Contributions to Statistics

Gilbert MacKenzie
Defen Peng Editors

Statistical
Modelling in
Biostatistics and
Bioinformatics
Selected Papers
Contributions to Statistics

For further volumes:

http://www.springer.com/series/2912
Gilbert MacKenzie • Defen Peng
Editors

Statistical Modelling
in Biostatistics
and Bioinformatics
Selected Papers

123
Editors
Gilbert MacKenzie
Defen Peng
Centre of Biostatistics
University of Limerick
Limerick
Ireland

ISBN 978-3-319-04578-8 ISBN 978-3-319-04579-5 (eBook)

DOI 10.1007/978-3-319-04579-5
Springer Cham Heidelberg New York Dordrecht London
Library of Congress Control Number: 2014938132

© Springer International Publishing Switzerland 2014

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

This volume contains an appreciation of John Nelder, FRS, inventor of Generalized

Linear Models (GLMs) and hierarchical generalized linear models HGLMs, and
14 papers on statistical modelling. The range of topics covered is diverse, but falls
squarely into the following classification: (a) survival and event history modelling,
(b) longitudinal and time series modelling, (c) statistical model development and
(d) applied statistical modelling. The themes are representative of modern statistical
model development and several celebrate John Nelder’s various contributions to the
subject.
The volume owes its genesis, in part, to the 3rd International Workshop on
Correlated Data Modelling held at the University of Limerick in Ireland (www3.
ul.ie/wcdm07) and in equal measure to Science Foundation Ireland’s Biostatistics
and Bioinformatics research programme, BIO-SI (www3.ul.ie/bio-si) based in the
Centre of Biostatistics, University of Limerick, Ireland. The combination of these
contributions has led to an interesting compilation of papers on statistical model
development. In particular we would like to thank Professor Roger Payne for
contributing a paper on the new class of Hierarchical Generalised Non-Linear
Models. There is also emphasis on important emerging areas such as Bioinformatics
and Statistical Genetics.
We are particularly grateful to members of the BIO-SI project without whose
various contributions this volume could not have been compiled. Professor Defen
Peng of Zhongnan University of Economics and Law, PRC and BIO-SI Research
Fellow, was invited to serve as a guest editor, and she has contributed an interesting
paper on the classical problem of reference subclass choice in categorical regression
models.
We are also pleased that this volume has been produced by Springer—a recog-
nized academic publishing house. This should have the effect of raising awareness
and encouraging participation in the WCDM workshops and in the BIO-SI project.
Accordingly, we have been fortunate to have attracted Springer’s interest and we
gratefully acknowledge Veronika Rosteck, Associate Editor, Statistics, Springer
Heidelberg, Germany, for her encouragement and assistance with this project.

v
vi Preface

Many people worked hard to complete the volume. Foremost amongst these we
must thank the team in the Centre of Biostatistics (Dr. Emma Holian, Dr. David
Ramsey and the students) who organised the WCDM workshop. Of course, this
would not have been possible without the sterling efforts of Alessandra Durio
and the late Ennio Isaia (see Postscript) in Torino. Professors John Hinde (NUIG,
Ireland) and Jianxin Pan (Manchester, UK) also gave valuable advice and support.
We have of course to thank all of the sponsors of the workshop for their support,
especially the Irish Statistical Association (www.istat.ie) and Science Foundation
Ireland (www.sfi.ie), which kindly supported the workshop with grants. The editors
also acknowledge the financial support of SFI’s BIO-SI research programme
(www3.ul.ie/bio-si.) and the Research Office in the University of Limerick.

Limerick, Ireland Gilbert MacKenzie

Limerick, Ireland Defen Peng
Autumn 2013
Acknowledgements: An Appreciation—John
Nelder, FRS

We pause to remember John Ashworth Nelder, FRS (Figs. 1 and 2), “John” to those
of us who knew him and admired his contribution to Statistical Modelling and more
widely to Statistical Science. John died on Saturday, August 7, 2010, aged 85 years.
He had retired from Rothamsted, Ronald Fisher’s stomping ground, aged 60, having
already had an outstanding career which included, inter alia, the development of the
family of Generalized Linear Models (GLMs) with his collaborator Robert MacLa-
gan Wedderburn, a (Scottish) colleague at the Experimental Research Station.
Inimitably, John’s research output in “statistical mathematics” continued undi-
minished from Imperial College, London, where he held a visiting Professorship.
With Professor Youngjo Lee (Seoul National University, South Korea) John built
on his earlier work on Generalised Linear Models and together they developed the
theory of Hierarchical Generalised Linear Models (HGLMs).
Their first paper was read to the Royal Statistical Society in 1996. No longer
were the random effects in mixed models confined to the Gaussian stable, but
were generalised to the Exponential Family: Exponential Family response (Y )
and Exponential Family error (u). Mixed models would never be the same again.
Inference was generalised to a hierarchical likelihood containing random effects
unlike standard likelihood methods advocated by Fisher. One consequence was that
a new method of estimating the dispersion parameters was required and this problem
was neatly solved by an appeal to the Cox–Reid adjustment method of creating
an adjusted profile likelihood. Later this technique would come to be viewed as a
nuisance parameter elimination technique of great utility.
In many ways the 1996 paper was a tour de force, but curiously it was not
well received. Doubts were cast on the technique’s ability to cope with matched
pairs (binary data) in small samples and the Bayesian fraternity did not fully
grasp the hierarchical arguments which allowed Lee and Nelder to treat random
parameters as fixed effects. Moreover, H-likelihood inference did not require the
use of a prior. And, in any case, the problem could always be cast in a Bayesian
mould. It was to be fully 5 years before the 2001 Biometrika paper, which first
introduced the notion of double hierarchical generalised linear models, DHGLMs,
began to redress the balance of opinion. It took another 5 years to the seminal 2006

vii
viii Acknowledgements: An Appreciation—John Nelder, FRS

Fig. 1 John Ashworth Nelder, FRS, July 2006

Fig. 2 John Nelder presenting prizes at WCDM, Torino, 2004

paper, entitled Double Hierarchical Generalised Linear Models, again read to the
Royal Statistical Society to demonstrate the power of the method. DHGLMs have a
HGLM for the location parameters and another linked HGLM for the dispersion
parameters, the latter being a natural extension of Nelder and Lee’s ideas on
structural dispersion. This paper dealt with many criticisms of the H-likelihood
method and so-called counter examples. These transpired to be largely bogus due to
various misunderstandings of the 1996 paper.
It is worth remarking that there are very close links with the development of
covariance modelling techniques. It is natural from an optimality perspective to wish
to model the mean and covariance structure with the same rigor. Unsurprisingly, it
turns out that many covariance models are related to DHGLMs, although typically,
Acknowledgements: An Appreciation—John Nelder, FRS ix

in the longitudinal data setting, the modelling is undertaken in the observation,

rather than in the latent, space.
Once again, the 2006 paper was not without its critics. However, that the
H-likelihood idea worked well, in more general classes of models than claimed
by Lee and Nelder in their 2006 paper, was known to statisticians working on
survival analysis, where many models lie beyond the censored Exponential Family
and involve frailty. Some of these findings appear in the now classic 2006 book
by Lee, Nelder and Pawitan and are due in large measure to the sterling work of
Professor Il Do Ha.
Younger statisticians may take heart from this story. Little of lasting value in
statistical model development is built overnight and it has taken Lee and Nelder
more than a decade to place the H-likelihood idea on a recognised footing. John
Nelder, never completely satisfied, was often heard to remark that the work was far
from complete and there was still much to be done. Such tenacity was part of John
Nelder’s nature, from which we can all draw inspiration.

Limerick, Ireland Gilbert MacKenzie

Limerick, Ireland Defen Peng
Herts, UK Roger Payne
Contents

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
Gilbert MacKenzie and Defen Peng

Part I Survival Modelling

Multivariate Interval-Censored Survival Data: Parametric,

Semi-parametric and Non-parametric Models . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
Philip Hougaard
Multivariate Survival Models Based on the GTDL . . . . . .. . . . . . . . . . . . . . . . . . . . 23
Gilbert MacKenzie and Il Do Ha
Frailty Models with Structural Dispersion . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
Joseph Lynch and Gilbert MacKenzie
Random Effects Ordinal Time Models for Grouped
Toxicological Data from a Biological Control Assay . . . . .. . . . . . . . . . . . . . . . . . . . 45
Marie-José Martinez and John P. Hinde

Part II Longitudinal Modelling and Time Series

Modelling Seasonality and Structural Breaks: Visitors to NZ

and 9/11 . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 61
John Haywood and John Randal
Forecasting the Risk of Insolvency Among Customers
of an Automotive Financial Service . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 81
Rita Allais and Marco Bosco
On Joint Modelling of Constrained Mean and Covariance
Structures in Longitudinal Data .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 93
Jing Xu and Gilbert MacKenzie

xi
xii Contents

Part III Statistical Model Development

Hierarchical Generalized Nonlinear Models . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111

Roger W. Payne
Comparing Robust Regression Estimators to Detect Data
Clusters: A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 125
Alessandra Durio and Ennio Isaia
Finite Mixture Model Clustering of SNP Data.. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 139
Norma Bargary, J. Hinde, and A. Augusto F. Garcia
Discrepancy and Choice of Reference Subclass in Categorical
Regression Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 159
Defen Peng and Gilbert MacKenzie

Part IV Applied Statistical Modelling

Statistical Methods for Detecting Selective Sweeps . . . . . .. . . . . . . . . . . . . . . . . . . . 187

David Ramsey
A Mixture Model and Bootstrap Analysis to Assess
Reproductive Allocation in Plants . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 213
Caroline Brophy, D. Gibson, P.W. Wayne, and J. Connolly
On Model Selection Algorithms in Multi-dimensional
Contingency Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 221
Susana Conde and Gilbert MacKenzie

Obituary: Professor Ennio Isaia .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 243

Contributors

Rita Allais Department of Social-Economic, Mathematical & Statistical Sciences,

University of Turin, Turin, Italy
Norma Bargary Department of Mathematics & Statistics, University of Limerick,
Limerick, Ireland
Marco Bosco Department of Management, University of Turin, Turin, Italy
Caroline Brophy Department of Mathematics & Statistics, National University of
Ireland Maynooth, Maynooth, Co Kildare, Ireland
UCD School of Mathematical Sciences, Environmental & Ecological Modelling
Group, University College Dublin, Dublin, Ireland
Susana Conde Department of Mathematics, Imperial College, London, UK
J. Connolly UCD School of Mathematical Sciences, Environmental & Ecological
Modelling Group, University College Dublin, Belfield, Dublin, Ireland
Alessandra Durio Department of Economics & Statistics “S. Cognetti de Martiis”,
University of Turin, Turin, Italy
A. Augusto F. Garcia Department of Genetics, Luiz de Queiroz College of
Agriculture, University of São Paulo, São Paulo, Brazil
D. Gibson Department of Plant Biology, Center for Ecology, Southern Illinois
University, Carbondale, IL, USA
Il Do Ha Faculty of Information Science, Daegu Haany University, Gyeongsan,
South Korea
John Haywood School of Mathematics, Statistics and Operations Research, Vic-
toria University of Wellington, Wellington, New Zealand
John P. Hinde School of Mathematics, Statistics and Applied Mathematics,
National University of Ireland, Galway, Ireland

xiii
xiv Contributors

Philip Hougaard Biostatistics, Lundbeck, Valby, Denmark

Ennio Isaia Department of Statistics & Applied Mathematics, University of Turin,
Turin, Italy
Joseph Lynch The Centre for Biostatistics, University of Limerick, Limerick,
Ireland
Gilbert MacKenzie The Centre for Biostatistics, University of Limerick, Limer-
ick, Ireland
Marie-José Martinez Team MISTIS, INRIA Rhône-Alpes & Laboratoire Jean
Kuntzmann, Montbonnot, Saint-Ismier Cedex, France
Roger W. Payne Department of Computational and Systems Biology, Rothamsted
Research, Harpenden, Herts, UK
VSN International, Hemel Hempstead, Herts, UK
Defen Peng The Centre for Biostatistics, University of Limerick, Limerick, Ireland
David Ramsey Department of Computer Science and Management, Wrocaw
University of Technology, Wrocaw, Poland
John Randal School of Economics and Finance, Victoria University of Welling-
ton, Wellington, New Zealand
P.W. Wayne Harvard Medical School, Osher Center for Integrative Medicine,
Boston, MA, USA
Jing Xu Department of Mathematics, Statistics & Economics, Birkbeck College,
London, UK
Introduction

Gilbert MacKenzie and Defen Peng

1 Preamble

The aim of this book is to provide the reader with an interesting selection of papers
on modern statistical model development in bio-statistics and bio-informatics. The
book also attempts to celebrate, in passing, the work of John Nelder who made
an enormous contribution to statistical model development. The papers presented
herein have been compiled from several sources. The majority contribution by num-
bers stems from Science Foundation Ireland’s BIO-SI project (www3.ul.ie/bio-si)
centred in the Universities of Limerick and Galway in Ireland. However, the volume
has also been exceptionally fortunate to attract papers from several distinguished
international statisticians who had participated in a Workshop on Correlated Data
Modelling held in the University of Limerick (www3.ul..ie/wcdm07). The various
papers offered represent a refreshing blend of experience and youth as the next
generation of researchers begin to contribute.
Accordingly, it is with great pleasure that the Editors introduce the material.

2 Contributions

The contributed papers have been arranged naturally into four groups, namely:
• Survival and event history analysis.
• Longitudinal analysis and time series.
• Statistical model development.
• Applied statistical modelling.

G. MacKenzie ()
The Centre for Biostatistics, University of Limerick, Limerick, Ireland
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 1

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__1,
© Springer International Publishing Switzerland 2014
2 G. MacKenzie and D. Peng

2.1 Survival and Event History Analysis

The survival section presents an interesting mixture of papers on this evergreen

topic. The first paper from Professor Philip Hougaard provides a thorough review
of recent developments in multivariate survival analysis covering the gamut of
modelling settings including parametric and semi-parametric models and right and
interval censoring schemes. The paper is non-technical and succeeds in emphasising
the key ideas while providing insight into known pitfalls. Altogether an excellent
introduction to the area and a most apt leading paper.
The next paper by MacKenzie and Ha deals inter alia with non-proportional haz-
ards (non-PH) parametric survival models introducing two new log-normal frailty
models via the Generalised Time-Dependent Logistic model (GTDL) (MacKenzie
1996). The authors employ the H-likelihood method of estimation which has the
advantage of conveniently avoiding intractable integrations. The paper picks up
an important theme based on the notion of a canonical scale. Lee et al. (2006)
argue that H-likelihood estimation requires the random effects and fixed effects to
be placed on the same scale (canonical), but MacKenzie and Ha show that this
requirement is strictly unnecessary as inference is similar in both GTDL frailty
models, even when the random effects are on a different scale. This paper illustrates
that the H-likelihood method works well beyond the envelope claimed by Lee et al.
(2006).
The third paper in this section explores the idea of structural dispersion. Lynch
and MacKenzie develop the idea of Gamma frailty models starting from Weibull
and GTDL basic hazards when analysing survival from a large number of incident
cases of breast cancer in 13 local health authorities in the West Midlands of the
UK. The resulting models are both non-PH and have structural dispersion models
of the subject-specific form i2 D exp.!i / D exp.xi0 ˇ/ for the frailty variances,
 2 . These models show superior fit compared to routine frailty models i2 D  2 for
i D 1; : : : ; n patients and compared to the basic survival models. The fact that the
variance of the random effects may be subject-specific, yet controlled by baseline
fixed effects, seems a plausible alternative to conventional schemes.
The last paper deals with a type of discrete survival analysis. Martinez and
Hinde argue that discrete survival times can be viewed as ordered multicategorical
data. They use a continuation-ratio model which is particularly appropriate when
the ordered categories represent a progression through different stages, such as
survival through various times. This particular model has the advantage of being a
simple decomposition of a multinomial distribution into a succession of hierarchical
binomial models. In a clustered data context, they incorporate random effects
into the linear predictor of the model to account for uncontrolled experimental
variation. Assuming a normal distribution for the random effects, an EM algorithm
with adaptive Gaussian quadrature is used to estimate the model parameters. This
approach is applied to the analysis of grouped toxicological data obtained from a
biological control assay. This paper illustrates the variety of methods available in
the field of survival analysis.
Introduction 3

2.2 Longitudinal and Time Series

The opening paper presents a thoughtful analysis of visitor arrivals in New Zealand.
Haywood and Randal demonstrate the poor performance, with seasonal data, of
existing methods for endogenously dating multiple structural breaks. Motivated
by iterative nonparametric techniques, they present a new approach for estimating
parametric structural break models that performs well. They suggest that iterative
estimation methods are a simple but important feature of this approach when
modelling seasonal data. The methodology is illustrated by simulation and is then
used to analyse a monthly short-term visitor arrival time series to New Zealand, in
order to assess the effect of the 9/11 terrorist attacks. While some historical events
had a marked structural effect on trends in those arrivals, they show that 9/11 did not.
Continuing the time series theme Allais and Bosco present an application of
Generalized Linear Models to the prevention of risks of insolvency of an automotive
financial service. In order to forecast the performance of instalment payments of the
customers, they resort to a logit multivariate regression model. Before fitting the
model, they resort to sample logits, Generalized Additive Models and univariate
logistic regression in order to identify the subset of best predictors and verify
the assumption of the statistical model. For the estimated model, they use Wald
statistics to assess the significance of the coefficients, the Likelihood Ratio to test
the goodness of fit of the model, and the Odds Ratio to interpret the estimated
coefficients. In order to verify the goodness of fit of the model, they employ
classification tables and the Receiver Operating Characteristic curve. Finally, they
validate the fitted model by means of a predictive-test on a training set.
Xu and MacKenzie return to the joint mean-covariance modelling theme in
a longitudinal setting where the mean is subject to some constraints. A data-
driven method for modelling the intra-subject covariance matrix is developed in the
context of constrained marginal models arising in longitudinal data. A constrained
iteratively re-weighted least squares estimation algorithm is applied. Some key
asymptotic properties of the constrained ML estimates are given. They analyze
a real data set in order to compare data-driven covariance modelling methods
with classical menu-selection-based modelling techniques under a constrained
mean model, extending the usual regression model for estimating generalized
autoregressive parameters. Finally, they demonstrate, via a simulation study, that
a correct choice of covariance matrix is required in order to minimise not only the
bias, but also the variance, when estimating the constrained mean component.

