growth of girls is on average slower than that of boys, and faster before this age.
Parents’ height has a strong main effect: for each centimeter of extra height of the parents, the
children are on average 0.263 cm taller at 15 years of age. Moreover, for every centimeter extra of
the parents, on average the children grow faster by 0.03 cm/year.

Table 15.11: Growth variability of 12–17-year-old children explained by

gender and parents’ height.

The intercept variance and the slope variances of f1 and f2 have decreased, compared to Table
15.10. The residual variance at level one has remained the same, which is natural since the effects
included explain differences between curves and do not yield better-fitting curves. The deviance went
down by 113.88 points (df = 5, p < 0.0001).
 15.2.4 Changing covariates
Individual-level variables such as the Zk of the preceding section are
referred to as constant covariates because they are constant over time. It is
also possible that changing covariates are available – changing social or
economic circumstances, performance on tests, mood variables, etc. Fixed
effects of such variables can be added to the model without problems, but
the neat forward model selection approach is disturbed because the
changing covariate normally will not be a linear combination of the
functions fh in (15.21). Depending on, for example, the primacy of the
changing covariates in the research question, one can employ the changing
covariates in the fixed part right from the start (i.e., add this fixed effect to
(15.19) or (15.22)), or incorporate them in the model at the point where the
constant covariates are also considered (add the fixed effect to (15.30)).
Example 15.12 Cortisol levels in infants.
This example reanalyzes data collected by de Weerth (1998, Chapter 3) in a study on stress in infants;
also see de Weerth and van Geert (2002). In this example the focus is not on discovering the shape of
the development curves, but on testing a hypothesized effect, while taking into account the
longitudinal design of the study. Because of the complexity of the data analysis, we shall combine
techniques described in various of the preceding chapters.
The purpose of the study was to investigate experimentally the effect of a stressful event on
experienced stress as measured by the cortisol hormone, and whether this effect is stronger in certain
hypothesized ‘regression weeks’. The experimental subjects were 18 normal infants with gestational
ages ranging from 17 to 37 weeks. Gestational age is age counted from the due day on which the
baby should have been born after a complete pregnancy. Each infant was seen repeatedly, the number
of visits per infant ranging from 8 to 13, with a total of 222 visits. The infants were divided randomly
into an experimental group (14 subjects) and a control group (4 subjects). Cortisol was measured
from saliva samples. The researcher collected a saliva sample at the start of the session; then a play
session between the mother and the infant followed. For the experimental group, during this session
the mother suddenly put the infant down and left the room. In the control group, the mother stayed
with the child. After the session, a saliva sample was taken again. This provided a pretest and a post-
test measure of the infant’s cortisol level. In this example we consider only the effect of the
experiment, not the effect of the ‘regression weeks’. Further information about the study and about
the underlying theories can be found in de Weerth (1998, Chapter 3).
The post-test cortisol level was the dependent variable. Explanatory variables were the pretest
cortisol level, the group (coded Z = 1 for infants in the experimental group and Z = 0 for the control
group), and gestational age. Gestational age is the time variable t. In order to avoid very small
coefficients, the unit is chosen as 10 weeks, so that t varies between 1.7 and 3.7. A preliminary
inspection of the joint distribution of pretest and post-test showed that cortisol levels had a rather
skewed distribution (as is usual for cortisol measurements). This skewness was reduced satisfactorily
by using the square root of the cortisol level, both for the pretest and for the post-test. These variables
are denoted by X and Y, respectively. The pretest variable, X, is a changing covariate.
The law of initial value proposed by Wilder in 1956 implies that an inverse relation is expected
between basal cortisol level and the subsequent response to a stressor (cf. de Weerth, 1998, p. 41). An
infant with a low basal level of cortisol will accordingly react to a stressor with a cortisol increase,
whereas an infant with a high basal cortisol level will react to a stressor with a cortisol decrease. The
basal cortisol level is measured here by the pretest. The average of X (the square root of the pretest
cortisol value) was 2.80. This implies that infants with a basal value x less than about 2.80 are
expected to react to stress with a relatively high value for Y whereas infants with x more than about
2.80 are expected to react to stress with a relatively low value of Y. The play session itself may also
lead to a change in cortisol value, so there may be a systematic difference between X and Y.
Therefore, the stress reaction was investigated by testing whether the experimental group has a more
strongly negative regression coefficient of Y on X – 2.80 than the control group; in other words, the
research hypothesis is that there is a negative interaction between Z and X - 2.80 in their effect on Y.
It appeared that a reasonable first model for the square root of the post-test cortisol level, if the
difference between the experimental and control group is not yet taken into account, is the model
where the changing covariate, defined as X – 2.80, has a fixed effect, while time (i.e., gestational age)
has a linear fixed as well as random effect:

The results are given as Model 1 in Table 15.12. This model can be
regarded as the null hypothesis model, against which we shall test the
hypothesis of the stressor effect.

Table 15.12: Estimates for two models for cortisol data.

The random effect of gestational age is significant. The model without this effect, not shown in
the table, has deviance 289.46, so that the deviance comparison yields χ2 = 8.89, df = 2. Using the
mixture of chi-squared distributions with 1 and 2 df, according to Section 6.2.1 and Table 6.2, we
obtain p < 0.05. The pretest has a strongly significant effect (t = 8.0, p < 0.0001): children with a
higher basal cortisol value tend also to have a higher post-test cortisol value. The fixed effect of
gestational age is not significant (t = 1.67, two-sided p > 0.05). The intercept variance is rather large
because the time variable is not centered and t = 0 refers to the due date of birth, quite an
extrapolation from the sample data.
To test the effect of the stressor, the fixed effect of the experimental group Z and the interaction
effect of Z × (X – 2.80) were added to the model. The theory, that is, Wilder’s law of initial value,
predicts the effect of the product variable Z × (X – 2.80) and therefore the test is focused on this
effect. The main effect of Z is included only because including an interaction effect without the
corresponding main effects can lead to errors of interpretation.
The result is presented as Model 2 in Table 15.12. The stressor effect (parameter γ03 in the table)
is significant (t = – 2.07 with many degrees of freedom because this is the effect of a level-one
variable, one-sided p < 0.025). This confirms the hypothesized effect of the stressor in the
experimental group.
The model fit is not good, however. The standardized level-two residuals defined by (10.10) were
calculated for the 18 infants. For the second infant (i = 2) the value was Si2 = 34.76 (df = ni = 13, p =
0.0009). With the Bonferroni correction which takes into account that this is the most significant out
of 18 values, the significance value still is 18 × 0.0009 = 0.016. Inspection of the data showed that
this was a child who had been asleep shortly before many of the play sessions. Being just awake is
known to have a potential effect on cortisol values. Therefore, for each session it had been recorded
whether the child had been asleep in the half hour immediately preceding the session. Subsequent
data analysis showed that having been asleep (represented by a dummy variable equal to 1 if the
infant had been asleep in the half hour preceding the session and equal to 0 otherwise) had an
important fixed effect, not a random effect, at level two, and also was associated with
heteroscedasticity at level one (see Chapter 8). Including these two effects led to the estimates
presented in Table 15.13.

Table 15.13: Estimates for a model controlling for having been asleep.
 The two effects of sleeping are jointly strongly significant (the comparison between the deviances
of Models 2 and 3 yields χ2 = 25.11, df = 2, p < 0.0001). The estimated level-one variance is 0.130
for children who had not slept and (using formula (8.1)) 0.362 for children who had slept. But the
stressor (Z × (X – 2.80) interaction) effect has now lost its significance (t = –1.42, n.s.). The most
significant standardized level-two residual defined by (10.10) for this model is obtained for the fourth
child (j = 4), with the value Sj2 = 29.86, df = nj = 13, p = 0.0049. Although this is a rather small p-
value, the Bonferroni correction now leads to a significance probability of 18 × 0.0049 = 0.09, which
is not alarmingly low. The fit of this model therefore seems satisfactory, and the estimates do not
support the hypothesized stressor effect. However, these results cannot be interpreted as evidence
against the stressor effect, because the parameter estimate does have the predicted negative sign and
the number of experimental subjects is not very large, so that the power may have been low.
It can be concluded that it is important to control for the infant having slept shortly before the
play session, and in this case this control makes the difference between a significant and a
nonsignificant effect. Having slept not only leads to a higher post-test cortisol value, controlling for
the pretest value, but also triples the residual variance at the occasion level.

15.3 Autocorrelated residuals

Of the relevant extensions to the hierarchical linear model, we wish briefly
to mention the hierarchical linear model with autocorrelated residuals. In
the fixed occasion design, the assumption that the level-one residuals Rti are
independent can be relaxed and replaced by the assumption of first-order
autocorrelation,

where are independent and identically distributed variables. The

parameter ρ is called the autocorrelation coefficient. This model represents
a type of dependence between adjacent observations, which quickly dies out
between observations which are further apart. The correlation between the
residuals is

Other covariance and correlation patterns are also possible. For variable
occasion designs, level-one residuals with the correlation structure (15.32)
also are called autocorrelated residuals, although they cannot be constructed
by the relations (15.31). These models are discussed in various textbooks,
such as Diggle et al. (2002), Goldstein (2011, Section 5.4), Verbeke and
Molenbergs (2000, Chapter 10), Singer and Willett (2003, Section 7.3), and
Hedeker and Gibbons (2006).
The extent to which time dependence can be modeled by random
slopes, or rather by autocorrelated residuals, or a combination of both,
depends on the phenomenon being modeled. This issue usually will have to
be decided empirically; most computer programs implementing the
hierarchical linear model allow the inclusion of autocorrelated residuals in
the model.

15.4 Glommary
Longitudinal data. Repeated measurements of the same characteristic for a
sample of individual subjects. Another name is panel data. The level-
one units here are the measurements (usually labeled as time points)
and the level-two units are individual subjects or respondents.
Fixed occasion designs. Data structures with a fixed set of measurement
occasions for all subjects, which may be conveniently labeled t = 1,. . .,
m. This does not exclude that data may be incomplete for some or all
of the subjects.

t-test for paired samples. This provides the simplest case of a longitudinal
design with fixed measurement occasions.

Random intercept model for repeated measures. This is also called the
compound symmetry model.

Random slope models. These models, with random slopes for time and
transformed time variables, can be used to represent time-dependent
variances as well as nonconstant correlations between different
measurements of the same subject.

Complete data vector. This can be defined for fixed measurement

occasions as the vector of data for all occasions. In this case the
specification of the random part can be regarded as the way to define
assumptions about the covariance matrix of the complete data vector.

Fully multivariate model. This is the model in which the covariance

matrix of the complete data vector is completely unrestricted. This
model can be used for longitudinal data with a fixed occasion design,
but also for multivariate regression analysis where the components of
the dependent variable are not necessarily repeated measures of the
same characteristic.

Proportion of explained variance in longitudinal models. This can be

defined as the proportional reduction in prediction error for individual
measurements, averaged over the m measurement occasions.

Variable occasion designs. In these designs the data are ordered according
to some underlying dimension such as time, and for each individual
data are recorded at some set of time points which is not necessarily
related to the time points at which the other individuals are observed.
Repeated measures according to variable occasion designs can be
regarded as observations on a population of curves.
Multilevel approach to populations of curves. This approach can
represent the average curve by fixed effects of individual-level
variables and of their cross-level interactions with time variables.
Individual deviations are represented by random effects of time
variables. In addition, changing covariates can be included to represent
further individual deviations.

Families of curves. Various families of curves were proposed for modeling

such a population of curves: polynomial functions, spline functions,
and piecewise linear functions.

Spline functions. Smooth piecewise polynomials. Splines are polynomial

functions between points called nodes, the polynomials being
smoothly linked at the nodes. Denoting the degree of the piecewise
polynomials by p, for p = 1 we obtain piecewise linear functions, for p
= 2 quadratic splines, and for p = 3 cubic splines. Spline functions are
generally more flexible in use than polynomial functions.

Autocorrelated residuals. A further way to represent correlation patterns

of residuals across time.

1
The notation with the parameters γ has nothing to do with the γ parameters used earlier in this
chapter in (15.4), but is consistent with the notation in Chapter 5.
2
In the fixed occasion design, the number of random effects, including the random intercept, cannot
be greater than the number of measurement occasions. In the variable occasion design this strict
upper bound does not figure, because the variability of time points of observations can lead to richer
information about intra-individual variations. But if one obtains a model with clearly more random
slopes than the maximum of all mi, this may mean that one has made an unfortunate choice of the
functions of time (polynomial or other) that constitute the random part at level two, and it may be
advisable to try and find another, smaller, set of functions of t for the random part with an equally
good fit to the data.
 16
Multivariate Multilevel Models

This chapter is devoted to the multivariate version of the hierarchical linear

model treated in Chapters 4 and 5. The term ‘multivariate’ refers here to the
dependent variable: there are assumed to be two or more dependent
variables. The model of this chapter is a three-level model, with variables as
level-one units, individuals as level-two units, and groups as level-three
units. One example is the nesting structure of Chapter 15, measurements
nested within individuals, combined with the nesting structure considered in
the earlier chapters, individuals within groups. This leads to three levels:
repeated measurements within individuals within groups. The model of this
chapter, then, is an extension of the fully multivariate model of Section
15.1.3 to the case where individuals are nested in groups. However, in the
present chapter, the dependent variables are not necessarily supposed to be
longitudinal measurements of the same characteristic (although they could
be).
As an example of multivariate multilevel data, think of pupils (level-two
units) i in classes (level-three units) j, with m variables, Y1, . . . , Ym, being
measured for each pupil if data are complete. The measurements are the
level-one units and could refer, for example, to test scores in various school
subjects but also variables measured in completely unrelated and
incommensurable scales such as attitude measurements.
The dependent variable is denoted

Yhij is the measurement on the h’th variable for individual i in group j.

It is not necessary that, for each individual i in each group j, an observation

of each of the m variables is available. (It must be assumed, however, that
missingness is at random, that is, the availability of a measurement should
be unrelated to its residual; cf. Chapter 9).
Just as in Chapter 15, the complete data vector can be defined as the
vector of data for the individual, possibly hypothetical, who does have
observations on all variables. This complete data vector is denoted by

If a researcher is interested in more than one dependent variable, it may still

not be necessary to analyze them simultaneously as a multivariate
dependent variable. Therefore, by way of introduction, we discuss some
arguments for making this additional effort. Subsequently, we discuss the
multivariate random intercept model and, briefly, the multivariate random
slope model.

16.1 Why analyze multiple dependent variables

simultaneously?
It is possible to analyze all m dependent variables separately. There are
several reasons why it may be sensible to analyze the data jointly, that is, as
multivariate data.

Conclusions can be drawn about the correlations between the

dependent variables – notably, the extent to which the unexplained
1. correlations depend on the individual and on the group level. Such
conclusions follow from the partitioning of the covariances between
the dependent variables over the levels of analysis.
2. The tests of specific effects for single dependent variables are more
powerful in the multivariate analysis. This will be visible in the form
of smaller standard errors. The additional power is negligible if the
dependent variables are only weakly correlated, but may be
considerable if the dependent variables are strongly correlated while
 at the same time the data are very incomplete, that is, the average
number of measurements available per individual is considerably less
than m.
Testing whether the effect of an explanatory variable on dependent
variable Y1 is larger than its effect on Y2, when the data on Y1 and Y2
3.
were observed (totally or partially) on the same individuals, is
possible only by means of a multivariate analysis.
If one wishes to carry out a single test of the joint effect of an
explanatory variable on several dependent variables, then a
4. multivariate analysis is also required. Such a single test can be useful,
for example, to avoid the danger of capitalization on chance which is
inherent in carrying out a separate test for each dependent variable.

A multivariate analysis is more complicated than separate analyses for each

dependent variable. Therefore, when one wishes to analyze several
dependent variables, the greater complexity of the multivariate analysis will
have to be balanced against the reasons listed above. Often it is advisable to
start by analyzing the data for each dependent variable separately.

16.2 The multivariate random intercept model

Suppose there are covariates X1, . . . , Xp, which may be individual-
dependent or group-dependent. The random intercept model for dependent
variable Yh is expressed by the formula

In words: for the hth dependent variable, the intercept is γ0h, the regression
coefficient on X1 is γ1h, the coefficient on X2 is γ2h, ...., the random part of
the intercept in group j is Uhj, and the residual is Rhij. This is just a random
intercept model like (4.5) and (4.9). Since the variables Y1, . . . , Ym are
measured on the same individuals, however, their dependence can be taken
into account. In other words, the Us and Rs are regarded as components of
vectors
Instead of residual variances at levels one and two, there are now residual
covariance matrices,

The covariance matrix of the complete observations, conditional on the

explanatory variables, is the sum of these,

To represent the multivariate data in the multilevel approach, three

nesting levels are used. The first level is that of the dependent variables
indexed by h = 1, . . . , m, the second level is that of the individuals i = 1, . .
. , nj, and the third level is that of the groups, j = 1, . . . ,N. So each
measurement of a dependent variable on some individual is represented by
a separate line in the data matrix, containing the values i, j, h, Yhij, x1ij, and
those of the other explanatory variables.
The multivariate model is formulated as hierarchical linear model using
the same trick as in Section 15.1.3. Dummy variables d1, . . . . , dm are used
to indicate the dependent variables, just as in formula (15.2). Dummy
variable dh is 1 or 0, depending on whether the data line refers to dependent
variable Yh or to one of the other dependent variables. Formally, this is
expressed by

With these dummies, the random intercept models (16.1) for the m
dependent variables can be integrated into one three-level hierarchical
linear model by the expression
All variables (including the constant) are multiplied by the dummy
variables. Note that the definition of the dummy variables implies that in
the sums over s = 1, . . . ,m only the term for s = h gives a contribution and
all other terms disappear. So this formula is just a complicated way of
rewriting formula (16.1).
The purpose of this formula is that it can be used to obtain a
multivariate hierarchical linear model. The variable-dependent random
residuals Rhij in this formula are random slopes at level two of the dummy
variables, Rsij being the random slope of ds, and the random intercepts Uhj
become the random slopes at level three of the dummy variables. There is
no random part at level one.
This model can be further specified, for example, by omitting some of
the variables Xk from the explanation of some of the dependent variables Yh.
This amounts to dropping some of the terms γks dshij xkij from (16.4).
Another possibility is to include variable-specific covariates, in analogy to
the changing covariates of Section 15.2.4. An example of this is a study of
school pupils’ performance on several subjects (these are the multiple
dependent variables), using the pupils’ motivation for each of the subjects
separately as explanatory variables.

Multivariate empty model

The multivariate empty model is the multivariate model without
explanatory variables. For m = 2 variables, this is just the model of Section
3.6.1. Writing out formulas (16.1) and (16.4) without explanatory variables
(i.e., with p = 0) yields the following formula, which is the specification of
the three-level multivariate empty model:

This empty model can be used to decompose the raw variances and
covariances into parts at the two levels. When referring to the multivariate
empty model, the covariance matrix

may be called the within-group covariance matrix, while

may be called the between-group covariance matrix. In the terminology of

Chapter 3, these are the population (and not observed) covariance matrices.
In Chapter 3 we saw that, if the group sizes nj all are equal to n, then the
covariance matrix between the group means is

(cf. equation (3.9)).

Section 3.6.1 presented a relatively simple way to estimate the within-
and between-group correlations from the intraclass coefficients and the
observed total and within-group correlations. Another way is to fit the
multivariate empty model (16.5) using multilevel software by the ML or the
REML method. For large sample sizes these methods will provide virtually
the same results, but for small sample sizes the ML and REML methods
will provide more precise estimators. The gain in precision will be
especially large if the correlations are high while there are relatively many
missing data (i.e., there are many individuals who provide less than m
dependent variables, but missingness is still at random).
Example 16.1 Language and arithmetic scores in elementary schools.
The example used throughout Chapters 4 and 5 is extended by analyzing not only the scores on the
language test but also those on the arithmetic test. So there are m = 2 dependent variables. In the
multivariate model there may be missing values on one of the dependent variables, but not both,
because each data point must make some contribution to the dependent variables. We now use a
version of the data set containing 3,773 pupils, 4 of whom are missing an arithmetic test score and 7
of whom are missing a language score.

Table 16.1: Parameter estimates for multivariate empty model.

 First the multivariate empty model, represented by (16.5), is fitted. The results are given in Table
16.1. The results for the language test may be compared with the univariate empty model for the
language test, presented in Table 4.1 of Chapter 4. The parameter estimates are slightly different,
which is a consequence of the fact that the estimation procedure is now multivariate. The extra result
of the multivariate approach is the estimated covariance at level three, cov(U1j, U2j) = 13.77, and at
level two, cov(R1ij, R2ij) = 28.51. The corresponding population correlation coefficients at the school
and pupil level are, respectively,

This shows that especially the random school effects for language and arithmetic are very
strongly correlated.
For the correlations between observed variables, these estimates yield a correlation between
individuals of (cf. (16.2))

Table 16.2: Parameter estimates for multivariate model for language and
arithmetic tests.
and, for groups of a hypothetical size n = 30, a correlation between group means (cf. (16.6)) of

Explanatory variables included are IQ, SES, the group mean of IQ, and the group mean of SES.
As in the examples in Chapters 4 and 5, the IQ measurement is the verbal IQ from the ISI test. The
correspondence with formulas (16.1) and (16.4) is that X1 is IQ, X2 is SES, X3 is the group mean of
IQ, X4 is the group mean of SES, X5 = X1 × X2 represents the interaction between IQ and SES, and X6
= X3 × X4 represents the interaction between group mean IQ and group mean SES. The results are
given in Table 16.2.
Calculating t-statistics for the fixed effects shows that for language, all effects are significant at
the 0.05 level. For arithmetic the fixed effect of mean SES is not significant, nor are the two
interaction effects, but the other three fixed effects are significant. The residual correlations are
ρ(U1j,U2j) = 0.83 at the school level and ρ(R1ij, R2ij) = 0.45 at the pupil level. This shows that taking
the explanatory variables into account has led to somewhat smaller, but still substantial residual
correlations. Especially the school-level residual correlation is large. This suggests that, also when
controlling for IQ, SES, mean IQ, and mean SES, the factors at school level that determine language
and arithmetic proficiency are the same. Such factors could be associated with school policy but also
with aggregated pupil characteristics not taken into account here, such as average performal IQ.
When the interaction effect of mean IQ with mean SES is to be tested for both dependent
variables simultaneously, this can be done by fitting the model from which these interaction effects
are excluded. In formula (16.1) this corresponds to the effects γ61 and γ62 of X6 on Y1 and Y2; in
formula (16.4) this corresponds to the effects γ61 and γ62 of d1X6 and d2X6. The model from which
these effects are excluded has a deviance of 46,570.5, which is 6.2 less than the model of Table 16.2.
In a chi-squared distribution with df = 2, this is a significant result (p < 0.05).

16.3 Multivariate random slope models

The notation is quite complex already, and therefore only the case of one
random slope is treated here. More random slopes are, in principle, a
straightforward extension.
Suppose that variable X1 has a random slope for the various dependent
variables. For the hth dependent variable, denote the random intercept by
U0hj and the random slope of X1 by U1hj. The model for the hth dependent
variable is then

The random slopes U1hj are uncorrelated between groups j but correlated
between variables h; this correlation between random slopes of different
dependent variables is a parameter of this model that is not included in the
hierarchical linear model of Chapter 5.
Just like the multilevel random intercept model, this model is
implemented by a three-level formulation, defined by

This means that again, technically, there is no random part at level one,
there are m random slopes at level two (of variables d1, . . . , dm) and 2m
random slopes at level three (of variables d1,. . . , dm and of the product
variables d1X1, . . . , dmX1). With this kind of model, an obvious further step
is to try and model the random intercepts and slopes by group-dependent
variables as in Section 5.2.

16.4 Glommary
Multivariate multilevel model. This is a three-level model, in which level
one consists of variables or measurements; level two consists of
individuals; and level three of groups.

Dummy variables. To specify this three-level model, dummy variables are

used to represent the variables. The dummy variables define the
variables represented by the level-one units, and have random slopes at
level two (individuals) to represent the between-individual within-
group variation, and random slopes at level three (groups) to represent
the between-group variation.

Multivariate random intercept model. This represents within-group and

between-group variances, covariances, and correlations, as also
discussed in Section 3.6.

Multivariate random slope model. This has, in addition, random slopes

for dependent variables Yh as a function of one or more explanatory
variables X1, etc. These random slopes are correlated between the
dependent variables.
 17
Discrete Dependent Variables

Up to now, it has been assumed in this book that the dependent variable has
a continuous distribution and that the residuals at all levels (U0j, Rij, etc.)
have normal distributions. This provides a satisfactory approximation for
many data sets. However, there also are many situations where the
dependent variable is discrete and cannot be well approximated by a
continuous distribution. This chapter treats the hierarchical generalized
linear model, which is a multilevel model for discrete dependent variables.