2.3 Statistical Model Development

In this section the emphasis is firmly on the development of new classes of model
and or statistical methods.
Professor Payne introduced Hierarchical generalized non-linear models. Hierar-
chical generalized linear models allow non-Normal data to be modelled in situations
4 G. MacKenzie and D. Peng

when there are several sources of error variation. He extends the familiar generalized
linear models to include additional random terms in the linear predictor. However,
he does not constrain these terms to follow a Normal distribution nor to have an
identity link, as is the case in the more usual generalized linear mixed model. He
thus provides a much richer set of models that may seem more intuitively appealing.
Another extension to generalized linear models allows nonlinear parameters to be
included in the linear predictor. The fitting algorithm for these generalized nonlinear
models operates by performing a nested optimization, in which a generalized
linear model is fitted for each evaluation in an optimization over the nonlinear
parameters. The optimization search thus operates only over the (usually relatively
few) nonlinear parameters, and this should be much more efficient than a global
optimization over the whole parameter space. This paper reviews the generalized
nonlinear model algorithm, and explains how similar principles can be used to
include nonlinear fixed parameters in the mean model of a hierarchical generalized
linear model, thus defining a hierarchical generalized nonlinear model.
Durio and Isaia argue that it is well known that in all situations involving the
study of large data sets where a substantial number of outliers or clustered data
are present, regression models based on M -estimators are likely to be unstable.
Resorting to the inherent properties of robustness of the estimates based on the
Integrated Square Error criterion, a technique of regression analysis which consists
in comparing the results arising from L2 estimates with the ones obtained by
applying some common M -estimators. The discrepancy between the estimated
regression models is measured by means of a new concept of similarity between
functions and a system of statistical hypothesis. A Monte Carlo Significance test
is introduced to verify the similarity of the estimates. Whenever the hypothesis of
similarity between models is rejected, a careful investigation of the data structure is
required in order to check for the presence of clusters, which can lead us to consider
a mixture of regression models. Concerning this, they show how the L2 criterion
can be applied in fitting a finite mixture of regression models. The theory is outlined
and the whole procedure is applied to a case study concerning the evaluation of the
risk of fire and the risk of electric shocks arising in electronic transformers.
Bargary et al. develop the idea of model-based clustering using orthogonal
regressions. They note that finite mixture models have been used extensively in clus-
tering applications, where each component of the mixture distribution is assumed to
represent an individual cluster. The simplest example describes each cluster in terms
of a multivariate Gaussian density with various covariance structures. However,
using finite mixture models as a clustering tool is highly flexible and allows for
the specification of a wide range of statistical models to describe the data within
each cluster. These include modelling each cluster using linear regression models,
mixed effects models, generalized linear models, etc. This paper investigates the use
of mixtures of orthogonal regression models to cluster biological data arising from
a study of the sugarcane plant.
Peng and MacKenzie consider the evergreen problem of choosing the reference
subclass in categorical regression models. They show how to choose the reference
Introduction 5

subclass optimally and note that this choice requires the use of secondary criteria.
They derive the optimal design allocation of observations to subclasses and provide
a statistic, based on the generalized variance and its distribution for measuring the
discrepancy between the optimal allocation and any observed allocation occurring
in an observational studies in the context of the general linear model. They then
extend their methods to generalized linear models. The focus is on techniques which
maximize the precision of the resulting estimators. They explore the general form of
the optimal design matrix for general linear models with categorical regressors, and
propose an algorithm to find the optimal design matrix for generalized linear models
when the design matrix is of high dimension. They note that the proposed statistic,
a measure of discrepancy, can be used to show if secondary criteria for the choice of
reference subclasses are needed in parametric categorical regression models. They
illustrate their methods by means of simulation studies and the analysis of a set of
lung cancer survival data.

2.4 Applied Statistical Analysis

The papers in this section derive their main thrust from applications but nevertheless
contain a substantial element of model development and/or review.
David Ramsey reviews tests for detecting selective sweeps. The emigration of
humankind from Africa and the adoption of agriculture have meant that the selective
pressures on humankind have changed in recent evolutionary times. A selective
sweep occurs when a positive mutation spreads through a population. For example, a
mutation that enables adults to digest lactase has spread through the Northern
European population, although it is very rare in the African population. Since neutral
alleles that are strongly linked to such a positive mutation also tend to spread through
the population, these sweeps leave a signature, a valley of low genetic variation. He
reviews the development of statistical tests for the detection of selective sweeps
using genomic data, particularly in the light of recent advances in genome mapping
and considers directions for future research.
Brophy et al. discuss issues that arise in the analysis of reproductive allocation
(RA) in plants when predicting from complex models. Communicating models
of RA requires predictions on the original scale of the data and this can present
challenges if transformations were used during the modelling. It is also necessary
to estimate without bias the mean level of RA as this may reflect a plant’s ability to
contribute in the next generation. Several issues can arise in modelling RA including
the occurrence of zero values and the clustering of plants in stands which can lead
to the need for more complex models. They present a two-component finite mixture
model framework for the analysis of RA data with the first component being a
censored regression model on the logarithmic scale and the second component being
a logistic regression model. Both components contained random error terms to allow
for potential correlation between grouped plants. They implement the models using
data from an experiment carried out to assess environmental factors on reproductive
6 G. MacKenzie and D. Peng

allocation and detail the issues that arose in predicting from the model. They also
present a bootstrap analysis to generate standard errors for predictions from the
model and to test for comparisons among predictions.
Conde and MacKenzie review methods for searching high-dimensional con-
tingency tables and present some recent developments. The review deals with
the main ideas and is relatively non-technical providing a natural book-end to
the volume. They present a review focussed on model selection in log-linear
models and contingency tables. The concepts of sparsity and high-dimensionality
have become more important nowadays, for example, in the context of high-
throughput genetic data. In particular, they describe recently developed automatic
search algorithms for finding optimal hierarchical log-linear models (HLLMs) in
sparse multi-dimensional contingency tables in R and some LASSO-type penalized
likelihood model selection approaches. The methods rely, in part, on a new result
which identifies and thus permits the rapid elimination of non-existent maximum
likelihood estimators in high-dimensional tables. They are illustrated using a set of
high-dimensional comorbidity data.
Overall, the papers contained in this volume represent a level of variety and
strength in statistical model development which would have no doubt pleased John
Nelder.

References

Lee, Y., Nelder, J. A., & Pawitan, Y. (2006). Generalised linear models with random effects:
Unified analysis via h-likelihood. London: Chapman and Hall.
MacKenzie, G. (1996). Regression models for survival data: The generalised time dependent
logistic family. Journal of the Royal Statistical Society, 45, 21–34.
 Part I
Survival Modelling
Multivariate Interval-Censored Survival Data:
Parametric, Semi-parametric
and Non-parametric Models

Philip Hougaard

Abstract Interval censoring means that an event time is only known to lie in
an interval (L,R], with L the last examination time before the event, and R the
first after. In the univariate case, parametric models are easily fitted, whereas for
non-parametric models, the mass is placed on some intervals, derived from the
L and R points. Asymptotic results are simple for the former and complicated
for the latter. This paper is a review describing the extension to multivariate data,
like eruption times for teeth examined at visits to the dentist. Parametric models
extend easily to multivariate data. However, non-parametric models are intrinsically
more complicated. It is difficult to derive the intervals with positive mass, and
estimated interval probabilities may not be unique. A semi-parametric model makes
a compromise, with a parametric model, like a frailty model, for the dependence
and a non-parametric model for the marginal distribution. These three models are
compared and discussed. Furthermore, extension to regression models is considered.
The semi-parametric approach may be sensible in many cases, as it is more flexible
than the parametric models, and it avoids some technical difficulties with the
non-parametric approach.

Keywords Bivariate survival • Frailty models • Interval censoring • Model

choice

1 Introduction

Interval-censored survival data refer to survival data, where the times of events are
not known precisely; they are only known to lie in given intervals. The event could,
for example, be HIV infection, or outbreak of a disease. Interval censoring typically

P. Hougaard ()
Biostatistics, Lundbeck, DK-2500 Valby, Denmark
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 9

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__2,
© Springer International Publishing Switzerland 2014
10 P. Hougaard

arises because the status of an individual is not known at all times, due to it being
necessary to make a specific measurement in order to find out whether the event has
happened. Such a measurement could be bioanalytical analysis of a blood sample
or an X-ray. It turns out that instead of knowing all examinations at all times, it is
sufficient to know that the event time is in an interval of the form .Li ; Ri , where
the left endpoint is the time last seen without the disease, and the right endpoint is
the first time seen with the disease. Subjects with the disease at the first examination
has Li D 0, and subjects that are never observed to have the disease have Ri D 1,
that is, are right-censored. For many diseases (for example, diabetes type II), the
natural time scale is age. Interval censoring is in contrast to the standard survival
data setup, where all event times are either known precisely, or they happen after
the end of observation (that is, right-censored data).
This paper will give a brief overview of the univariate case (Sect. 2), but the real
purpose is to consider models for multivariate data. An introduction to modelling
multivariate data is in Sect. 3, without reference to how data are collected. The inter-
val censoring way of observing data is then described in Sect. 4. This paper describes
and compares three different ways of modelling multivariate interval-censored
data. The fully parametric models are described in Sect. 5, the non-parametric
models in Sect. 6, and the semi-parametric models in Sect. 7. Regression models
are considered in Sect. 8. The various approaches are compared in Sect. 9.
Further reading on interval censoring are Sun (2006) and Hougaard (2014). In
particular, the latter includes a chapter that gives more details on the issues discussed
in this paper. Further reading on multivariate survival data is Hougaard (2000).

2 Univariate Interval-Censored Data

For each subject one or more examinations are made over time to find out if/when
the subject gets the disease, or, more generally, experiences the event. It is assumed
that the event studied (like outbreak of a disease) induces a permanent change
that can be measured without error. For example, an infectious disease leads to
antibodies, and in the interval censoring frame, it is assumed that these are detectable
at any time thereafter.
If all subjects are studied at the same times, the data are grouped and can easily
be analysed. So, we will consider the more general case of individual inspection
times. The inspection times are supposed to be chosen independently of the response
process and not informative of the parameters governing the response process. The
likelihood then has the following form
Y
fSi .Li /  Si .Ri /g; (1)
i

where Si ./ is the survivor function for the time to the event, for subject i .
Strictly speaking, this function is only the probability of the observations, when
Multivariate Interval-Censored Survival Data 11

the inspection times are fixed beforehand. If, instead the inspection times are chosen
randomly, but independent of the failure time process, the density of this distribution
enters as a factor. However, as this factor does not influence the maximization,
we may omit that term. That is, Eq. (1) can still be used as the likelihood of the
observations.
However, the inspection times could easily depend on the failure time process,
for example, if we study patients coming to a general practitioner. The chance that a
person will go to a doctor must in general be higher when the subject suffers from a
disease than for a completely healthy subject. This case is not treated in the present
paper.
A parametric model is easily handled just by inserting the relevant expression
for Si ./ into Eq. (1). If there are many different potential inspection times, and the
model is reasonable, the distribution is identifiable, and the estimated parameters can
be found by Newton-Raphson iteration and one can derive asymptotic results in the
usual way, that is, using the observed information (minus the second derivative of the
log likelihood function), which,
p inter alia, implies that the asymptotic distribution
has an asymptotic order of n, where n is the number of observations.
A non-parametric model is somewhat more complicated. If the inspection times
are chosen from a finite set of times, say the birthdays of the subjects, we can
only identify period probabilities, in the birthday case, 1-year probabilities. The
asymptotic calculations will still follow standard principles. If on the other hand,
the inspection times are chosen randomly from a continuous distribution, we can
asymptotically identify the survivor function at all times (within the range of the
inspection time distribution). However, the estimated survivor function will follow
a slower than standard asymptotic order, namely n1=3 . This result has been proven
under either an assumption of only one measurement for each subject (current
status data) or under an assumption of several measurements per subject with the
requirement that two measurements on the same subject cannot happen arbitrarily
close (that is, the interval between measurements must be larger than some  > 0).
For a given data set, we cannot determine the full distribution. Indeed, the
resolution cannot exceed the intervals generated by the left and right endpoints
(L and R). That is, all that can be determined are the masses of these intervals.
These interval probabilities are uniquely estimated. Actually, most of the intervals
will have estimate 0. Peto (1973) suggested a simple procedure that could identify
some of the intervals with estimate 0 and thus simplify the estimation problem to
find the probabilities of the remaining intervals. Various software can be used to find
these estimates. For the parametric models, SAS includes the procedure lifereg to
cover selected models, and the procedure nlmixed, which can handle more general
models. In the latter case, it is necessary to code the survivor functions, but no
derivatives (neither with respect to time, nor with respect to the parameters) are
needed. For the non-parametric case, SAS has a macro called ICE that covers the
one-sample case.
The subscript i on the survivor function is introduced in order to allow for
regression models, either parametric or semi-parametric. In the semi-parametric
case, the models have to be hazard-based, which could either be a proportional
12 P. Hougaard

hazards model, or a more complex model, like a frailty model. Accelerated failure
time models are markedly more complicated in the semi-parametric case and not
covered by the present paper. Regression models are considered in more detail in
Sect. 8.

3 Multivariate Data

Multivariate data can either be family data, or multivariate data for a single subject.
A common example of the second type is the age at eruption of teeth. All teeth are
then checked at visits to a dentist at one or more times. Indeed, most of the examples
in the literature have been of the second type. An example of the first type of data
could be pairs of twins studied for the time to outbreak of a disease that could only
be detected by some sort of bioanalytical measurement.
Due to the complexity, this paper will be formulated for bivariate data rather
than general multivariate data. A bivariate distribution can be formulated in several
different ways. The most general is just to formulate the bivariate survivor function,
that is, S.t1 ; t2 / D P rfT1 > t1 ; T2 > t2 g. Here, the subscript i has been neglected.
However, in some cases, it is an advantage to parametrize the model by means of
the marginal distributions S1 .t/, respectively, S2 .t/, and then use a copula model to
obtain the joint distribution as S.t1 ; t2 / D C.S1 .t1 /; S2 .t2 //. The copula C.; / is
a bivariate distribution on the unit square, which is used to model the dependence.
When studying a single bivariate distribution, this is just another way to express
the distribution. Indeed, for continuous distributions, there is a unique relationship
between these quantities. However, when we consider a family of distributions in a
statistical model, for example, a regression model, the two formulations can lead to
different models. Frailty models are of the copula type.
Some methods will require the survivor function to be known, whereas other
methods may work with less explicit expressions, like an integral.

3.1 Purpose

The purpose of studying multivariate data could be to study the dependence (as was
the case in the dental application). As a consequence to this dependence, we may use
the information available at some time point for predicting later events. For example,
if one twin is found to suffer from an inherited disease, we may suspect that the
partner either has or will develop the same disease. This risk can be quantified by
calculating the conditional distribution given information on the first twin or by a
more symmetric bivariate measure, like Kendall’s  or Spearman’s . Of course,
we could also be interested in the dependence just for understanding the underlying
processes.
Multivariate Interval-Censored Survival Data 13

Another purpose could be to study the effect of some covariates. For this
purpose, it is important whether the regression coefficients are shared, because
if they are not shared, it may be sufficient (and certainly simpler) to study the
marginal distributions separately. When the coefficients are shared for the various
components, we may gain precision by considering them simultaneously, and the
multivariate approach is necessary in order to evaluate the precision correctly. Also
for a comparison within groups, a multivariate approach is necessary. To be slightly
more precise, a multivariate approach is not always necessary for a comparison
within groups. A classical example of this is the paired t-test. However, there is no
analogue of this test for interval censoring, so in practice, it is difficult to handle such
a comparison without modelling the full bivariate distribution. When the aim is to
study the effect of covariates, it might not be necessary to quantify the dependence.

3.2 Frailty Models

One way to obtain a model for multivariate data is to use a frailty model. This
is basically a random effects model, where the frailty refers to the shared random
effect. The conditional hazard formulated with the hazard of Tij conditional on the
frailty Yi as

Yi ij .t/;

so that the frailty Yi is common for the i -th group and the hazard ij .t/ may be a
parametrized function of i , j and t. In the non-parametric case, the function may
depend only on t or only on j and t. Furthermore, it is assumed that the times
are independent given the frailty, or in other words, it is the frailty that creates the
dependence. A frailty model may offer simple expressions for the survivor function,
because it is simple to integrate the frailty out. In general, the survival function
(neglecting the subscript i ) is found as
Z 1
S.t1 ; t2 / D g.y/ expŒyfM1 .t1 / C M2 .t2 /gdy D L.M1 .t1 / C M2 .t2 //; (2)
0

where g.y/ is the density of Y Rand L.s/ D E exp.sY / is the Laplace transform
t
of the frailty. Finally Mj .t/ D 0 j .u/d u is the integrated conditional hazard.
As a specific example, consider a gamma frailty model, which has Laplace
transform L.s/ D .s C /ı = ı . We may restrict the scale parameter by letting
 D ı, which secures the mean frailty is 1. With Weibull conditional hazards, the
survivor function is

S.t1 ; t2 / D 1=f1 C . 1 t1 C 2 t2 /=ıgı :

14 P. Hougaard

The value of Kendall’s  is 1=.1 C 2ı/. The Spearman correlation depends only on
ı, and is most naturally derived by numerical integration. So it is relatively simple
to quantify the dependence as it does not depend on the hazard function.
Describing the gamma frailty model is enough to illustrate the idea, but there are
other possibilities, where the most interesting are distributions with 1–2 parameters
having an explicit Laplace transform.

4 Multivariate Interval-Censored Data

Based on examinations at various time points, which may be common or individual

for the components, we derive interval observations. The observations will be of the
form .Lij ; Rij , where i refers to the group and j to the individual within the group.
Handling probabilities is more complicated in the bivariate case than in the
univariate. In the univariate case, the probability of an interval .a; b, that is,
P r.T 2 .a; b/, is found as S.b/S.a/, but in the bivariate case, the corresponding
formula is

P r.T1 2 .a1 ; b1 ; T2 2 .a2 ; b2 / D S.b1 ; b2 /  S.a1 ; b2 /  S.b1 ; a2 / C S.a1 ; a2 /:

(3)

This formula is easily generalized to higher dimensions, but the number of terms
quickly increases, being 2k for k-dimensional data, and therefore, it may be
preferable to keep the dimension low.
For setting up the likelihood, we then have to insert the L and R values in Eq. (3)
and multiply over subjects. We may also have covariates, which in general will
depend on both i and j , and thus be formulated as vectors zij .