OVERVIEW OF THE CHAPTER

After a brief overview of the reasons for employing hierarchical generalized
linear models, this chapter first focuses on multilevel modeling for
dichotomous outcome variables (with only two values). The basic idea of
logistic regression is presented, and then the random intercept and random
slope logistic models are treated. It is shown how intraclass correlation
coefficients and measures of proportion of explained variance can be
defined here in slightly different ways than for continuous outcomes.
Subsequently a hierarchical generalized linear model is treated for ordered
categorical variables (with a small number of ordered categories, for
example, 1, 2, 3, 4); and for counts (with natural numbers as values: 0, 1, 2,
3,. . .).

17.1 Hierarchical generalized linear models

Important instances of discrete dependent variables are dichotomous
variables (e.g., success versus failure of whatever kind) and counts (e.g., in
the study of some kind of event, the number of events happening in a
predetermined time period). It is usually unwise to apply linear regression
methods to such variables, for two reasons.
The first reason is that the range of such a dependent variable is
restricted, and the usual linear regression model might take its fitted value
outside this allowed range. For example, dichotomous variables can be
represented as having the values 0 and 1. A fitted value of, say, 0.7, can still
be interpreted as a probability of 0.7 for outcome 1 and a probability of 0.3
for outcome 0. But what about a fitted value of −0.21 or 1.08? A
meaningful model for outcomes that have only the values 0 or 1 should not
allow fitted values that are negative or greater than 1. Similarly, a
meaningful model for count data should not lead to negative fitted values.
The second reason is of a more technical nature, and is the fact that for
discrete variables there is often some natural relation between the mean and
the variance of the distribution. For example, for a dichotomous variable Y
that has probability p for outcome 1 and probability 1 – p for outcome 0, the
mean is

and the variance is

Thus, the variance is not a free parameter but is determined by the mean;
and the variance is not constant: there is heteroscedasticity. In terms of
multilevel modeling, this could lead to a relation between the parameters in
the fixed part and the parameters in the random part.
This has led to the development of regression-like models that are more
complicated than the usual multiple linear regression model and that take
account of the nonnormal distribution of the dependent variable, its
restricted range, and the relation between mean and variance. The best-
known method of this kind is logistic regression, a regression-like model
for dichotomous data. Poisson regression is a similar model for count data.
In the statistical literature, such models are known as generalized linear
models; see McCullagh and Nelder (1989) or Long (1997).
 The present chapter gives an introduction to multilevel versions of some
generalized linear models; these multilevel versions are aptly called
hierarchical generalized linear models or generalized linear mixed models.

17.2 Introduction to multilevel logistic regression

Logistic regression (e.g., Hosmer and Lemeshow, 2000; Long, 1997;
McCullagh and Nelder, 1989; Ryan, 1997) is a kind of regression analysis
for dichotomous, or binary, outcome variables, that is, outcome variables
with two possibly values such as pass/fail, yes/no, or like/dislike. The
reader is advised to study an introductory text on logistic regression if he or
she is not already acquainted with this technique.

17.2.1 Heterogeneous proportions

The basic data structure of two-level logistic regression is a collection of N
groups (‘units at level two’), with, in group j (j = 1,…, N), a random sample
of nj level-one units (‘individuals’). The outcome variable is dichotomous
and denoted by Yij for level-one unit i in group j. The two outcomes are
coded 0 for ‘failure’ and 1 for ‘success’. The total sample size is denoted by
M = ∑j nj. If one does not (yet) take explanatory variables into account, the
probability of success is regarded as constant in each group. The success
probability in group j is denoted by Pj. In a random coefficient model (cf.
Section 4.3.1), the groups are considered as being taken from a population
of groups and the success probabilities in the groups, Pj, are regarded as
random variables defined in this population. The dichotomous outcome can
be represented as the sum of this probability and a residual,

In words, the outcome for individual i in group j, which is either 0 or 1, is

expressed as the sum of the probability (average proportion of successes) in
this group plus some individual-dependent residual. This residual has (like
all residuals) mean zero but for these dichotomous variables it has the
peculiar property that it can assume only the values – Pj and 1 – Pj, since
(17.2) must be 0 or 1. A further peculiar property is the fact that, given the
value of the probability Pj, the variance of the residual is

in accordance with formula (17.1).

Equation (17.2) is the dichotomous analog of the empty (or
unconditional) model defined for continuous outcomes in equations (3.1)
and (4.6). Section 3.3.1 remains valid for dichotomous outcome variables,
with Pj taking the place of μ + Uj, except for one subtle distinction. In the
empty model for continuous outcome variables it was assumed that the
level-one residual variance was constant. This is not adequate here because,
in view of formula (17.3), the groups have different within-group variances.
Therefore the parameter σ2 must be interpreted here as the average residual
variance, that is, the average of (17.3) in the population of all groups. With
this modification in the interpretation, the formulas of Section 3.3.1 still are
valid. For example, the intraclass correlation coefficient is still defined by
(3.2) and can be estimated by (3.12). Another definition of the intraclass
correlation is also possible, however, as is mentioned below in Section
17.3.2.
Since the outcome variable is coded 0 and 1, the group average

is now the proportion of successes in group j. This is an estimate of the

group-dependent probability Pj. Similarly, the overall average

here is the overall proportion of successes.

Testing heterogeneity of proportions

To test whether there are indeed systematic differences between the groups,
the well-known chi-squared test can be used. The test statistic of the chi-
squared test for a contingency table is often given in the familiar form ∑(O
– E)2/E, where O is the observed and E the expected count in a cell of the
contingency table. In this case it can be written also as

It can be tested against the chi-squared distribution with N – 1 degrees of

freedom. This chi-squared distribution is an approximation valid if the
expected numbers of successes and of failures in each group,
respectively, are all at least 1 while 80% of them are at least
5 (cf. Agresti, 2002). This condition will not always be satisfied, and the
chi-squared test may then be seriously in error. For a large number of
groups the null distribution of X2 can then be approximated by a normal
distribution with the correct mean and variance (Haldane, 1940; McCullagh
and Nelder, 1989, p. 244), or an exact permutation test may be used.
Another test of heterogeneity of proportions was proposed by
Commenges and Jacqmin (1994). The test statistic is

Large values of this statistic are an indication of heterogeneous proportions.

This statistic can be tested against a standard normal distribution.
The fact that the numerator contains a weight of nj2 whereas the chi-
squared test uses the weight nj shows that these two tests combine the
groups in different ways: larger groups count more strongly in the
Commenges–Jacqmin test. When the group sizes nj are different, it is
possible that the two tests lead to different outcomes. A similar test for the
intercept variance, also controlling for fixed effects, was proposed by
Commenges et al. (1994).
The advantage of test (17.7) over the chi-squared test is that, when
group sizes differ, it has a higher power to test randomness of the
observations against the empty model treated below, that is, against the
alternative hypothesis represented by (17.10) with A further practical
advantage is that it can be applied whenever there are many groups, even
with small group sizes, provided that no single group dominates. A rule of
thumb for the application of this test is that there should be at least N = 10
groups, the biggest group should not have a relative share larger than nj/M =
0.10, and the ratio of the largest group size to the 10th largest group size
should not be more than 10.

Estimation of between- and within-groups variance

The true variance between the group-dependent probabilities, that is, the
population value of var(Pj), can be estimated by formula (3.11), repeated
here,

where ñ is defined as in (3.7) by

For dichotomous outcome variables, the observed between-groups variance

is closely related to the chi-squared test statistic (17.6). They are connected
by the formula

The within-groups variance in the dichotomous case is a function of the

group averages, namely,

Example 17.1 Religiosity across the world.

This example is about differences in religiosity between countries, as
explained by individual- and country-level characteristics. The example
follows the paper by Ruiter and van Tubergen (2009), which is about the
question why people in some nations attend religious meetings more
frequently than people in others. They use data from the European Values
Surveys/World Values Surveys collected in 1990–2001 for 136,611
individual respondents in 60 countries.
The dependent variable, religious attendance, is whether the respondent
attends religious services at least once a week. The number of respondents
per country ranges from 95 (Ghana) to 6,501 (Russian Federation). In the
data set as a whole, 23.8% of the respondents reported that they attended
religious services at least once a week. Ruiter and van Tubergen (2009)
formulate various different theories about religiosity and test hypotheses
derived from these theories.
A two-level structure is used with the country as the second-level unit.
This is based on the idea that there may be differences between countries
that are not captured by the explanatory variables and hence may be
regarded as unexplained variability within the set of all countries. Figure
17.1 shows for each of the 60 countries the proportion of people in the
sample who attend religious services at least once a week. With these large
group sizes, a statistical test is not needed to conclude that these
proportions, ranging from 0.025 (Denmark) to 0.874 (Nigeria), are very
different indeed. The chi-squared test (17.6) for equality of these 60
proportions yields X2 = 29,732.52, df = 59, significant at any meaningful
significance level.

Figure 17.1: Proportion of religious attendance.

The estimated true variance between the country-dependent proportions

calculated from (3.11) is thus the estimated true between-country
standard deviation is The groups are so large that is
hardly less than S2between. With an average probability of 0.238, this
standard deviation is quite large and points to considerable skewness of the
distribution, as we can see in Figure 17.1.

17.2.2 The logit function: Log-odds

It can be relevant to include explanatory variables in models for
dichotomous outcome variables. For example, in the example above,
individuals with religious parents will tend to have a higher religious
attendance than others. When explanatory variables are included to model
probabilities, one problem (mentioned in Section 17.1) is that probabilities
are restricted to the domain between 0 and 1, whereas a linear effect for an
explanatory variable could take the fitted value outside this interval.
Instead of the probability of some event, one may consider the odds: the
ratio of the probability of success to the probability of failure. When the
probability is p, the odds are p/(1 – p). In contrast to probabilities, odds can
assume any value from 0 to infinity, and odds can be considered to
constitute a ratio scale.

Figure 17.2: The logit function.

The logarithm transforms a multiplicative to an additive scale and

transforms the set of positive real numbers to the whole real line. Indeed,
one of the most widely used transformations of probabilities is the log-odds,
defined by
where ln(x) denotes the natural logarithm of the number x. The logit
function, of which a graph is shown in Figure 17.2, is an increasing
function defined for numbers between 0 and 1, and its range is from minus
infinity to plus infinity. Figure 17.3 shows in a different way in which
probability values are transformed to logit values.
For example, p = 0.269 is transformed to logit(p) =–1 and p = 0.982 to
logit(p) = 4. The logit of p = 0.5 is exactly 0.
The logistic regression model is a model where logit(p) is a linear
function of the explanatory variables. In spite of the attractive properties of
the logit function, it is by no means the only suitable function for
transforming probabilities to arbitrary real values. The general term for such
a transformation function is the link function, as it links the probabilities (or
more generally, the expected values of the dependent variable) to the
explanatory variables. The probit function (which is the inverse cumulative
distribution function of the standard normal distribution) also is often used
as a link function for dichotomous variables (see also Section 17.3.2). A
generalized linear model for a dichotomous outcome with the probit link
function is called a probit regression model. For still other link functions
see, for example, Long (1997) or McCullagh and Nelder (1989).

Figure 17.3: Correspondence between p and logit(p).

The choice of link function has to be guided by the empirical fit of the
model, ease of interpretation, and convenience (e.g., availability of
computer software). In this chapter the link function for a probability will
be denoted by f (p), and we shall concentrate on the logit link function.

17.2.3 The empty model

The empty two-level model for a dichotomous outcome variable refers to a
population of groups (level-two units) and specifies the probability
distribution for the group-dependent probabilities Pj in (17.2), without
taking further explanatory variables into account. Several such
specifications have been proposed.
We focus on the model that specifies the transformed probabilities f (Pj)
to have a normal distribution. This is expressed, for a general link function f
(p), by the formula

where is the population average of the transformed probabilities and U0j

the random deviation from this average for group j. If f (p) is the logit
function, then f (Pj) is just the log-odds for group j. Thus, for the logit link
function, the log-odds have a normal distribution in the population of
groups, which is expressed by

For the deviations U0j it is assumed that they are independent random
variables with a normal distribution with mean 0 and variance
This model does not include a separate parameter for the level-one
variance. This is because the level-one residual variance of the dichotomous
outcome variable follows directly from the success probability, as indicated
by equation (17.3).
Denote by the probability corresponding to the average value as
defined by
For the logit function, this means that is the so-called logistic transform
of defined by

Here exp(γ0) = eγ0 denotes the exponential function, where e is the basis of
the natural logarithm. The logistic and logit functions are mutual inverses,
just like the exponential and the logarithmic functions. Figure 17.4 shows
the shape of the logistic function. This π0 is close (but not quite equal) to
the average value of the probabilities Pj in the population of groups.
Because of the nonlinear nature of the link function, there is no simple
relation between the variance of these probabilities and the variance of the
deviations U0j. There is an approximate formula, however, valid when the
variances are small. The approximate relation (valid for small τ02) between
the population variances is

Figure 17.4: The logistic function.

For the logit function, this yields

When τ02 is not so small, the variance of the probabilities will be less than
the right-hand side of (17.13). (Note that these are population variances and
not variances of the observed proportions in the groups; see Section 3.3 for
this distinction.)

Example 17.2 Empty model for the religious attendance data.

For the further analyses of the religious attendance data, for one of the
countries (Turkey) the income data seemed incorrectly transformed, and
therefore this country was omitted from the data set. This leaves 135,508
respondents in 59 countries. Fitting the empty model with normally
distributed log-odds to the data of religious attendance in 59 countries
yields the results presented in Table 17.1. Note that the table gives the level-
two standard deviation, not the variance1. The model was estimated using
the Laplace approximation (see Section 17.2.5).

Table 17.1: Estimates for empty logistic model.

The estimated average log-odds is –1.447, which corresponds

(according to equation (17.11)) to a probability of π0 = 0.19, different from
the overall fraction of 0.24 because of the nonlinearity of the link function
and the weighting of the countries inherent in the multilevel model. The
variance approximation formula (17.13) yields var(Pj) ≈ 0.0451, not too far
from the nonparametric estimate calculated from (3.11), which is 0.0405.
The difference between these values is caused by the facts that formula
(17.13) is only an approximation and that the estimation methods are
different.
Figure 17.5 presents the observed log-odds of the 59 countries with the
fitted normal distribution. These observed log-odds are the logit
transformations of the proportions depicted in Figure 17.1. It can be seen
that the skewness to the right in Figure 17.1 has more or less disappeared
because of the log-odds transformation. There now is one outlier to the left,
which diminishes the quality of the normal approximation. However,
nonnormality of the level-two residuals is not a serious issue at the stage
where we are fitting the empty model, of which we know that it is likely to
be misspecified because no predictor variables are included. The normality
assumption will become important mainly when considering residuals in a
model with predictor variables.

Figure 17.5: Observed log-odds and estimated normal distribution of

population log-odds of religious attendance.

17.2.4 The random intercept model

In logistic regression analysis, linear models are constructed for the log-
odds. The multilevel analog, random coefficient logistic regression, is based
on linear models for the log-odds that include random effects for the groups
or other higher-level units. As mentioned above, Yij denotes the
dichotomous outcome variable for level-one unit i in level-two unit j. There
are nj level-one units in the jth level-two unit. The outcome Yij is coded as 0
or 1, also referred to as ‘failure’ and ‘success’. We shall use the terms
‘individual’ and ‘group’ to refer to the level-one and level-two units.
We now assume that there are variables which are potential explanations
for the observed success or failure. These variables are denoted by X1, . . .,
Xr. The values of Xh (h = 1, . . ., r) are indicated in the usual way by xhij.
Since some (or all) of these variables could be level-one variables, the
success probability is not necessarily the same for all individuals in a given
group. Therefore the success probability now depends on the individual as
well as on the group, and is denoted by Pij. Accordingly, equation (17.2)
expressing how the outcome is split into an expected value and a residual
now is replaced by

The logistic random intercept model expresses the log-odds, that is, the
logit of Pij, as a sum of a linear function of the explanatory variables2 and a
random group-dependent deviation U0j:

Thus, a unit difference between the Xh-values of two individuals in the

same group is associated with a difference of γh in their log-odds, or
equivalently, a ratio of exp(γh) in their odds. The deviations U0j are assumed
to have zero means (given the values of all explanatory variables) and
variances equal to τ02 Formula (17.15) does not include a level-one residual
because it is an equation for the probability Pij rather than for the outcome
Yij. The level-one residual is already included in formula (17.14).

Example 17.3 Random intercept model for religious attendance.

Continuing the study of religious attendance in 59 countries, we use some
of the explanatory variables also considered by Ruiter and van Tubergen
(2009). These authors test a number of hypotheses derived from several
different theories about religiosity. For these theories and hypotheses, we
refer to their paper. Here we use the following variables, some of which
have been approximately centered. At the individual level:

Educational level, measured as the age at which people left school (14–
* 21 years), minus 14.This variable was centered within countries. The
within-country deviation variable has mean 0,standard deviation 2.48.
Income, standardized within country; mean −0.03, standard deviation
*
0.99.
Employment status, 1 for unemployed, 0 for employed; mean 0.19,
*
standard deviation 0.39.
* Sex, 1 for female, 0 for male; mean 0.52, standard deviation 0.50.
Marital status, 1 for single/divorced/widowed, 0 for married/cohabiting;
*
mean 0.23,standard deviation 0.42.
Divorce status, 1 for divorced, 0 for other; mean 0.06, standard deviation
*
0.25.
Widowed, 1 for widowed, 0 for other; mean 0.08, standard deviation
*
0.27.
Urbanization, the logarithm of the number of inhabitants in the
* community or town of residence, truncated between 1,000 and
1,000,000, minus 10; mean 0.09, standard deviation 2.18.
At the country level, the averages of some individual-level variables
representing social and cultural differences between countries were used,
with one other country-level variable. Using the definitions above, this
leads to the following list of variables:

* Average educational level: mean 0.82, standard deviation 0.94.

* Average unemployment: mean 0.19, standard deviation 0.08.
* Average divorce status: mean 0.06, standard deviation 0.03.
Gini coefficient measuring income inequality, minus 35. Mean 0.10,
*
standard deviation 9.58.

Parameters were estimated by the Laplace algorithm implemented in R

package lme4 (see Section 17.2.5). The results are shown in Table 17.2.

Table 17.2: Logistic random intercept model for religious attendance in 59

countries.
 All the individual-level variables have significant effects, except for
unemployment, which only has an effect at the country level. For education
and divorce, both of which have negative effects on religious attendance,
the between-country regression coefficients are stronger than the within-
country coefficients. The fact that regression coefficients of the country
average variables are larger in size than the coefficients of individual
variables must be seen in the light that these are unstandardized
coefficients, and the country averages have considerably smaller variability
than the individual-level variables.

17.2.5 Estimation
Parameter estimation in hierarchical generalized linear models is more
complicated than in hierarchical linear models. Inevitably some kind of
approximation is involved, and various kinds of approximation have been
proposed. Good reviews are given in Demidenko (2004, Chapter 8),
Skrondal and Rabe-Hesketh (2004, Chapter 6), Tuerlinckx et al. (2006), and
Rodríguez (2008). We mention some references and some of the terms used
without explaining them. The reader who wishes to study these algorithms
is referred to the literature cited. More information about the computer
programs mentioned in this section is given in Chapter 18.
Currently the best and most often used methods are two frequentist
ones: the Laplace approximation and adaptive numerical quadrature as
algorithms for approximating the maximum likelihood estimates; and
Bayesian methods.
The Laplace approximation for hierarchical generalized linear models
was proposed by Raudenbush et al. (2000). Because of its good quality and
computational efficiency this is now one of the most frequently used
estimation algorithms for generalized linear mixed models. It is
implemented in the software package HLM and in various packages of the
statistical system R, such as lme4, glmmADMB and glmmML.
Numerical integration is an approach for dealing with the random
coefficients which is straightforward in principle, but poses various
difficulties in implementation. Its use for hierarchical generalized linear
models was initially developed by Stiratelli et al. (1984), Anderson and
Aitkin (1985) and Gibbons and Bock (1987); later publications include
Longford (1994), Hedeker and Gibbons (1994), Gibbons and Hedeker
(1997), and Hedeker (2003). Pinheiro and Bates (1995) and Rabe-Hesketh
et al. (2005) developed so-called adaptive quadrature methods of numerical
integration; see also Skrondal and Rabe-Hesketh (2004). Adaptive
quadrature has considerable advantages over nonadaptive numerical
integration. It is implemented in SAS proc NLMIXED, the Stata package,
and the software gllamm which combines with Stata. Nonadaptive
numerical integration is implemented in the program MIXOR/SuperMix
and its relatives.
Much work has been done on Bayesian methods (cf. Section 12.1),
mainly used in Markov chain Monte Carlo (MCMC) methods, which are
computationally intensive, because they are based on repeated simulation of
draws from the posterior distribution. Zeger and Karim (1991) and Browne
and Draper (2000) are important milestones for Bayesian inference for
multilevel models, and this approach is now being increasingly used. An
overview is given by Draper (2008) and an extensive treatment in Congdon
(2010). For hierarchical generalized linear models, Bayesian methods
perform very well, and they also have good frequentist properties (i.e.,
small mean squared errors of estimators and good coverage probabilities of
confidence intervals); see Browne and Draper (2006). Bayesian MCMC
methods are implemented in MLwiN, WinBUGS, and BayesX. A recent
development is the use of the Laplace approximation for Bayesian
inference, which circumvents the computationally intensive MCMC
methods; see Rue et al. (2009) and Fong et al. (2010). This is implemented
in the R package INLA.
Before these methods had been developed, the main methods available
were those based on first- or second-order Taylor expansions of the link
function. When the approximation is around the estimated fixed part, this is
called marginal quasi-likelihood (MQL), when it is around an estimate for
the fixed plus the random part it is called penalized or predictive quasi-
likelihood (PQL) (Breslow and Clayton, 1993; Goldstein, 1991; Goldstein
and Rasbash, 1996). For estimating and testing fixed effects these methods
are quite adequate, especially if the cluster sizes nj are not too small, but
they are not satisfactory for inference about random effects. The first-order
MQL and PQL estimates of the variance parameters of the random part
have an appreciable downward bias (Rodríguez and Goldman, 1995;
Browne and Draper, 2006; Rodríguez, 2008; Austin, 2010). The second-
order MQL and PQL methods produce parameter estimates with less bias
but, it seems, a higher mean squared error. The biases of MQL and PQL in
the estimation of parameters of the random effects can be diminished by
bootstrapping (Kuk, 1995; van der Leeden et al., 2008), but this leads to
quite computer-intensive procedures. These methods are implemented in
MLwiN, HLM, R packages MASS and nlme, and SAS proc GLIMMIX.
Several other methods have been proposed for parameter estimation in
hierarchical generalized linear models. McCulloch (1997), Ng et al. (2006),
and Jank (2006) gave various algorithms for estimation by simulated
maximum likelihood. Another computer-intensive method is the method of
simulated moments. This method is applied to these models by Gouriéroux
and Montfort (1996, Section 3.1.4 and Chapter 5), and an overview of some
recent work is given by Baltagi (2008, Section 11.2). A method based on
the principle of indirect inference was proposed by Mealli and Rampichini
(1999).
Various simulation studies have been done comparing different
estimation procedures for hierarchical generalized linear models, mostly
focusing on multilevel logistic regression: Rodríguez and Goldman (1995,
2001), Browne and Draper, (2006), Rodríguez, (2008), and Austin (2010).
These are the references on which the statements in this section about the
qualities of the various estimation procedures are based.