4.1 Inspection Pattern

The inspection pattern becomes even more important in the multivariate case than
it was in the univariate case. In simple cases, like the dental example, all teeth of
a subject are considered simultaneously and the time of examination is determined
by external factors like a whole school being examined each year at the same time.
This is also called univariate monitoring times.
When there is only one such inspection for each subject, this is called multivariate
current status data (Jewell et al. 2005). In the case of multivariate current status data,
it is not in the non-parametric case possible to identify the full bivariate distribution
even asymptotically. More precisely, it is asymptotically possible to identify the two
marginal distributions S.t; 0/ and S.0; t/ and the diagonal S.t; t/, see Jewell et al.
(2005). So, in this case, one should be cautious with a non-parametric model. One
can either be satisfied with this limited identifiability, or use more restricted models,
Multivariate Interval-Censored Survival Data 15

like those presented in Sects. 5 or 7. More generally, we can say the inspection
process needs to be sufficiently rich in order to be able to determine the whole
distribution.
Also in this case, non-independent inspection is possible. For example, if one
subject in a couple is found to have an infectious disease, one would immediately
examine the other and due to the dependence this would not be independent
inspection. This case is not treated in the present paper.

5 Parametric Models

Here we consider a parametric model for the multivariate set of times to events, and
it is then assumed that we have interval-censored observations from this distribution.
One standard multivariate model is the normal distribution (for the logarithm
to the times). It is, however, more difficult to argue for its relevance to survival
data, and therefore, it has not been used much for such data. Besides, an explicit
expression for the survivor function is not available.
As we are considering survival times, a better alternative might be a frailty
model as described in Sect. 3.2. We will then have a range of choices for the hazard
function, like Weibull, Gompertz, piecewise constant hazards, or other standard or
non-standard distributions with a preference for distributions, which have explicit
hazard functions as well as explicit integrated hazard functions. There is also a range
of choices for the frailty distributions, with a preference for distributions with an
explicit Laplace transform. Above, only the gamma distribution has been described,
but there are other possibilities.
When the dimension is above two, the normal distribution is more flexible with
respect to dependence as the correlations can vary freely, whereas the standard
(shared) frailty models will have only a single dependence parameter; that is, in
spirit like the compound symmetry model for normally distributed data.
Estimation can follow two lines. One approach is to work directly with the
bivariate distribution, that is, the estimation is based on inserting the relevant
expressions in Eq. (3), with insertion of the right hand expression in Eq. (2). After
taking the logarithm, summing over subjects and differentiating, one can use the
Newton-Raphson method to find the estimate. Another approach, which is useful,
when the multivariate model is formulated by means of conditional independence,
like, for example, the frailty model, is to set up the likelihood as a product of
individual (univariate) contributions conditional on the frailty and then integrating
out over the frailty. This corresponds to using the middle expression in Eq. (2).
In some cases, this integration can be done automatically by the software. In
particular, it may be sensible, when the dimension of the data is high (in which
case, the generalization of Eq. (3) has too many terms), or when the Laplace
transform is not explicitly known (in which case the bivariate survivor function is
not explicitly known). An example of the latter type is a frailty model, where the
frailty follows a lognormal distribution. SAS proc nlmixed can accommodate both
lines of estimation.
16 P. Hougaard

6 Non-parametric Models

This approach assumes a completely general multivariate distribution. This is the

natural generalization of the non-parametric approach from the univariate case. This
implies, for example, that no covariates are accounted for. The purpose of using
this method is to avoid any parametric assumptions, which may be an advantage,
when we want to evaluate the goodness-of-fit of a specific parametric or semi-
parametric model. That is, the results of a parametric model can be compared to
the non-parametric estimate. It turns out that in this case, there are many difficulties
compared to the univariate case. From a practical point of view, it is difficult to
handle a multivariate non-parametric survivor function, and the practical problems
become worse with the dimension of the data.
Compared to the univariate case, the estimation method runs with exactly the
same steps. However, all of these steps are more complicated than in the univariate
case. First, one must select the intervals, where there can be positive mass. This
is possible, but it is more complicated than in the univariate case. For finding the
intervals, there is no longer a simple explicit approach, like the Peto approach.
Second, one must optimise the probabilities, and the likelihood in Eq. (1) should
be modified to account for Eq. (3). The likelihood still includes sums of interval
probabilities and thus the estimation procedure is formally the same as in the
univariate case. However, it is not in all cases possible to determine the interval
probabilities uniquely. This is a new problem compared to the univariate case,
where the interval probabilities are uniquely estimable. An example is shown in
Fig. 1, where there are four bivariate observations, .2; 3  .1; 6, .4; 5  .1; 6,
.1; 6  .2; 3, and .1; 6  .4; 5. The interval selection procedure leads to the
following four intervals that potentially can have mass, .2; 3  .2; 3, .2; 3  .4; 5,
.4; 5  .2; 3, and .4; 5  .4; 5, say they have probabilities p1 ; p2 ; p3 ; p4 . The
likelihood becomes .p1 C p2 /.p3 C p4 /.p1 C p3 /.p2 C p4 /. So even though these
are sums of interval probabilities, the terms are more complicated than for the
likelihood functions seen in the univariate case, because the subscripts to p are not
necessarily consecutive. For example, a term like .p1 C p3 / would not be possible
in the univariate case. This implies that the likelihood is not guaranteed to offer
a unique solution. The example illustrates the problem. The maximum likelihood
is obtained when p1 D p4 2 Œ0; 0:5 and p2 D p3 D 0:5  p1 . This will give
each observation a likelihood contribution of 1/2 and thus the likelihood value is
24 . If for some data, this problem occurs, it is easily detected because the matrix
of second derivatives is no longer negative definite. Yu et al. (2000) present the
example described above and suggest a procedure to select one of the parameter
values among those with maximal likelihood. They also show that asymptotically
any choice is satisfactory. That is, even though the problem is present also for large
samples, the probability that might be moved between sets becomes smaller.
Third, it can be difficult to illustrate the survivor function estimate graphically.
Even in the bivariate case, it is difficult to report the estimate in a table or figure,
and this makes the approach less attractive in many cases. This is a problem already
Multivariate Interval-Censored Survival Data 17

Fig. 1 Artificial data on four

bivariate observations,
illustrating non-identifiability.
Observations (dashed):
(1,6](2,3], (1,6] (4,5],
(2,3](1,6], and (4,5](1,6].
Sets with possible mass
shown in solid

in the right-censored case, as illustrated in Hougaard (2000, Chap. 14), which also
compares the various graphical approaches to illustrate a non-parametric bivariate
survivor function. The interval probabilities estimates could be reported in a (long)
table, but reporting it as a bivariate survivor function would be very difficult. The
problem is that the survivor function cannot be determined in all points, because we
need the probability of sets of the form fT1  t1 ; T2  t2 g, and when these sets cut
through a bivariate interval, which has positive estimate, we cannot determine the
probability of the set.
Furthermore, the completely non-parametric model does not give a quantifiable
expression for the dependence in terms of Kendall’s  and Spearman’s , for
example, and for the marginal distribution, there might be large sets, where it is
not identifiable.
Due to these problems and the technical problems, it may in many cases be
preferable to select another approach, like fitting a parametric or semi-parametric
model.

7 Semi-parametric Models

Semi-parametric modelling implies that some quantities are considered in a para-

metric way, whereas other are considered in a non-parametric way. This section
considers a non-parametric model for the hazard as function of time, but a
parametric model for the dependence between the coordinates. Thus, the section
18 P. Hougaard

does not allow for covariates. This model could be seen as a compromise between
the two previous suggestions. It extends the parametric models of Sect. 5 and it is a
submodel of the completely non-parametric model described in Sect. 6. This offers
technical advantages like a simpler way of finding the intervals with positive mass
and improved efficiency, but also advantages in interpretation, as it becomes easier
to quantify the dependence.
This could be done by describing the dependence by means of a frailty model
(with just a single or maybe two parameters to describe the dependence), and letting
the non-parametric component describe the univariate distributions. In practice, this
means that the frailty Y as defined above is shared between the family members,
respectively the teeth of an individual. The advantage of this model is that we
have a simple explicit expression for the bivariate survivor function and at the
same time it can allow for rather flexible hazard functions. The interval finding
procedure is reduced to the simple Peto procedure described above and Kendall’s 
and Spearman’s  can be estimated being functions of the frailty distribution. This
will also simplify the model in case of including covariates in a regression model.
Thus, one can see this method as a way of reducing the dimension of the problem,
as it reduces the non-parametric aspect from being multivariate to being univariate.
Of course, it can also be seen as a disadvantage, by making the model less flexible.
The purpose of such a study could be to study the dependence or to find
the marginal distribution accounting for dependence. To be more specific, the semi-
parametric model has the advantage that the dependence is easily quantifiable,
at least for some models. It is easy to reduce the problem to the intervals with
potentially positive mass. It is also easier to report the estimate as the non-parametric
component is only one-dimensional. However, we do not completely avoid the
problem of non-uniqueness of the estimate, but the problem will occur less
frequently.
The model will still suffer from the slow convergence (asymptotic order n1=3 )
for the survival time distribution, but the dependence parameter will show standard
results (asymptotic order n1=2 ), see Wang and Ding (2000).

8 Regression Models

The advantage of using a regression model is that the dependence on some

covariates is described in a simple and interpretable way that allows for testing of
the effect on the one hand, and on the other hand, if there is a significant effect, it
gives a quantifiable expression for the effect. Indeed, the purpose of such a study is
often to find these regression coefficients and to test whether a specified covariate
has an influence. The covariates are given as a vector, denoted z.
This paper will mention three types of regression models, which are first
described for the univariate case. The accelerated failure time model, which assumes
Sz .t/ D S0 .t= exp.0 z//, where S0 ./ denotes the survivor function for a subject with
0 value of all covariates, and  is the vector of regression coefficients. This model is
Multivariate Interval-Censored Survival Data 19

very easy to apply for parametric models. However, it is much more complicated
when the hazard is described non-parametrically, both for right-censored and
interval-censored data, and therefore not applied in practice. The proportional
0
hazards model assumes Sz .t/ D S0 .t/exp.ˇ z/ , where we have used a different
symbol for the regression coefficients, in order not to confuse the two models.
This model is very easy to apply both for parametric models and non-parametric
models for the hazard. Indeed, it is the standard model in the non-parametric case.
The final model is the conditional proportional hazards model, which assumes that
0
conditional on the frailty, Sz .t j Y / D S0 .t/Y exp.ˇ z/ . When Y is gamma distributed,
this makes an extension of the proportional hazards model, so that the hazards in the
marginal distribution (when Y is integrated out) are non-proportional. The case of
proportional hazards occurs on a boundary of the parameter set. The Weibull model
is both an accelerated failure time model and a proportional hazards model, and it
is the only model, where the two regression models coincide.
In the multivariate parametric case, we have a choice between these models, with
the first two models always applicable and the third model applicable, when the
dependence is given by a frailty model.
In the multivariate non-parametric case, the bivariate survivor function is not
defined without a description of the dependence structure. Without covariates, we
could let it be arbitrary, but with a covariate, we need to specify the relation between
the covariate and the dependence. An assumption of the dependence being of the
copula type would be sufficient to make the model well defined. Alternatively,
one could use a GEE (generalized estimating equation) type model, where the
dependence structure is not explicitly formulated.
In the multivariate semi-parametric case, we have a non-parametric model for
the hazard as function of time, but a parametric model for the dependence between
the coordinates and for the effect of covariates. Again, this model is well defined
for the first two regression models, whereas the last model is only defined when the
dependence is created by a frailty model. To make these considerations practical,
we assume that the dependence is indeed created by a frailty model. However, the
accelerated failure time model is not easy to fit with right-censored or interval-
censored data, so it is less useful in practice. For the other two models, where we
assume either proportional hazards in the marginal distributions, or proportional
hazards conditional on the frailty, fitting is possible, also with interval-censored
data. The probability will be concentrated on a set of intervals, which can be found
by the simple Peto procedure applied separately on each coordinate (if the hazards
are not shared over coordinates) or the combination of all marginal data (if the
hazards are shared). Furthermore, the dependence can be quantified by Kendall’s
 or Spearman’s , which are both functions of the frailty distribution. All in all,
this is quite operational.
The purpose of such a study could be to find the effect of covariates accounting
for dependence, or to evaluate the dependence accounting for the effect of selected
covariates. When the aim is to find the effect of a covariate, one should always
consider whether the same research question could be considered by means of
univariate data, because it will make it technically so much easier to handle.
20 P. Hougaard

9 Comparison and Discussion

First, we note that in the univariate case, the combination of interval censoring
and a non-parametric model is much more complicated than either alone (that is,
a parametric model for interval-censored data; or a non-parametric model for right-
censored data). These complexities carry over to the multivariate case.
A parametric model for multivariate interval-censored data is easy to handle,
when the dimension is low. When the dimension is high, it becomes more com-
plicated, because the number of terms in the generalization of Eq. (3) grows. Such
models will have the usual disadvantages of parametric models, namely that we need
to specify a sensible model, and the results might be more or less misleading, in the
case when the model is not satisfied. The assumptions for the parametric models will
have two directions, the (marginal) distributions for the times, and the dependence.
A non-parametric model for multivariate interval-censored data suffers from a
number of technical issues, finding intervals, handling intervals, potential parameter
non-uniqueness and difficulties in reporting the estimates. The latter problem
includes both reporting the survivor function and quantifying the dependence.
Overall, this makes it intrinsically more complicated than non-parametric models
for univariate interval-censored data. Therefore, this model is less attractive.
A compromise between these models is a semi-parametric model for multivariate
interval-censored data, where the dependence is generated by a frailty model
(with 1–2 parameters for the dependence). This will be more complicated than
the univariate non-parametric model for interval-censored data, but it will not
be intrinsically more complicated, and therefore tractable. In popular terms, the
non-parametric issues are reduced to dimension one. On the other hand, this model
is more flexible than parametric models, by allowing for non-parametric hazards.
We will still have to formulate models for the dependence, so the non-parametric
component only refers to assuming a general hazard over time.
All types of regression models are easily introduced into parametric models. For
the semi-parametric model, it is easy to handle the hazard-based models, whereas
the accelerated failure time model is complicated. The non-parametric model does
not allow for introducing regression effects, without restricting the dependence
structure (which implies that it is no longer a non-parametric model).
Weighing all the evidence above, the practical conclusion is that when we insist
on using the accelerated failure time model, only a parametric model makes sense,
whereas without this restriction, the semi-parametric approach is overall the most
attractive.
Finally, it should be said that interval censoring for multivariate data is so
complicated that we should always consider whether the conclusion can be drawn
based on univariate data only rather than using multivariate data.
Multivariate Interval-Censored Survival Data 21

References

Hougaard, P. (2000). Analysis of multivariate survival data. New York: Springer.

Hougaard, P. (2014). Analysis of interval-censored survival data. Book manuscript (in prepara-
tion).
Jewell, N., Van der Laan, M., & Lei, X. (2005). Bivariate current state data with univariate
monitoring times. Biometrika, 92, 847–862.
Peto, R. (1973). Experimental survival curves for interval-censored data. Applied Statistics, 22,
86–91.
Sun, J. (2006). The statistical analysis of interval-censored failure time data. New York: Springer.
Wang, W., & Ding, A. (2000). On assessing the association for bivariate current status data.
Biometrika, 87, 879–883.
Yu, Q., Wong, G., & He, Q. (2000). Estimation of a joint distribution function with multivariate
interval-censored data when the nonparametric mle is not unique. Biometrical Journal, 42,
747–763.
Multivariate Survival Models Based
on the GTDL

Gilbert MacKenzie and Il Do Ha

Abstract Correlated survival times may be modelled by introducing a random

effect, or frailty component, into the hazard function. For multivariate survival
data we extend a non-PH model, the generalized time-dependent logistic (GTDL)
survival model (MacKenzie, J R Stat Soc 45:21–34, 1996; MacKenzie, Stat Med
16:1831–1843, 1997), to include random effects. The extension leads to two
different, but related, non-PH models according to the method of incorporating
the random effects in the hazard function. The h-likelihood procedures of Ha et al.
(Biometrika 88:233–243, 2001) and Ha and Lee (J Comput Graph Stat 12:663–681,
2003), which obviate the need for marginalization (over the random effect distribu-
tion) are derived for these extended models and their properties discussed. The new
models are used to analyze two practical examples in the survival literature and the
results are compared with those obtained from fitting PH and PH frailty models.

Keywords Frailty models • Generalized time-dependent logistic • h-likelihood •

Non-PH model • Random effect

1 Introduction

Proportion hazards (PH) models (Cox 1972), extended to incorporate a frailty

component, are frequently used to analyze multivariate survival data which may
arise when recurrent or multiple event times are observed on the same subject.
However, the assumption of proportionality of the basic hazard function can
sometimes be untenable.

G. MacKenzie ()
The Centre for Biostatistics, University of Limerick, Limerick, Ireland
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 23

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__3,
© Springer International Publishing Switzerland 2014
24 G. MacKenzie and I.D. Ha

Accordingly, we investigate the utility of two flexible non-PH random-effect

models based on the generalized time-dependent logistic (GTDL) survival model,
MacKenzie (1996). The GTDL generalizes the relative risk (RR) in Cox’s semi-
parametric PH model to time-dependent form. The model, a wholly parametric
competitor for the PH model, has several interesting properties including a mixture
interpretation. In particular, by retaining Cox’s constant of proportionality as the
leading term in the time-dependent relative risk, the model is not only capable of
representing data which conform to the PH assumption, but can also accommodate
a wider class of survival data in which the assumption of proportionality does not
hold.
We thus extend the GTDL model to the multivariate survival data setting in
two ways, adopting the hierarchical-likelihood (h-likelihood) method of Ha et al.
(2001) and Ha and Lee (2003) for inference. In general, the h-likelihood approach
provides a unified inferential framework and a numerically efficient fitting algorithm
for various random-effect models (Lee and Nelder 1996, 2001) and, novelly, allows
us to consider time-dependent frailties. We use the new models to analyze two well-
known practical data sets from the literature and compare the results with those
obtained by fitting PH and PH frailty models.
The paper is organized as follows. In Sect. 2 we review the GTDL model briefly,
while in Sect. 3 we formulate the two extended models based on the GTDL. The
h-likelihood approach to inference is developed in Sect. 4, and in Sect. 5 the models
are used to analyze two well-known data sets which have appeared in the survival
literature, the results being compared with those obtained from the corresponding
PH models. Finally, some further discussion is given in Sect. 6.

2 The GTDL Regression Models

A non-PH model, the GTDL regression model (MacKenzie 1996) is defined by the
hazard function:

.tI x/ D 0 p.tI x/; (1)

where 0 > 0 is a scalar, p.tI x/ D exp.t˛ C x T ˇ/=f1 C exp.t˛ C x T ˇ/g is a linear

logistic function in time, ˛ is a scalar measuring the effect of time and ˇ is a p  1
vector of regression parameters associated with fixed covariates x D .x1 ; : : : ; xp /T .
The time-dependent relative risk, RR(t), the ratio of hazard rates for two subjects
with different covariate vectors, x1 and x2 , is given by

.tI x1 ; x2 / D .tI x1 /= .tI x2 / D expf.x1  x2 /T ˇg .tI x1 ; x2 /; (2)

Multivariate Survival Models 25

where

f1 C exp.t˛ C x2T ˇ/g

.tI x1 ; x2 / D :
f1 C exp.t˛ C x1T ˇ/g

The leading term on the right hand side of (2) is Cox’s constant of proportionality
(the RR in a PH model) and thus in the GTDL model this constant is moderated
by ./, a function of both time and covariates, demonstrating, unequivocally, that
the model is non-PH. Moreover, it should be noted that (2) does not depend on the
parameter 0 . When ˛ D 0 the relative risk does not depend on time and from (1)
the resulting model is PH—an Exponential with .tI x/ D 0 p.x/, i.e., a multiple
of the usual multiple logistic function (Cox 1970).
From (1), the cumulative hazard function is given by
Z  
t
0 1 C exp.t˛ C x T ˇ/
ƒ.tI x/ D .sI x/ds D log : (3)
0 ˛ 1 C exp.x T ˇ/

Under non-informative censoring the ordinary censored-data likelihood, which

depends on (1) and (3), is constructed and the maximum likelihood estimators for
the parameters can be obtained using numerical methods such as Newton–Raphson:
for more details see MacKenzie (1996, 1997).
It is often convenient to take 0 D 1. MacKenzie (1996, 1997) has shown that
this model is satisfactory in a variety of applications and leads to a formulation in
which the covariates act linearly on the log-odds hazard scale, rather than on the log
hazard scale, as in the PH model. For more details, including the role of this model
in the analysis of strata, see MacKenzie (1996).

3 Extended GTDL Models

The correlation between survival times, which arises in recurrent or multiple event
times on the same subject, may be modelled by introducing a frailty, or random
effect. We thus extend the non-PH model (1) to include random effects in two
different ways.
First we define the multivariate P data structures as follows. Let Tij .i D
1; : : : ; q; j D 1; : : : ; ni ; n D i ni / be the survival time for j th observation
of the i th subject and Cij be the corresponding censoring time. Let the observable
random variables be Yij D min.Tij ; Cij / and ıij D I.Tij  Cij /, where I./
is the indicator function. Denote by Ui the continuous random variable denot-
ing the unobserved frailty (or random effect) for the i th subject, with density
g.:j/.
26 G. MacKenzie and I.D. Ha

3.1 Non-PH Frailty Model

Our first extension includes a frailty term acting multiplicatively on the individual
hazard rate of (1). The GTDL non-PH frailty model is then defined as follows. Given
Ui D ui , the conditional hazard function of Tij takes the form

ij .tjui / D ij .t/ui ; (4)

where ij .t/ D 0 Œexp.tij ˛ C xijT ˇ/=f1 C exp.tij ˛ C xijT ˇ/g is given in (1)
and xij D .xij1 ; : : : ; xijp /T . The frailties Ui are assumed to be independent and
identically distributed random variables with a density function depending on the
frailty parameter . Note that if in the model (4) ij .t/ D 0 .t/ exp.xijT ˇ/ with
unspecified baseline hazard function 0 .t/, it becomes a semiparametric PH frailty
model, an extension of Cox’s model to allow frailty: see, for example, McGilchrist
and Aisbett (1991) and Ha et al. (2001).

3.2 Non-PH Random Effects Model

Secondly, we consider another natural extension of model (1), by including a

random component in the linear predictor, t˛ C x T ˇ, of (1). Let Vi D log Ui . The
GTDL non-PH random effect model is then defined as follows. Given Vi D vi , the
conditional hazard function of Tij is then of the form

exp.tij ˛ C xijT ˇ C vi /
ij .tjvi / D 0 : (5)
1 C exp.tij ˛ C xijT ˇ C vi /

The Vi are assumed to be independent and identically distributed random variables

with a density function depending on the frailty parameter .
Models (4) and (5) are similar, but the manner in which the random effect is
allowed to influence the basic GTDL hazard function differs in each case. That is,
(4) assumes that the random effects vi D loge .ui / act additively on the log hazard
scale while (5) assumes they are additive on a generalized loge -odds hazard scale,
which is the usual loge -odds hazard scale when 0 D 1. Note that model (4) is a
conventional frailty model, but (5) is not since the random effect does not operate
multiplicatively on the GTDL hazard function.
It may be observed that in model (5) the conditional log-odds hazard given vi is
linear in t. In general we have:
 
.tjv/
log D t˛ C x T ˇ C v
1  .tjv/
Multivariate Survival Models 27

which is a direct generalization of model (1) when 0 D 1. Thus, we note that in (5)
the quantity exp.t˛ C x T ˇ C v/ has a conditional odds ratio interpretation whereas
in the PH model it is a relative risk. However, in model (4) exp.t˛ C x T ˇ C v/ may
be shown to have a generalized conditional odds-ratio interpretation.
In practice, the choice of g.:j/ may be important. For h-likelihood inference,
the choice of parametric forms is wide and testable, since marginalization is not
required. In this paper we shall adopt the log-Normal distribution for h-likelihood
inference—a choice to which inference on ˇ is robust (Ha et al. 2001; Ha and Lee
2003). An alternative choice for model (4) is the Gamma distribution, see Blagojevic
et al. (2003) for a marginal likelihood approach in this scenario. For the use of other
frailty distributions see Hougaard (2000) (Chap. 7). Alternatively, one may adopt a
non-parametric mixture model.
Because of the convenience mentioned in Sect. 2, hereafter we adopt 0 D 1 in
(4) and (5), which gives a reduced form of each model.

4 H-Likelihood Estimation

4.1 Estimation of Non-PH Frailty Model

Estimation in the Non-PH GTDL family has been considered in detail by Ha

and MacKenzie (2010). However, for the sake of self-sufficiency and because we
consider two different parametrizations of the Non-PH GTDL, we reproduce the
main line below.
From Lee and Nelder (1996) and Ha et al. (2001), the h-likelihood for model (4),
denoted by h, is defined by
X X
h D h.˛; ˇ; / D `1ij C `2i ; (6)
ij i

where

`1ij D `1ij .˛; ˇI yij ; ıij jui /

D ıij log .yij jui /  ƒ.yij jui /
D ıij .log pij C vi / C ui ˛ 1 log.qij gij /

is the logarithm of the conditional density function for Yij and ıij given Ui D ui ,
and `2i D `2i .I vi / is the logarithm of the density function for Vi D log Ui
with parameter . For many models this device places the random effect, exactly
(or approximately) on the same scale as the linear predictor—see Lee and Nelder
(1996) for further details. With these arrangements the conditional hazard becomes
.yij jui / D pij ui where

pij D pij .˛; ˇ/ D exp.yij ˛ C xijT ˇ/=f1 C exp.yij ˛ C xijT ˇ/g

28 G. MacKenzie and I.D. Ha

and the conditional cumulative hazard ƒ.yij jui / D ui ˛ 1 log.qij gij / with qij D
1  pij and gij D gij .ˇ/ D 1 C exp.xijT ˇ/.
The maximum h-likelihood (MHL) joint estimating equations of  D
.˛; ˇ T ; vT /T with v D .v1 ; : : : ; vq /T are given by

@h=@ D 0:

The score equations are then:

X
@h=@˛ D fıij qij yij  .ui =˛/pij yij  .ui =˛ 2 / log.qij gij /g;
ij
X
@h=@ˇk D fıij qij C .ui =˛/.rij  pij /gxijk .k D 1; : : : ; p/; (7)
ij
X
@h=@vi D fıij C .ui =˛/ log.qij gij /g C @`2i =@vi .i D 1; : : : ; q/;
j

where ui D exp.vi / and rij D rij .ˇ/ D exp.xijT ˇ/=f1 C exp.xijT ˇ/g. The first two
equations depend on  only through v D .v1 ; : : : ; vq /T as does the first member of
the third equation. However, the second member of the third equation is, in general,
a function of the frailty parameter . Given , the estimating equations (7) are easily
solved using the Newton–Raphson method and Lee and Nelder (1996, 2001) and Ha
et al. (2001) have shown that the asymptotic covariance matrix for O   is given by
the inverse of H D @2 h=@ 2 .
In order to estimate the frailty parameter, , we adopt the restricted likelihood
(Lee and Nelder 1996, 2001) (or adjusted profile h-likelihood) approach which
yields a likelihood hP ./ for , after eliminating , defined by

hP ./ D hA j DO ; (8)

where hA D h C 12 logfdet.2H 1 /g and O D ./.

O Given estimates of , Lee and
Nelder (2001) REML (restricted maximum likelihood) estimating equation for ,
maximizing hP , is given by

@hA =@j DO D 0: (9)

4.2 Estimation of Non-PH Random Effect Model

The corresponding h-likelihood for the model (5) is given by

X X
h D h.˛; ˇ; / D `1ij C `2i ; (10)
ij i
Multivariate Survival Models 29

where

`1ij D `1ij .˛; ˇI yij ; ıij jvi /

D ıij log .yij jvi /  ƒ.yij jvi /
D ıij log pij0 C ˛ 1 log.qij0 gij0 /

is the logarithm of the conditional density function for Yij and ıij given Vi D vi , and
`2i D `2i .I vi / is the logarithm of the density function for Vi with parameter .
Here, the conditional hazard is .yij jvi / D pij0 where

pij0 D pij0 .˛; ˇ; vi / D exp.yij ˛ C xijT ˇ C vi /=f1 C exp.yij ˛ C xijT ˇ C vi /g

and the conditional cumulative hazard ƒ.yij jvi / D ˛ 1 log.qij0 gij0 / with qij0 D 1 
pij0 and gij0 D gij0 .ˇ; vi / D 1 C exp.xijT ˇ C vi /. In this case the score equations for
 D .˛; ˇ T ; vT /T are given by
X
@h=@˛ D fıij qij0 yij  .1=˛/pij0 yij  .1=˛ 2 / log.qij0 gij0 /g;
ij
X
@h=@ˇk D fıij qij0 C .1=˛/.rij0  pij0 /gxijk .k D 1; : : : ; p/;
ij
X
@h=@vi D fıij qij0 C .1=˛/.rij0  pij0 /g C @`2i =@vi .i D 1; : : : ; q/;
j

where rij0 D rij0 .ˇ; vi / D exp.xijT ˇ C vi /=f1 C exp.xijT ˇ C vi /g.

With the h-likelihood (10) and the score equations given above, the fitting
procedure outlined in the previous section can be applied directly to model (5).

5 Examples

We illustrate the use of the proposed models by analyzing two well-known

multivariate survival data sets. For the purposes of comparison, we include two
classical PH models—Cox’s PH and PH frailty model. For the reasons given above,
we have chosen the distribution of the random effect to be log-Normal. Accordingly,
we consider the following five models:
• M1 (PH): Cox’s PH model without frailty component.
• M2 (PHF): Cox’s PH model with frailty component.
• M3 (NPH): GTDL non-PH model without frailty component.
• M4 (NPHF): GTDL non-PH with frailty component.
• M5 (NPHR): GTDL non-PH with random-effect component.
Here M1 and M2 are PH models, whereas M3, M4 and M5 are non-PH models.
30 G. MacKenzie and I.D. Ha

Table 1 Analyses from various models for the kidney infection data
M1(PH) M1(PHF) M3(NPH) M4(NPHF) M5(NPHR)
Variable Est. SE Est. SE Est. SE Est. SE Est. SE
Intercept 3.069 0.527 2.065 0.783 2.099 0.773
Sex 0.830 0.297 1.368 0.427 0.940 0.288 1.542 0.445 1.521 0.438
Time (˛) 0.001 0.001 0.002 0.001 0.002 0.001
 0.509 0.656 0.663
2hP 370.0 364.7 688.0 679.9 680.0
Est., estimate; SE, standard error;  , variance of random-effect distribution; M1(PH), Cox’s
PH model without frailty component; M2(PHF), Cox’s PH model with frailty component;
M3(NPH), GTDL non-PH model without frailty component; M4(NPHF), GTDL non-PH with
frailty component; M5(NPHR), GTDL non-PH with random-effect component; hP , the restricted
likelihood, given in (8)

5.1 Kidney Infection Data

McGilchrist and Aisbett (1991) presented a small data set, which describes times to
the first and second recurrences of infection in 38 kidney patients being treated by
portable dialysis. Infections may occur at the location where the catheter is inserted.
The catheter is removed if infection occurs and can be removed for other reasons, in
which case, the observation is censored. Thus, survival time is defined as the time
from insertion of the catheter to infection. Survival times measured on the same
patient may be correlated because they share the same patient characteristics. For
the purpose of illustration, we consider fitting a single fixed covariate—sex of the
patient, coded as 1 for female and as 0 for male. The results of fitting the five models
are summarized in Table 1.
Overall, the results from models M1 and M3, which do not contain random
components, are similar, as are the results from models M2, M4 and M5, in which
random components have been incorporated. In particular, the non-PH random-
effect models M4 and M5 lead to similar results, despite the fact that in model M4
the random effect is not quite on the same scale as the linear predictor. Overall, this is
reassuring given the similarity of the functional forms involved in their formulation.
For testing the absence of a random component (i.e. H0 W  D 0), we use
the deviance (2hP in Table 1) based upon the restricted likelihood hP (Lee and
Nelder 2001). Because such a hypothesis is on the boundary of the parameter space,
the critical value is 22 for a size test. This value results from the fact that the
asymptotic distribution of likelihood ratio test is a 50:50 mixture of 20 and 21
distributions (Chernoff 1954; Self and Liang 1987): for application to random-effect
models, see Stram and Lee (1994) and Vu and Knuiman (2002) and Ha and Lee
(2005).
Firstly, for the PH models, M1 and M2, testing H0 W  D 0 yields a difference in
deviances of 5.28, indicating that the frailty component is necessary, i.e.  > 0.
In other words, the correlation between survival times is significant at the 5 %
level ( 21;0:10 D 2:71). Moreover, we conclude that M2 with frailty is superior to
Multivariate Survival Models 31

M1 without it. Secondly, for the three non-PH models (M3, M4 and M5), testing
H0 W  D 0 yields differences 8.09 and 7.96, from M3 and M4 and M3 and M5,
respectively. Thus, models M4 and M5 also explain the correlation and are to be
preferred to M3. In particular, when comparing M4 and M5 with M3, the sign of
the time coefficient ˛ changes, from negative to positive, suggesting that the frailty
has been taken into account successfully, i.e., the decreasing time-trend in hazard
has been abolished. A similar finding is sometimes observed in M3 when important
missing covariates are added—see MacKenzie (1996) for further details.
In all five models the sex effect is highly statistically significant indicating that
the female patients are at lower risk of infection compared with males. However, the
absolute magnitude of the fixed effect (and its SE) in both M1 and M3 is smaller
than in M2, M4 and M5, presumably, because the former models fail to account for
the correlation between survival times on the same patient—see Ha et al. (2001) for
more detail. In relation to models M2, M4 and M5 the results are similar, both with
respect to the magnitude of the fixed effect, ˇO (and its SE) and with respect to ,
O the
estimate of the frailty parameter.
Table 1 suggests that models M2, M4 and M5, all of which incorporate random
components, provide reasonable explanations for the correlated survival data.
However, models M4 and M5 cannot be formally justified, on this occasion, since
˛O is not significantly different from zero.

5.2 CGD Data

Fleming and Harrington (1991) provide more extensive multivariate survival data
on a placebo-controlled randomized trial of gamma interferon (
-IFN) in chronic
granulomatous disease (CGD). The aim of the trial was to investigate the effec-
tiveness of the
-IFN in reducing the rate of serious infections in CGD patients.
In this study, 128 patients were followed for approximately 1 year. Out of the 63
patients in the treatment group, 14 patients experienced at least one infection and
a total of 20 infections were recorded. In the placebo group, 30 out of 65 patients
experienced at least one infection, with a total of 56 infections being recorded. Here,
the survival times are the times between repeated CGD infections on each patient
(i.e., gap times). Censoring occurred at the last observation on all patients, except
one, who experienced a serious infection on the date he left the study. The recurrent
infection times for each patient are likely to be correlated as in the kidney infection
data study.
We fitted the same set of fixed covariates considered by Yau and McGilchrist
(1998), namely: treatment (0 D placebo, 1 D
-IFN), pattern of inheritance
(0 D autosomal recessive, 1 D X-linked); age (in years); height (in cm); weight
(in kg); using corticosteroids at time of study entry (0 D no, 1 D yes); using
prophylactic antibiotics at time of study entry (0 D no, 1 D yes); sex (0 D male,
1 D female), hospital region (0 D U.S., 1 D Europe), and a longitudinal variable
representing the accumulated time from the first infection in years. The rationale
32 G. MacKenzie and I.D. Ha

Table 2 Analyses from various models for the CGD data

M1(PH) M1(PHF) M3(NPH) M4(NPHF) M5(NPHR)
Variable Est. SE Est. SE Est. SE Est. SE Est. SE
Intercept 5.551 1.080 5.706 1.298 5.707 1.299
Inheritance 0.634 0.287 0.650 0.358 0.666 0.288 0.684 0.359 0.684 0.359
Age 0.082 0.037 0.085 0.043 0.084 0.037 0.087 0.043 0.087 0.043
Height 0.007 0.010 0.008 0.013 0.007 0.010 0.008 0.013 0.008 0.013
Weight 0.012 0.016 0.010 0.020 0.012 0.016 0.011 0.020 0.011 0.020
Corticosteroids 1.793 0.594 1.962 0.795 1.851 0.593 2.039 0.801 2.034 0.797
Prophylactic 0.542 0.324 0.662 0.421 0.550 0.325 0.660 0.422 0.660 0.421
Sex 0.709 0.394 0.751 0.496 0.752 0.394 0.796 0.498 0.795 0.497
Hospital region 0.636 0.318 0.688 0.377 0.655 0.318 0.712 0.377 0.711 0.377
Longitudinal 1.442 0.467 0.914 0.505 1.605 0.459 1.062 0.500 1.067 0.501
T (˛) 0.003 0.001 0.004 0.002 0.004 0.002
 0.508 0.511 0.513
2hP 694.6 690.5 1,074.2 1,069.9 1,069.8
Est., estimate; SE, standard error;  , variance of random-effect distribution; M1(PH), Cox’s
PH model without frailty component; M2(PHF), Cox’s PH model with frailty component;
M3(NPH), GTDL non-PH model without frailty component; M4(NPHF), GTDL non-PH with
frailty component; M5(NPHR), GTDL non-PH with random-effect component; hP , the restricted
likelihood, given in (8)

under-pinning this variable is that the infection rate may increase over time
following the first infection (Yau and McGilchrist 1998). The results pertaining to
the five models are given in Table 2.
Overall, the results are similar to those for the kidney infection data, except for
the findings in relation to time-dependent effects. In particular, the interpretation of
the beneficial effect of treatment is unequivocal in all models. The results for models
M4 and M5 are very similar in these data, the time parameter ˛O is statistically
significant in each, showing that the PH-class is formally inappropriate. However,
the magnitude of the dependence of the hazard on time is so small as to be almost
immaterial in this case. However, the non-PH random-effect models M4 and M5
suggest that the longitudinal effect is on the borderline of statistical significance.
Models M1 and M3 also support the existence of the effect, but rather exaggerate
its significance, especially M3. On the other hand, further evidence for its existence,
may be adduced from the analysis by Yau and McGilchrist (1998), which identified a
longitudinal effect using a more complicated PH model with time-dependent AR(1)
frailties. This analysis highlights the advantages of fitting a variety of different
models to the same dataset, rather than drawing conclusions from a single model
class (typically PH).
Multivariate Survival Models 33

6 Discussion

The Cox model with frailty is based on a conditional PH hazard function. In

general, the marginal survival model, obtained by integrating out the frailty from
the conditional model, will be non-PH (Hougaard 2000). Exactly how the marginal
model deviates from proportionality is unknown for many frailty distributions,
including the log-Normal distribution adopted in this paper. Accordingly, there is
considerable scope for expanding the class of non-PH models available in this
setting. This consideration led us to two new non-PH models based on the GTDL.
Both models were easily implemented in the h-likelihood framework and have
proved flexible tools for analyzing PH or non-PH correlated data. An interesting
point is that the similarity of the conclusions suggests that the use of the canonical
scale (Lee et al. 2006) for random effects and fixed effects is not strictly necessary,
thereby extending the envelope within which h-likelihood methods work.
It remains to be seen which of the two models proves the more successful in
practice. Overall, the findings suggest that the models are capable of representing
data which are more non-PH than the datasets analyzed in this paper (i.e., when
˛ 6 0/.
It should be noted that both of these models, are wholly parametric competitors
for the PH frailty model, and are more convenient computationally, as all of the
time-dependent quantities of interest (hazard trend, odds-ratios and relative risks)
may be readily derived.
From Tables 1 and 2 we have observed that the values of deviance (2hP )
in PH models (M1 and M2) are very different from those in non-PH models
(M3, M4 and M5). The reason is that the former models are semiparametric with
nonparametric baseline hazard, while the latter models are wholly parametric. Thus,
developing a suitable measure for model selection among these PH and non-PH
models would be an interesting piece of future work.