17.2.6 Aggregation
If the explanatory variables assume only few values, then it is advisable to
aggregate the individual 0–1 data to success counts, depending on the
explanatory variables, within the level-two units. This will improve the
speed and stability of the algorithm and reduce memory use. This is carried
out as follows.
For a random intercept model with a small number of discrete
explanatory variables X1, . . . , Xr, let L be the total number of combinations
of values (x1, . . . , xr). All individuals with the same combination of values
(x1, . . . , xr) are treated as one subgroup in the data. They all have a
common success probability, given by (17.15). Thus, each level-two unit
includes L subgroups, or fewer if some of the combinations do not occur in
this level-two unit. Aggregation is advantageous if L is considerably less
than the average group size nj.
Denote by

the number of individuals in group j with the values x1,. . ., xr on the

respective explanatory variables, and denote by

the number of individuals among these who yielded a success, that is, for
whom Yij = 1. Then Yj+(x1,. . ., xr) has the binomial distribution with
binomial denominator (‘number of trials’) nj+(x1,. . ., xr) and success
probability given by (17.15), which is the same for all individuals i in this
subgroup. The multilevel analysis is now applied with these subgroups as
the level-one units. Subgroups with nj+(x1,. . ., xr) = 0 can be omitted from
the data set.
 17.3 Further topics on multilevel logistic
regression

17.3.1 Random slope model

The random intercept logistic regression model of Section 17.2.4 can be
extended to a random slope model just as in Chapter 5. We only give the
formula for one random slope; the extension to several random slopes is
straightforward. The remarks made in Chapters 5 and 6 remain valid, given
the appropriate changes, and will not be repeated here.
As in the random intercept model, assume that there are r explanatory
variables X1,. . ., Xr. Assume that the effect of X1 is variable across groups,
and accordingly has a random slope. Expression (17.15) for the logit of the
success probability is extended with the random effect U1j x1ij, which leads
to

There now are two random group effects, the random intercept U0j and the
random slope U1j. It is assumed that both have a zero mean. Their variances
are denoted, respectively, by τ02 and τ12 and their covariance is denoted by
τ01.
Example 17.4 Random slope model for religious attendance.
Continuing the example of religious attendance in 59 countries, Table 17.3 presents a model which
has the same fixed part as the model of Table 17.2, and random slopes for income and education.
Recall that income is standardized within countries and education is a within-country deviation
variable. Estimating a model with a random slope for education was successful for this relative
measure of education, but not when the grand-mean-centered education variable was used; this may
be a signal of misspecification of the latter model.
The deviance goes down by 115,969.9 – 115,645.7 = 324.2, a huge improvement for five extra
parameters. Also when each of the random slope variances is considered separately, the gain in
deviance is considerable (results not shown here). With such large sample sizes per country, the
regression coefficients within each country are estimated very precisely and it is not surprising to
find evidence for slope heterogeneity.
 For education as well as income, the slope standard deviation is larger in absolute magnitude than
the estimated fixed effect, which is negative for both variables. This implies that for most countries
the regression coefficient per country is negative, but for a nonnegligible number of countries it is
positive. For example, the regression coefficient of income is which has an estimated
normal distribution with mean −0.074 and standard deviation 0.063, which leads to a probability of
0.12 of a randomly chosen country having a positive coefficient.
The standard errors of the fixed effects of education and income are considerably larger here than
in the random intercept model of Table 17.2. The random slope model must be considered to be more
realistic, and therefore the larger standard errors are a better representation of the uncertainty about
these fixed parameters, which are the estimated average effects in the population of all countries.
This suggests that for other level-one variables a random slope could, or should, also be considered,
and their estimated coefficients might then also get higher standard errors. For a data set with very
large level-two units such as this one, the two-step approach based on country-by-country equations
of the type (3.38)–(3.40) – but of course with a larger number of predictor variables – may be more
suitable than the approach by a hierarchical generalized linear model, because less stringent
assumptions are being made about the similarity of many parameters across the various different
countries (cf. Achen, 2005).

Table 17.3: Logistic random intercept model for religious attendance in 59

countries.
 17.3.2 Representation as a threshold model
The multilevel logistic regression can also be formulated as a so-called
threshold model. The dichotomous outcome Y, ‘success’ or ‘failure’, is then
conceived as the result of an underlying non-observed continuous variable.
When Y denotes passing or failing some test or exam, the underlying
continuous variable could be the scholastic aptitude of the subject; when Y
denotes whether the subject behaves in a certain way, the underlying
variable could be a variable representing benefits minus costs of this
behavior; etc. Denote the underlying variable by . Then the threshold
model states that Y is 1 if is larger than some threshold, and 0 if it is less
than the threshold. Since the model is about unobserved entities, it is no
restriction to assume that the threshold is 0. This leads to the representation

For the unobserved variable , a usual random intercept model is assumed:

To represent a logistic regression model, the level-one residual of the

underlying variable must have a logistic distribution. This means that,
when the level-one residual is denoted by Rij, the cumulative distribution
function of Rij must be the logistic function,

defined in (17.11). This is a symmetric probability distribution, so that also

Its mean is 0 and its variance is π2/3 = 3.29. When it is assumed that Rij has
this distribution, the logistic random intercept model (17.15) is equivalent
to the threshold model defined by (17.17) and (17.18).
To represent the random slope model (17.16) as a threshold model, we
define
where Rij has a logistic distribution. It then follows that

Since the logit and the logistic functions are mutual inverses, the last
equation is equivalent to (17.16).
If the residual Rij has a standard normal distribution with unit variance,
then the probit link function is obtained. Thus, the threshold model which
specifies that the underlying variable has a distribution according to the
hierarchical linear model of Chapters 4 and 5, with a normally distributed
level-one residual, corresponds exactly to the multilevel probit regression
model. Since the standard deviation of Rij is for the logistic
and 1 for the probit model, the fixed estimates for the logistic model will
tend to be about 1.81 times as large as for the probit model and the variance
parameters of the random part about times as large (but see Long
(1997, p. 48), who notes that in practice the proportionality constant for the
fixed estimates is closer to 1.7).

17.3.3 Residual intraclass correlation coefficient

The intraclass correlation coefficient for the multilevel logistic model can
be defined in at least two ways. The first definition is by applying the
definition in Section 3.3 straightforwardly to the binary outcome variable
Yij. This approach was also mentioned in Section 17.2.1. It was followed,
for example, by Commenges and Jacqmin (1994).
The second definition is by applying the definition in Section 3.3 to the
unobserved underlying variable Since the logistic distribution for the
level-one residual implies a variance of this implies that for a
two-level logistic random intercept model with an intercept variance of τ02,
the intraclass correlation is

These two definitions are different and will lead to somewhat different
outcomes. For example, for the empty model for the religious attendance
data presented in Table 17.1, the first definition yields
0.0404/(0.0404+0.1414) = 0.22, whereas the second definition leads to the
value 1.896/(1.896+3.290) = 0.37. In this case, the difference is large,
underscoring the rather arbitrary nature of the definition of these
coefficients.
An advantage of the second definition is that it can be directly extended
to define the residual intraclass correlation coefficient, that is, the intraclass
correlation which controls for the effect of explanatory variables. The
example can be continued by moving to the model in Table 17.2. The
residual intraclass correlation controlling for the variables in this model is
1.177/(1.177+3.290) = 0.26, lower than the raw intraclass correlation
coefficient.
For the multilevel probit model, the second definition for the intraclass
correlation (and its residual version) leads to

since this model fixes the level-one residual variance of the unobservable
variable to to 1.

17.3.4 Explained variance

There are several definitions of the explained proportion of variance (R2) in
single-level logistic and probit regression models. Reviews are given by
Hagle and Mitchell (1992), Veall and Zimmermann (1992), and Windmeijer
(1995). Long (1997, Section 4.3) presents an extensive overview. One of
these definitions, the R2 measure of McKelvey and Zavoina (1975), which
is based on the threshold representation treated in Section 17.3.2, is
considered very attractive in each of these reviews. In this section we
propose a measure for the explained proportion of variance which extends
McKelvey and Zavoina’s measure to the logistic and probit random
intercept model.
It is assumed that it makes sense to conceive of the dichotomous
outcomes Y as being generated through a threshold model with underlying
variable . In addition, it is assumed that the explanatory variables Xh are
random variables. (This assumption is always made for defining the
explained proportion of variance; see the introduction of Section 7.1.)
Therefore the explanatory variables are, as in Chapter 7, indicated by
capital letters. For the underlying variable , equation (17.18) gives the
expression

Denote the fixed part by

This variable is also called the linear predictor for Y. Its variance is denoted
by The intercept variance is and the level-one residual
variance is denoted by Recall that for the
logistic and to 1 for the probit model.
For a randomly drawn level-one unit i in a randomly drawn level-two
unit j, the X-values are randomly drawn from the corresponding population
and hence the total variance of is equal to

The explained part of this variance is σF2 and the unexplained part is τ02 +
σR2. Of this unexplained variation, τ02 resides at level two and σR2 at level
one. Hence the proportion of explained variation can be defined by
 The corresponding definition of the residual intraclass correlation,

was also given in Section 17.3.3.

Estimating (17.22) is done in three steps. First one computes, as a new
transformed variable, the linear predictor Ŷij defined in (17.21) using the
estimated coefficients . . ., . Next the variance σF2 is estimated by
computing the observed variance of this new variable. Finally, this value is
plugged into (17.22) together with the estimated intercept variance and
the fixed value of σR2. The example below gives an illustration.
In the interpretation of this R2-value it should be kept in mind that such
values are known for single-level logistic regression to be usually
considerably lower than the OLS R2 values obtained for predicting
continuous outcomes.
Example 17.5 Explained variance for religious attendance.
For the random intercept model in Table 17.2, the linear predictor is

using the symbols X1,…,X12 for the predictor variables. This constructed
variable has variance (a value obtained by calculating this new
variable and computing its variance). Equation (17.22) yields the proportion
of explained variance R2dicho = 1.049/(1.049 + 1.082 + 3.29)= 0.19.
Example 17.6 Taking a science subject in high school.
This example continues the analysis of the data set of Example 8.3 about the cohort of pupils entering
secondary school in 1989, studied by Dekkers et al. (2000). The focus now is on whether the pupils
chose at least one science subject for their final examination. The sample is restricted to pupils in
general education (excluding junior vocational education), and to only those who progressed to their
final examination (excluding drop-outs and pupils who repeated grades once or twice). This left
3,432 pupils distributed over 240 secondary schools. There were 736 pupils who took no science
subjects, 2,696 who took one or more.
 A multilevel logistic regression model was estimated by R package lme4 and the MIXOR
program. Both gave almost the same results; the MIXOR results are presented. Explanatory variables
are gender (0 for boys, 1 for girls) and minority status (0 for children of parents born in industrialized
countries, 1 for other countries). The results are shown in Table 17.4.

Table 17.4: Estimates for probability of taking a science subject.

The linear predictor for this model is

and the variance of this variable in the sample is . Therefore the explained proportion of
variation is

In other words, gender and minority status explain about 13% of the variation in whether the pupil
takes at least one science subject for the high school exam.
The unexplained proportion of variation, 1−0.13 = 0.87, can be written as

which represents the fact that 11% of the variation is unexplained variation at the school level and
76% is unexplained variation at the pupil level. The residual intraclass correlation is ρI =
0.481/(0.481 + 3.29) = 0.13.
 17.3.5 Consequences of adding effects to the model
When a random intercept is added, to a given logistic or probit regression
model, and also when variables with fixed effects are added to such a
model, the effects of earlier included variables may change. The nature of
this change, however, may be different from such changes in OLS or
multilevel linear regression models for continuous variables.
This phenomenon can be illustrated by continuing the example of the
preceding section. Table 17.5 presents three other models for the same data,
in all of which some elements were omitted from Model 1 as presented in
Table 17.4.

Table 17.5: Three models for taking a science subject.

Models 2 and 3 include only the fixed effect of gender; Model 2 does
not contain a random intercept and therefore is a single-level logistic
regression model, while Model 3 does include the random intercept. The
deviance difference (χ2 = 3,345.15 – 3,251.86 = 93.29, df = 1) indicates that
the random intercept variance is significantly positive. But the sizes of the
fixed effects increase in absolute value when adding the random intercept to
the model, both by about 8%. Gender is evenly distributed across the 240
schools, and one may wonder why the absolute size of the effect of gender
increases when the random school effect is added to the model.
Model 4 differs from Model 1 in that the effect of gender is excluded.
The fixed effect of minority status in Model 4 is −0.644, whereas in Model
1 it is −0.727. The intercept variance in Model 4 is 0.293 and in Model 1 it
is 0.481. Again, gender is evenly distributed across schools and across the
majority and minority pupils, and the question is how to interpret the fact
that the intercept variance, that is, the unexplained between-school
variation, rises, and that also effect of minority status becomes larger in
absolute value, when the effect of gender is added to the model.
The explanation can be given on the basis of the threshold
representation. If all fixed effects γh and also the random intercept U0j and
the level-one residual Rij were multiplied by the same positive constant c,
then the unobserved variable would also be multiplied by c. This
corresponds also to multiplying the variances and σR2 by c2. However, it
follows from (17.17) that the observed outcome Yij would not be affected
because when is positive then so is c . This shows that the regression
parameters and the random part parameters of the multilevel logistic and
probit models are meaningful only because the level-one residual variance
σR2 has been fixed to some value; but this value is more or less arbitrary
because it is chosen merely by the convention of σR2 = π2/3 for the logistic
and σR2 = 1 for the probit model.
The meaningful parameters in these models are the ratios between the
regression parameters γh, the random effect standard deviations τ0 (and
possibly τ1, etc.), and the level-one residual standard deviation σR. Armed
with this knowledge, we can understand the consequences of adding a
random intercept or a fixed effect to a logistic or probit regression model.
When a single-level logistic or probit regression model has been
estimated the random variation of the unobserved variable in the
threshold model is When subsequently a random intercept is added this
random variation becomes σR2 + τ02. For explanatory variables that are
evenly distributed between the level-two units, the ratio of the regression
coefficients to the standard deviation of the (unexplained) random variation
will remain approximately constant. This means that the regression
coefficients will be multiplied by about the factor
In the comparison between Models 2 and 3 above, this factor is
1.08. This is indeed approximately the number by which
the regression coefficients were multiplied when going from Model 2 to
Model 3.
It can be concluded that, compared to single-level logistic or probit
regression analysis, including random intercepts tends to increase (in
absolute value) the regression coefficients. For single-level models for
binary variables, this was discussed by Winship and Mare (1984). In the
biostatistical literature, this is known as the phenomenon where population-
averaged effects (i.e., effects in models without random effects) are closer
to zero than cluster-specific effects (which are the effects in models with
random effects). Further discussions can be found in Neuhaus et al. (1991),
Neuhaus (1992), Diggle et al. (2002, Section 7.4), and Skrondal and Rabe-
Hesketh (2004, Section 4.8). Bauer (2009) proposes a way to put the
estimates from different models on a common scale.
Now suppose that a multilevel logistic or probit regression model has
been estimated, and the fixed effect of some level-one variable Xr+1 is
added to the model. One might think that this would lead to a decrease in
the level-one residual variance σR2. However, this is impossible as this
residual variance is fixed so that instead the estimates of the other
regression coefficients will tend to become larger in absolute value and the
intercept variance (and slope variances, if any) will also tend to become
larger. If the level-one variable Xr+1 is uncorrelated with the other included
fixed effects and also is evenly distributed across the level-two units (i.e.,
the intraclass correlation of Xr+1 is about nil), then the regression
coefficients and the standard deviations of the random effects will
all increase by about the same factor. Correlations between Xr + 1 and other
variables or positive intraclass correlation of Xr + 1 may distort this pattern
to a greater or lesser extent.
This explains why the effect of minority status and the intercept
variance increase when going from Model 4 to Model 1. The standard
deviation of the random intercept increases by a larger factor than the
regression coefficient of minority status, however. This might be related to
an interaction between the effects of gender and minority status and to a
very even distribution of the sexes across schools (cf. Section 7.1).
 17.4 Ordered categorical variables
Variables that have as outcomes a small number of ordered categories are
quite common in the social and biomedical sciences. Examples of such
variables are outcomes of questionnaire items (with outcomes such as
‘completely disagree’, ‘disagree’, ‘agree’, ‘completely agree’), a test scored
by a teacher as ‘fail’, ‘satisfactory’, or ‘good’, etc. This section is about
multilevel models where the dependent variable is such an ordinal variable.
When the number of categories is two, the dependent variable is
dichotomous and Section 17.2 applies. When the number of categories is
rather large (5 or more), it may be possible to approximate the distribution
of the residuals by a normal distribution and apply the hierarchical linear
model for continuous outcomes. The main issue in such a case is the
homoscedasticity assumption: is it reasonable to assume that the variances
of the random terms in the hierarchical linear model are constant? (The
random terms in a random intercept model are the level-one residuals and
the random intercept, Rij and U0j in (4.8).) To check this, it is useful to
investigate the skewness of the distribution. If in some groups, or for some
values of the explanatory variables, the dependent variable assumes
outcomes that are very skewed toward the lower or upper end of the scale,
then the homoscedasticity assumption is likely to be violated.
If the number of categories is small (3 or 4), or if it is between 5 and,
say, 10, and the distribution cannot well be approximated by a normal
distribution, then statistical methods for ordered categorical outcomes can
be useful. For single-level data such methods are treated, for example, in
McCullagh and Nelder (1989) and Long (1997).
It is usual to assign numerical values to the ordered categories, taking
into account that the values are arbitrary. To have a notation that is
compatible with the dichotomous case of Section 17.2, the values for the
ordered categories are defined as 0, 1 , . . . , c – 1, where c is the number of
categories. Thus, in the four-point scale mentioned above, ‘completely
disagree’ would get the value 0, ‘disagree’ would be represented by 1,
‘agree’ by 2, and ‘completely agree’ by the value 3. The dependent variable
for level-one unit i in level-two unit j is again denoted Yij, so that Yij now
assumes values in the set {0, 1 , . . . , c – 1}.
 A very useful model for this type of data is the multilevel ordered
logistic regression model, also called the multilevel ordered logit model or
the multilevel proportional odds model; and the closely related multilevel
ordered probit model. These models are discussed, for example, by Agresti
and Natarajan (2001), Gibbons and Hedeker (1994), Hedeker (2008),
Hedeker and Gibbons (2006, Chapter 10), Rabe-Hesketh and Skrondal
(2008, Chapter 7), and Goldstein (2011). A three-level model was discussed
by Gibbons and Hedeker (1997).
These models can be formulated as threshold models as in Section
17.3.2, now with c –1 thresholds rather than one. The real line is divided by
the thresholds into c intervals (of which the first and the last have infinite
length), corresponding to the c ordered categories. The first threshold is 0
= 0, the higher thresholds are denoted Threshold k defines
the boundary between the intervals corresponding to observed outcomes k
and k + 1 (for k = 0, 1 , . . . , c – 2). The assumed unobserved underlying
continuous variable is again denoted by and the observed categorical
variable Y is related to by the ‘measurement model’ defined as

For c = 2 categories this reduces to just the two-category threshold

representation in (17.17).
The random intercept ordered category model with explanatory
variables X1, . . . , Xr is based on the ‘structural model’ for the unobserved
underlying variable,

The structural model and the measurement model (17.24) together

determine the distribution of Yij. If Rij has the logistic distribution (17.19)
this results in the multilevel ordered logistic regression model, also called
the multilevel ordered logit model or multilevel proportional odds model. If
Rij has the standard normal distribution, this leads to the multilevel ordered
probit model. The differences between these two models are minor and the
choice between them is a matter of fit and convenience.
The parameters of the structural model can be interpreted in principle
just as in the hierarchical linear model. The intraclass correlation coefficient
can be defined as in Section 17.3.3. This definition referred only to the
model for the underlying variable , and therefore can be applied
immediately to the multi-category case. Similarly, the proportion of
explained variance can be defined, as in Section 17.3.4, by the formula

where σF2 is the variance of the fixed part (or the linear predictor) while σR2
is π2/3 = 3.29 for the logistic model and 1 for the probit model.
The threshold parameters are usually of secondary importance and
reflect the marginal probabilities of the outcome categories: if category k
has a low probability then θk – 1 will be not much less than θk. For more
discussion about the interpretation of the fixed parameters we refer to the
literature on the single-level version of this model, such as Long (1997).
The model can be extended with a random slope in a straightforward
manner. However, estimation algorithms for these models are less stable
than for the standard hierarchical linear model, and it is not uncommon that
it is impossible to obtain converging parameter estimates for models with
even only one random slope.
What was said in Section 17.3.5 about the effect of adding level-one
variables to a multilevel logistic regression model is valid also for the
multilevel ordered logit and probit models. When some model has been
fitted and an important level-one variable is added to this model, this will
tend to increase the level-two variance parameters (especially if the newly
added variable explains mainly within-group variation), the threshold
parameters, and the absolute sizes of the regression coefficients (especially
for variables that are uncor-related with the newly added variable). This is
also discussed in Fielding (2004b) and Bauer (2009).
These models can be estimated by the methods discussed above in
Section 17.2.5. Various of these procedures are implemented in the
programs MLwiN, HLM, MIXOR, Stata, and SAS (see Chapter 18). Some
computer programs do not put the first threshold equal to 0, but the
intercept. Thus, γ0 = 0 and this parameter is not estimated, but instead θ0 is
not equal to 0 and is estimated. This is a reparametrization of the model and
yields parameters which can simply be translated into one another, since it
follows from (17.24) and (17.25) that subtracting the same number of all
thresholds as well as of γ0 yields the same distribution of the observed
variables Yij.

Dichotomization of ordered categories

Models for ordered categorical outcomes are more complicated to fit and to
interpret than models for dichotomous outcomes. Therefore it may make
sense also to analyze the data after dichotomizing the outcome variable. For
example, if there are three outcomes, one could analyze the dichotomization
1 versus {2, 3} and also {1, 2} versus 3. Each of these analyses separately
is of course based on less information, but may be easier to carry out and to
interpret.
Suppose that the multilevel c-category logistic or probit model, defined
by (17.24) and (17.25), is indeed valid. Then for each of the
dichotomizations, the population parameters of the structural model (17.25),
except for the fixed intercept γ0, are also the population parameters of the
multilevel logistic regression model. (The fixed intercept in the analysis of
the dichotomized outcomes depends on the fixed intercept for the
multicategory outcomes together with the threshold θk for this
dichotomization.) This implies that the analyses of the dichotomized
outcomes also provide insight into the fit of the model for the c categories.
If the estimated regression parameters γ1, . . . , γr depend strongly on the
dichotomization point, then it is likely that the multilevel multicategory
logistic or probit model does not fit well.
Example 17.7 The number of science subjects taken in high school.
This example continues the analysis of Example 17.5, but now analyzes the number of science
subjects instead of only whether this number was larger than 0.
The number of science subjects ranges from 0 to 3 (mathematics, chemistry, physics). There were
736 pupils who took no science subjects, 1,120 who took one, 873 who took two, and 703 who took
all three.
The multilevel four-category logistic regression model was estimated by the MLwiN and MIXOR
programs, with gender (0 for boys, 1 for girls), socio-economic status (an ordered scale with values
1–6, from which the mean, 3.68, was subtracted), and minority status (0 for children of parents born
in industrialized countries, 1 for other countries) as explanatory variables. Both programs produced
very similar results. The MIXOR results are presented so that the deviance is available.
Table 17.6 shows the result for the empty model (which only has the thresholds and the intercept
variance as parameters) and for the model with these three explanatory variables. The deviance is
minus twice the log-likelihood reported by MIXOR. Further, MIXOR reports standard deviations of
random effects rather than variances. To obtain consistency with the other tables in this book, we
report the intercept variance instead. The standard error of the variance estimate is obtained from the
formula S.E. , obtained by rearranging equation (6.2).
In the empty model the intercept variance is 0.243. In the model with explanatory variables this
parameter has increased to 0.293. This increase is in accordance with what was said on p. 309 about
the effect of including level-one variables in the model. The same effect is responsible for the fact
that the threshold parameters have increased. All three fixed effects in Model 2 are significant: girls
tend to take considerably fewer science subjects than boys, the number of science subjects is an
increasing

Table 17.6: Multilevel four-category logistic regression model for the

number of science subjects.

function of socio-economic status, and minority pupils tend to take fewer

science subjects than nonminority pupils. The proportion of explained
variance can be calculated from equation (17.26). The linear predictor is
given here by

−1.680 X1 + 0.117X2−0.514X3,

and has variance 0.745. Hence the explained variance is

 The thresholds are approximately equidistant (recall that the first
threshold is θ0 = 0), with a difference between them in Model 2 of about
1.6. The gender effect has about the same size in absolute value, −1.68, so
girls tend to take about one science subject fewer than boys. The effect of
minority status, −0.514, amounts to an average difference of about 0.3
subjects. The random intercept standard deviation is , so the
difference between schools is quite large (as the span from schools with the
few percent lowest U0j to schools with the few percent highest U0j is four
standard deviations). The fact that the intercepts (fixed coefficients) are so
different between Models 1 and 2 is because, of the explanatory variables,
only SES is centered: for Model 1 the intercept corresponds to the average
value of all pupils, whereas for Model 2 it corresponds to the average
value of the nonminority boys.