Acknowledgements The work in this paper was supported by the Science Foundation Ireland’s
(SFI, www.sfi.ie) project grant number 07/MI/012 (BIO-SI project, www3.ul.ie/bio-si) and by
KOSEF, Daejeon, South Korea.

References

Blagojevic, M., MacKenzie, G., & Ha, I. (2003). A comparison of non-ph and ph gamma frailty
models. In Proceedings of the 18th International Workshop on Statistical Modelling, Leuven,
Belgium (pp. 39–44).
Chernoff, H. (1954). On the distribution of the likelihood ratio. The Annals of Mathematical
Statistics, 25, 573–578.
Cox, D. R. (1970). The analysis of binary data. London: Methuen.
Cox, D. R. (1972). Regression models and life-tables (with discussion). Journal of the Royal
Statistical Society, 34, 187–220.
34 G. MacKenzie and I.D. Ha

Fleming, T. R., & Harrington, D. P. (1991). Counting processes and survival analysis. New York:
Wiley.
Ha, I. D., & Lee, Y. (2003). Estimating frailty models via poisson hierarchical generalized linear
models. Journal of Computational and Graphical Statistics, 12, 663–681.
Ha, I. D., & Lee, Y. (2005). Multilevel mixed linear models for survival data. Lifetime Data
Analysis, 11, 131–142.
Ha, I. D., Lee, Y., & Song, J. (2001). Hierarchical likelihood approach for frailty models.
Biometrika, 88, 233–243.
Ha, I. D., & MacKenzie, G. (2010). Robust frailty modelling using non-proportional hazards
models. Statistical Modelling, 10(3), 315–332.
Hougaard, P. (2000). Analysis of multivariate survival Data. New York: Springer.
Lee, Y., Nelder, J. A. (1996). Hierarchical generalized linear models (with discussion). Journal of
the Royal Statistical Society, 58, 619–678.
Lee, Y., & Nelder, J. A. (2001). Hierarchical generalised linear models: A synthesis of generalised
linear models, random-effect models and structured dispersions. Biometrika, 88, 987–1006.
Lee, Y., Nelder, J. A., & Pawitan, Y. (2006). Generalised linear models with random effects: Unified
analysis via h-likelihood. London: Chapman and Hall.
MacKenzie, G. (1996). Regression models for survival data: The generalised time dependent
logistic family. Journal of the Royal Statistical Society, 45, 21–34.
MacKenzie, G. (1997). On a non-proportional hazards regression model for repeated medical
random counts. Statistics in Medicine, 16, 1831–1843.
McGilchrist, C. A., & Aisbett, C. W. (1991). Regression with frailty in survival analysis.
Biometrics, 47, 461–466.
Self, S. G., & 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.
Stram, D. O., & Lee, J. W. (1994). Variance components testing in the longitudinal mixed effects
model. Biometrics, 50, 1171–1177.
Vu, H. T. V., & Knuiman, M. W. (2002). A hybrid ml-em algorithm for calculation of maximum
likelihood estimates in semiparametric shared frailty models. Computational Statistics and Data
Analysis, 40, 173–187.
Yau, K. K. W., & McGilchrist, C. A. (1998). Ml and reml estimation in survival analysis with time
dependent correlated frailty. Statistics in Medicine, 17, 1201–1213.
Frailty Models with Structural Dispersion

Joseph Lynch and Gilbert MacKenzie

Abstract In observational survival studies unmeasured covariates are the norm. In

univariate survival data we allow for their influence by means of frailty models.
Below we develop two Gamma frailty models with parametric basic hazard
functions: one with a proportional basic hazard function and the other with a
non-proportional basic hazard function. It transpires that both of the resulting
marginal frailty models have non-proportional hazard functions. We use these
models to analyse survival from breast cancer in the local health authorities in the
West Midlands of England in the UK and show that North Staffordshire is at, or near
to, the bottom of the resulting league tables, confirming the findings of other workers
who analysed earlier data at a National level. We also introduce and explain the
notion of structural dispersion which generalises the frailty variance to a regression-
based, subject-specific form and show that the fit of the structural dispersion models
is superior to the classical Gamma frailty models, but that the applied results are
unchanged.

Keywords Frailty models • Generalized time-dependent logistic • Non-PH

model • Random effect • Structural dispersion • Weibull

1 Introduction

The proportional hazards (PH) model implies that the hazard function is fully
determined by the observed fixed-effect covariates. Because we are all biologically
different, the natural course of a given disease and the effect of its treatment
vary from person to person. When modelling survival, some of this heterogeneity

J. Lynch ()
The Centre for Biostatistics, University of Limerick, Limerick, Ireland
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 35

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__4,
© Springer International Publishing Switzerland 2014
36 J. Lynch and G. MacKenzie

may be explained in terms of the fixed-effects, i.e., the covariates. However, in many
settings not all of the relevant covariates may be measured. Indeed, from a scientific
perspective, the set of relevant covariates may not even be known. This provides an
intrinsic justification for always making allowance for unmeasured covariates. One
approach is by using the concept of frailty whereby an unobservable, non-negative,
random effect is introduced which multiplies on the usual basic hazard function,
as in

.tjx; u/ D u: .tjx/;

where u is the multiplicative random effect and .tjx/ is the basic hazard function.
A convenient choice for u is the Gamma distribution which we adopt below.
When the data are univariate, as in the time to a single event, the frailty, describes
the effect of unobserved covariates in the model for each individual. Thus, the
survival of an individual with u D 1 follows the basic hazard function, while
for individuals with u > 1, the basic hazard is increased and early failure is then
more likely. In multivariate settings with recurrent events, u is common to several
individuals and generates the correlation between the survival times.
In this short paper only univariate data are studied and we have elected to
present results for two wholly parametric frailty models with different basic hazard
functions, PH and non-PH. We generalise the parametric PH Weibull model to a
Gamma frailty model in order to illustrate our methods. However, not all survival
data are PH and a flexible non-PH model is the Generalised Time-Dependent
Logistic (GTDL) described by MacKenzie (1996). Accordingly, we generalise this
model to frailty form by using a multiplicative frailty term. Here, too, we assume
the random effect follows a Gamma distribution, whence both the resulting frailty
models studied have closed forms after marginalisation and are non-PH (Hougaard
2000).
Finally, we generalise both frailty models to have structural dispersion (Lee and
Nelder 2001) and investigate the performance of the models when analysing survival
in a large study of incident cases of breast cancer in the West Midlands of the UK.

2 Regression Models with Frailty

The failure time density f .tj; ˇ/ is the product of the hazard function .tj:/ and
the survivor function S.tj:/ where  is a vector valued parameter and ˇ is a p  1
vector of regression parameters. Frailty is represented by the random variable U
with density g.;  2 /, such that E.U / D 1 and Var.U / D  2 .
The marginal density is found by integrating out the random effects and the
marginal likelihood function Lf is obtained from the joint probability of this
marginal density for i D 1; : : : ; n subjects with individual survival times ti ,
covariates xi , censoring indicators ıi and frailty components ui ,
Frailty Models 37

n Z
Y 1
Lf .; ˇ;  2 / D .ti jui ; ; ˇ/ıi .ti jui ; ; ˇ/g.ui j 2 /dui
i D1 0

and this forms a natural vehicle for inference since it retains the properties of a
conventional (Fisherian) likelihood.

2.1 Weibull Model

The Weibull model has a hazard function given by

.tjx/ D .t /1 exp.x 0 ˇ/;

where 0  t < 1. ˇ is a p  1 vector of regression parameters associated with

fixed covariates x 0 D .x1 ; : : : ; xp /. The scale parameter and the shape parameter
 are both > 0. The censored log-likelihood is

X
n
 
l. ; ; ˇ/ D ıi .loge  C loge C .  1/loge ti C xi0 ˇ/  . ti / exp.xi0 ˇ/ :
i D1

For the Weibull-gamma frailty model, we assume the random component U

has a multiplicative effect on the hazard, such that .t; x; u/ D u .t; x/. .1;  2 /
distribution, then
1
u  2 1 exp. u
2
/
g.uj / D
2
1
:
2
. 12 / 2 

Integrating out the random effect, we get the frailty survival distribution,
1
Sf .tjx/ D Œ1 C  2 .t / exp.xi0 ˇ/  2 :

The marginal hazard is

.t /1 exp.xi0 ˇ/
f .tjx/ D :
1 C  2 .t / exp.xi0 ˇ/

The log likelihood is

n h
X i
ıi loge .  ti exp.xi0 ˇ//  .ıi C  2 /loge . 2 .ti / exp.xi0 ˇ/ C 1/ :
1
l. ; ; ˇ;  2 / D
i D1
38 J. Lynch and G. MacKenzie

2.2 GTDL Model

The GTDL model has a hazard function given by

exp.t˛ C x 0 ˇ/
.tjx/ D ;
1 C exp.t˛ C x 0 ˇ/

where ˛ is a time parameter, ˇ is a p  1 vector of regression parameters associated

with fixed covariates, x 0 D .x1 ; : : : ; xp / and > 0 is a scalar.
The corresponding survival function is:
  
1 C exp.t˛ C x 0 ˇ/ ˛
S.tjx/ D :
1 C exp.x 0 ˇ/

The censored log-likelihood is

X
n 
exp.ti ˛ C x 0 ˇ/
l. ; ˛; ˇ/ D Œıi loge C ıi loge C .loge gi C loge qi /;
i D1
1 C exp.ti ˛ C xi0 ˇ/ ˛

where
1
qi D ;
1 C exp.ti ˛ C xi0 ˇ/
gi D 1 C exp.xi0 ˇ/ :

Using similar techniques as for the Weibull model, the log likelihood for the frailty
distribution is

X n  
 2
l. ; ; ˇ;  2 / D ıi loge .pi /  .ıi C  2 /loge .1  loge .qi gi // :
i D1
˛

In evaluating the marginal frailty distribution for the GTDL model, we again
assume the frailty component U has a .1;  2 / distribution. The marginal survival
distribution is
  12
 2 
Sf .tjx/ D 1  loge .qi gi /
˛

and the marginal hazard is

pi
f D  2
:
1 ˛
loge .qi gi /
Frailty Models 39

See also Blagojevic et al. (2003) and Blagojevic and MacKenzie (2004) for other
details.

2.3 Goodness-of-Fit Measures

We compared the fit of the Weibull and GTDL models, without frailty, with frailty
and with structural dispersion using a range of techniques, i.e., AIC, BIC and
a modified 2 measure of goodness-of-fit. The latter test statistic is the sum of
the squared difference between the Kaplan Meier (KM) and the model survivor
functions estimates for each subject, divided by the subject’s model survivor
function estimates

X
n
fSKM .ti /  SModel .ti /g2
2 D ;
i D1
SModel .ti /

for i D 1; : : : ; n D 15;516 subjects, where  is unknown. This corresponds to a

quasi Chi-squared goodness-of-fit statistic in which the role of the observed values
is played by the KM survivor estimates and the role of the expected values are played
by the model survivor estimates. The statistic can be presented in graphical form as
the partial sum (y-axis) plotted against time (x-axis) enabling one to see where any
lack of fit emerges.

3 The Concept of Structural Dispersion

In many regression models we are used to modelling the mean structure in a

relatively sophisticated way, and, in survival analysis, the same is true for the hazard
function. This is very natural as the hazard function is the defining component of
the survival distribution.
On the other hand, MacKenzie and Pan (2007) have highlighted the importance
of joint mean-covariance modelling in longitudinal data setting. These ideas spill-
over to joint hazard-dispersion modelling in the frailty survival modelling context,
especially in multi-component frailty models. Moreover, they extend further into the
arena of model selection, e.g., Ha et al. (2007).
Meanwhile, let us consider why it is important to model the dispersion jointly
with the hazard. In many physical problems additional random components are often
required to accommodate departures from standard models (e.g., over-dispersion
in the Poisson model, random trajectories over time in longitudinal studies, etc.).
Put another way, these components account in some sense for variation which is
unexplained in the standard model. However, we may enquire where this variation
comes from and how it is structured in terms of intrinsic factors, such as baseline
40 J. Lynch and G. MacKenzie

covariates, in the original physical problem. Taking this view is more likely to be
persuasive, since “plucking” multiple random effects from the ether to improve the
fit of models often seems unrealistic in scientific terms. On the other hand, if it can
be shown that the random effects are required and that their variance is controlled
by factors intrinsic to the physical problem, the model gains additional scientific
credence. Thus, viewing the variance of the random effect as being controlled
by fixed effects may sometimes make sense and contribute to a greater scientific
understanding.
If we consider the simplest univariate frailty case, covariates which influence the
hazard function via the regression x 0 ˇ, may or may not, also influence the frailty
variance,  2 . Thus, just as we can assess the influence of factors on the hazard,
we can also assess the influence of the same or other factors, on the amount of
unexplained variation, via another regression, x 0 ˇ  . We can estimate ˇ and ˇ 
by their joint MLE .ˇ; O ˇO  / and hence obtain cov( Ǒ ; ˇO  ) by standard methods.
Accordingly, the univariate frailty models can be generalised by replacing  2 by
i2 , where

i2 D exp.xi0 ˇ  /:

Thus, in general, the frailty variance is not constant, as in the classical case, but is
subject-specific. The additional linear predictor, xi0 ˇ  , always contains an intercept
term, x0i D 1; 8i , so that when the frailty variance does not depend on the variable
components in xi we automatically recover the classical model. On the other hand,
the larger any particular component, the more that particular variable contributes to
the unexplained variation. Thus, when there is significant structure we can assess
the independent contribution of the individual components. It should be noted that
the structural dispersion model is not a random coefficient model.

4 Analysis

Using a relative survival approach, Coleman (1998) reported that North Stafford-
shire Local Health Authority (LHA) was ranked last of 99 LHAs in England &
Wales with respect to breast cancer survival. The method makes no allowance for
case-mix. Accordingly, we re-analyse an augmented dataset from the West Midlands
of England, including North Staffordshire, by more traditional methods and report
on the resulting case-mix adjusted league tables.

4.1 Data for Analysis

The data set obtained from the West Midlands Cancer Intelligence Unit, comprised
15,516 women diagnosed with incident breast cancer between 1991 and 1995
Frailty Models 41

Table 1 Five-year adjusted survival league tables

LHA SOG LHA SOGCGF LHA SOW LHA SOW CGF LHA SOGCSD LHA SOW CSD
Unknown 0.79 Unk 0.85 Unk 0.78 Unk 0.79 Unk 0.88 Unk 0.83
Birm’ham 0.74 Cov 0.76 Bir 0.75 Bir 0.74 Bir 0.71 Bir 0.71
Coventry 0.73 Bir 0.76 Her 0.74 Her 0.73 Cov 0.71 Her 0.71
Hereford 0.72 Her 0.75 Cov 0.74 Cov 0.72 Her 0.71 Cov 0.70
Solihull 0.72 Wolv 0.73 Warw 0.72 Wolv 0.67 Wolv 0.69 Sand 0.69
Shropsh 0.71 Sol 0.73 Sol 0.72 Sand 0.67 Shrop 0.69 Wolv 0.68
Warwick 0.71 Sand 0.73 Sand 0.72 Sol 0.67 Warw 0.69 Warw 0.68
Sandwell 0.70 Warw 0.73 Shrop 0.71 Worc 0.66 Sol 0.68 Shrop 0.68
Wolverhm 0.70 Shrop 0.72 Wolv 0.71 Warw 0.65 Sand 0.68 Sol 0.67
Worcester 0.69 Dudl 0.72 Worc 0.71 Shrop 0.65 Worc 0.68 Worc 0.67
Dudley 0.69 Worc 0.72 Dudl 0.70 Dudl 0.65 Dudl 0.68 Dudl 0.66
Walsall 0.68 Wals 0.70 Wals 0.70 Wals 0.6 Wals 0.66 NSt 0.66
NStaff 0.67 SSt 0.69 NSt 0.69 SSt 0.62 NSt 0.65 Wals 0.65
SStaff 0.67 NSt 0.69 SSt 0.68 NSt 0.62 SSt 0.64 SSt 0.64
G D GTDL, W D Weibull, GF D Gamma frailty, SD D structural dispersion

in the West Midlands. They were followed up over a 10-year period from 1991
to 2001. Nine categorical covariates were initially studied in each patient. They
were age, diagnosis basis, stage, morphology, screening, Townsend score, year of
diagnosis, LHA and treatment. Age refers to age at diagnosis. Diagnosis basis refers
to whether the diagnosis of breast cancer was made on clinical, histological or
cytological grounds. Stage refers to how far the tumour has spread from its site
of origin at the time of diagnosis. Morphology refers to the histological tissue type
of tumour. Screening detection refers to whether or not the patient had her breast
screened prior to referral. Townsend score is a social deprivation index. Years of
diagnosis range from 1991 to 1995, so earlier years had a longer follow up. There
were 13 different LHAs and one unknown category to which 94 patients were
allocated. Treatment includes the various combinations of surgery, chemotherapy,
radiotherapy and hormone therapy which each patient received during the course of
the study.

4.2 League Tables

Initial analysis using the Cox model showed that the PH assumption did not hold
up for some of the covariates studied. Accordingly, the survivor functions for
each model, with and without frailty, were computed at the mean of the baseline
covariates in all 14 LHAs and the 5-year survival figures were ranked (Table 1).
The results are broadly similar from model to model. Goldstein and Spiegelhalter
(1996) caution that league tables based on the rank of Health Authorities should
be interpreted with caution as small quantitative differences may result in large
changes in rank. However, it seems that, overall, Coleman’s National findings are
42 J. Lynch and G. MacKenzie

Table 2 Goodness-of-fit measures for the six models fitted

Model `.O / Parameters AIC BIC
Weibull 21,427.95 62 42,979.90 43,456.18
GTDL 21,430.64 62 42,985.28 43,459.56
Weibull Frailty 21,627.42 63 43,380.84 43,862.77
GTDL Frailty 21,300.73 63 42,727.46 43,209.39
Weibull SD 21,058.83 123 42,363.66 43,304.56
GTDL SD 20,977.94 123 42,201.88 43,142.78

Fig. 1 Comparison of 2 goodness-of-fit estimates for the six fitted models. Panels: left—basic
survival curves, right—frailty models and centre—frailty with structural dispersion. Weibull
(dashed), GTDL (solid) and Index = ordered time

corroborated by our more detailed regional analysis. Survival in North Staffordshire

is poor even after taking the effect of the measured and unmeasured covariates into
account by means of frailty. It should be noted that Birmingham, which is a regional
centre, dealing with the most difficult cases, comes top after adjusting for case-mix
variables.
On the whole, structural dispersion models fitted better than frailty models,
which, in turn were better fitting than the models without frailty (Table 2). The
cumulative quasi chi-squared measures of fit are presented in Fig. 1 which also
shows that the structural dispersion models fit best. However, these goodness-of-
fit measures are not concordant with the AIC and BIC information criteria in
Frailty Models 43

Table 2 when comparing Weibull and GTDL models. For example, in relation to
the structural dispersion models, Table 2 suggests that the GTDL model fits best but
Fig. 1 (centre panel) suggests that the Weibull model fits best. These inconsistencies
require further research pending which we prefer to rely on the classical information
criteria.

5 Discussion

The applied findings of this study largely confirm Coleman’s original conclusions
and will disappoint clinicians working on breast cancer care in North Staffordshire
in the West Midlands of the United Kingdom. In 1999, these clinicians were
skeptical about Coleman’s finding arguing that they were already out of date and
that, in the interim, many improvements to the service in North Staffordshire had
been made. However, it seems that our more sophisticated analysis of the local
cancer registry data over a longer period of time shows that survival in North
Staffordshire remains poor and that this LHA is at, or is near, the bottom of the
regional league table in terms of 5-year survival rates.
Our analyses take nine major covariates into account as well as allowing for
the non-PH nature of some of the covariates. We have also allowed for the effect
of unmeasured covariates by generalising the basic models to incorporate Gamma
frailty extensions. Usually, these models provide a better fit to the data than the more
basic models. Moreover, we have shown the value of extending the Gamma frailty
models to incorporate structural dispersion. It seems that there are real gains to be
made in the standard model selection criteria using this method. While this is an
interesting technical improvement, it seems that the applied findings are unchanged.
There are several additional analyses which remain to be carried out, but at the
time of writing, further studies aimed at improving survival in North Staffordshire
need to be designed and undertaken.

Acknowledgements The work in this paper was supported by IRCSET, Ireland, and by the
Science Foundation Ireland’s (SFI, www.sfi.ie) project grant number 07/MI/012 (BIO-SI project,
www3.ul.ie/bio-si).

References

Blagojevic, M., & MacKenzie, G. (2004). A comparison of ph and non-ph frailty models. In
Correlated data modelling (pp. 239–249). Milano: Franco Angeli.
Blagojevic, M., MacKenzie, G., & Ha, I. (2003). A comparison of non-ph and ph gamma
frailty models. In Proceedings of the 18th International Workshop on Statistical Modelling
(pp. 39–44). Statistical, Modelling Society, Leuven.
Coleman, M. (1998). Cancer survival in the health authorities of England up to 1998. A report
prepared for the national health services executive.
44 J. Lynch and G. MacKenzie

Goldstein, H., & Spiegelhalter, D. (1996). League tables and their limitations. Journal of the Royal
Statistical Society, 59, 385–343.
Ha, I., Lee, Y., & MacKenzie, G. (2007). Model selection for multi-component frailty models.
Statistics in Medicine, 26(26), 4790–4807.
Hougaard, P. (2000). Analysis of multivariate survival data. New York: Springer.
Lee, Y., & Nelder, J. A. (2001). Hierarchical generalised linear models: A synthesis of generalised
linear models, random-effect models and structured dispersions. Biometrika, 88, 987–1006.
MacKenzie, G. (1996). Regression models for survival data: The generalised time dependent
logistic family. Journal of the Royal Statistical Society, 45, 21–34.
MacKenzie, G., & Pan, J. (2007). Optimal covariance modelling. In Correlated data modelling.
Milano: Franco Angeli.
Random Effects Ordinal Time Models
for Grouped Toxicological Data from
a Biological Control Assay

Marie-José Martinez and John P. Hinde

Abstract Discrete survival times can be viewed as ordered multicategorical data.

Here a continuation-ratio model is considered, which is particularly appropriate
when the ordered categories represent a progression through different stages, such
as survival through various times. This particular model has the advantage of being a
simple decomposition of a multinomial distribution into a succession of hierarchical
binomial models. In a clustered data context, random effects are incorporated
into the linear predictor of the model to account for uncontrolled experimental
variation. Assuming a normal distribution for the random effects, an EM algorithm
with adaptive Gaussian quadrature is used to estimate the model parameters. This
approach is applied to the analysis of grouped toxicological data obtained from a
biological control assay. In this assay, different isolates of the fungus Beauveria
bassiana were used as a microbial control for the Heterotermes tenuis termite, which
causes considerable damage to sugarcane fields in Brazil. The aim is to study the
pathogenicity and the virulence of the fungus in order to determine effective isolates
for the control of this pest population.

Keywords Adaptive quadrature • Clustered data • Multicategorical data • Ordi-

nal regression • Random effects

1 Introduction

Discrete survival times can be viewed as ordered multicategorical data. In the

ordinal data modelling context, a variety of multinomial regression models can
be used, including the baseline-category logit model, the cumulative logit model, the

John P. Hinde ()

School of Mathematics, Statistics and Applied Mathematics, National University of Ireland,
Galway, University Road, Galway, Ireland
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 45

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__5,
© Springer International Publishing Switzerland 2014
46 M.-J. Martinez and J.P. Hinde

adjacent-category logit model or the continuation-ratio logit model. This last model
has been given some attention in the literature (Agresti 2002). Such a model form is
useful when the ordered categories represent a progression through different stages,
such as survival through various times. This particular model has the advantage of
being a simple decomposition of a multinomial distribution into a succession of
hierarchical binomial models. The property of conditional independence enables us
to fit it by adapting the methods available for binary response data (Agresti 2002).
When one has ordered replicated data, random effects can be incorporated into
the linear predictor to account for uncontrolled experimental variation (this may be
apparent through overdispersion of multinomial responses across the replicates).
An increasing number of papers are concerned with random effects models for
ordered categorical responses. For example, Stram et al. (1988) fit a cumulative
model separately for each time in a repeated measures context. In a psychological
study, Ten Have and Uttal (1994) present subject-specific and population-averaged
continuation ratio logit models for multivariate discrete time survival data. In their
paper, they consider a modified Gibbs sampling algorithm to estimate the parameters
of the subject-specific model. Tutz and Hennevogl (1996) consider the general
case of an ordinal cumulative model including random effects and develop three
alternative estimation procedures based on the EM-algorithm.
Here, we focus on random effects continuation-ratio models. We consider a
continuation-ratio model and include a random intercept into the linear predictor in
order to analyse grouped toxicological data with overdispersion. More specifically,
the data considered here have been obtained from a biological control assay
carried out by the Insect Pathology Laboratory of ESALQ-USP, Sao Paulo, Brazil
(De Freitas 2001). In this assay, different isolates of the fungus Beauveria bassiana
are used as a microbial control for the Heterotermes tenuis termite which causes
much damage to sugarcane fields in Brazil. In this context, experiments have been
carried out to study the pathogenicity and the virulence of the fungus in order to
determine effective isolates for the control of this pest population. The full data set
compares 142 separate isolates of the fungus. A solution of each isolate was applied
to groups of 30 termites, with five independent replicates for each isolate, and the
cumulative mortality in each group is measured daily during an eight-day period
after the application of the fungus. Figure 1 displays the cumulative proportions
of dead termites for 30 different isolates showing the five replicates. It clearly
shows different isolate efficacies and different degrees of variability among the
replicates within the different isolates. The aim of this study is to determine effective
isolates for use in the field taking into account the replicated data structure.
To fit the proposed random effects models, we consider here different approaches
to maximum likelihood estimation. As in the general case of generalized linear
mixed models, the likelihood of the observed data is obtained by integrating out
the random effects. Unfortunately, this marginal likelihood does not generally have
a closed form expression and some form of approximation is needed for parameter
estimation. Classical methods commonly used include penalized quasi-likelihood
(PQL) or Gaussian quadrature (GQ) methods. The first method is based on a
decomposition of the data into the mean and an appropriate error term using a linear
Discrete Survival Times 47

2 4 6 8 2 4 6 8 2 4 6 8

25 26 27 28 29 30
1.0
0.8
0.6
0.4
0.2
0.0
19 20 21 22 23 24
1.0
0.8
0.6
cumulative proportions

0.4
0.2
0.0
13 14 15 16 17 18
1.0
0.8
0.6
0.4
0.2
0.0
7 8 9 10 11 12
1.0
0.8
0.6
0.4
0.2
0.0
1 2 3 4 5 6
1.0
0.8
0.6
0.4
0.2
0.0
2 4 6 8 2 4 6 8 2 4 6 8

time

Fig. 1 Cumulative proportions of dead termites for 30 isolates with five replicates of each

Taylor expansion of the mean around current estimates of the fixed and random
effects. We refer to Breslow and Clayton (1993) for more details. The second
method is a standard approach that evaluates the likelihood numerically using
Gauss-Hermite quadrature. An improvement on Gaussian quadrature is adaptive
Gaussian quadrature (AGQ), which essentially consists of shifting and scaling the
quadrature locations to adapt them to the integrand. Assuming normal distributions
for the random effects, we propose here to use AGQ methods combined with an
EM-algorithm to estimate the model parameters.
The paper is organized as follows. The random effects continuation-ratio models
proposed to fit the grouped toxicological data set are developed in Sect. 2. The pro-
cedure considered to estimate the model parameters is given in Sect. 3. The results
of fitting the different proposed mixed models are presented in Sect. 4, while Sect. 5
discusses the advantages and the limits of this methodology.

2 Model Specification

Suppose the cumulative mortality is measured over d consecutive days. For the
j th replicate of isolate i , j D 1; : : : ; r and i D 1; : : : ; n, we denote the initial
number of insects by mij . Let Yij;k denote the number of dead insects on day k,
48 M.-J. Martinez and J.P. Hinde

P
k D 1; : : : ; d and Yij;d C1 D mij  dkD1 Yij;k the number of insects still alive on
day d . The probability of an insect dying on day k for isolate i and replicate j is
denoted by ij;k . For each isolate and replicate, we treat the counts in the d C 1
categories, Yij D .Yij;1 ; : : : ; Yij;d C1 /, as independent multinomials with probabilities
P C1
.ij;1 ; : : : ; ij;d C1 / where dkD1 ij;k D 1. These probabilities can then be modelled,
in terms of isolate and time covariates, through some suitably defined link function.
Here, we use the continuation-ratio logits defined as
!
ij;k
ij;k D log ;
ij;kC1 C    C ij;d C1

since these have a simple survival model interpretation, as we show below. Note that
other forms of ordered multinomial logits and links could also be used, see Agresti
(2002).
We now consider various model specifications for the linear predictor ij;k
containing isolate specific factors and time-related covariates in order to model the
time dependency. In addition, the variability observed among the replicates for some
isolates leads us to introduce an additive random effect into the linear predictor
as in Hinde (1982) and Nelder (1985). We use this random effect to account for
an additional component of variability between the experimental units, namely the
replicates for each isolate.
We first consider a model with a common arbitrary day effect, giving a baseline
cumulative mortality pattern, an isolate effect and a random effect with a linear
predictor defined by

Model I: ij;k D ˛i C ˇk C  ij ; i D 1; : : : ; n; j D 1; : : : ; r; k D 1; : : : ; d:

Here ˛i is the effect of isolate i , ˇk the time effect of day k,  is the random effect
scale parameter, ij  N .0; 1/ and the ij ’s are assumed independent. Note that we
could also have formulated the random effect term as uij  N .0;  2 / as in Brillinger
and Preisler (1983), however, the above formulation has the advantage of passing the
random effect parameter into the linear predictor and having a completely specified
(parameter-free) distribution for the unobserved variable ij . The two formulations
in fact lead to different EM algorithms with different convergence rates, see Hinde
(1997), although the basic computational approaches are the same.
Next we consider a second model in which we add an isolate specific time trend
to the previous linear predictor leading to

Model II: ij;k D ˛i C ˇk C

i tk C  ij ;

where tk D k is a quantitative variable giving a linear trend over days, and

i is
the linear time effect for isolate i allowing variation from the common baseline
cumulative mortality pattern.
Discrete Survival Times 49

The third model considered is a simplification that attempts to model the time
dependency purely in terms of an isolate specific linear time trend and excludes
the arbitrary common baseline pattern.

Model III: ij;k D ˛i C

i tk C  ij :

Finally, two additional models will be considered using quadratic rather than
linear isolate specific time trends. This is an attempt to capture the overall variation
in the response patterns over time.

Model IV: ij;k D ˛i C ˇk C

i tk C ıi tk2 C  ij ;

Model V: ij;k D ˛i C
i tk C ıi tk2 C  ij :

Note that all coefficients ˛i , ˇk ,

i and ıi , i D 1; : : : ; n and k D 1; : : : ; d , in the
five models are assumed to be constant over replicates.
These five models are all random intercept models in which the introduction
of an additive random effect allows a random location shift in the linear predictor
for each replicate of each isolate in an attempt to capture the appreciable within
isolate replicate variability. In this work, we consider a logit link leading to random
effects continuation-ratio logit models. Other link functions can also be used and
another common choice is the complementary log–log link yielding the so-called
proportional hazards model.
An advantage of the continuation ratio model is that it can be viewed as a
sequence of binomial models by considering the conditional probability, wij;k , that
an insect dies on day k given that it has survived up to this day, for isolate i and
replicate j . This conditional probability is defined by
ij;k
wij;k D Pd C1 :
k 0 Dk ij;k 0

Let b.mI yI w/ denote the binomial probability of obtaining y successes out

of m trials with probability w for each trial. The multinomial probability for
.yij;1 ; : : : ; yij;d C1 / can be expressed in the form

b.mij I yij;1 I wij;1 /b.mij yij;1 I yij;2 I wij;2 /  b.mij yij;1 : : :yij;d 1 I yij;d I wij;d /:

Thus, the multinomial model can be expressed as a succession of hierarchical

binomial models, see McCullagh and Nelder (1989). The continuation-ratio logits
are then defined as
!
ij;k
w 
ij;k
ij;k D log D log D logit.wij;k /;
ij;kC1 C    C ij;d C1 1  wij;k
50 M.-J. Martinez and J.P. Hinde

Table 1 Rearrangement of the data for the j th replicate of the i th isolate

Day No. at risk No. of deaths Prob(death)
1 mij yij;1 ij;1
ij;2
2 mij  yij;1 yij;2 wij;2 D 1 ij;1
ij;3
3 mij  yij;1  yij;2 yij;3 wij;3 D 1ij;1 ij;2
::
:
ij;d
d mij  yij;1      yij;d 1 yij;d wij;d D Pd 1
1 k 0 D1 ij;k 0

and are ordinary logits of the conditional probabilities wij;k . Clearly, a main
advantage of the continuation-ratio model is that it can be fitted using methods
for binomial logit models merely by a rearrangement of the data. Thus, to fit the
different logit models, we require a derived data structure as presented in Table 1, a
binomial model specification and appropriately defined linear predictors.