17.5 Multilevel event history analysis

Event history analysis is the study of durations until some event occurs,
such as recovery from a disease, entry into the labor market, or marriage;
see, for example, Singer and Willett (2003) and Mills (2011). When events
are considered that can only happen once, and an approach is followed
where time proceeds in discrete steps (days, years, etc.), a standard
procedure is to transform the data into the so-called person–period format,
which is a two-level format with periods nested within individuals, and with
a binary outcome which is 0 if the event has not yet occurred, and 1 if it has
occurred; and where for every person for whom the event has occurred
during the observation period, only the first time step after the event is
included in the data set. This transforms the outcome variable into a binary
one, so that methods for binary response variables can be used. A single-
level event history analysis can be modeled, for example, by logistic
regression of such a data file. In spite of the two-level data structure this
should be a single-level logistic regression without additional random
effects, because the data for each person are restricted to being either a
sequence of zeros if the event has never occurred, or a sequence of zeros
followed by a single one if it has occurred. A multilevel nesting structure,
with persons nested in groups (higher-level units), leads to a three-level data
structure which can be modeled, for example, by multilevel logistic
regression, with random effects for the groups, which then are the units at
level three.
The logit link function leads here to the so-called proportional odds
model for discrete time multilevel event history analysis, while the
complementary log-log link function yields the proportional hazards model.
These models, and the choice between them, are treated in the literature
cited.
Multilevel models for event history analysis that fit within this
framework are discussed by, for example, Steele (2008, Section 3), Rabe-
Hesketh and Skrondal (2008, Chapter 8), Goldstein (2011), and Mills (2011,
Chapter 8). The literature contains many examples; one of these, Windzio
(2006), focuses on the incorporation of time-changing contextual variables.
Other models for multilevel event history analysis were presented by
Glidden and Vittinghoff (2004).

17.6 Multilevel Poisson regression

Another important type of discrete data is count data. For example, for a
population of road crossings, one might count the number of accidents in
one year; or, for a population of doctors, one could count how often in one
year they are confronted with a certain medical condition. The set of
possible outcomes of count data is the set of natural numbers: 0, 1, 2, 3, 4,
... The standard distribution for counts is the Poisson distribution – see
textbooks on probability theory, or, for example, Long (1997, Section 8.1).
A nice introduction to the analysis of count data on the basis of the Poisson
distribution in cases where there are no explanatory variables can be found
in in Box et al. (1978, Section 5.6). An extensive treatment of Poisson and
other models for count data is given by Cameron and Trivedi (1998).
The Poisson distribution is an approximation to the binomial
distribution for the situation that the number of trials is large and the
success probability is low. For the road accident example, one could
consider a division of the year into 8,760 one-hour periods, and assume that
the probability is negligible that more than one accident will happen in any
single hour, that the occurrence of accidents is independent for different
hours, and that the probability of an accident in any given hour, say, p, is
very low. The total number of accidents would then be binomially
distributed with parameters 8,760 and p, and this distribution is extremely
close to the Poisson distribution with mean 8,760 × p. Even when the
‘success’ probabilities are variable over trials (the accident probability is
variable from one hour to the next), the number of ‘successes’ will still
have approximately a Poisson distribution.
Just as we described in Section 17.1 for the binomial distribution, there
is also for the Poisson distribution a natural relation between the mean and
variance: they are equal. If Y has a Poisson distribution with mean λ, then

If the counts tend to be large, their distribution can be approximated by

a continuous distribution. If all counts are large enough (say, more than 8),
then it is advisable to use the square root of the counts, , as the
dependent variable and apply the hierarchical linear model (with the usual
assumption checks). The reason why this is a good approach resides in the
fact that the square root transformation succeeds very well in transforming
the Poisson distribution into an approximately homoscedastic normal
distribution – the square root is the so-called variance-stabilizing
transformation for the Poisson distribution (see Box et al., 1978, p. 144). An
even better transformation to a homoscedastic distribution is the Freeman–
Tukey transformation defined by (see Bishop et al., 1975,
Section 14.6.2).
If all or some of the counts are small, a normal distribution will not be
satisfactory and a hierarchical generalized linear model can be considered.
This model is analogous to the multilevel logistic regression model.
However, instead of the probability of success of Section 17.2 we now
model the expected value of the count. Denote by Yij the count for level-one
unit i in group j, and by Lij the expected (i.e., population average) count,
given that unit i is in group j and given the values of the explanatory
variables (if any). Then Lij is necessarily a nonnegative number, which
could lead to difficulties if we considered linear models for this value (cf.
Section 17.2.2). Just as the logit function was used as the link function for
probabilities in Section 17.2.2, so the natural logarithm is mostly used as
the link function for expected counts. For single-level data this leads to the
Poisson regression model (Long 1997, Chapter 8) which is a linear model
for the natural logarithm of the counts, ln(Lij). For multilevel data,
hierarchical linear models are considered for the logarithm of Lij. The
random intercept Poisson regression model is thus formulated, analogously
to (17.15), as a regression model plus a random intercept for the logarithm
of the expected count:

The variance of the random intercept is again denoted by τ02. This model is
treated in Diggle et al. (2002, Section 9.4), Hedeker and Gibbons (2006,
Chapter 12), Goldstein (2011, Section 4.5), and Skrondal and Rabe-Hesketh
(2004, Chapter 11).
The multilevel Poisson regression model can be estimated by various
multilevel software packages (see Section 17.2.5 and Chapter 18), including
gllamm, HLM, MIX-PREG/SuperMix, MLwiN, various R packages
including lme4 and glmmADMB, SAS, Stata, and Latent Gold. The
estimation methods, as with those for multilevel logistic regression, include
numerical integration (in the version of Gaussian quadrature, with or
without adaptive node placement; see Rabe-Hesketh et al., 2005), Laplace
approximation (Raudenbush et al., 2000), and various Taylor series
approximations (Goldstein, 2011). The numerical integration method and
the Laplace approximation provide not only parameter estimates but also a
deviance statistic which can be used for hypothesis testing.
To transform the linear model back to the expected counts, the inverse
transformation of the natural logarithm must be used, which is the
exponential function exp(x)= ex. This function has the property that it
transforms sums into products:

Therefore the explanatory variables and the level-two random effects in the
(additive) multilevel Poisson regression model have multiplicative effects
on the expected counts. For example, if there is only r = 1 explanatory
variable, equation (17.28) is equivalent to

Therefore, each additional unit of X1 will have the effect of multiplying the
expected count by eγ1. Similarly, in a group with a high intercept, for
example, two standard deviations so that U0j = 2 τ0, the expected count will
be e2τ0 times as high as in a group with an average value, U0j= 0, of the
intercept.
In models for counts it is quite usual that there is a variable D that is
known to be proportional to the expected counts. For example, if the count
Yij is the number of events in some time interval of nonconstant length dij, it
often is natural to assume that the expected count is proportional to this
length of the time period. If in the example of counts of medical problems
there are several doctors each with his or her own population of patients,
then D could be the size of the patient population of the doctor. In view of
equation (17.29), in order to let the expected count be proportional to D,
there should be a term ln(dij) in the linear model for ln(Lij), with a
regression coefficient fixed to 1. Such a term is called an offset in the linear
model (see McCullagh and Nelder, 1989). Goldstein (2011, Section 4.5)
suggests that such offset variables be centered to improve the numerical
properties of the estimation algorithm.
Example 17.8 Memberships in voluntary organizations.
Memberships in voluntary organizations may be regarded as a measurement of social activity and a
dimension of social capital. In this example a subset is used of the data also analyzed in an
international comparative study by Peter and Drobnicˇ (2010). Here only the data for the Netherlands
are studied. They were collected in the Dutch part of the European Social Survey (ESS), in the
‘Citizenship, Involvement and Democracy’ module, during 2002–2003.
The dependent variable is the number of types of association of which the respondent is a
member. Twelve types of organizations are mentioned, ranging from sports clubs through political
parties to consumer organizations. Thus, the count can range from 0 to 12. This data set has a
regional identifier which ranges over 40 regions for the Netherlands. Respondents aged between 25
and 65 years were selected, and a very small number with missing values were discarded. This left a
total of 1,738 respondents. The nesting structure is respondents nested in regions.
The number of memberships ranges from 0 to 10, with a mean of 2.34 and a standard deviation of
1.78. The variance is 3.16. For the Poisson distribution according to (17.27) in the population the
mean and variance are equal. The fact that here dispersion is greater than would be expected for the
Poisson distribution points to heterogeneity, which may have to do with individual differences and/or
differences between regions. The Poisson distribution can assume all nonnegative values while the
outcome variable here is limited by the questionnaire to a maximum of 12, but in view of the low
mean and standard deviation this in itself is not a problem for applying the Poisson distribution.
Deviating from our practice in earlier chapters, here we report for the random part the standard
deviations and correlation parameters, rather than the variances and covariance. There are two
reasons for this. One is the fact that the numbers reported are smaller. The other is the use of the R
packages lme4 and glmmADMB for the computations, which use these parameters for reporting.
The empty model is reported as Model 1 in Table 17.7. Explorations showed that religion (coded
as Protestant versus other), gender, and age had effects, and the effect of age was well described by a
quadratic function. Furthermore, there appeared to be a random slope for gender and an interaction
between age and gender. This is reported as Model 2. Age is centered at 40 years. To obtain
parameters that are not too small, age is measured in decades: for example, 30 years is coded as –1,
40 years as 0, 50 years as +1.

Table 17.7: Two Poisson models for number of memberships.

A random intercept model with the same fixed effects as Model 2 has
deviance 6531.9. The deviance difference is 6531.9 – 6522.6 = 9.3. Against
a chi-squared distribution with 2 degrees of freedom this has p < < 0.01,
even without the more powerful test (cf. Section 6.2.1). This shows there is
clear evidence for a random slope for gender: the effect of gender on
number of membership differs between regions.
We see that Protestants have more memberships on average. To assess
the contribution of gender, we must take into account that there is an
interaction between gender and age, as well as a random slope for gender.
Since age is centered at 40 years, the main effect of gender as well as the
intercept variance refer to age 40. It can be concluded that on average,
females have fewer memberships at 40 years. The total contribution of
being female is

which is 0 for age 28 if U1j = 0. It can be concluded that in an average region, females have on
average the same number of memberships as males at age 28, and the average increase by age is
smaller for females than for males. But the male–female difference depends considerably on region,
as the between-region standard deviation is , larger than the main effect of
gender. Therefore there are many regions where at age 40 females have more memberships than
males.

In linear regression we are accustomed to not being very concerned

about deviations from normality for the distribution of residuals, as long as
there are no serious outliers. For Poisson regression this robustness to
deviations from the distributional assumption does not hold. The reason for
the difference is that the normal distribution has two separate parameters for
the mean and variance, whereas the Poisson distribution has just one
parameter, and in formula (17.27) we saw that the mean and variance are
equal. It is not unusual, however, for counts to have a distribution where the
variance differs from the mean; in most cases the variance will then be
larger than the mean, which is called overdispersion with respect to the
Poisson distribution. To obtain robustness for such deviations from the
Poisson distribution, two approaches can be used.
One is to add overdispersion to the model without precisely specifying
the resulting probability distribution (McCullagh and Nelder, 1989). Denote
the overdispersion parameter by φ; this must be positive. Given that the
mean of the distribution is parametrized by λ just like the Poisson
distribution, the mean and variance of the distribution are then given by

Thus,φ = 1 represents the Poisson distribution, φ > 1 represents

overdispersion, and φ < 1 underdispersion, which is possible but less
common.
The other approach is to replace the Poisson distribution by the so-
called negative binomial distribution (see, for example, Long, 1997). The
negative binomial distribution has two parameters, and The
only property we need to know of this distribution is that if the random
variable Y has a negative binomial distribution with parameters λ and α then
its mean and variance are given by

Thus, the parameter α determines the amount of overdispersion; the smaller

it is, the greater the overdispersion; and as α grows very large the Poisson
distribution is obtained as a boundary case.
The two mean–variance relationships (17.30) and (17.31) are different,
because the latter can represent only overdispersion and the variance-to-
mean ratio also increases as a function of the mean λ, whereas the former
can represent underdispersion as well, and has a constant variance-to-mean
ratio. Which of the two relationships is more suitable is an empirical matter,
but the practical interpretation of the results for an overdispersed Poisson
model and for a negative binomial model will usually be very similar.
Thus, for a dependent variable which is a count, or more generally a
nonnegative integer, one can choose between Poisson regression and
negative binomial regression. The latter is more robust and has Poisson
regression as a boundary case, and is therefore preferable unless the
researcher is quite confident that the dependent variable, given the predictor
variables and given the random effects, indeed has a Poisson distribution.
Poisson regression with an overdispersion parameter often leads to quite
similar conclusions to negative binomial regression; it has the disadvantage
that it is not based on an explicit probability distribution, but the advantage
that it also allows underdispersion.
Further treatments of Poisson, overdispersed Poisson, and negative
binomial regression are given by Gelman and Hill (2007, Chapter 15) and
Berk and MacDonald (2008). Berk and MacDonald warn that it is important
especially to work toward a good specification of the predictors and
covariance structure of the model, rather than resorting too easily to
overdispersed modeling as quick fix. A mathematical treatment is given by
Demidenko (2004), who also provides a test for overdispersion (Section
7.5.10).
Example 17.9 Negative binomial regression for memberships in voluntary organizations.
We continue the example on memberships in voluntary organizations, now applying negative
binomial regression. The same models are estimated as in Table 17.7, but now for negative binomial
regression. The results are presented in Table 17.8.
 Table 17.8: Two negative binomial models for number of memberships.
Comparing the two empty models, Model 1 (Table 17.7) and Model 3, we see that Model 3 fits
better (deviance difference 6,658.3 −6,588.6 = 69.7, df = 1, p < 0.001), the negative binomial
parameter is rather large indicating a moderate degree of overdispersion, the estimated
intercept parameter is virtually the same, and the intercept standard deviation is slightly less in Model
3 than in Model 1. The overdispersion occurs at level one, but in Model 1 it cannot be represented
anywhere at level one and therefore adds to the dispersion of the level-two random effects.
Comparing the two substantive models, Model 2 and Model 4, we see again that Model 4 fits
better (deviance difference 6,522.6 – 6,483.3 = 379.3, df = 1, p < 0.001), all regression coefficients
are practically the same, the standard deviations of the intercept and slope are smaller in Model 4
than in Model 2, and the slope―intercept correlation is larger. The standard errors of the fixed effects
are all slightly larger in Model 4 than in Model 2. The negative binomial parameter has increased
compared to Model 3 which shows that apparently the explanatory variables capture
some, but not all, of what appears to be overdispersion in the empty model.
The larger standard errors in Model 4 are definitely more trustworthy than those in Model 2,
given our observations above about the greater robustness of overdispersed models. Similarly, the
different parameters for the level-two random effects in Model 4 are more trustworthy than those in
Model 2. In this case the differences between the overdispersed and the Poisson models are
moderate, but in other data sets they may be more important.

17.7 Glommary
Dichotomous variables. Also called binary variables, these are variables
with two possible values, such as ‘yes’ and ‘no’; often the values are
formally called ‘success’ and ‘failure’. For such dependent variables, a
 hierarchical linear model with homoscedastic normally distributed
residuals is not appropriate, in the first place because the residual
variance of a binary dependent variable will depend on the predicted
value. The same holds for dependent variables with a small number of
ordered numerical categories, because the residual variance must
become small when the predicted value approaches either of the
extremes of the range of values; and for nonnegative integer (count)
data, because the residual variance must become small when the
predicted value approaches 0.

Hierarchical generalized linear model. This is similar to the hierarchical

linear model, but for dependent variables that are not normally
distributed given the explanatory variables. For dichotomous variables,
the best-known hierarchical generalized linear model is the multilevel
logistic regression model. In this model, the success probability is the
logistic transform of the linear predictor – the fixed part of the model –
plus the contribution made by the random effects. The latter will
usually comprise a random intercept to which random slopes may be
added; as in the regular hierarchical linear model, these random
coefficients vary over the level-two units.

Logistic function. This is given by logistic(x) = exp(x) / (1 + exp(x)), and

maps the set of real numbers on to the interval (0,1).

Logit function. This is given by logit(x) = ln (p / ( 1 —p)), and maps the

interval (0,1) on to the set of real numbers. It is also called the log-
odds. It is used as the link function in logistic regression, linking the
linear predictor to the expected value of the dependent variable. The
logit function is the inverse of the logistic function, that is,
logistic(logit(p)) = p and logit(logistic(x)) = x.

Empty model for dichotomous data. This is a logistic model for

dichotomous data where the linear predictor is constant (consisting of
only the fixed intercept) and the logit of the probability of success is a
random intercept, that is, the constant term plus a random residual
depending on the level-two unit.
Random intercept model for dichotomous data. This is a logistic model
for dichotomous data where the linear predictor consists of a linear
combination of explanatory variables, and in addition there is a
random intercept.

Random slope model for dichotomous data. In addition to the linear

predictor and the random intercept, this comprises a random slope for
one or more level-one variables.

Multilevel logistic regression model. This is a logistic model with random

coefficients, also called the multilevel logit model.

Multilevel probit model. This is quite analogous to the multilevel logistic

regression model, but now the link function, transforming the linear
predictor plus random coefficients to the probability of success, is the
cumulative distribution function of the standard normal distribution.

Threshold representation. This is a nice way of interpreting the multilevel

logit and probit models. It postulates a continuous variable,
unobserved, of which the distribution is the linear predictor plus a
‘residual’ having, respectively, a logistic or a standard normal
distribution, and of which only the sign (positive = success, negative =
failure) is observed.

Intraclass correlation for binary variables. This can be defined in two

ways. First, it can be defined just as for other variables, as in Chapter
3. Second, it can be defined in terms of the underlying continuous
variable in the threshold representation. The second definition has the
advantage that it corresponds to an elegant definition of the proportion
of explained variance (see below).

Testing that the intraclass correlation for binary variables is 0. This can
be done by applying the well-known chi-squared test. Another test was
also mentioned, which can always be applied when there are many
groups (even if they are small, which may lead to difficulties for the
chi-squared test) as long as there is not a very small number of groups
making up almost all of the data.
Proportion of explained variance for dichotomous dependent variables.
This can be defined as the proportion of explained variance for the
underlying continuous variable in the threshold representation.

Adding effects of level-one variables. Doing this to a multilevel logistic

regression model can increase the random intercept variance and the
effects of uncorrelated level-one variables; this was explained by
considering the threshold representation.

Multilevel ordered logistic regression model. This is a hierarchical

generalized linear model for dependent variables having three or more
ordered categories. It can be represented as a threshold model just like
the multilevel logistic regression model for dichotomous outcome
variables, with c ― 1 thresholds between adjacent categories, where c
is the number of categories.

Event history analysis. The study of durations until some event occurs.
One approach is to transform data to the person–period format, which
is a two-level format with time periods nested within individuals, and
with a binary outcome which is 0 if the event has not yet occurred, and
1 if it has occurred; including only the first (if any) 1 response per
individual. A multilevel nesting structure, with persons nested in
groups (higher-level units), is then represented by a three-level data
structure (periods within persons within groups) which can be modeled
by multilevel logistic regression, with random effects at level three.

Multilevel Poisson regression. A hierarchical generalized linear model for

count data, in which the dependent variable has possible values 0, 1,
2,. . .. Conditional on the explanatory variables and the random effects,
under this model the dependent variable has a Poisson distribution.
This is the most often used probability distribution for counts. A
particular property of the Poisson distribution is that the variance is
equal to the expected value. If the expected values are large (say, all
larger than 8) then an alternative approach is to apply a square root
transformation to the dependent variable and use the regular
hierarchical linear model.
Overdispersion. For count data, overdispersion means that, given the
explanatory variables and random effects, the residual variance is
larger than the expected value. If this is the case, the use of the Poisson
distribution may lead to erroneous inferences – in particular,
underestimated standard errors. Then it is advisable to use negative
binomial regression. The negative binomial distribution is another
distribution for count data, but it has an extra parameter to
accommodate variances larger than the mean. Another option is to use
a multilevel Poisson regression model to which an overdispersion
parameter has been added that does not correspond to a particular
probability distribution.

Estimation methods. For hierarchical generalized linear models,

estimation is more complicated than for the hierarchical linear model
with normal distributions. Various algorithms (i.e., procedures for
calculating estimates) are used. One type of estimation procedure is
maximum likelihood; for these models approximations to the
maximum likelihood estimator have been developed using adaptive or
nonadaptive quadrature (i.e., numerical integration) and the Laplace
approximation. Bayesian methods have been developed based on
simulation (Markov chain Monte Carlo) and on the Laplace
approximation. Marginal quasi-likelihood and penalized quasi-
likelihood are estimation procedures which were developed earlier and
which are less effective compared to maximum likelihood estimation
by numerical integration or to Bayesian methods. For the estimation of
fixed effects when group sizes are not too small, however, these
methods still can perform quite well.

1
This is because in further modeling of this data set it will be preferable to present standard
deviations rather than variances of random slopes.
2
Rather than the double-subscript notation γhk used earlier in Chapters 4 and 5, we now use – to
obtain a relatively simple notation – a single-subscript notation γh analogous to (5.15).
 18
Software

Almost all procedures treated in this book can be carried out by standard
software for multilevel statistical models. This of course is intentional,
since this book covers those parts of the theory of the multilevel model that
can be readily applied in everyday research practice. However, things
change rapidly. Some of the software discussed in the previous edition of
this book is no longer available. On the other hand, new software packages
are shooting up like mushrooms. The reader is therefore advised to keep
track of the changes that can be found at the various websites we mention in
this chapter – although we have to add that some of these websites and even
internet addresses tend to change. But then again, search engines will
readily guide the curious multilevel researcher to the most recent sites.
Currently most details on the specialized multilevel software packages
can be found via the links provided on the homepage of the Centre for
Multilevel Modelling at the University of Bristol; see
http://www.bristol.ac.uk/cmm/. At this site one can also find reviews of the
multilevel software packages. Chapter 18 of Goldstein (2011) also provides
a list of computer programs for multilevel analysis.

OVERVIEW OF THE CHAPTER

We will provide a brief review of the multilevel software packages,
organized into three sections. Section 18.1 is devoted to special purpose
programs, such as HLM and MLwiN, specifically designed for multilevel
modeling. Section 18.2 treats modules in general-purpose software
packages, such as R, SAS, Stata, and SPSS, that allow for multilevel
modeling. Finally, Section 18.3 mentions some specialized software
programs, built for specific research purposes.

18.1 Special software for multilevel modeling

Multilevel software packages are aimed at researchers who specifically seek
to apply multilevel modeling techniques. Other statistical techniques are not
available in these specific multilevel programs (with some exceptions).
Each of the programs described in this section was designed by pioneers in
the field of multilevel modeling. We only review HLM, MLwiN and the
MIXOR suite, although there are other packages as well. The interested
reader may take a look at
http://www.bristol.ac.uk/cmm/learning/mmsoftware/ to find descriptions
and reviews of other multilevel software programs such as aML, EGRET,
and GENSTAT, stemming from the fields of econometrics, epidemiology,
and agriculture, respectively.

18.1.1 HLM
HLM was originally written by Bryk et al. (1996), and the theoretical
background behind most applications can be found in Raudenbush and
Bryk (2002). The main features of HLM are its interactive operation
(although one can also run the program in batch mode) and the fact that it is
rather easy to learn. Therefore it is well suited for undergraduate courses
and for postgraduate courses for beginners. The many options available also
make it a good tool for professional researchers. Information is obtainable
from the website, http://www.ssicentral.com/hlm/, which also features a
free student version. West et al. (2007) give an introduction to hierarchical
linear modeling with much attention to implementation in HLM.
Input consists of separate files for each level in the design, linked by
common identifiers. In a simple two-level case, for example, with data
about students in schools, one file contains all the school data with a school
identification code, while another file contains all the student data with the
school identification code for each student. The input can come from
system files of SPSS, SAS, SYSTAT or Stata, or may be given in the form
of ASCII text files.
Once data have been read and stored into a sufficient statistics file (a
kind of system file), there are three ways to work with the program. One
way is to run the program interactively (answering questions posed by the
program). Another is to run the program in batch mode. Batch and
interactive modes can also be combined. Finally, one can make full use of
the graphical interface. In each case the two-step logic of Section 6.4.1 is
followed. HLM does not allow for data manipulation, but both the input and
output can come from, and can be fed into, SPSS, SAS, SYSTAT, or Stata.
HLM does not go beyond four levels. It can be used for practically all the
analyses presented in this book, with the exception of multiple membership
models. Almost all examples in this book can be reproduced using HLM
version 7. Some interesting features of the program are the ability to test
model assumptions directly, for example, by test (10.5) for level-one
heteroscedasticity, and the help provided to construct contrast tests.
Furthermore, the program routinely asks for centering of predictor
variables, but the flipside of the coin is – if one opts for group mean
centering – that group means themselves must have been calculated outside
HLM, if one wishes to use these as level-two predictor variables.
A special feature of the program is that it allows for statistical meta-
analysis (see Section 3.7) of research studies that are summarized by only
an effect size estimate and its associated standard error (called in
Raudenbush and Bryk, 2002, the ‘V-known problem’). Other special
features are the analysis of data where explanatory variables are measured
with error (explained in Raudenbush and Sampson, 1999a), and the analysis
of multiply imputed data as discussed in Chapter 9. Note, however, that the
imputations have to be done with other specialized software packages. Next
to that HLM also offers facilities for the analysis of a special case of
multiply imputed data, namely the multilevel analysis of plausible values.
These are imputed data on the dependent variable for incomplete
designs,such as designs with booklet rotations as used in the PISA studies.
As a special feature it allows for the modeling of dependent random effects,
a topic not treated in this book.