3 Parameter Estimation

In this section, parameter estimation for the random effects continuation-ratio mod-
els defined previously is considered based on the EM algorithm. This algorithm is a
powerful computational technique for maximizing likelihoods including unobserved
variables. However, as with the binary model, the non-conjugate normal distribution
for  means that the marginal likelihood cannot be obtained analytically. Indeed,
assuming that '.:/ denotes the standard normal density function, the likelihood of
the kth replicate of the i th isolate is given by

Z C1 Y
d
Lij .; / D f .yij;k j; ; ij / '.ij I 0; 1/ d ij
1 kD1
Z C1 Y
d Pk1
y
D ij;k
wij;k .1  wij;k /mij  k 0 D1 yij;k0
'.ij I 0; 1/ d ij
1 kD1
Z " #yij;k
C1 Y
d
exp.ij;k /
D
1 1 C exp.ij;k /
kD1
" #mij Pkk1
0 D1 yij;k 0
1
 '.ij I 0; 1/ d ij :
1 C exp.ij;k /

Clearly, this likelihood function has no closed form and has to be evaluated
numerically before being maximized as a function of the fixed effect parameters
 and the random effect parameter . Note that, for model IV for instance,  is the
Discrete Survival Times 51

vector of .˛i ; ˇk ;
i ; ıi / for all i and k. In this section, we consider two integration
methods for approximating the likelihood which will be then combined with an EM
algorithm for the maximization step.
First, we consider classical Gaussian quadrature to evaluate numerically this like-
lihood integral. Gauss-Hermite quadrature methods have already been used in the
important case of ordered categorical data. For instance, they have been considered
by Jansen (1990) for the case of shifted thresholds and by Hinde (1982) or Anderson
and Hinde (1988) for the binary case. The dimension of the integral determining
the likelihood function depends on the random effect structure. When the random
effects are assumed normally distributed and the dimension is small, as in the
integral defined above, Gaussian-Hermite quadrature methods can approximate the
likelihood function. In fact, Gauss-Hermite quadrature approximates the integral
of a function f .:/ multiplied by another function having the shape of a normal
density by a finite weighted sum of the function evaluated at a set of points called
the quadrature points. Thus, the likelihood is approximated by

( )
X
Q
Y
d
Lij .; /  vq f .yij;k j; ; zq / ;
qD1 kD1

with weights vq and quadrature points zq that are tabulated. Note that the approxima-
tion improves as the number Q of quadrature points increases. However, in practice,
a large number of quadrature points is often required to approximate correctly
the likelihood. Moreover, the approximation can be poor for large random effect
variances or can fail for small cluster sizes. We refer to Crouch and Spiegelman
(1990); Lesaffre and Spiessens (2001) or Albert and Follmann (2000), who point
out some of these problems.
To solve these problems associated with ordinary quadrature, we then consider
AGQ methods. An adaptive version of the Gauss-Hermite quadrature shifts and
scales the quadrature points to place them under the peak of integrand. Note that
after normalization with respect to ij , the integrand is the posterior density of ij
given the response and can be approximated for large sample sizes by a normal
density '.ij I ij ; ij2 / with mean ij and variance ij2 . In this version, the normal
density '.ij I ij ; ij2 / approximating the posterior density is treated as the weight
function. The integral is now written as
Z C1
Qd
f .yij;k j; ; ij / '.ij I 0; 1/
Lij .; / D '.ij I ij ; ij2 / kD1
d ij ;
1 '.ij I ij ; ij2 /

and applying the standard quadrature rules, the integral is now approximated by
( )
X
Q
Y
d
Lij .; /  vij;q f .yij;k j; ; zij;q / :
qD1 kD1
52 M.-J. Martinez and J.P. Hinde

Hence, the adaptive quadrature points are given by zij;q D ij zq C ij with
p z2
corresponding weights vij;q D 2ij exp. 2q /'.zij;q /vq .
Essentially, the posterior density is here approximated by a normal density with
the same mean and variance. However, the posterior mean and variance required
in this approach are not known and have to be computed. As in Rabe-Hesketh
et al. (2005), we obtain these posterior moments using adaptive quadrature leading
to an iterative integration. This approach is similar to the method described in
Naylor and Smith (1988) in a different context. Note that an alternative way to
approximate the posterior moments ij and ij2 is proposed by Liu and Pierce (1994)
by using the mode and the curvature at the mode. More precisely, this approach
centers the quadrature points with respect to the mode of the integrand and scales
them according to the estimated curvature at the mode.
In this work, quadrature methods are thus used to evaluate numerically the
integral in calculating the marginal likelihood. In general, the higher the order
Q, the better the approximation is. Typically, AGQ needs less quadrature points
than classical quadrature. On the other hand, AGQ is more time consuming since
the quadrature points and weights used in this approach depend on the unknown
parameters and hence will need to be updated in each step of the iterative estimation
procedure. The differences between ordinary and adaptive quadrature are discussed
in particular in Lesaffre and Spiessens (2001) or Rabe-Hesketh et al. (2002).
Finally, once the marginal likelihood is evaluated numerically for given param-
eter values, it has to be maximized with respect to  and . Several methods
for maximizing the likelihood can be considered and combined with the two
integration methods presented above. Rabe-Hesketh et al. (2005) use, for instance,
a Newton–Raphson algorithm where the Hessian matrix is obtained by numerical
differentiation. In this work, we consider an EM-algorithm which is easy to
implement compared to other optimization methods (Anderson and Hinde 1988).

4 Results

We now consider the analysis of the data of interest presented previously. The results
of fitting the models described in Sect. 2 are presented here. For simplicity, we
only consider a subset of 30 isolates. Obviously, results for all 142 isolates can
be obtained in the same way. Table 2 shows disparity and  values for models I,
II, III, IV, and V using ordinary Gaussian quadrature with 3, 5, 10, 20, 40, and 60
quadrature points and AGQ using 3, 5, and 10 quadrature points. For each of these
models, we also fit the associated fixed model obtained by dropping the random
effect i .
As expected, the different results show that those obtained by AGQ are the same
from using 10 quadrature points and more. For as few as 3 quadrature points, only
very small differences can be observed. On the other hand, the results, in particular
the variance estimates, change considerably using ordinary Gaussian quadrature.
Discrete Survival Times 53

Table 2 Disparity and  values for models I, II, III, IV and V using ordinary and adaptive
quadrature methods
GQ(3) GQ(5) GQ(10) GQ(20) GQ(40) GQ(60)
Model I 2 logL 4,937.8 4,899.9 4,924.4 4,933.6 4,940.0 4,940.2
 0.972 0.976 1.084 0.971 0.839 0.836
Model II 2 logL 4,269.2 4,269.7 4,274.7 4,287.8 4,294.8 4,294.9
 0.931 1.014 1.143 1.027 0.852 0.850
Model III 2 logL 5,202.6 5,197.8 5,189.2 5,195.0 5,204.8 5,205.0
 0.930 1.012 1.066 1.077 0.908 0.893
Model IV 2 logL 3,991.88 3,988.96 3,988.70 4,006.33 4,011.36 4,011.44
 0.910 0.984 1.162 0.990 0.840 0.837
Model V 2 logL 4,069.82 4,065.57 4,068.77 4,082.00 4,087.36 4,087.44
 0.913 0.989 1.112 1.007 0.844 0.841

AGQ(3) AGQ(5) AGQ(10) Fixed model

Model I 2 logL 4,941.0 4,940.2 4,940.2 5,651.8
 0.831 0.836 0.836 –
Model II 2 logL 4,295.7 4,294.9 4,294.9 5,014.5
 0.846 0.850 0.850 –
Model III 2 logL 5,206.0 5,205.1 5,205.1 6,029.7
 0.887 0.893 0.893 –
Model IV 2 logL 4,012.2 4,011.4 4,011.4 4,697.5
 0.833 0.837 0.837 –
Model V 2 logL 4,088.2 4,087.4 4,087.4 4,783.1
 0.837 0.841 0.841 –

Clearly, we need to increase the number of quadrature points to 40 for ordinary

quadrature in order to get similar results. Therefore, it is clear that we may be able
to achieve good accuracy with a smaller number of quadrature points using AGQ
instead of ordinary Gaussian quadrature for the different models I, II, III, IV or V.
Table 3 displays deviance, AIC and BIC values for models I, II, III, IV and
V using AGQ with 10 quadrature points. Deviance, AIC and BIC values for the
associated simple GLMs are also displayed. These fixed effect models led to an
AIC value of 5,725.8, 5,146.5, 6,149.7, 4,887.5 and 4,963.1 for models I, II, III,
IV and V, respectively. Introducing a random effect into the linear predictors of
these models improved these results, yielding, respectively, AIC values of 5,016.2,
4,430.9, 5,327.1, 4,207.4 and 4,269.4. Thus, the large variability in the data has been
captured by these random effects models.
As noted previously, fitting random effects models yields the AIC value 5,016.2
for Model I, 4,430.9 for Model II, and 5,327.1 for Model III. Model III which
imposes more structure on the linear predictor and implies a more specific form
of the response over time appears not convincing in this particular case. Clearly, this
linear trend model does not seem to capture all of the structure in the data. Adding
the time factor effect ˇj into the linear predictor makes the structure over time more
flexible reproducing an overall common baseline response pattern and allowing
54 M.-J. Martinez and J.P. Hinde

Table 3 Deviance, AIC and BIC values for Models I, II, III, IV, V and their associated fixed
models
Model I Model II Model III Model IV Model V
Deviance 3,329.9 2,684.6 3,594.8 2,401.2 2,477.2
AIC 5,016.2 4,430.9 5,327.1 4,207.4 4,269.4
BIC 5,209.6 4,777.0 5,637.6 4,706.3 4,732.6
Fixed I Fixed II Fixed III Fixed IV Fixed V
Deviance 4,041.6 3,404.2 4,419.4 3,087.2 3,172.8
AIC 5,725.8 5,146.5 6,149.7 4,887.5 4,963.1
BIC 5,914.2 5,482.4 6,455.1 5,371.1 5,421.2

isolate-specific effects to capture departures from this in terms of the location and
steepness of the response. In comparison with Model I, the AIC value of Model II
which is defined by adding a time trend into the linear predictor decreases by 585.3
on 30 df.
Concerning Model IV and Model V, it seems that adding a quadratic term in
time allows us to get near to reproducing the general overall pattern over all isolates
and all replicates. The AIC value is 4,207.4 for Model IV and 4,269.4 for Model V.
Thus, Model IV with the smallest AIC value seems to be the model among the five
continuation-ratio random effects logit models considered in this paper that best fits
the data. From now, all results presented are obtained from Model IV.
For each isolate, Fig. 2 shows the fitted replicate-specific evolutions and the
marginal average evolution implied by Model IV. The posterior quantities of interest
are the random effects and the corresponding model random linear predictors. As
noted by Aitkin (1996), one nice feature of using numerical integration via the EM-
algorithm is that we can easily calculate these quantities from the estimated posterior
distribution of the random effects. For example, using ordinary Gaussian quadrature
methods, the posterior distribution of ij is provided by
Qd
vq kD1 f .yij;k j; ; zq /
f .zq jyij / D PQ Qd D pij;q ; q D 1; : : : ; Q:
lD1 vl kD1 f .yij;k j; ; zl /

These posterior probabilities pijq that the unobserved ij takes the value zq corre-
spond to the weights at the final iteration of the EM-algorithm, and they provide
the posterior distribution of the ij in the empirical Bayes sense by replacing the
unknown parameters by their ML estimates. In Model IV, for instance, the linear
predictors are defined as

wij;k
log D ij;k D ˛i C ˇk C
i tk C ıi tk2 C ij ;
1  wij;k

exp.ij;k /
and the corresponding means as wij;k D .
1 C exp.ij;k /
Discrete Survival Times 55

Fig. 2 Graphical representation of the predicted replicate-specific evolutions (dotted lines) and
the marginal average evolution (solid line) obtained from Model IV using adaptive Gaussian
quadrature

In this case, the empirical Bayes predictions are calculated by:

Qd
O
vq kD1 f .yij;k j; ;
O zq /
pOij;q D PQ Qd ;
O O zl /
lD1 vl kD1 f .yij;k j; ;

X
Q
Oij D pOij;q zq ;
qD1

X
Q
O ij;k D pOij;q O ij;k;q with O ij;k;q D ˛O i C ˇOk C
Oi tk C ıbi tk2 C z
O q;
qD1

exp.O ij;k /
wO ij;k D :
1 C exp.O ij;k /

Note that a similar approach is used when using AGQ. Finally, the fitted prob-
abilities of real interest O ij;k are directly obtained from the empirical Bayes
predictions wO ij;k .
Concerning the marginal average evolution, note that it can be derived from
averaging the conditional means over the random effects ij . Again, this can be done
using numerical integration methods or based on numerical averaging by sampling
a large number of random effects from their fitted distribution (Molenberghs and
56 M.-J. Martinez and J.P. Hinde

Fig. 3 Marginal LT50 6

Variance of the random effects posterior estimates

2.0
against the variance of the 15
10

random effects posterior

estimates obtained from
Model IV using adaptive 12
1.5
Gaussian quadrature
27 30
4
2

13

1.0
26
18

29 16
21
0.5 24
11 25
19
28
1

2 3 4 5 6 7 8
Marginal median lethal time

Verbeke 2005). In this work, we derive the marginal average evolution based on the
second method by sampling 1,000 random effects ij from their fitted distribution.
Note that the approach which consists of plotting the profile for an “average”
replicate, i.e., a replicate with random intercept ij D 0 rather than the marginal
average usually results in different fitted average trends (Molenberghs and Verbeke
2005).
Finally, one of the aims of this study is to determine the effective isolates. In this
context, one quantity commonly used is the lethal time LTp which is the time
required to obtain p % mortality. This quantity can be easily used to summarize
and to rank the different isolate effectiveness. For the 30 isolates, Fig. 3 shows the
marginal median lethal time (LT50 ) over the replicates obtained from Model IV
using AGQ. More precisely, for each isolate, we plot the marginal LT50 when it is
smaller than 8 against the variance of the posterior estimates of the random effect
to account for variability among the replicates. Clearly, effective isolates are those
with both low lethal time and low replicate variability. Using the K-means clustering
method on the two-dimensional data matrix formed by the marginal median lethal
times and the variances of the posterior estimates of the random effect, the different
isolates are grouped here into three clusters: strongly, intermediate and weakly
virulent. This provides some indication of which isolates to use in the field.

5 Discussion

In this paper, we have proposed to use random effects continuation-ratio models

to model discrete survival times by considering them as ordered multicategorical
data. We have seen that this particular model can be easily fitted using the methods
Discrete Survival Times 57

available for binary response data by a rearrangement of the data. The use of
this specific model also makes possible the generalization of these approaches to
replicate measures. In this work, we have introduced a random intercept into the
linear predictor to model the variability observed among the replicates within the
different isolates. A possible extension of this work is to include both a random
intercept and a random slope. By including a random coefficient into the linear
predictor, we allow the time effect to vary between the replicates of each isolate.
In this case, the random effects could be assumed to have a bivariate normal
distribution with unknown covariance matrix.
In the different models considered in this work, the random effects are assumed to
be sampled from a normal distribution. This assumption reflects the prior belief that
the random effects are drawn from one homogeneous population. However, if we
look in detail at the results obtained using ordinary Gaussian quadrature in Table 2,
we can observe that the disparity does not decrease monotonically as we increase the
number of quadrature points. In other words, bad approximations can give better fits.
This behaviour, observed for instance in Lesaffre and Spiessens (2001), suggests
that maybe the normality assumption is not really convincing in this case. To relax
this assumption, it would be interesting in future work to use heterogeneity models
as defined by Molenberghs and Verbeke (2005). This extension consists of replacing
the normality assumption by a mixture of normal distributions. Thus, the model
will reflect the prior belief of the presence of unobserved heterogeneity among
the replicates. The use of this particular model would be an interesting extension
since it relaxes the classical normality assumption and is also perfectly suitable for
classification purposes.

Acknowledgements This work was part-funded under Science Foundation Ireland’s Research
Frontiers Programme (07/RFP-MATF448) and Science Foundation Ireland’s BIO-SI project, grant
number 07/MI/012.

References

Agresti, A. (2002). Categorical data analysis (2nd ed.). New York: Wiley.
Albert, P., & Follmann, D. (2000). Modeling repeated count data subject to informative dropout.
Biometrics, 56, 667–677.
Anderson, D., & Hinde, J. (1988). Random effects in generalized linear models and the EM
algorithm. Communications in Statistics—Theory and Methods, 17, 3847–3856.
Aitkin, M. (1996). Empirical Bayes shrinkage using posterior random effect means from non-
parametric maximum likelihood estimation in general random effect models. In Proceedings
of the 11th International Workshop on Statistical Modelling (pp. 87–94). Orvieto: Statistical
Modelling Society.
Breslow, N., & Clayton, D. (1993). Approximate inference in generalized linear mixed model.
Journal of the American Statistical Association, 88, 9–25.
Brillinger, D., & Preisler, M. (1983). Maximum likelihood estimation in a latent variable problem.
In S. Karlin, T. Amemiya, & L. Goodman (Eds.), Studies in econometrics, time series and
multivariate statistics (pp. 31–35). New York: Academic.
58 M.-J. Martinez and J.P. Hinde

R
Crouch, E., & Spiegelman, D. (1990). The evaluation of integrals of the form f .t / exp.t 2 /dt :
Application to logistic-normal models. Journal of the American Statistical Association, 85, 464–
469.
De Freitas, S. (2001). Modelos para Proporções com Superdispersão proveniente de Ensaios
Toxicológicos no Tempo (Ph.D. thesis). ESALQ, Universidade de São Paulo.
Hinde, J. (1982). Compound Poisson regression models. In R. Gilchrist (Ed.), GLIM 82. New York:
Springer.
Hinde, J. (1997). Contribution to discussion of The EM algorithm - An old folk-song to a fast new
tune by Meng, X-L. and Van Dyk, D. Journal of the Royal Statistical Society, Series B, 59,
549–550.
Jansen, J. (1990). On the statistical analysis of ordinal data when extravariation is present. Applied
Statistics, 39, 74–85.
Lesaffre, E., & Spiessens, B. (2001). On the effect of the number of quadrature points in a logistic
random-effects model: An example. Applied Statistics, 50, 325–335.
Liu, Q., & Pierce, D. (1994). A note on Gauss-Hermite quadrature. Biometrika, 81, 624–629.
McCullagh, P., & Nelder, J. (1989). Generalized linear models. London: Chapman & Hall.
Molenberghs, G., & Verbeke, G. (2005). Models for discrete longitudinal data. New York:
Springer.
Naylor, J., & Smith, A. (1988). Econometric illustrations of novel numerical integration strategies
for Bayesian inference. Journal of Econometrics, 38, 103–125.
Nelder, J. (1985). Quasi-likelihood and GLIM. In R. Gilchrist, B. Francis, & J. Whittaker (Eds.),
Generalized linear models: Proceedings of the GLIM 85 conference. New York: Springer.
Rabe-Hesketh, S., Skrondal, A., & Pickles, A. (2002). Reliable estimation of generalized linear
mixed models using adaptive quadrature. The Stata Journal, 2, 1–21.
Rabe-Hesketh, S., Skrondal, A., & Pickles, A. (2005). Maximum likelihood estimation of limited
and discrete dependent variable models with nested random effects. Journal of Econometrics,
128, 301–323.
Stram, D., Wei, L., & Ware, J. (1988). Analysis of repeated categorical outcomes with possibly
missing observations and time-dependent covariates. Journal of the American Statistical
Association, 83, 631–637.
Ten Have, T., & Uttal, D. (1994). Subject-specific and population-averaged continuation ratio logit
models for multiple discrete time survival profiles. Applied Statistics, 43, 371–384.
Tutz, G., & Hennevogl, W. (1996). Random effects in ordinal regression models. Computational
Statistics and Data Analysis, 22, 537–557.
 Part II
Longitudinal Modelling and Time Series
Modelling Seasonality and Structural Breaks:
Visitors to NZ and 9/11

John Haywood and John Randal

Abstract We demonstrate the poor performance, with seasonal data, of existing

methods for endogenously dating multiple structural breaks. Motivated by iterative
nonparametric techniques, we present a new approach for estimating parametric
structural break models that perform well. We suggest that iterative estimation
methods are a simple but important feature of this approach when modelling
seasonal data. The methodology is illustrated by simulation and then used for an
analysis of monthly short-term visitor arrival time series to New Zealand, to assess
the effect of the 9/11 terrorist attacks. While some historical events had a marked
structural effect on trends in those arrivals, we show that 9/11 did not.