18.1.2 MLwiN
MLwiN is the most extensive multilevel package, written by researchers
currently working at the Centre for Multilevel Modelling at the University
of Bristol (Rasbash and Wood-house, 1995; Goldstein et al., 1998; Rasbash
et al., 2009). Current information, including a wealth of documentation, can
be obtained from the website http://www.bristol.ac.uk/cmm/. Almost all the
examples in this book can be reproduced using MLwiN, which allows for
standard modeling of up to five levels. For heteroscedastic models (Chapter
8), the term used in the MLwiN documentation is ‘complex variation’. For
example, level-one heteroscedasticity is complex level-one variation. Next
to the standard IGLS (ML) and RIGLS (REML) estimation methods,
MLwiN also provides bootstrap methods (based on random drawing from
the data set or on random draws from an estimated population distribution),
and extensive implementation of Markov chain Monte Carlo methods
(Browne, 2009) for Bayesian estimation (see Section 12.1). With MLwiN
one may use the accompanying package REALCOM-Impute to impute
missing data in a multilevel model, and analyze these subsequently in
MLwiN. The program MLPowSim (Section 18.3.3) may be used to create
MLwiN macros for simulation-based power analysis.
A nice feature of MLwiN is that the program was built on NANOSTAT,
a statistical environment which allows for data manipulation, graphing,
simple statistical computations, file manipulation (e.g., sorting), etc. Data
manipulation procedures include several handy procedures relating to the
multilevel data structure. Input for MLwiN may be either an ASCII text file
or a system file from Minitab, SPSS, or Stata that contains all the data,
including the level identifiers. MLwiN can even be called directly from
Stata using the runmlwin command (Leckie and Charlton, 2011). Since the
data are available in one file, this implies that all the level-two and higher-
level data are included in disaggregated form at level one. The data are read
into a worksheet, a kind of system file. This worksheet, which can also
include model specifications, variable labels, and results, can be saved and
used in later sessions. The data can also be exported to Minitab, SPSS, or
Stata.
The most obvious way to work with the program is interactively using
the graphical interface. One can, however, also use the ‘command menu’ or
give a series of commands in a previously constructed macro.
MLwiN is the most flexible multilevel software package, but it may
take some time to get acquainted with its features. It is an excellent tool for
professional researchers and statisticians. The macro facilities in particular
provide experienced researchers ample opportunities for applications such
as meta-analysis, multilevel factor analysis, multilevel item response theory
modeling, to name just some examples.

18.1.3 The MIXOR suite and SuperMix

Hedeker and Gibbons (1996a, b) have constructed several modules for
multilevel modeling, focusing on special models that go beyond the basic
hierarchical linear model. DOS and Windows versions of these programs
are freely available with manuals and examples at
http://tigger.uic.edu/~hedeker/mixwin.html. These programs have no
facilities for data manipulation and use ASCII text files for data input. The
algorithms are based on numerical integration (see Section 17.2.5).
MIXREG (Hedeker and Gibbons, 1996b) is a computer program for
mixed effects regression analysis with autocorrelated errors. This is suitable
especially for longitudinal models (see Chapter 15). Various correlation
patterns for the level-one residuals Rti in (15.32) are allowed such as
autocorrelation or moving average dependence.
MIXOR (Hedeker and Gibbons, 1996a) provides estimates for
multilevel models for dichotomous and ordinal discrete outcome variables
(cf. Chapter 17). It allows probit, logistic, and complementary log-log link
functions. These models can include multiple random slopes. It also permits
right-censoring of the ordinal outcome (useful for analysis of multilevel
grouped-time survival data), nonproportional odds (or hazards) for selected
covariates, and the possibility for the random effect variance terms to vary
by groups of level-one or level-two units. This allows estimation of many
types of item response theory models (where the variance parameters vary
by the level-one items) as well as models where the random effects vary by
groups of subjects (e.g., males versus females).
MIXNO implements a multilevel multinomial logistic regression model.
As such, it can be used to analyze two-level categorical outcomes without
an ordering. As in MIXOR, the random effect variance terms can also vary
by groups of level-one or level-two units. This program has an extensive
manual with various examples.
 MIXPREG is a program for estimating the parameters of the multilevel
Poisson regression model (see Section 17.6). This program also has an
extensive manual with examples.
Finally, the program SuperMix combines all these packages with the
multilevel model for continuous data as an add-on, and moreover provides
ample opportunities for data manipulation, importing of files from
statistical packages or Excel, and allows the user to specify three-level
models using a graphical interface. This package is available at
http://www.ssicentral.com/supermix/, where one also can download a free
student version with limited capacity.
In the interpretation of the output of these programs, it should be noted
that the parametrization used for the random part is not the covariance
matrix but its Cholesky decomposition. This is a lower triangular matrix C
with the property that CC’ = Σ, where Σ is the covariance matrix. The
output does also give the estimated variance and covariance parameters, but
(in the present versions) not their standard errors. If you want to know these
standard errors, some additional calculations are necessary.
If the random part only contains the random intercept, the parameter in
the Cholesky decomposition is the standard deviation of the intercept. The
standard error of the intercept variance can be calculated with formula (6.2).
For other parameters, the relation between the standard errors is more
complicated. We treat only the case of one random slope. The level-two
covariance matrix is then

with Cholesky decomposition denoted by .

The correspondence between these matrices is

Denote the standard errors of the estimated elements of C by s00, s01, and
s11, respectively, and the correlations between these estimates by r00,01,
r00,11, and r01,11. These standard errors and correlations are given in the
output of the programs. Approximate standard errors for the elements of the
covariance matrix of the level-two random part are then given by

For the second and further random slopes, the formulas for the standard
errors are even more complicated. These formulas can be derived with the
multivariate delta method explained for example, in Bishop et al. (1975,
Section 14.6.3).

18.2 Modules in general-purpose software

packages
Several of the main general-purpose statistical packages have incorporated
modules for multilevel modeling. These are usually presented as modules
for mixed models, random effects, random coefficients, or variance
components. As it is, most researchers may be used to one of these
packages and may feel reluctant to learn to handle the specialized software
discussed above. For these people especially the threshold to multilevel
modeling may be lowered when they can stick to their own software. We
provide a brief introduction. For further details the interested reader may
turn to Twisk (2006), who discusses the possibilities of SPSS, Stata, SAS,
and R, and also gives some examples of setups and commands to run
multilevel models. And of course the website of the Centre for Multilevel
Modelling contains reviews of the options the general-purpose packages
provide for multilevel analysis as well; see
http://www.bristol.ac.uk/cmm/learning/mmsoftware/. West et al. (2007)
give a treatment of the hierarchical linear model (but not of generalized
linear versions) with much attention to the use of HLM, SAS, SPSS, and
Stata.

18.2.1 SAS procedures VARCOMP, MIXED, GLIMMIX, and

NLMIXED
The SAS procedure MIXED has been up and running since 1996 (Littell et
al., 2006). For experienced SAS users a quick introduction to multilevel
modules in general-purpose software packages modeling using SAS can be
found in Singer (1998a, 1998b). Verbeke and Molenberghs (1997) present a
SAS-oriented practical introduction to the hierarchical linear model. The
book on generalized hierarchical linear models for longitudinal data by
Molenberghs and Verbeke (2006) also offers much explanation of the
implementation in SAS. An introduction to using SAS proc MIXED for
multilevel analysis, specifically geared at longitudinal data, is given by
Singer (2002).
Unlike some other general-purpose packages, SAS allows quite general
variance components models to be fitted with VARCOMP, and in MIXED
one can fit very complex multilevel models and calculate all corresponding
statistics (such as deviance tests). MIXED is a procedure oriented toward
general mixed linear models, and can be used to analyze practically all the
hierarchical linear model examples for continuous outcome variables
presented in this book. The general mixed model orientation has the
advantage that crossed random coefficients can be easily included, but the
disadvantage is that this procedure does not provide the specific efficiency
for the nested random coefficients of the hierarchical linear model that is
provided by dedicated multilevel programs. Also available are the
procedures GLIMMIX and NLMIXED, which can be used to fit the models
for discrete outcome variables described in Chapter 17.

18.2.2 R
R is an open source language and environment for statistical computing and
graphics similar to the S language. It is highly flexible and freely
downloadable at http://cran.r-project.org/. Due to the open source character
of R one can use procedures that have been developed by others (Ihaka and
Gentleman, 1996). Procedures in R are organized into so-called packages. R
is operated by a command language, and the commands are collected in
scripts. In the initial phase it may require some effort to learn the command
language, but once mastered this has the advantage of flexibility and
reproducibility of results.
There are several packages in R implementing multilevel models. The
main ones are nlme (Pinheiro and Bates, 2000) and lme4 (Bates, 2010; see
also Doran et al., 2007). The nlme package has extensive possibilities for
linear and nonlinear models with normally distributed residuals. The lme4
package is currently still under vigorous development and, in spite of its
title (‘Linear mixed-effects models using S4 classes’) also estimates
hierarchical generalized linear models, using Laplace and other methods.
With both of these packages one can perform most of the analyses treated in
this book. Some texts introducing multilevel analysis using R are, next to
the two books just mentioned, Maindonald and Braun (2007, Chapter 10),
Bliese (2009), Wright and London (2009), Berridge and Crouchley (2011),
and, specifically for multilevel models for discrete data, Thompson (2009,
Chapter 12).
There are several packages that can be used for more limited, but very
useful, purposes. The R packages mlmmm and mice allow multiple
imputation of missing data (see Chapter 9) under a two-level model. The
pamm package contains procedures for simulation-based power analysis.
The program MLPowSim (Section 18.3.3) creates R scripts for simulation-
based power analysis. Various new methods for hierarchical generalized
linear models (Section 17.1) have been made available in R; some of these
packages are HGLMMM, glmml, glmmADMB, and INLA. Package
multilevel contains some procedures that are especially useful for those
working with multi-item scales. Using packages WinBUGS or
glmmmBUGS gives access to the WinBUGS program (Section 18.3.7).

18.2.3 Stata
Stata (StataCorp, 2009) contains some modules that permit the estimation
of certain multilevel models (see Rabe-Hesketh and Skrondal, 2008).
Module loneway (‘long oneway’) gives estimates for the empty model. The
xt series of modules are designed for the analysis of longitudinal data (cf.
Chapter 15), but can be used to analyze any two-level random intercept
model. Command xtreg estimates the random intercept model, while xtpred
calculates posterior means. Commands xtpois and xtprobit, respectively,
provide estimates of the multilevel Poisson regression and multilevel probit
regression models (Chapter 17). These estimates are based on the so-called
generalized estimating equations method. A special feature of Stata is the
so-called sandwich variance estimator, also called the robust or Huber
estimator (Section 12.2). This estimator can be applied in many Stata
modules that are not specifically intended for multilevel analysis. For
statistics calculated in a single-level framework (e.g., estimated OLS
regression coefficients), the sandwich estimator, when using the keyword
‘cluster’, computes standard errors that are asymptotically correct under
two-stage sampling. In terms of our Chapter 2, this solves many instances
of ‘dependence as a nuisance’, although it does not help to get a grip on
‘interesting dependence’.
Within Stata one can use the procedure GLLAMM – an acronym for
General Linear Latent And Mixed Models – which was developed by Rabe-
Hesketh et al. (2004, 2005) and which analyzes very general models indeed,
including those of Chapter 17 but also much more general models with
latent variables, such as multilevel structural equation modeling and
multilevel item response theory models. The algorithms use adaptive
quadrature (Section 17.2.5). The package is available at
http://www.gllamm.org, where one can also find the necessary
documentation.
As mentioned above, from Stata one may call MLwiN through the
command runmlwin.

18.2.4 SPSS, commands VARCOMP and MIXED

The simplest introduction to multilevel modeling for SPSS users is the
module (in the language of SPSS, ‘command’) VARCOMP. One can get no
further, however, than the random intercept and random slope model, as
described in Chapters 4 and 5 of this book, with the possibility also of
including crossed random effects (Chapter 13). The SPSS module MIXED
goes further and provides ample opportunities for three-level models with
random slopes and complex covariance structures. A detailed introduction
into multilevel modeling with SPSS is provided by Heck et al. (2010).

18.3 Other multilevel software

There are other programs available for special purposes, and which can be
useful to supplement the software mentioned above.

18.3.1 PinT
PinT is a specialized program for calculations of Power in two-level
designs, implementing the methods of Snijders and Bosker (1993). This
program can be used for a priori estimation of standard errors of fixed
coefficients. This is useful in the design phase of a multilevel study, as
discussed in Chapter 11. Being shareware, it can be downloaded with the
manual (Bosker et al., 2003) from
http://www.stats.ox.ac.uk/~snijders/multilevel.htm#progPINT.

18.3.2 Optimal Design

Optimal Design is a flexible program for designing multilevel studies and
was developed by Spybrook et al. (2009). The designs considered include
cluster randomized trials, multisite trials, and repeated measures. The
program is very flexible and even allows for power calculations for single-
level randomized trials, cluster randomized trials with the outcome of
interest at the cluster level, three-level experiments, and meta-analysis. The
graphical interface is very user-friendly, and within a short while one can
manipulate significance levels, number of clusters, number of observations
per cluster, intraclass correlations, associations between covariates and
outcomes, and infer effects on the resulting power. The software and the
manual can be downloaded from
http://www.wtgrantfoundation.org/resources/overview/research_tools.
 18.3.3 MLPowSim
MLPowSim (Browne et al., 2009) is a program for power analysis and
determination of sample sizes for quite general random effect models.
Standard errors and power are computed based on Monte Carlo simulations
for which either R or MLwiN can be used. MLPowSim creates R command
scripts and MLwiN macro files which, when executed in those respective
packages, employ their simulation facilities and random effect estimation
engines to perform these computations.

18.3.4 Mplus
Mplus is a program with very general facilities for covariance structure
analysis (Muthén and Muthén, 2010). Information about this program is
available at http://www.StatModel.com. This program allows the analysis of
univariate and multivariate two-level data not only with the hierarchical
linear model but also with path analysis, factor analysis, and other structural
equation models. Introductions to this type of model are given by Muthén
(1994) and Kaplan and Elliott (1997).

18.3.5 Latent Gold

Latent Gold (Vermunt and Magidson, 2005a, 2005b) is a program primarily
designed for analyzing latent class models, including the multilevel latent
class models discussed in Section 12.3. It can also estimate the more regular
multilevel models treated in other chapters. This program is available at
http://www.statisticalinnovations.com/products/latentgold_v4.html.

18.3.6 REALCOM
At the time of writing the people at the Centre for Multilevel Modelling are
developing new software for, as they call it, ‘realistic multilevel modeling’.
Such modeling includes measurement errors in variables, simultaneous
outcomes at various levels in the hierarchy, structural equation modeling,
modeling with imputed data for missing values, and modeling with
misclassifications. The interested reader is referred to
http://www.bristol.ac.uk/cmm/software/realcom for more details and recent
developments.

18.3.7 WinBUGS
A special program which uses the Gibbs sampler is WinBUGS (Lunn et al.,
2000), building on the previous BUGS program (Gilks et al., 1996). Gibbs
sampling is a simulation-based procedure for calculating Bayesian
estimates (Section 12.1). This program can be used to estimate a large
variety of models, including hierarchical linear models, possibly in
combination with models for structural equations and measurement error. It
is extremely flexible, and used, for example, as a research tool by
statisticians; but it can also be used for regular data analysis. Gelman and
Hill (2007) and Congdon (2010) give extensive attention to the use of
WinBUGS for multilevel analysis. The WinBUGS example manuals also
contain many examples of hierarchical generalized linear models.
The program is available with manuals from http://www.mrc-
bsu.cam.ac.uk/bugs/. From R (Section 18.2.2) it is possible to access
WinBUGS by using the R package WinBUGS or glmmmBUGS.
References

Abayomi, K., Gelman, A., and Levy, M. (2008) ‘Diagnostics for

multivariate imputations’. Applied Statistics, 57, 273–291.
Achen, C.H. (2005) ‘Two-step hierarchical estimation: Beyond
regression analysis’. Political Analysis, 13, 447–456.
Agresti, A. (2002) Categorical Data Analysis, 2nd edn. New York:
Wiley.
Agresti, A., and Natarajan, R. (2001) ‘Modeling clustered ordered
categorical data’. International Statistical Review, 69, 345–371.
Aitkin, M., Anderson, D., and Hinde, J. (1981) ‘Statistical modelling of
data on teaching styles’. Journal of the Royal Statistical Society, Series
A, 144, 419–461.
Aitkin, M., and Longford, N. (1986) ‘Statistical modelling issues in
school effectiveness studies’ (with discussion). Journal of the Royal
Statistical Society, Series A, 149, 1–43.
Alker, H.R. (1969) ‘A typology of ecological fallacies’. In: M. Dogan
and S. Rokkan (eds), Quantitative Ecological Analysis in the Social
Sciences, pp. 69–86. Cambridge, MA: MIT Press.
Alonso, A., Litiére, S., and Laenen, A. (2010) ‘A note on the
indeterminacy of the random-effects distribution in hierarchical
models’. The American Statistician, 64, 318–324.
Anderson, D., and Aitkin, M. (1985) ‘Variance components models with
binary response: Interviewer variability’. Journal of the Royal
Statistical Society, Series B, 47, 203–210.
Atkinson, A.C. (1985) Plots, Transformations, and Regression. Oxford:
Clarendon Press.
Asparouhov, T. (2006) ‘General multi-level modeling with sampling
weights’. Communications in Statistics – Theory and Methods, 35,
 439–460.
Austin, P.C. (2010) ‘Estimating multilevel logistic regression models
when the number of clusters is low: A comparison of different
statistical software procedures’. International Journal of Biostatistics,
6, Article 16.
Baltagi, B.H. (2008) Econometric Analysis of Panel Data, 4th edn.
Chichester: Wiley.
Bates, D.M. (2010) lme4: Mixed-effects modeling with R. In preparation.
http://lme4.r-forge.r-project.org/book/
Bauer, D.J. (2009) ‘A note on comparing the estimates of models for
cluster-correlated or longitudinal data with binary or ordinal
outcomes’. Psychometrika 74, 97–105.
Bauer, D.J., and Curran, P.J. (2005) ‘Probing interactions in fixed and
multilevel regression: Inferential and graphical techniques’.
Multivariate Behavioral Research, 40, 373–400.
Bell, R.M., and McCaffrey, D.F. (2002) ‘Bias reduction in standard errors
for linear regression with multi-stage samples’. Survey Methodology,
28(2), 169–179.
Beretvas, S.N. (2011) ‘Cross-classified and multiple-membership
models’. In: J.J. Hox and J.K. Roberts (eds), Handbook of Advanced
Multilevel Analysis, pp. 313–334. New York: Routledge.
Berk, R., and MacDonald, J. (2008) ‘Overdispersion and Poisson
regression’. Journal of Quantitative Criminology, 24, 269–284.
Berkhof, J., and Kampen, J.K. (2004) ‘Asymptotic effect of
misspecification in the random part of the multilevel model’. Journal
of Educational and Behavioral Statistics, 29, 201–218.
Berkhof, J., and Snijders, T.A.B. (2001) ‘Variance component testing in
multilevel models’. Journal of Educational and Behavioral Statistics,
26, 133–152.
Berridge, D.M., and Crouchley, R. (2011) Multivariate Generalized
Linear Mixed Models Using R. London: Psychology Press, Taylor and
Francis.
Bertolet, M. (2008) To weight or not to weight? Incorporating sampling
designs into model-based analyses. PhD dissertation, Carnegie Mellon
University. http://gradworks.umi.com/3326665.pdf
Bertolet, M. (2010) To Weight or Not To Weight? A survey sampling
simulation study. Submitted for publication.
Bishop, Y.M.M., Fienberg, S.E., and Holland, P.W. (1975) Discrete
Multivariate Analysis: Theory and Practice. Cambridge, MA: MIT
Press.
Blakely, T.A., and Woodward, A.J. (2000) ‘Ecological effects in multi-
level studies’. Journal of Epidemiology and Community Health 54,
367–374.
Bliese, P.D. (2009) Multilevel Modeling in R (2.3). http://cran.r-
project.org/doc/contrib/Bliese_Multilevel.pdf
Bohrenstein, M., Hedges, L.V., Higgins, J.P.T., and Rothstein, H.R.
(2009) Introduction to Meta-Analysis. New York: Wiley.
Bosker, R.J., Snijders, T.A.B., and Guldemond, H. (2003) PinT (Power in
Two-level designs). Estimating Standard Errors of Regression
Coefficients in Hierarchical Linear Models for Power Calculations.
User’s Manual Version 2.1. Groningen: University of Groningen.
http://www.stats.ox.ac.uk/~snijders/multilevel.htm#progPINT
Box, G.E.P., and Cox, D.R. (1964) ‘An analysis of transformations’ (with
discussion). Journal of the Royal Statistical Society, Series B, 26, 211–
252.
Box, G.E.P., Hunter, W.G., and Hunter, J.S. (1978) Statistics for
Experimenters. New York: Wiley.
Brandsma, H.P. and Knuver, J.W.M. (1989) ‘Effects of school and
classroom characteristics on pupil progress in language and
arithmetic’. International Journal of Educational Research, 13, 777–
788.
Breslow, N.E., and Clayton, D.G. (1993) ‘Approximate inference in
generalized linear mixed models’. Journal of the American Statistical
Association, 88, 9–25.
Browne, W.J. (2004) ‘An illustration of the use of reparameterisation
methods for improving MCMC efficiency in crossed random effect
models’. Multilevel Modelling Newsletter, 16(1), 13–25.
Browne, W.J. (2009) MCMC Estimation in MLwiN (Version 2.13).
Bristol: Centre for Multilevel Modelling, University of Bristol.
Browne W.J., and Draper, D. (2000) ‘Implementation and performance
issues in the Bayesian and likelihood fitting of multilevel models’.
Computational Statistics, 15, 391–420.
Browne, W.J., and Draper, D. (2006) ‘A comparison of Bayesian and
likelihood methods for fitting multilevel models’ (with discussion).
 Bayesian Analysis, 1, 473–550.
Browne, W.J., Goldstein, H. and Rasbash, J. (2001) ‘Multiple
membership multiple classification (MMMC) models’. Statistical
Modelling, 1, 103–124.
Browne, W.J., Lahi, M.G., and Parker, R.M.A. (2009) A Guide to Sample
Size Calculations for Random Effects Models via Simulation and the
MLPowSim Software Package. Bristol: University of Bristol.
Browne, W.J., McCleery, R.H., Sheldon, B.C., and Pettifor, R.A. (2007)
‘Using cross-classified multivariate mixed response models with
application to life history traits in great tits (Parus major)’. Statistical
Modelling, 7, 217–238.
Bryk, A.S., Raudenbush, S.W., and Congdon, R.T. (1996) HLM.
Hierarchical Linear and Nonlinear Modeling with the HLM/2L and
HLM/3L Programs. Chicago: Scientific Software International.
Burr, D., and Doss, H. (2005) ‘A Bayesian semiparametric model for
random-effects meta-analysis’. Journal of the American Statistical
Association, 100, 242–251.
Burstein, L., Linn, R.L., and Capell, F.J. (1978) ‘Analyzing multilevel
data in the presence of heterogeneous within-class regressions’.
Journal of Educational Statistics, 3, 347–383.
Cameron, A.C., Gelbach, J.B., and Miller, D.L. (2008) Bootstrap-based
improvements for inference with clustered errors. Review of
Economics and Statistics, 90, 414–427.
Cameron, A.C., and Trivedi, P.K. (1998) Regression Analysis of Count
Data. Cambridge: Cambridge University Press.
Carle, A.C. (2009) ‘Fitting multilevel models in complex survey data
with design weights: Recommendations’. BMC Medical Research
Methodology, 9:49.
Carpenter, J.R., and Kenward, M.G. (2008) Missing Data in Randomised
Controlled Trials – A Practical Guide. Birmingham: National Institute
for Health Research, Publication RM03/JH17/MK.
http://www.hta.ac.uk/nihrmethodology/reports/1589.pdf
Chen, J., Zhang, D., and Davidian, M. (2002) ‘A Monte Carlo EM
algorithm for generalized linear mixed models with flexible random-
effects distribution’. Biostatistics, 3, 347–360.
Chow, G.C. (1984) ‘Random and changing coefficient models’. In: Z.
Griliches and M.D. Intriligator (eds), Handbook of Econometrics,
 Volume 2. Amsterdam: North-Holland.
Cleveland, W.S. (1979) ‘Robust locally weighted regression and
smoothing scatterplots’. Journal of the American Statistical
Association, 74, 829–836.
Cochran, W.G. (1977) Sampling Techniques, 3d edn. New York: Wiley.
Cohen, J. (1988) Statistical Power Analysis for the Behavioral Sciences.
2nd edn. Hillsdale, NJ: Erlbaum.
Cohen, J. (1992) ‘A power primer’. Psychological Bulletin, 112, 155–
159.
Cohen, M. (1998) ‘Determining sample sizes for surveys with data
analyzed by hierarchical linear models’. Journal of Official Statistics,
14, 267–275.
Cohen, M.P. (2005) ‘Sample size considerations for multilevel surveys’.
International Statistical Review, 73, 279–287.
Coleman, J.S. (1990) Foundations of Social Theory. Cambridge, MA:
Harvard University Press.
Commenges, D. and Jacqmin, H. (1994) ‘The intraclass correlation
coefficient: Distribution-free definition and test’. Biometrics, 50, 517–
526.
Commenges, D., Letenneur, L., Jacqmin, H., Moreau, Th., and Dartigues,
J.-F. (1994) ‘Test of homogeneity of binary data with explanatory
variables’. Biometrics, 50, 613–620.
Congdon, P.D. (2010) Applied Bayesian Hierarchical Methods. Boca
Raton, FL: Chapman & Hall/CRC.
Cook, R.D., and Weisberg, S. (1982) Residuals and Influence in
Regression. New York and London:Chapman & Hall.
Cook, R.D., and Weisberg, S. (1999) Applied Regression Including
Computing and Graphics. New York: Wiley.
Copt, S., and Victoria-Feser, M.-P. (2006) ‘High-breakdown inference for
mixed linear models’. Journal of the American Statistical Association
101, 292–300.
Cornfield, J., and Tukey, J.W. (1956) ‘Average values of mean squares in
factorials’, Annals of Mathematical Statistics, 27, 907–949.
Cox, D.R. (1990) ‘Role of models in statistical analysis’. Statistical
Science, 5, 169–174.
Croon, M.A., and van Veldhoven, M.J. (2007) ‘Predicting group-level
outcome variables from variables measured at the individual level: A
 latent variable multilevel model’. Psychological Methods, 12, 45–57.
Davidian, M., and Giltinan, D.M. (1995) Nonlinear Models for Repeated
Measurement Data.London: Chapman & Hall.
Davis, J.A., Spaeth, J.L., and Huson, C. (1961) ‘A technique for
analyzing the effects of group composition’. American Sociological
Review, 26, 215–225.
de Leeuw, J., and Kreft, I. (1986) ‘Random coefficient models for
multilevel analysis’. Journal of Educational Statistics, 11 (1), 57–85.
de Leeuw, J. and Meijer, E. (eds) (2008a) Handbook of Multilevel
Analysis. New York: Springer.
de Leeuw, J. and Meijer, E. (2008b) ‘Introduction to multilevel analysis’.
In: J. de Leeuw and E. Meijer (eds), Handbook of Multilevel Analysis,
pp. 1–75.New York: Springer.
de Weerth, C. (1998) Emotion-related behavior in infants. PhD thesis,
University of Groningen, The Netherlands.
de Weerth, C., and van Geert, P. (2002) ‘A longitudinal study of basal
cortisol in infants: Intra-individual variability, circadian rhythm and
developmental trends’. Infant Behavior and Development, 25, 375–
398.
Dekkers, H.P.J.M., Bosker, R.J., and Driessen, G.W.J.M. (2000)
‘Complex inequalities of educational opportunities. A large-scale
longitudinal study on the relation between gender, SES, ethnicity, and
school success’. Educational Research and Evaluation, 6, 59–82.
Demidenko, E. (2004) Mixed Models. Theory and Applications.
Hoboken, NJ: Wiley.
Diez-Roux, A. (1998) ‘Bringing context back into epidemiology:
Variables and fallacies in multilevel analysis’. American Journal of
Public Health 88, 216–222.
Diez-Roux, A. (2000) ‘Multilevel analysis in public health research.’
Annual Review of Public Health,21, 171–192.
Diggle, P.J., Heagerty, P.K., Liang, K.Y., and Zeger, S.L. (2002) Analysis
of Longitudinal Data (2nd edn). Oxford: Oxford University Press.
Dogan, M., and Rokkan, S. (eds) (1969) Quantitative Ecological
Analysis in the Social Sciences.Cambridge, MA: MIT Press.
Donner, A. (1986) ‘A review of inference procedures for the intraclass
correlation coefficient in the one-way random effects model’.
International Statistical Review, 54, 67–82.
Donner, A., and Klar, N. (2000) Design and analysis of cluster
randomization trials in health research. London: Arnold.
Donner, A., and Wells, G. (1986) ‘A comparison of confidence interval
methods for the intraclass correlation coefficient’. Biometrics, 42,
401–412.
Doolaard, S. (1999) Schools in change or schools in chains? Enschede:
Twente University Press.
Doran, H., Bates, D., Bliese, P. and Dowling, M. (2007) ‘Estimating the
multilevel Rasch model: With the lme4 package’. Journal of Statistical
Software, 20(2).
Dorman, J.P. (2008) ’The effect of clustering on statistical tests: An
illustration using classroom environment data’, Educational
Psychology, 28, 583–595.
Draper, D. (2008) ‘Bayesian multilevel analysis and MCMC’. In: J. de
Leeuw and E. Meijer (eds),Handbook of Multilevel Analysis, pp. 77–
139.New York: Springer.
Dueck, A., and Lohr, S. (2005) ‘Robust estimation of multivariate
covariance components’.Biometrics, 61, 162–169.
DuMouchel, W.H., and Duncan, G.J. (1983) ‘Using sample survey
weights in multiple regression analysis of stratified samples’. Journal
of the American Statistical Association, 78, 535–543.
Duncan, O.D., Curzort, R.P., and Duncan, R.P. (1961) Statistical
Geography: Problems in Analyzing Areal Data. Glencoe, IL: Free
Press.
Eberly, L.E., and Thackeray, L.M. (2005) ‘On Lange and Ryan’s plotting
techniques for diagnosing non-normality of random effects’. Statistics
and Probability Letters, 75, 77–85.
Efron, B., and Morris, C.N. (1975) ‘Data analysis using Stein’s estimator
and its generalizations’. Journal of the American Statistical
Association, 74, 311–319.
Eicker, F. (1963) ‘Asymptotic normality and consistency of the least
squares estimator for families of linear regressions’. Annals of
Mathematical Statistics, 34, 447–456.
Eisenhart, C. (1947) ‘The assumptions underlying the analysis of
variance’. Biometrics, 3, 1–21.
Enders, C.K. and Tofighi, D. (2007) ‘Centering predictor variables in
cross-sectional multilevel models: a new look at an old issue’.
 Psychological Methods, 12(2), 121–138.
Fielding, A. (2004a) ‘The role of the Hausman test and whether higher
level effects should be treated as random or fixed’. Multilevel
Modelling Newsletter, 16(2), 3–9.
Fielding, A. (2004b) ‘Scaling for residual variance components of
ordered category responses in generalised linear mixed multilevel
models’. Quality and Quantity 38, 425–433.
Firebaugh, G. (1978) ‘A rule for inferring individual-level relationships
from aggregate data’.American Sociological Review, 43, 557–572.
Fisher, R.A. (1924) ‘On a distribution yielding the error functions of
several well-known statistics’.Proceedings of the International
Mathematical Congress, Toronto, pp. 805–813.
Fisher, R.A. (1958) Statistical Methods for Research Workers, 13th edn.
London: Hafner Press.
Fisher, R.A., and Yates, F. (1963) Statistical Tables for Biological,
Agricultural, and Medical Research. Edinburgh: Oliver & Boyd.
Fitzmaurice, G.M., Laird, N.M., and Ware, J.H. (2004) Applied
Longitudinal Analysis. Hoboken, NJ: Wiley.
Fong, Y., Rue, H., and Wakefield, J. (2010) ‘Bayesian inference for
generalized linear mixed models’.Biostatistics 11, 397–412.
Fox, J. (2008) Applied Regression Analysis and Generalized Linear
Models, 2nd edn. Thousand Oaks, CA: Sage.
Freedman, D.A. (2006) ‘On the so-called “Huber Sandwich estimator”
and “robust standard errors”’.The American Statistician, 60, 299–302.
Fuller, W.A. (2009) Sampling Statistics. Hoboken, NJ: Wiley.
Gardiner, J.C., Luo, Z.H., and Roman, L.A. (2009) ‘Fixed effects,
random effects and GEE: What are the differences?’. Statistics in
Medicine, 28, 221–239.
Gelman, A. (2007) ‘Struggles with survey weighting and regression
modeling’. Statistical Science,22, 153–164.
Gelman, A., Carlin, J.B., Stern, H.S., and Rubin, D.B. (2004) Bayesian
Data Analysis, 2nd edn. Boca Raton, FL: Chapman & Hall/CRC.
Gelman, A., and Hill, J. (2007) Data Analysis Using Regression and
Multilevel/Hierarchical Models.Cambridge: Cambridge University
Press.
Gibbons, R.D., and Bock, R.D. (1987) ‘Trends in correlated proportions’.
Psychometrika, 52,113–124.
Gibbons, R.D., and Hedeker, D. (1994) ‘Application of random-effects
probit regression models’.Journal of Consulting and Clinical
Psychology, 62, 285–296.
Gibbons, R.D., and Hedeker, D. (1997) ‘Random effects probit and
logistic regression models for three-level data’. Biometrics, 53, 1527–
1537.
Gilks, W.R., Richardson, S., and Spiegelhalter, D.J. (1996) Markov
Chain Monte Carlo in Practice.London: Chapman & Hall.
Glass, G.V., and Stanley, J.C. (1970) Statistical Methods in Education
and Psychology. Englewood Cliffs, NJ: Prentice Hall.
Glidden, D.V., and Vittinghoff, E. (2004) ‘Modelling clustered survival
data from multicentre clinical trials’. Statistics in Medicine, 23, 369–
388.
Goldstein, H. (1986) ‘Multilevel mixed linear model analysis using
iterative generalized least squares’. Biometrika, 73, 43–56.
Goldstein, H. (1991) ‘Nonlinear multilevel models with an application to
discrete response data’.Biometrika, 78, 45–51.
Goldstein, H. (2011) Multilevel Statistical Models. 4th edn. London:
Edward Arnold.
Goldstein, H., Carpenter, J. Kenward, M.G., and Levin, K.A. (2009)
‘Multilevel models with multivariate mixed response types’. Statistical
Modelling, 9, 173–197.
Goldstein, H., and Healy, M.J.R. (1995) ‘The graphical presentation of a
collection of means’.Journal of the Royal Statistical Society, Series A,
158, 175–177.
Goldstein, H., and Rasbash, J. (1996) ‘Improved approximations for
multilevel models with binary responses’. Journal of the Royal
Statistical Society, Series A, 159, 505–513.
Goldstein, H., Rasbash, J., Plewis, I., Draper, D., Browne, W., Yang, M.,
Woodhouse, G., and Healy, M. (1998) A user’s guide to MLwiN.
London: Multilevel Models Project, Institute of Education, University
of London.
Gouriéroux, C., and Montfort, A. (1996) Simulation-Based Econometric
Methods. Oxford: Oxford University Press.
Graham, J.W. (2009) ‘Missing data analysis: Making it work in the real
world’. Annual Review of Psychology, 60, 549–576.
Graham, J.W., Olchowski, A.E., and Gilreath, T.D. (2007) ‘How many
imputations are really needed? Some practical clarifications of
multiple imputation theory’. Prevention Science, 8, 206–213.
Greene, W. (2008) Econometric Analysis, 6th edn. Upper Saddle River,
NJ: Prentice Hall.
Grilli, L., and Pratesi, M. (2004) ‘Weighted estimation in multi-level
ordinal and binary models in the presence of informative sampling
designs’. Survey Methodology, 30, 93–103.
Guldemond, H. (1994) Van de kikker en de vijver [About the frog and the
pond]. PhD thesis, University of Amsterdam.
Hagle, T.M., and Mitchell II, G.E. (1992) ‘Goodness-of-fit measures for
probit and logit’. American Journal of Political Science, 36, 762–784.
Haldane, J.B.S. (1940) ‘The mean and variance of χ2, when used as a test
of homogeneity, when expectations are small’. Biometrika, 31, 346–
355.
Hamaker, E.L., and Klugkist, I. (2011) ‘Bayesian estimation of multilevel
models’. In: J.J. Hox and J.K. Roberts (eds), Handbook of Advanced
Multilevel Analysis, pp. 137–161.New York: Routledge.
Hand, D., and Crowder, M. (1996) Practical Longitudinal Data Analysis.
London: Chapman & Hall.
Harker, R., and Tymms, P. (2004) ‘The effects of student composition on
school outcomes’. School Effectiveness and School Improvement, 15,
177–199.
Hauser, R.M. (1970) ‘Context and consex: A cautionary tale.’ American
Journal of Sociology, 75, 645–654.
Hauser, R.M. (1974) ‘Contextual analysis revisited’. Sociological
Methods and Research, 2, 365–375.
Hausman, J.A. (1978) ‘Specification tests in econometrics’.
Econometrica, 46, 1251–1271.
Hausman, J.A., and Taylor, W.E. (1981) ‘Panel data and unobservable
individual effects’. Econometrica, 49, 1377–1398.
Hays, W.L. (1988) Statistics. 4th edn. New York: Holt, Rinehart and
Winston.
Heck, R.H., Thomas, S.L., and Tabata, L.N. (2010) Multilevel and
Longitudinal Modeling with IBM SPSS. New York: Routledge.
Hedeker, D. (2003) ‘A mixed-effects multinomial logistic regression
model’. Statistics in Medicine, 22, 1433–1446.
Hedeker, D. (2008) ‘Multilevel models for ordinal and nominal
variables’. In: J. de Leeuw and E.Meijer (eds), Handbook of Multilevel
Analysis, pp. 237–274.New York: Springer.
Hedeker, D., and Gibbons, R.D. (1994) ‘A random effects ordinal
regression model for multilevel analysis’. Biometrics, 50, 933–944.
Hedeker, D., and Gibbons, R.D. (1996a) ‘MIXOR: A computer program
for mixed-effects ordinal regression analysis’. Computer Methods and
Programs in Biomedicine, 49, 157–176.
Hedeker, D., and Gibbons, R.D. (1996b) ‘MIXREG: A computer
program for mixed-effects regression analysis with autocorrelated
errors’. Computer Methods and Programs in Biomedicine, 49, 229–
252.
Hedeker, D., Gibbons, R.D., and Waternaux, C. (1999) ‘Sample size
estimation for longitudinal designs with attrition: Comparing time-
related contrasts between two groups.’ Journal of Educational and
Behavioral Statistics, 24, 70–93.
Hedeker, D., and Gibbons, R.D. (2006) Longitudinal Data Analysis.
Hoboken, NJ: Wiley.
Hedges, L.V. (1992) ‘Meta-analysis’. Journal of Educational Statistics,
17, 279–296.
Hedges, L.V., and Hedberg, E.C. (2007) ‘Intraclass correlation values for
planning group-randomized trials in education’. Educational
Evaluation and Policy Analysis, 29, 60–87.
Hedges, L.V., and Olkin, I. (1985) Statistical Methods for Meta-analysis.
New York: Academic Press.
Hilden-Minton, J.A. (1995) Multilevel diagnostics for mixed and
hierarchical linear models. PhD dissertation, Department of
Mathematics, University of California, Los Angeles.
Hill, P.W., and Goldstein, H. (1998) ‘Multilevel modeling of educational
data with cross-classification and missing identification for units’.
Journal of Educational and Behavioral Statistics, 23, 117–128.
Hodges, J.S. (1998) ‘Some algebra and geometry for hierarchical linear
models, applied to diagnostics’. Journal of the Royal Statistical
Society, Series B, 60, 497–536.
Hogan, J., Roy, J. and Korkontzelou, C. (2004) ‘Handling drop-out in
longitudinal studies’. Statistics in Medicine, 23, 1455–1497.
Hosmer, D.W., and Lemeshow, S. (2000) Applied Logistic Regression,
2nd edn. New York: Wiley.
Hox, J.J. (2010) Multilevel Analysis: Techniques and Applications, 2nd
edn. Mahwah, NJ: Erlbaum.
Hsiao, C. (1995) ‘Panel analysis for metric data’. In: G. Arminger, C.C.
Clogg, and M.E. Sobel (eds),Handbook of Statistical Modeling for the
Social and Behavioral Sciences, pp. 361–400. New York: Plenum
Press.
Huber, P.J. (1967) ‘The behavior of maximum likelihood estimates under
non-standard conditions’. In: L. LeCam and J. Neyman (eds),
Proceedings of the Fifth Berkeley Symposium on Mathematical
Statistics and Probability, pp. 221–233.Berkeley: University of
California Press.
Hüttner, H.J.M. (1981) ‘Contextuele analyse’ [Contextual analysis]. In:
M. Albinski (ed.), Onderzoekstypen in de Sociologie, pp. 262–
288.Assen: Van Gorcum.
HÜttner, H.J.M., and van den Eeden, P. (1995) The Multilevel Design. A
Guide with an Annotated Bibliography, 1980–1993. Westport, CT:
Greenwood Press.
Ibrahim, J.G., and Molenberghs, G. (2009) ‘Missing data methods in
longitudinal studies: a review’.Test, 18, 1–43.
Ihaka, R., and Gentleman, R. (1996) ’R: A language for data analysis and
graphics’. Journal of Computational and Graphical Statistics, 5, 299–
314.
Jackman, S. (2009) Bayesian Analysis for the Social Sciences.
Chichester: Wiley.
Jang, M.J., Lee, Y., Lawson, A.B., and Browne, W.J. (2007) ’A
comparison of the hierarchical likelihood and Bayesian approaches to
spatial epidemiological modelling’. Environmetrics, 18, 809–821.
Jank, W. (2006) ‘Implementing and diagnosing the stochastic
approximation EM algorithm’. Journal of Computational and
Graphical Statistics, 15, 1–27.
Johnson, N.L., Kotz, S., and Balakrishnan, N. (1995) Continuous
Univariate Distributions, Volume 2, 2nd edn. New York: Wiley.
Kaplan, D., and Elliott, P.R. (1997) ‘A didactic example of multilevel
structural equation modeling applicable to the study of organizations’.
Structural Equation Modeling, 4, 1–24.
Kasim, R.M., and Raudenbush, S.W. (1998) ‘Application of Gibbs
sampling to nested variance components models with heterogeneous
within-group variance’. Journal of Educational and Behavioral
Statistics, 23, 93–116.
Kelley, T.L. (1927) The Interpretation of Educational Measurements.
New York: World Books.
King, G. (1997) A Solution to the Ecological Inference Problem.
Reconstructing Individual Behavior from Aggregate Data. Princeton,
NJ: Princeton University Press.
Kish, L. (1992) ‘Weighting for unequal Pi’. Journal of Official Statistics,
8, 183–200.
Knuver, A.W.M. and Brandsma, H.P. (1993) ‘Cognitive and affective
outcomes in school effectiveness research’. School Effectiveness and
School Improvement, 4, 189–204.
Korn, E.L., and Graubard, B.I. (2003) ‘Estimating variance components
by using survey data’.Journal of the Royal Statistical Society, Series B,
65, 175–190.
Kovačević, M., Rong, H., and You, Y. (2006) ‘Bootstrapping for variance
estimation in multi-level models fitted to survey data’. ASA
Proceedings of the Survey Research Methods Section, pp. 3260–3269.
http://www.amstat.org/sections/srms/proceedings/y2006/Files/JSM200
6-000286.pdf
Kreft, I.G.G., and de Leeuw, J. (1998) Introducing Multilevel Modeling.
London: Sage Publications.
Kreft, I.G.G., de Leeuw, J., and Aiken, L. (1995) ‘The effect of different
forms of centering in hierarchical linear models’. Multivariate
Behavioral Research, 30, 1–22.
Kuk, A.Y.C. (1995) ‘Asymptotically unbiased estimation in generalized
linear models with random effects’. Journal of the Royal Statistical
Society, Series B, 57, 395–407.
Kuyper, H., and Van der Werf, M.P.C. (2003) VOCL’99-1: Technisch
rapport. Groningen: GION.
LaHuis, D.M., and Ferguson, M.W. (2009) ‘The accuracy of significance
tests for slope variance components in multilevel random coefficient
models’. Organizational Research Methods, 12, 418–435.
Laird, N. (1978) ‘Nonparametric maximum likelihood estimation of a
mixture distribution’. Journal of the American Statistical Association,
 73, 805–811.
Laird, N.M., and Ware, J.H. (1982) ‘Random-effects models for
longitudinal data’. Biometrics, 38, 963–974.
Lange, N., and Ryan, L. (1989) ‘Assessing normality in random effects
models’. Annals of Statistics, 17, 624–642.
Langford, I.H., and Lewis, T. (1998) ‘Outliers in multilevel data’.
Journal of the Royal Statistical Society, Series A, 161, 121–160.
Lazarsfeld, P.F., and Menzel, H. (1971) ‘On the relation between
individual and collective properties’. In: A. Etzioni (ed.), A
Sociological Reader on Complex Organizations, p. 499–516. New
York: Holt, Rinehart & Winston.
Leckie, G. and Charlton, C. (2011) runmlwin: Stata module for fitting
multilevel models in the MLwiN software package. Bristol: Centre for
Multilevel Modelling, University of Bristol.
Lehmann, E.L., and Romano, J.R. (2005) Testing Statistical Hypotheses,
3rd edn. New York: Springer.
Lesaffre, E., and Verbeke, G. (1998) ‘Local influence in linear mixed
models’. Biometrics, 54, 570–582.
Liang, K.Y., and Zeger, S.L. (1986) ‘Longitudinal data analysis using
generalized linear models’. Biometrika, 73, 13–22.
Littell, R.C., Milliken, G.A., Stroup, W.W., Wolfinger, R.D., and
Schabenberger, B. (2006) SAS®System for Mixed Models, 2nd edn.
Cary, NC: SAS Institute.
Little, R.J. (2004) ‘To model or not to model? Competing modes of
inference for finite population sampling’ Journal of the American
Statistical Association, 99, 546–556.
Little, R.J.A., and Rubin, D.B. (2002) Statistical Analysis with Missing
Data, 2nd edn. Hoboken, NJ: Wiley.
Lohr, S.L. (2010) Sampling: Design and Analysis, 2nd edn. Boston:
Brooks/Cole.
Long, J.S. (1997) Regression Models for Categorical and Limited
Dependent Variables. Thousand Oaks, CA: Sage Publications.
Longford, N.T. (1987) ‘A fast scoring algorithm for maximum likelihood
estimation in unbalanced mixed models with nested random effects’.
Biometrika, 74(4), 812–827.
Longford, N.T. (1993) Random Coefficient Models. New York: Oxford
University Press.
Longford, N.T. (1994) ‘Logistic regression with random coefficients’.
Computational Statistics and Data Analysis, 17, 1–15.
Longford, N.L. (1995) ‘Random coefficient models’. In: G. Arminger,
C.C. Clogg, and M.E. Sobel (eds), Handbook of Statistical Modeling
for the Social and Behavioral Sciences, pp. 519–577. New York:
Plenum Press.
Lord, F.M., and Novick, M.R. (1968) Statistical Theory of Mental Test
Scores. Reading, MA: Addison-Wesley.
Lukočienė, O., and Vermunt, J.K. (2007) ‘A Comparison of multilevel
logistic regression models with parametric and nonparametric random
intercepts’. Submitted for publication.
Lukočienė, O., and Vermunt, J.K. (2009) ‘Logistic regression analysis
with multidimensional random effects: A comparison of three
approaches’. Submitted for publication.
Lüdtke, O., Marsh, H.W., Robitzsch, A., Trautwein, U., Asparouhov, T.,
and Muthén, B. (2008) ‘The multilevel latent covariate model: A new,
more reliable approach to group-level effects in contextual studies’.
Psychological Methods, 13, 203–229.
Lunn, D.J., Thomas, A., Best, N., and Spiegelhalter, D. (2000)
’WinBUGS – a Bayesian modelling framework: Concepts, structure,
and extensibility’. Statistics and Computing, 10, 325–337.
Maas, C.J.M., and Hox, J.J. (2004) ‘The influence of violations of
assumptions on multilevel parameter estimates and their standard
errors’. Computational Statistics & Data Analysis, 46, 427–440.
Maas, C.J.M., and Hox, J.J. (2005) ‘Sufficient sample sizes for multilevel
modeling’. Methodology, 1, 86-92.
Maas, C.J.M., and Snijders, T.A.B. (2003) ‘The multilevel approach to
repeated measures for complete and incomplete data’. Quality and
Quantity, 37, 71–89.
MacKinnon, J.G., and White, H. (1985) ‘Some heteroskedasticity-
consistent covariance matrix estimators with improved finite sample
properties’. Journal of Econometrics, 29, 305–325.
Maddala, G.S. (1971) ‘The use of variance component models in pooling
cross section and time series data’. Econometrica, 39, 341–358.
Maindonald, J., and Braun, J. (2007) Data Analysis and Graphics Using
R, 2nd edn. Cambridge: Cambridge University Press.
Mancl, L.A., and DeRouen, T.A. (2001) ‘A covariance estimator for GEE
with improved small-sample properties’. Biometrics, 57, 126–134.
Manor, O. and Zucker, D.M. (2004) ‘Small sample inference for the fixed
effects in the mixed linear model’. Computational Statistics and Data
Analysis, 46, 801–817.
Mason, A., Richardson, S., Plewis, I. and Best, N. (2010) Strategy for
modelling non-random missing data mechanisms in observational
studies using Bayesian methods. The BIAS project, London.
Mason, W.M., Wong, G.M., and Entwisle, B. (1983) ‘Contextual analysis
through the multilevel linear model’. In: S. Leinhardt (ed.),
Sociological Methodology – 1983–1984, pp. 72–103. San Francisco:
Jossey-Bass.
Maxwell, S.E., and Delaney, H.D. (2004) Designing Experiments and
Analyzing Data, 2nd edn. Mahwah, NJ: Lawrence Erlbaum
Associated.
McCullagh, P., and Nelder, J.A. (1989) Generalized Linear Models. 2nd
edn. London: Chapman & Hall.
McCulloch, C.E. (1997) ‘Maximum likelihood algorithms for
generalized linear mixed models’.Journal of the American Statistical
Association, 92, 162–170.
McCulloch, C.E., and Searle, S.R. (2001) Generalized, Linear, and
Mixed Models. New York: Wiley.
McKelvey, R.D., and Zavoina, W. (1975) ‘A statistical model for the
analysis of ordinal level dependent variables’. Journal of
Mathematical Sociology, 4, 103–120.
Mealli, F., and Rampichini, C. (1999) ‘Estimating binary multilevel
models through indirect inference’. Computational Statistics and Data
Analysis, 29, 313–324.
Miller, J.J. (1977) ‘Asymptotic properties of maximum likelihood
estimates in the mixed model of the analysis of variance’. Annals of
Statistics, 5, 746–762.
Mills, M. (2011) Introducing Survival and Event History Analysis.
Thousand Oaks, CA, and London: Sage.
Moerbeek, M. (2004) ‘The consequence of ignoring a level of nesting in
multilevel analysis’. Multivariate Behavioral Research, 39, 129–149.
Moerbeek, M. (2005) ‘Randomization of clusters versus randomization
of persons within clusters: which is preferable?’. The American
 Statistician 59, 173–179.
Moerbeek, M.N., and Teerenstra, S. (2011) ’Optimal design in multilevel
experiments’. In: J.J. Hox and J.K. Roberts (eds), Handbook of
Advanced Multilevel Analysis, pp. 257–281. New York: Routledge.
Moerbeek, M., van Breukelen, G.J.P., and Berger, M.P.F. (2003) A
comparison between traditional methods and multilevel regression for
the analysis of multicenter intervention studies. Journal of Clinical
Epidemiology, 56, 341–350.
Mok, M. (1995) ‘Sample size requirements for 2-level designs in
educational research’. Multilevel Modeling Newsletter, 7 (2), 11–15.
Molenberghs, G., and Kenward, M.G. (2007) Missing Data in Clinical
Studies. Chichester: Wiley.
Molenberghs, G., and Verbeke, G. (2006) Models for Discrete
Longitudinal Data. New York: Springer.
Molenberghs, G., and Verbeke, G. (2007) ‘Likelihood ratio, score, and
Wald tests in a constrained parameter space’. The American
Statistician 61, 22–27.
Mosteller, F., and Tukey, J.W. (1977) Data Analysis and Regression.
Reading, MA: Addison-Wesley.
Muthén, B.O. (1994) ‘Multilevel covariance structure analysis’.
Sociological Methods & Research, 22, 376–398.
Muthén, B., and Asparouhov, T. (2009) ‘Multilevel regression mixture
analysis’. Journal of the Royal Statistical Society, Series A, 172, 639–
657.
MuthÉn, L.K. and MuthÉn, B.O. (2010) Mplus User’s Guide, 6th edn.
Los Angeles: MuthÉn and MuthÉn.
Neuhaus, J.M. (1992) ‘Statistical methods for longitudinal and cluster
designs with binary responses’. Statistical Methods in Medical
Research, 1, 249–273.
Neuhaus, J.M., and Kalbfleisch, J.D. (1998) ‘Between- and within-cluster
covariate effects in the analysis of clustered data’. Biometrics, 54,
638–645.
Neuhaus, J.M., Kalbfleisch, J.D., and Hauck, W.W. (1991) ‘A
comparison of cluster-specific and population-averaged approaches for
analyzing correlated binary data’. International Statistical Review, 59,
25–35.
Ng, E.S.W., Carpenter, J.R., Goldstein, H., and Rasbash, J. (2006)
‘Estimation in generalised linear mixed models with binary outcomes
by simulated maximum likelihood’. Statistical Modelling, 6, 23–42.
Nordberg, L. (1989) ‘Generalized linear modeling of sample survey
data’. Journal of Official Statistics, 5, 223–239.
O’Campo, P. (2003) ‘Invited commentary: Advancing theory and
methods for multilevel models of residential neighborhoods and
health’. American Journal of Epidemiology, 157, 9–13.
OECD (2009) PISA 2006 Technical Report.
http://www.oecd.org/dataoecd/0/47/42025182.pdf
OECD (2010) PISA 2009 Results: What Makes a School Successful?
Resources, Policies and Practices (Volume IV).
http://dx.doi.org/10.1787/9789264091559-en
Opdenakker, M.C., and Van Damme, J. (1997) ‘Centreren in multi-level
analyse: implicaties van twee centreringsmethoden voor het bestuderen
van schooleffectiviteit’. Tijdschrift voor Onderwijsresearch, 22, 264–
290.
Pan, H., and Goldstein, H. (1998) ‘Multi-level repeated measures growth
modelling using extended spline functions’. Statistics in Medicine, 17,
2755–2770.
Pan, W., and Wall, M. (2002) ‘Small-sample adjustments in using the
sandwich variance estimator in generalized estimating equation’.
Statistics in Medicine, 21, 1429–1441.
Pedhazur, E.J. (1982) Multiple Regression in Behavioral Research. 2nd
edn. New York: Holt, Rinehart and Winston.
Peter, S., and Drobnič, S. (2010) ‘Women and their memberships: A
multilevel cross-national study on gender differentials in associational
involvement’. Submitted for publication.
Pfeffermann, D. (1993) ‘The role of sampling weights when modeling
survey data’. International Statistical Review, 61, 317–337.
Pfeffermann, D., Moura, F.A.D.S., and Silva, P.L.D.N. (2006) ‘Multi-
level modelling under informative sampling’. Biometrika, 93, 943–
959.
Pfeffermann, D., Skinner, C. J., Holmes, D. J., Goldstein, H., and
Rasbash, J. (1998) ‘Weighting for unequal selection probabilities in
multi-level models’. Journal of the Royal Statistical Society, Series B,
60, 23–56.
Pinheiro, J.C., and Bates, D.M. (1995) ‘Approximations to the log-
likelihood function in nonlinear mixed-effects models’. Journal of
Computational and Graphical Statistics, 4, 12–35.
Pinheiro, J.C., and Bates, D.M. (2000) Mixed-Effects Models in S and S-
Plus. New York: Springer.
Pinheiro, J.C., Liu, C., and Wu, Y.N. (2001) ‘Efficient algorithms for
robust estimation in linear mixed-effects models using the multivariate
t-distribution. Journal of Computational and Graphical Statistics, 10,
249–276.
Potthoff, R. F., Woodbury, M. A., and Manton, K. G. (1992) “‘Equivalent
sample size” and “equivalent degrees of freedom” refinements for
inference using survey weights under superpopulation models’.
Journal of the American Statistical Association, 87, 383–396.
Preacher, K.J., Curran, P.J., and Bauer, D.J. (2006) ‘Computational tools
for probing interactions in multiple linear regression, multilevel
modeling, and latent curve analysis’. Journal of Educational and
Behavioral Statistics, 31, 437–448.
Pregibon, D. (1981) ‘Logistic regression diagnostics’. Annals of
Statistics, 9, 705–724.
Rabe-Hesketh, S., and Skrondal, A. (2006) ‘Multilevel modelling of
complex survey data’. Journal of the Royal Statistical Society, Series
A, 169, 805–827.
Rabe-Hesketh, S., and Skrondal, A. (2008) Multilevel and Longitudinal
Modeling Using Stata, 2nd edn. College Station, TX: Stata Press.
Rabe-Hesketh, S., Skrondal, A., and Pickles, A. (2004) ‘Generalized
multilevel structural equation modelling’. Psychometrika, 69, 167–
190.
Rabe-Hesketh, S., Skrondal, A., and Pickles, A. (2005) ‘Maximum
likelihood estimation of limited and discrete dependent variable
models with nested random effects.’ Journal of Econometrics, 128,
301–323.
Raghunathan, T.E., Lepkowski, J.M., Van Hoewyk, J., and Solenberger,
P. (2001) ‘A multivariate technique for multiply imputing missing
values using a sequence of regression models’. Survey Methodology,
27.1, 85–95.
Rasbash, J. and Browne, W.J. (2001) ‘Modeling non-hierarchical
structures’. In: A.H. Leyland and H. Goldstein (eds), Multilevel
 Modeling of Health Statistics, pp. 93–105. Chichester: Wiley.
Rasbash, J. and Browne, W.J. (2008) ‘Non-hierarchical multilevel
models’. In: J. de Leeuw and E. Meijer (eds), Handbook of Multilevel
Analysis, pp. 301–334. New York: Springer.
Rasbash, J., and Goldstein, H. (1994) ‘Efficient analysis of mixed
hierarchical and cross-classified random structures using a multilevel
model.’ Journal of Educational and Behavioral Statistics, 19, 337–
350.
Rasbash, J., Steele, V., Browne, W.J., and Goldstein, H. (2009) A User’s
Guide to MLwiN. Bristol: Centre for Multilevel Modelling, University
of Bristol.
Rasbash, J., and Woodhouse, G. (1995) MLn: Command Reference.
London: Multilevel Models Project, Institute of Education, University
of London.
Raudenbush, S.W. (1993) ‘A crossed random effects model for
unbalanced data with applications in cross sectional and longitudinal
research’. Journal of Educational Statistics, 18, 321–349.
Raudenbush, S.W. (1995) ‘Hierarchical linear models to study the effects
of social context on development’. In: Gottman, J.M. (ed.), The
Analysis of Change, pp. 165–201. Mahwah, NJ: Lawrence Erlbaum
Ass.
Raudenbush, S.W. (1997) ‘Statistical analysis and optimal design for
cluster randomized trials’. Psychological Methods, 2, 173–185.
Raudenbush, S.W., and Bryk, A.S. (1986) ‘A hierarchical model for
studying school effects’. Sociology of Education, 59, 1–17.
Raudenbush, S.W., and Bryk, A.S. (1987) ‘Examining correlates of
diversity’. Journal of Educational Statistics, 12, 241–269.
Raudenbush, S.W., and Bryk, A.S. (2002) Hierarchical Linear Models:
Applications and Data Analysis Methods, 2nd edn. Thousand Oaks,
CA: Sage.
Raudenbush, S.W., and Liu, X. (2000) ‘Statistical power and optimal
design for multisite randomized trials’. Psychological Methods, 5,
199–213.
Raudenbush, S.W., and Liu, X. (2001) ‘Effects of study duration,
frequency of observation, and sample size on power in studies of
group differences in polynomial change’. Psychological Methods, 6,
387–401.
Raudenbush, S.W., Martinez, A., and Spybrook, J. (2007) ‘Strategies for
improving precision in group-randomized experiments’. Educational
Evaluation and Policy Analysis, 1, 5–29.
Raudenbush, S.W., and Sampson, R. (1999a) ‘Assessing direct and
indirect effects in multilevel designs with latent variables’.
Sociological Methods and Research, 28, 123–153.
Raudenbush, S.W., and Sampson, R. (1999b) “‘Ecometrics”: Toward a
science of assessing ecological settings, with application to the
systematic social observation of neighborhoods’. Sociological
Methodology, 29, 1–41.
Raudenbush, S.W., Yang, M.-L., and Yosef, M. (2000) ‘Maximum
likelihood for generalized linear models with nested random effects via
high-order, multivariate Laplace approximation’. Journal of
Computational and Graphical Statistics, 9, 141–157.
Rekers-Mombarg, L.T.M., Cole, L.T., Massa, G.G., and Wit, J.M. (1997)
‘Longitudinal analysis of growth in children with idiopathic short
stature.’ Annals of Human Biology, 24, 569–583.
Richardson, A.M. (1997) ‘Bounded influence estimation in the mixed
linear model’. Journal of the American Statistical Association, 92,
154–161.
Robinson, W.S. (1950) ‘Ecological correlations and the behavior of
individuals’. American Sociological Review, 15, 351–357.
Rodríguez, G. (2008) ‘Multilevel generalized linear models’. In: J. de
Leeuw and E. Meijer (eds), Handbook of Multilevel Analysis, pp. 335–
376.New York: Springer.
Rodríguez, G. and Goldman, N. (1995) ‘An assessment of estimation
procedures for multilevel models with binary responses’. Journal of
the Royal Statistical Society, Series A, 158, 73–89.
Rodríguez, G. and Goldman, N. (2001) ‘Improved estimation procedures
for multilevel models with binary response: A case study’. Journal of
the Royal Statistical Society, Series A, 164, 339–355.
Rosenthal, R. (1991) Meta-analytic Procedures for Social Research, rev.
edn. Newbury Park, CA: Sage Publications.
Rotondi, M.A. and Donner, A. (2009) ‘Sample size estimation in cluster
randomized educational trials: An empirical Bayes approach. Journal
of Educational and Behavioral Statistics, 34, 229–237.
Roy, J., and Lin, X. (2005) ‘Missing covariates in longitudinal data with
informative dropout: Bias analysis and inference’. Biometrics, 61,
837–846.
Rubin, D.B. (1976) ‘Inference and missing data’. Biometrika, 63, 581–
592.
Rubin, D.B. (1987) Multiple Imputation for Nonresponse in Surveys.
New York: Wiley.
Rue, H., Martino, S., and Chopin, N. (2009) ‘Approximate Bayesian
inference for latent Gaussian models using integrated nested Laplace
approximations’ (with discussion). Journal of the Royal Statistical
Society, Series B, 71, 319–392.
Ruiter, S., and van Tubergen, F. (2009) ‘Religious attendance in cross-
national perspective: A multilevel analysis of 60 countries’. American
Journal of Sociology, 115, 863–895.
Ryan, T.P. (1997) Modern Regression Methods. New York: Wiley.
Sampson, R.J., Morenoff, J.D., and Gannon-Rowley, T. (2002)
‘Assessing “neighborhood effects”:social process and new directions
in research’. Annual Review of Sociology, 28, 443–478.
Schafer, J.L. (1997) Analysis of Incomplete Multivariate Data. New
York: Chapman & Hall.
Schafer, J.L., and Graham, J.W. (2002) ‘Missing data: Our view of the
state of the art’. Psychological Methods, 7, 147–177.
Schafer, J.L. and Yucel, R.M. (2002) ‘Computational strategies for
multivariate linear mixed-effects models with missing values’. Journal
of Computational and Graphical Statistics, 11, 437–457.
Searle, S.R. (1956) ‘Matrix methods in components of variance and
covariance analysis’. Annals of Mathematical Statistics, 29, 167–178.
Searle, S.R., Casella, G., and McCulloch, C.E. (1992) Variance
Components. New York: Wiley.
Seber, G.A.F., and Wild, C.J. (1989) Nonlinear Regression. New York:
Wiley.
Self, G.S., and Liang, K.-Y. (1987) ‘Asymptotic properties of maximum
likelihood estimators and likelihood ratio tests under nonstandard
conditions’. Journal of the American Statistical Association, 82, 605–
610.
Seltzer, M.H. (1993) ‘Sensitivity analysis for fixed effects in the
hierarchical model: A Gibbs sampling approach’. Journal of
 Educational Statistics, 18, 207–235.
Seltzer, M., and Choi, K. (2002) ‘Model checking and sensitivity analysis
for multilevel models’. In: N. Duan and S. Reise (eds), Multilevel
Modeling: Methodological Advances, Issues, and Applications.
Hillsdale, NJ: Lawrence Erlbaum.
Seltzer, M.H., Wong, W.H.H., and Bryk A.S. (1996) ‘Bayesian analysis
in applications of hierarchical models: Issues and methods’. Journal of
Educational Statistics, 21, 131–167.
Shadish, W.R., Cook, T.D., and Campbell, D.T. (2002) Experimental and
Quasi-experimental Designs for Generalized Causal Inference.
Boston: Houghton Mifflin.
Shavelson, R.J., and Webb, N.M. (1991) Generalizability Theory: A
Primer. Newbury Park, CA: Sage Publications.
Siddiqui, O., Hedeker, D., Flay, B.R., and Hu, F. (1996) ‘Intraclass
correlation estimates in a school-based smoking prevention study’.
American Journal of Epidemiology, 144, 425–433.
Singer, J.D. (1998a) ‘Using SAS PROC MIXED to fit multilevel models,
hierarchical models, and individual growth models’. Journal of
Educational and Behavioral Statistics, 23, 323–355.
Singer, J.D. (1998b) ‘Fitting multilevel models using SAS PROC
MIXED.’ Multilevel Modeling Newsletter, 10(2), 5–9.
Singer, J.D. (2002) ‘Fitting individual growth models using SAS PROC
MIXED’. In: D.S. Moskowitz and S.L. Hershberger (eds), Modeling
intraindividual variability with repeated measures data, pp. 135–170.
Hillsdale, NJ: Lawrence Erlbaum.
Singer, J.D., and Willett, J.B. (2003) Applied Longitudinal Data Analysis.
New York: Oxford University Press.
Skinner, C.J. (1989) ‘Domain means, regression and multivariate
analysis’. In: C.J. Skinner, D. Holt, and T.M.F. Smith (eds), Analysis of
Complex Surveys, pp. 59–87. New York: Wiley.
Skrondal, A., and Rabe-Hesketh, S. (2004) Generalized Latent Variable
Modeling: Multilevel, Longitudinal, and Structural Equation Models.
Boca Raton, FL: Chapman & Hall/CRC.
Smith, T.M.F. (1994) ‘Sample surveys 1975-1990: An age of
reconciliation?’ (with discussion). International Statistical Review, 62,
5–34.
Snijders, J.Th., and Welten, V.J. (1968) The I.S.I. School Achievement
and Intelligence Tests, Form I and II. Groningen: Wolters-Noordhoff.
Snijders, T.A.B. (1996) ‘Analysis of longitudinal data using the
hierarchical linear model’. Quality & Quantity, 30, 405–426.
Snijders, T.A.B. (2001) ’Sampling’. In: A. Leyland and H. Goldstein
(eds), Multilevel Modelling of Health Statistics, pp. 159-174.
Chichester: Wiley.
Snijders, T.A.B. (2005) ’Power and sample size in multilevel linear
models’. In: B.S. Everitt and D.C. Howell (eds), Encyclopedia of
Statistics in Behavioral Science. Volume 3, pp. 1570–1573. Chichester:
Wiley.
Snijders, T.A.B., and Berkhof, J. (2008) ‘Diagnostic checks for
multilevel models’. In: J. de Leeuw and E. Meijer (eds), Handbook of
Multilevel Analysis, pp. 141–175. New York: Springer.
Snijders, T.A.B., and Bosker, R.J. (1993) ‘Standard errors and sample
sizes for two-level research’. Journal of Educational Statistics, 18,
237–259.
Snijders, T.A.B., and Bosker, R.J. (1994) ‘Modeled variance in two-level
models’. Sociological Methods & Research, 22, 342–363.
Spybrook, J., Raudenbush, S.W., Congdon, R. and Martinez, A. (2009)
Optimal Design for Longitudinal and Multilevel Research:
Documentation for the “Optimal Design” Software, Version 2.0. Ann
Arbor, MI: University of Michigan.
Stapleton, L. (2002) ‘The incorporation of sample weights into multilevel
structural equation models’. Structural Equation Modeling, 9, 475–
502.
StataCorp (2009) Stata Statistical Software: Release 11. College Station,
TX: StataCorp.
Staudenmayer, J., Lake, E.E., and Wand, M.P. (2009) ‘Robustness for
general design mixed models using the t-distribution’. Statistical
Modelling, 9, 235–255.
Steele, F. (2008) ‘Multilevel models for longitudinal data’. Journal of the
Royal Statistical Society, Series A, 171, 5–19.
Sterba, S.K. (2009) ‘Alternative model-based and design-based
frameworks for inference from samples to populations: From
polarization to integration’. Multivariate Behavioral Research, 44,
711–740.
Stevens, J. (2009) Applied Multivariate Statistics for the Social Sciences,
5th edn. New York: Routledge.
Stiratelli, R., Laird, N.M., and Ware, J.H. (1984) ‘Random-effects models
for serial observations with binary response’. Biometrics, 40, 961–971.
Stram, D.O., and Lee, J.W. (1994) ‘Variance components testing in the
longitudinal mixed effects model’. Biometrics, 50, 1171–1177.
Correction (1995) in vol. 51, p. 1196.
Swamy, P.A.V.B. (1971) Statistical Inference in Random Coefficient
Regression Models. New York: Springer.
Tacq, J. (1986) Van Multiniveau Probleem naar Multiveau Analyse.
Rotterdam: Department of Research Methods and Techniques,
Erasmus University.
Tanner, M.A., and Wong, W.H. (1987) ‘The calculation of posterior
distributions by data augmentation’ (with discussion). Journal of the
American Statistical Association, 82, 528–550.
Thompson, L. A. (2009) ‘R (and S-PLUS) manual to accompany
Agresti’s Categorical Data Analysis (2002), 2nd edition’.
Timmermans, A.C., Snijders, T.A.B., and Bosker, R.J. (forthcoming) ‘In
search of value added in case of complex school effects’.
Tranmer, M., and Steel, D.G. (2001) ‘Ignoring a level in a multilevel
model: Evidence from UK census data’. Environment and Planning A,
33, 941–948.
Tuerlinckx, F., Rijmen, F., Verbeke, G., and De Boeck, P. (2006)
‘Statistical inference in generalized linear mixed models: A review’.
British Journal of Mathematical and Statistical Psychology, 59, 225–
255.
Twisk, J.W.R. (2006) Applied Multilevel Analysis. A Practical Guide.
Cambridge: Cambridge University Press.
van Buuren, S. (2007) ‘Multiple imputation of discrete and continuous
data by fully conditional specification’. Statistical Methods in Medical
Research, 16, 219–242.
van Buuren, S. (2011) ‘Multiple imputation of multilevel data’. In: J.J.
Hox and J.K. Roberts (eds), Handbook of Advanced Multilevel
Analysis, pp. 173–196. New York: Routledge.
van Buuren, S., Brand, J.P.L., Groothuis-Oudshoorn, C.G.M., and Rubin,
D.B. (2006) ‘Fully conditional specification in multivariate
 imputation’. Journal of Statistical Computation and Simulation, 76,
1049–1064.
van den Noortgate, W., Opdenakker, M.-C., and Onghena, P. (2005) ‘The
effects of ignoring a level in multilevel analysis’. School Effectiveness
and School Improvement, 16, 281–303.
van der Leeden, R., Meijer, E., and Busing, F.M.T.A. (2008) ‘Resampling
multilevel models’. In: J. de Leeuw and E. Meijer (eds), Handbook of
Multilevel Analysis, pp. 401–433. New York: Springer.
van Mierlo, H., Vermunt, J.K., and Rutte, C.G. (2009) ‘Composing
group-level constructs from individual-level survey data’.
Organizational Research Methods, 12, 368–392.
van Yperen, N.W., and Snijders, T.A.B. (2000) ‘Multilevel analysis of the
demands-control model’. Journal of Occupational Health Psychology,
5, 182–190.
Veall, M.R., and Zimmermann, K.F. (1992) ‘Pseudo-R2’s in the ordinal
probit model’. Journal of Mathematical Sociology, 16, 333–342.
Verbeke, G., and Lesaffre, E. (1996) ‘A linear mixed-effects model with
heterogeneity in the random-effects population’. Journal of the
American Statistical Association, 91, 217–221.
Verbeke, G., and Lesaffre, E. (1997) ‘The effect of misspecifying the
random-effects distribution in linear mixed models for longitudinal
data’. Computational Statistics and Data Analysis, 23, 541–556.
Verbeke, G., and Molenberghs, G. (1997) Linear Mixed Models in
Practice. A SAS-oriented Approach. Lecture Notes in Statistics, 126.
New York: Springer.
Verbeke, G., and Molenberghs, G. (2000) Linear Mixed Models for
Longitudinal Data. New York: Springer. Verbeke, G., and
Molenberghs, G. (2003) ‘The use of score tests for inference on
variance components’. Biometrics, 59, 254–262.
Vermeulen, C.J., and Bosker, R.J. (1992) De Omvang en Gevolgen van
Deeltijdarbeid en Volledige Inzetbaarheid in het Basisonderwijs.
Enschede: University of Twente.
Vermunt, J.K. (2003) ‘Multilevel latent class models’. Sociological
Methodology, 33, 213–239.
Vermunt, J.K. (2008) ‘Latent class and finite mixture models for
multilevel data sets’. Statistical Methods in Medical Research, 17, 33–
51.
Vermunt, J.K. (2011) ‘Mixture models for multilevel data sets’. In: J.J.
Hox and J.K. Roberts (eds), Handbook of Advanced Multilevel
Analysis, pp. 59–81. New York: Routledge.
Vermunt, J.K. and Magidson, J. (2005a) Latent GOLD 4.0 User’s Guide.
Belmont, MA: Statistical Innovations Inc.
Vermunt, J.K., and Magidson, J. (2005b) Technical Guide for Latent
GOLD 4.0: Basic and Advanced. Belmont, MA: Statistical Innovations
Inc.
Waternaux, C., Laird, N.M., and Ware, J.H. (1989) ‘Methods for analysis
of longitudinal data: Blood lead concentrations and cognitive
development.’ Journal of the American Statistical Association, 84, 33–
41.
Wedel, M. and DeSarbo, W. (2002) ‘Mixture regression models’. In: J.
Hagenaars and A. McCutcheon (eds), Applied Latent Class Analysis,
pp. 366–382. Cambridge: Cambridge University Press.
West, B.T., Welch, K.B. and GaŁecki, A.T. (2007) Linear Mixed Models:
A Practical Guide Using Statistical Software. Boca Raton, FL:
Chapman & Hall/CRC.
White, H. (1980) ‘A heteroskedasticity-consistent covariance matrix
estimator and a direct test for heteroskedasticity’. Econometrica 48,
817–830.
Wilson, D.B., and Lipsey, M.W. (2001) Practical Meta-analysis.
Thousand Oaks: Sage.
Windmeijer, F.A.G. (1995) ‘Goodness-of-fit measures in binary choice
models’. Econometric Reviews, 14, 101–116.
Windzio, M. (2006) ‘The problem of time dependent explanatory
variables at the context-level in discrete time multilevel event history
analysis: A comparison of models considering mobility between local
labour markets as an example’. Quality and Quantity, 40, 175–185.
Winship, C., and Mare, R.D. (1984) ‘Regression models with ordinal
variables’. American Sociological Review, 49, 512–525.
Wittek, R., and Wielers, R. (1998) ‘Gossip in organizations’.
Computational and Mathematical Organization Theory, 4, 189–204.
Wright, D.B., and London, K. (2009) ‘Multilevel modelling: Beyond the
basic applications’. British Journal of Mathematical and Statistical
Psychology, 62, 439–456.
Wu, L. (2010) Mixed Effects Models for Complex Data. Boca Raton, FL:
Chapman & Hall/CRC.
Yau, K.K.W., and Kuk, A.Y.C. (2002) ‘Robust estimation in generalized
linear mixed models’. Journal of the Royal Statistical Society, Series
B, 64, 101–117.
Yuan, K.-H., and Bentler, P.M. (2002) ‘On normal theory based inference
for multilevel models with distributional violations’. Psychometrika,
67, 539–562.
Yuan, Y., and Little, R.J.A. (2007) ‘Parametric and semiparametric
model-based estimates of the finite population mean for two-stage
cluster samples with item nonresponse’. Biometrics, 63, 1172–1180.
Yucel, R. (2008) ‘Multiple imputation inference for multivariate
multilevel continuous data with ignorable non-response’.
Philosophical Transactions of the Royal Society, Series A, 366, 2389–
2403.
Zaccarin, S., and Donati, C. (2009) ‘The use of sampling weights in
multilevel analysis of PISA data’. In: Proceedings, First Italian
Conference on Survey Methodology (ITACOSM09) (Siena, June 10–12,
2009), pp. 161–164. Siena: Tipografia Sienese.
Zeger, S.L., and Karim, M.R. (1991) ‘Generalized linear models with
random effects: A Gibbs sampling approach’. Journal of the American
Statistical Association, 86, 79–86.
Zeger S.L., Liang, K., and Albert, P. (1988) ‘Models for longitudinal
data: A generalised estimating equation approach’. Biometrics, 44,
1049–1060.
Zhao, E., and Yucel, R.M. (2009) ‘Performance of sequential imputation
method in multilevel applications’. In American Statistical Association
Proceedings of the Survey Research Methods Section. (American
Statistical Association, Alexandria, VA), pp. 2800–2810.
Zucker, D.M., Lieberman, O., and Manor, O. (2000) ‘Improved small
sample inference in the mixed linear model: Bartlett correction and
adjusted likelihood’. Journal of the Royal Statistical Society, Series B,
62, 827–838.
Index

adaptive quadrature, 300, 322

aggregation, 11, 14–16, 26, 33, 39, 55, 83, 102, 106, 154, 155, 157
of dichotomous data, 301
algorithms, 61, 89, 300–301, 322
alternative hypothesis, 192
analysis of covariance, 23, 30, 45, 47, 71, 100, 155
analysis of variance, 17, 22, 49, 100
analytic inference, 217, 222, 245
analytical weights, 220, 245
ANCOVA, see analysis of covariance
ANOVA, see analysis of variance
assumptions, 83, 107, 152, 153, 173
of hierarchical linear model, 75, 153–154
of random intercept model, 51, 55
autocorrelated residuals, 280, 281

Bayesian information criterion (BIC), 202

Bayesian statistics, see empirical Bayes, 63, 137, 138, 144, 146, 194–197,
203, 244, 300, 322, 325, 331
between-group correlation, 32–35, 40, 285–288
between-group covariance matrix, 285
between-group regression, 1, 15, 27–31, 40, 56–60, 72, 83, 88–89, 95, 96,
102, 106, 145, 149, 151, 154, 173, 234, 258
in three-level model, 70
between-group variance, 18, 20, 21, 43
for dichotomous data, 292, 296
between-subjects variables, 251
BIC, see Bayesian information criterion
binary variables, see dichotomous variables
Bonferroni correction, 170, 175, 278
bootstrap, 325
bootstrap standard errors, 244
Box–Cox transformation, 157, 174
budget constraint, 183, 186, 188, 189, 191

categorical variable, 96, 97, 108, 124

ordered, see ordered categorical variable
centering, 58, 81, 87–89, 93, 156
changing covariates, 276, 281
chi-squared test, 100, 291, 292, 321
Cholesky decomposition, 326
classification graph, 206, 211, 213, 215
cluster randomized trial, 184, 187, 193
cluster-robust standard error, see sandwich estimator
clusters, 6, 7, 13, 24, 39, 42, 198–200, 223, 227, 228, 237
combination of estimates, 36, 40
combination of tests, 36, 40, 229
comparative standard error, see standard error, comparative
complete data vector, 252, 281, 282
components of variance, see variance components
compound symmetry model, 249, 256, 261, 280
confidence interval, 65
for random part parameters, 100–101, 108
contextual analysis, 1–2
contextual effects, 56, 154, 173
contextual variable, 155
contingency table, 291
Cook’s distance, 168, 175
correlates of diversity, 127, 129
correlation ratio, 31
cost function, 182, 193
count data, 289, 314–320, 322
covariance matrix, 76, 90, 252–259, 281
cross-classification, 206, 215
cross-classified models, 205
cross-level inference, 17
cross-level interaction, 11, 13, 16, 43, 55, 72, 81–83, 92, 95, 103, 104, 106,
116, 154, 185, 251, 275, 281
cross-validation, 126, 154, 163
crossed random effects, 206–210, 215

data augmentation, 136, 138

degrees of freedom, 60, 94, 97, 108
deletion diagnostics, 167, 169, 170, 175
deletion standardized multivariate residual, 175
delta method, 100, 327
dependent variable, 9, 27, 42, 71
descriptive inference, 217, 245
design effect, 17, 23–24, 39, 179, 187, 192, 221, 224, 246
design of multilevel studies, 7, 133, 150, 176–193
design variables, 219, 222, 223, 226, 228, 234, 237, 238, 245, 246
design-based inference, 2, 218, 222, 224, 245
deviance, 97
deviance test, 50, 60, 89, 96–99, 108, 316
for multiple imputations, 141
deviation score, 58, 70, 95, 110, 154
diagnostic standard error, see standard error, diagnostic
dichotomous variables, 289–309, 320
disaggregation, 16–17, 39
discrete outcome variables, 228
discrete outcome variables(, 289
discrete outcome variables), 322
dummy variables, 96, 98, 120, 124
for longitudinal models, 249, 256
for multivariate models, 284, 288
duration analysis, 313–314

ecological fallacy, 1, 15, 17, 39, 59, 83, 258

effect size, 53, 109–114, 177–179, 192
effective sample size, see design effect, 221, 224, 245
EM algorithm, 61, 89
emergent proposition, 11, 13
empirical Bayes estimation, 62–63, 65, 73, 89, 93, 147, 196
empirical Bayes residuals, 165, 174
empty model, 17, 49–51, 62, 72
for dichotomous data, 291, 295–297, 320
for longitudinal data, 250, 263
multivariate, 257, 285
estimation methods, 60–61, 89–90, 300–301, 322
event history analysis, 263, 313–314, 321
exchangeability, 46, 47, 75
expected value, 5
explained variance, 109–114, 118, 156, 173
at level one, 112
for discrete dependent variable, 305, 311, 321
in longitudinal models, 261–262, 281
in random slope models, 113, 118
in three-level models, 113, 118
explanatory variables, 42, 51, 54, 71, 74, 82, 87, 152, 250, 264, 293, 320

F-test, 22, 23, 100, 108

Fisher scoring, 61, 89, 156, 168
Fisher’s Z transformation, 229
Fisher’s combination of p -values, 36, 229
fixed coefficients, 46, 71
fixed effects models, 46–48, 71
fixed occasion designs, 248–262, 280
fixed or random effects, 44–48
fixed part, 55, 57, 75, 87, 153, 156–158
frequency data, 314, 322
frequency weights, 220, 245
frequentist statistics, 195, 203
full information maximum likelihood, 134, 150, 258
fully multivariate model, 256, 261, 281, 282

generalizability theory, 25
generalized estimating equations, 198, 204
generalized linear models, 290
Gibbs sampling, 138, 173, 331
gllamm, 202, 300, 315
goodness of fit, 174
Greek letters, 5
group size, 56, 61, 188, 189
groups, 17, 42, 74

Hausman specification test, 57, 95

heteroscedasticity, 37, 75–76, 92, 119–129, 153, 173, 197, 221, 228, 244,
253, 258, 290, 320, 324, 325
at level one, 119–129, 159–161, 174, 279
at level two, 128–129
test of level-one, 159–161
hierarchical generalized linear models, 238, 289–320, 328, 331
hierarchical linear model, 2, 3, 41, 49, 72, 74–93, 116
assumptions, 75, 153–154
hierarchical model, 102
hierarchies
imperfect, 205–215
history of multilevel analysis, 1–2
HLM, 37, 104, 207, 238, 300, 301, 311, 315, 324–325
homoscedasticity, 43, 71, 119, 153, 159, 163, 173, 174, 197, 221, 232, 233

IGLS, see iterated generalized least squares

imputation, 135–151
chained equations, 144–148, 150
inclusion probabilities, 216, 245–246
independent variable, 42, see
explanatory variables, 71
influence, 161, 172, 175
influence diagnostics, 167–171, 175
informative design, 222, 223, 225–231, 246
interaction
random, 75
interactions, see cross-level interactions, 89, 102, 105, 129, 146, 151, 154,
155, 173
intercept, 27, 43, 44, 71
 random, 46, 49
intercept variance, 49, 51, 52, 72, 155, 190
intercept–slope covariance, 77, 79, 103
intercepts as outcomes, 1, 80
intraclass correlation, 17–35, 39, 50, 52, 72, 100, 104, 155, 181, 188, 192
for binary data, 291, 304–306, 321
for ordered categorical data, 311
for three-level model, 68
residual, see residual intraclass correlation
iterated generalized least squares, 61, 89, 156, 168

Kelley estimator, 63

Laplace approximation, 300, 322

latent class models, 173, 201–204, 330
Latent Gold, 202, 315, 330
latent variables, 62, 75, 80
level, 7, 8, 13, 155, 173
omission of, 155
level of randomization, 187
level-one residuals, see residuals, level-one
level-one unit, 8, 13, 17, 25, 42, 67, 74, 182
level-one variance, 190
level-two random coefficients, 75, 153, 174
level-two residuals, see residuals, level-two
level-two unit, 8, 13, 17, 25, 42, 67, 74, 182
leverage, 169–171, 174
likelihood ratio test, see deviance test
linear model, 71
link function, 294, 295, 300, 304, 314, 315, 320, 321
LISREL, 238
listwise deletion, 130
loess, see lowess smoother
log-odds, 293–298, 320
logistic distribution, 304
logistic function, 296, 304, 320
logistic regression, 133, 145, 149, 161, 290, 294, 297
logit function, 293–296, 320
longitudinal data, 2, 9, 247–282
lowess smoother, 162, 166, 174

macro–micro relations, 10, 11, 13, 15

macro-level propositions, 11, 13
macro-level relations, 26
macro-unit, see level-two unit
MAR, see missingness at random
marginal quasi-likelihood, 300, 322
Markov chain Monte Carlo, 138, 195–197, 203, 300, 322
maximum likelihood, 22, 35, 50, 51, 60, 72, 89, 97, 108, 156, 195, 197–
200, 202, 322
MCAR, see missingness completely at random
MCMC
see Markov chain Monte Carlo, 138
membership weights, 207, 211, 215
meta-analysis, 36, 40
random effects model for, 37
micro–macro relations, 11, 13
micro-level propositions, 10, 13, 15
micro-level relations, 15, 26
micro-unit, see level-one unit
missing data, 56, 130–151, 248, 257–259, 285
listwise deletion, 130
missingness at random, 132–134, 150
missingness completely at random, 132–133, 149
missingness indicators, 131–133, 149
missingness not at random, 132–133, 150
misspecification, 152, 156, 159, 162, 173, 198–200, 222
mixed model, 1, 41–93
MIXOR, 300, 311, 312, 325, 326
MIXPREG, 315
mixture models, 201–204
ML, see maximum likelihood
MLPowSim, 191, 325, 328, 330
MLwiN, 37, 153, 191, 207, 238, 301, 311, 312, 315, 325, 330
MNAR, see missingness not at random
model of interest, 134, 136, 139, 140, 143, 144, 148–150, 222, 246
model specification, 56–60, 76–77, 85, 87–90, 93, 102–108, 152, 154–175
of three-level model, 90
model-based inference, 2, 3, 48, 217–219, 222–224, 244, 245
Mplus, 202, 238, 330
multilevel logistic regression, 290–309, 320
Multilevel Models Project, 2, 323
multilevel negative binomial regression, 318, 322
multilevel ordered logistic regression model, 161, 310–313, 321
multilevel ordered probit model, 310, 311
multilevel Poisson regression, 314–320, 322
multilevel probit regression, 304, 305, 321
multilevel proportional odds model, 310
multiple classification, 207, 213–215
multiple correlation coefficient, 109
multiple imputation, 328
multiple membership models, 205, 207, 210–215
multiple membership multiple classification, 207, 213–215
multiple regression, 43
multisite trial, 187, 193
multivariate analysis, 283
multivariate empty model, 285
multivariate multilevel model, 282–288
multivariate regression analysis, 260, 281
multivariate residual, 169, 175

negative binomial regression, 318, 322

nested data, 3, 6–9
nonlinear models, 268, 289–320
normal probability plot, 163, 164, 166, 174
notational conventions, 5
null hypothesis, 36, 94, 96–98, 191
numerical integration, 300, 322, 326

observed between-group correlation, 32

observed variance, 18–20
odds, 293
offset, 316
OLS, 1, 23, 27, 44, 46, 52, 54, 62, 71, 87, 100, 104, 111, 159, 162, 198, 228
one-sided test, 94, 98
one-step estimator, 156, 168
Optimal Design, 330
ordered categorical variable, 289, 310, 321
ordinary least squares, see OLS
outliers, 165, 174
overdispersion, 318–320, 322

panel data, 2, 247, 280

parallel test items, 180
parameter estimation, see estimation methods
penalized quasi-likelihood, 300, 322
permutation test, 292
piecewise linear function, 268, 281
PinT, 181, 329
plausible values, 136, 325
Poisson distribution, 314, 322
Poisson regression, 290, 314
polynomial function, 265–267, 281
polynomial random part, 265
polynomial trend analysis, 253
population, 17, 46, 48–49, 217
population between-group variance, 18
population of curves, 263, 281
population within-group variance, 18
posterior confidence intervals, 64–67, 73
posterior distribution, 195, 203
posterior intercept, 63
posterior mean, 62–67, 73, 147, 165, 174, 180, 195, 203
posterior slope, 89, 93
posterior standard deviation, 65, 73, 195
power, 82, 103, 106, 132, 177–179, 192–193
power analysis, 193, 325, 328–330
precision weights, 220, 245
predictor variable, 42, 69, 71
prior distribution, 195, 203
probability model, 2–3, 217, 218, 237
probability weighted estimate, 220, 245
probit regression, 304
profile likelihood, 100
pseudo-cluster randomized trial, 187, 193
pseudo-maximum likelihood, 238–243, 246
psychological test theory, 25, 63

R, 37, 138, 153, 191, 207, 300, 301, 315, 328, 330
R2, see explained variance
random coefficient model, 1
random coefficient models, 75
random coefficients, 2, 41, 46, 47, 71, 75
random effects, 1, 2, 17, 18, 47, 49
comparison of intercepts, 66
discrete, 201–203
random effects ANOVA, 49
random intercept, 46, 49, 54, 72, 74
test, 97–98, 108
random intercept model, 41–73, 114, 249, 280
comparison with OLS model, 54
for dichotomous data, 297–299, 320
multivariate, 283–288
three-level, 67, 284
random part, 55, 75, 87, 153, 158–161
test, 97–101, 108
random slope, 75, 77, 82, 92, 114, 156, 174
explanation, 80–85, 92
test, 97–101, 106, 108, 156
random slope model, 74–93, 116
for dichotomous outcome variable, 302, 320
for longitudinal data, 253, 280
multivariate, 288
random slope variance, 75, 78
interpretation, 77
REALCOM, 138, 330
reliability, 25–26, 39, 62, 180
REML, see residual maximum likelihood
repeated measures, 180, 247, 280
residual intraclass correlation, 52, 61, 75, 187, 252
residual iterated generalized least squares, 61, 89
residual maximum likelihood, 22, 35, 51, 60, 72, 89, 97, 108, 257
residual variance, 51, 76, 80, 153, 160
for dichotomous data, 291
nonconstant, 120, 123, 127
residuals, 27, 43, 46, 48, 51, 71, 75, 80, 83
level-one, 153, 161–165, 170, 174, 175
level-two, 165–171, 174
multivariate, see multivariate residual
non-normal distributions, 199
nonnormal distributions, 172
restricted maximum likelihood, see residual maximum likelihood
RIGLS, see residual iterated generalized least squares
robust estimators, 175
robust standard errors, 197–200, 204

sample
cluster, 216, 231, 245
multistage, 6, 7, 13, 177
simple random, 6
stratified, 216, 231, 245
two-stage, 7, 17, 23, 39, 179, 180, 183, 192, 216, 231, 244
sample size, 23, 176–193
for estimating fixed effects, 180–187
for estimating group mean, 180
for estimating intraclass correlation, 188–190
for estimating population mean, 179–180
for estimating variance parameter, 190–191
sample surveys, 216–246
sampling designs, 7
sampling probabilities, 216, 245–246
sampling weights, 216–246
 scaling of, 225, 246
sandwich estimator, 173, 175, 197–200, 204, 220, 238, 246
SAS, 260, 300, 301, 312, 315, 327
Satterthwaite approximation, 95
shift of meaning, 15, 39, 59
shrinkage, 63–64, 66, 67, 73, 89
significance level, 177, 192
simple random sample, 216, 219, 224, 232, 245
slope variance, 92
slopes as outcomes, 1, 80, 92
software, 153, 300, 315, 323–331
spline function, 121, 129, 158, 174, 253, 270, 281
cubic, 270, 272
quadratic, 271
SPSS, 260, 329
standard error, 23, 37, 141, 178–179, 181
comparative, 65, 73
diagnostic, 65, 73, 165
of empirical Bayes estimate, 63, 65
of fixed coefficients, 155, 181
of intercept variance, 191
of intraclass correlation, 21, 188
of level-one variance, 191
of population mean, 24, 179
of posterior mean, 63, 65
of random part parameters, 90, 100, 108, 327
of standard deviation, 90, 100, 327
of variance, 90, 100, 327
standardized coefficients, 53
standardized multivariate residual, see multivariate residual
standardized OLS residuals, 174
Stata, 238, 300, 311, 315, 328–329
success probability, 290, 293, 295, 298, 314
SuperMix, 300, 315, 326
superpopulation model, 3, 236, 246
survey, 216
survey weights, see sampling weights
t-ratio, 59, 94, 108
t-test, 59, 94, 108
for random slope, 156
paired samples, 257, 280
test, 94–101, 163, 177
for random intercept, see random intercept
for random slope, see random slope
test of heterogeneity of proportions, 292, 321
textbooks, 4
three-level model, 67–71, 73, 90–93, 113, 282, 284, 321
for multivariate multilevel model, 282–288
threshold model, 304, 308, 310, 321
total variance, 18
total-group correlation, 32–35
total-group regression, 27–31
transformation
for count data, 315, 322
of dependent variable, 161
of explanatory variables, 157, 173
true score, 25, 63, 80, 180
two-stage sample, see sample, two-stage
type I error, 178, 192, 196, 198
type II error, 178, 192

unexplained variation, 41, 46, 48, 51, 71, 75, 80, 111, 306

V-known problem, 38, 324

variable occasion designs, 263–279, 281
variance components, 18, 109, 114–118, 188

Wald test, 94, 96, 108, 141

weighted least squares, 220–222, 245
weights, see membership weights, see sampling weights, see precision
weights
WinBUGS, 202, 331
within-group centering, 58, 87, 156
within-group correlation, 32–35, 40, 285, 286, 288
within-group covariance matrix, 285
within-group deviation score, see deviation score
within-group regression, 1, 15, 27–31, 40, 56–60, 72, 83, 88–89, 95, 96,
102, 106, 145, 149, 151, 154, 173, 234, 258
in three-level model, 70
within-group variance, 18, 21
for dichotomous data, 293
within-subjects design, 249
within-subjects variable, 251

zero variance estimate, 21, 61, 84, 85, 89