Keywords Break dates • Endogenous dating of structural changes • Iterative

fitting • Multiple breaks • Trend extraction

1 Introduction

The economic importance of tourism to New Zealand is high and has increased con-
siderably in recent years. As Pearce (2001) noted in his review article, international
visitor arrivals increased by 65 % over the period 1990–1999, and foreign exchange
earnings increased by 120 % (in current terms). More recently, for the year ended
March 2004 tourism expenditure was $17.2 billion (Statistics New Zealand 2005).
In that year, the tourism industry made a value-added contribution to GDP of 9.4 %,
while 5.9 % of the total employed workforce had work directly engaged in tourism.
Further, tourism’s 18.5 % contribution to exports was greater than that of all other

J. Haywood ()
School of Mathematics, Statistics and Operations Research, Victoria University of Wellington,
PO Box 600, Wellington, New Zealand
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 61

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__6,
© Springer International Publishing Switzerland 2014
62 J. Haywood and J. Randal

industries including dairy products, which in turn was greater than the contributions
from meat and meat products, wood and wood products, and seafood.
The time series of monthly short-term visitor arrivals to New Zealand is one
direct and easily recorded measurement of the international tourist contribution to
the New Zealand economy. A useful precursor to development of tourism policy or
business strategy is an understanding of the dynamic behaviour of these seasonal
data. A classical time series decomposition includes unobserved components
representing an evolving trend, a seasonal encapsulating regular deviation from the
trend on a within-year basis, and an irregular, which is the residual or unexplained
variation in the data. There are various ways to estimate these components, using
both parametric and nonparametric approaches; see for example Harvey (1989),
Hamilton (1994), Findley et al. (1998), Franses (1998) and Makridakis et al. (1998).
Such a decomposition then allows an interpretation of the dynamic behaviour of
visitor arrivals in terms of the estimated components.
There seems little doubt that the terrorist attacks of 11 September 2001 have
had a pronounced influence on world events since that time. For example, see US
Department of State (2004), for a summary of 100 editorial opinions from media
in 57 countries around the world, commenting on the 3 years following September
2001. Those terrorist events and their subsequent effects have been used to explain
apparent movements in many time series, and in this paper we concentrate on a
particular example: the number of short-term visitor arrivals to New Zealand.
Our focus is to detect any longer term, or structural, changes in trend or
seasonal components of the arrivals as a result of the 9/11 events. We also wish
to compare the magnitude of any 9/11 effects with those due to other causes.
Consequently we do not wish to specify the dates of any structural changes, but
rather estimate the number and position of these endogenously. To achieve this we
use Bai and Perron’s (1998, 2003) procedures for estimating multiple structural
changes in a linear model. Their approach permits periods of stable dynamic
behaviour between relatively infrequent but significant changes to the parameters
of the model. However, there is clearly no empirical requirement that changes
in the trend and seasonal components occur simultaneously. As we demonstrate,
for the visitor arrivals data changes typically occur more frequently in the trend.
In contrast, direct application of Bai and Perron’s (1998, 2003) methodology fits
components simultaneously and yields a relatively poor decomposition, as we show
via a simulation study and analysis of the visitor arrivals. We propose a new iterative
fitting procedure for seasonal data, based on Bai and Perron (1998, 2003) and using
existing R packages (R Development Core Team 2007), which gives much improved
performance in terms of flexibility of fitted trends (via more appropriate placement
of breaks) and lack of residual serial correlation.
Throughout the paper the term “trend” (or trend component) is used to describe
the evolving, underlying behaviour of a time series. That underlying behaviour
reflects both long-term movements and medium-term cyclical fluctuations, where
long term and medium term are in relation to the (shorter) period of the evolving
seasonal component that we also consider explicitly. This notion of trend agrees
with that used by many national statistical offices; e.g., see Australian Bureau of
Modelling Seasonality and Structural Breaks 63

Statistics (2003, Sect. 2.1). Certainly we agree with Busetti and Harvey (2008),
that the strong but common assumption in the econometrics literature of a constant
underlying slope when testing for a trend is often implausible. Since our focus is on
breaks in the structure of the trend and seasonal components, we choose to model
those components as piecewise linear, with endogenously estimated changes in the
trend slope and/or seasonal pattern corresponding to identified structural changes.
We find there is actually little to suggest that the September 11 incidents had
much effect on New Zealand visitor arrivals, when viewed in the context of
“normal” historically observed movements. In contrast, we identify some other
historical events which do appear to have affected visitor arrivals to New Zealand
quite markedly. We make no attempt to forecast the arrivals data using structural
break models; we suggest other approaches, such as ARIMA modelling (Box and
Jenkins 1976), would be more suitable if prediction was the aim. In fact Haywood
and Randal (2004) used that approach to demonstrate that the 9/11 events did not
significantly affect New Zealand visitor arrivals, by showing that the observations
post-9/11 were contained within out of sample prediction intervals computed using
a seasonal ARIMA (“airline”) model, fitted to arrivals data up to 9/11. In this paper
though, the focus is explicitly on identifying structural changes if they exist in the
arrivals data, in trend and/or seasonal components.
In Sect. 2 we present an exploratory data analysis (EDA) of New Zealand
visitor arrivals and a discussion of some apparent sources of variability in the data.
Section 3 motivates and presents the iterative estimation of a parametric model that
allows separate structural changes in the trend and seasonal components. Simulated
data is used to illustrate the good performance of the new methodology. In Sect. 4
we use our iterative approach to model the arrivals data and in Sect. 5 we give some
concluding comments.

2 EDA of Short-Term Visitor Arrivals to New Zealand

We consider 25 complete years of monthly short-term visitor arrival series from

January 1980 to December 2004. The arrivals are from the seven most important
countries of origin, ranked by current proportion of the total: Australia, UK, USA,
Japan, Korea, China, Germany, as well as a residual series from “Other” origins.
We analyse these series individually along with their aggregate, denoted “Total”
(Fig. 1).
As seen in Fig. 1 a “U”-shaped seasonal pattern is common, with visitor numbers
reaching a local maximum in the summer months December to February, and a
local minimum in the winter months June and July. Further, it is apparent that the
amplitude of the seasonal variation tends to increase with the level of the series,
indicating a multiplicative relationship between trend and seasonal components.
Australian and UK arrivals appear to be growing at a relatively steady rate.
In contrast, a large downturn in arrivals from the USA is evident in the late 1980s,
a period which immediately followed the stock market crash of October 1987. The
64 J. Haywood and J. Randal

40000 20000 60000 100000

8000
Australia

China
4000
10000
Germany
6000
20000
UK

0 2000
25000

60000
Other
USA
15000

20000
5000
15000

200000
Total
Japan
15000 0 5000

50000
1980 1985 1990 1995 2000 2005
Korea
5000
0

1980 1985 1990 1995 2000 2005

Fig. 1 Monthly short-term visitor arrivals to New Zealand, by origin, from January 1980 to
December 2004. The vertical scales are not equal

trend in Japanese arrivals levels off over the last 15 years. The effect of the Asian
financial crisis of 1997 is evident especially in the Korean data, with visitor numbers
dramatically reduced just after this event. Arrivals from China contain perhaps the
most visible short-term effect in these series, which is due to the SARS epidemic
that virtually eliminated international travel by Chinese nationals during May and
June 2003. German arrivals show a clear change from exponential growth prior to
the early 1990s to a more stable pattern in recent times. The Other arrivals show
a SARS effect much less prominent than that seen in the Chinese arrivals, as do
some further series including Total arrivals. One of the more obvious shifts in the
aggregate Total series appears to be linked to the Korean downturn, which can be
attributed to the Asian financial crisis.
The Asian financial crisis of 1997–1998 markedly affected stock markets and
exchange rates in several Asian countries and regions, including: Hong Kong,
Indonesia, Japan, Korea, Malaysia, Philippines, Singapore, Taiwan and Thailand.
See Kaminsky and Schmukler (1999) for a chronology of the crisis in those
locations, from the official onset marked by the devaluation of the Thai baht on
2 July, 1997 up to the resignation of Indonesian President Suharto in May 1998.
Kaminsky and Schmukler (1999) suggest the presence of important contagion
effects in those markets, based on an analysis of identified market jitters. More
recent analysis by Dungey et al. (2004) suggests, however, that increased exchange
rate volatility observed in Australia and New Zealand around that time was not due
to contagion from Asian countries, or unanticipated factors, but rather to common
(anticipated) world factors such as trade linkages. This is one context in which
Modelling Seasonality and Structural Breaks 65

Table 1 Summary statistics for the monthly proportion of visitors to New Zealand, by origin
Min LQ Median UQ Max 80–04 80–84 00–04
Australia 21:8 30:0 35:9 41:6 58:8 33:8 44:9 33:3
UK 3:4 6:3 8:0 10:6 18:3 9:8 7:6 11:8
USA 6:3 10:3 13:0 16:3 29:4 12:4 16:7 10:0
Japan 2:8 7:1 9:1 11:0 17:8 9:2 5:9 7:8
Korea 0:0 0:2 1:0 4:3 10:5 3:4 0:2 4:8
China 0:0 0:2 0:4 1:2 4:7 1:4 0:1 3:1
Germany 0:8 1:5 2:2 3:4 7:5 2:9 1:8 2:6
Other 17:9 23:4 26:1 28:6 34:3 27:0 22:8 26:7
The final three columns give proportions of the Total for the entire 25 year sample period, and the
5-year periods 1980–1984 and 2000–2004, respectively

changes in short-term visitor arrivals to New Zealand from Asian countries around
1997–1998 can be viewed, since tourism has become such an important sector of the
New Zealand economy, as noted above. In particular, Korea is one of the five source
countries with the largest recent (2000–2004) proportion of visitors to New Zealand
(Table 1).
Table 1 shows that Australia is by far the biggest single source of visitors to
New Zealand, accounting for almost exactly one-third of visitors in the 2000–
2004 5-year period and slightly more over the entire data period. The maximum
proportion in a month from Australia was 58.8 % in June 1985, and the minimum
was 21.8 % in February 1997. An Australian influence is notable in the Total
arrivals, because as the nearest neighbour to a geographically isolated country,
arrivals from Australia exhibit variation not seen in the remaining data. As seen
in Fig. 1, the Australian data has a regular seasonal pattern which is quite different
from that of any other country. A closer examination indicates three peaks per year
before 1987 and four thereafter; we discuss this further in Sect. 4.
One way of estimating unobserved trend and seasonal components is to use
a robust, nonparametric technique such as STL (Cleveland et al. 1990); here we
use STL as implemented in R (R Development Core Team 2007). This procedure
consists of an iterated cycle in which the data is detrended, then the seasonal
is updated from the resulting detrended seasonal subseries, after which the trend
estimate is updated. At each iteration, robustness weights are formed based on
the estimated irregular component and these are used to down-weight outlying
observations in subsequent calculations. A typical STL decomposition is shown
in Fig. 2 for the natural logarithm of the Total arrivals. The log transformation is
commonly used to stabilise a seasonal pattern which increases with the level of the
series, and effectively transforms a multiplicative decomposition into an additive
one.
Figure 2 shows an evolving seasonal pattern, an upward trend with several
changes in slope, and a relatively small irregular component. A vertical line is added
to indicate September 2001. There is no obvious (structural) change in the trend at
or about this month, although there is a reduction in the slope of the trend nearer
66 J. Haywood and J. Randal

12.0
log(Arrivals)

11.0

10.0
0.6
0.4
Seasonal

0.2
0.0

−0.4
12.0

11.5
Trend

11.0

10.5

0.1
Irregular

0.0

−0.1

−0.2
1980 1985 1990 1995 2000 2005

Fig. 2 The STL decomposition of the log aggregate monthly visitor arrivals to New Zealand from
January 1980 to December 2004. The vertical grey line is at September 2001, and the solid bars
on the right hand side of the plot are all the same height, to aid comparisons

the start of 2001, which we discuss further in Sect. 5. More prominent is a cluster
of negative irregulars immediately following 9/11, the largest of which is the third
largest negative irregular in the sample period. Jointly though, these irregulars are
smaller and less persistent than those occurring at the time of the SARS outbreak in
2003. Our exploratory analysis with STL thus suggests that while the events of 9/11
may have had a moderate short-term (irregular) effect, there is nothing to suggest
that a longer term (structural) effect occurred. We investigate this hypothesis more
formally in Sect. 4.

3 Iterative Break Estimation for Seasonal Data

Bai and Perron (1998, 2003) present a methodology for fitting a linear model with
structural breaks, in which the break points, i.e. the times at which the parameters
change, are determined optimally. The optimal positions of m break points are
determined by minimising the residual sum of squares, for each positive integer
m  mmax . The optimal number of break points (0  m  mmax ) may then
be determined by, for example, minimising an information criterion such as BIC
(Schwarz 1978). Given a sample of T observations, the selected break points are
estimated consistently, with rate T convergence of the estimated break fractions
(that is, the proportions of the data between consecutive breaks).
The maximum number of break points, mmax , is determined by the number of
observations relative to the number of parameters in the model. In general, for a
Modelling Seasonality and Structural Breaks 67

model with m breaks and q parameters, at least q observations are needed between
each pair of break points, requiring at least T  .m C 1/q observations in total.
Clearly if the model has many parameters, fewer break points can be estimated
from a given sequence of observations.
We consider implementing this approach for a time series of the form

Y t D T t C S t C It ; t D 1; : : : ; T;

where Yt are the observed data (transformed if necessary), Tt is an unobserved trend

component, St is an unobserved seasonal component with seasonal period s, and It
is an unobserved irregular component. Many observed time series do not follow
an additive decomposition, including the NZ visitor arrivals as noted in Sect. 2;
however, we assume a suitable stabilizing transformation can be applied (e.g., see
Sect. 4). In the following model, the evolution of trend and seasonal components
is explicitly modelled as structural changes occurring at endogenously identified
break points. Short-term, random changes may also occur, but these are modelled
by the irregular component It . We assume that between two break points tj1 and
tj .j D 1; : : : ; m C 1/, the trend Tt is linear,

Tt D ˛j C ˇj t; t D tj1 C 1; : : : ; tj

and, again between break points, the seasonal component is fixed,

X
s1
St D ıi;j Di;t ; t D tj1 C 1; : : : ; tj ;
i D1

where Di;t are seasonal dummies. We use the convention that t0 D 0 and tmC1

DT
(Bai and Perron 1998). Under these assumptions, we note that for daily or monthly
data (with s D 7 and s D 12 respectively), and for quarterly data (with s D 4) to
a lesser extent, the trend component will be parsimonious relative to the seasonal
component.
Bai and Perron’s (1998, 2003) methodology offers two alternatives for estimating
the unknown break points, tj .j D 1; : : : ; m/, in such a model. The first is that
the coefficients of one component are fixed over the entire sample period (a partial
structural change model); the second is that parameters in both components should
have the same break points (a pure structural change model). We demonstrate below
that neither of these options is satisfactory for the type of data examined in this
paper, i.e. seasonal time series with evolving trends, and large s (in this case,
s D 12).
When considering trend extraction and assuming that structural breaks will be
required, in general we wish to allow break points in the seasonal component, which
is inconsistent with a partial structural change model. Conversely, we would not
necessarily wish to constrain any seasonal break points to occur at the same places
as the trend break points, as required in a pure structural change model. On the face
68 J. Haywood and J. Randal

of it, this requirement is not necessarily restrictive, since the parameter estimates
of one component are not forced to change from one segment of the data to the
next. However, when selecting the optimal number of break points using a penalised
likelihood criterion, e.g. BIC, this compromises the ability to detect break points in
the data, i.e. the selected number of breaks may be too low. One example of where
these issues may be important is in arrivals from Australia. As noted in Sect. 2, the
Australian arrivals seem to have a seasonal break point in 1987 (changing from three
peaks to four), with no apparent change in trend.
To address this concern we estimate the trend and seasonal components sepa-
rately, using a new iterative approach motivated by the Macaulay cycle seasonal
decomposition method (Macaulay 1931) and the iterative technique of STL. This
allows more flexible structural break estimation than fitting both components
simultaneously. As above, we assume that the time series can be decomposed
into a piecewise linear time trend and a piecewise constant seasonal pattern. Each
component is then estimated using the methodology of Bai and Perron (1998,
2003), implemented in R (R Development Core Team 2007) using the strucchange
package of Zeileis et al. (2002). We employ the default method of selecting the
number of breaks, which uses BIC.
The trend of the data Yt is estimated using a piecewise linear model for the
seasonally adjusted time series Vt D Yt  SOt , i.e.,

Vt D ˛j C ˇj t C t ; t D tj1 C 1; : : : ; tj

for j D 1; : : : ; m C 1, where t is a zero-mean disturbance and tj , j D 1; : : : ; m,

are the unknown trend break points. For the first iteration, we set SOt D 0 for all t.
Once the trend has been estimated, we estimate the seasonal component of Yt
using a piecewise seasonal dummy model for the detrended data Wt D Yt  TOt , i.e.,

X
s1
Wt D ı0;j C ıi;j Di;t C t ; t D tj0 1 C 1; : : : ; tj0
i D1

for j D 1; : : : ; m0 C 1, where Di;t are the seasonal dummies, t is a zero-mean

disturbance and tj0 , j D 1; : : : ; m0 , are the unknown seasonal break points. As
before, we take t00 D 0 and tm0 0 C1 D T . The estimates ıOi;j are adjusted at the end
of each iteration so that they add to zero within each full seasonal cycle (between
seasonal breaks), to prevent any change in trend appearing as a result of a seasonal
break happening “mid-year”. That is,

X
s1
ıOi;j D 0 for all j:
i D0

This estimation process is then iterated to convergence of the estimated break points.
We are thus able to estimate a trend which, due to its parsimonious
representation, is able to react to obvious shifts in the general movement of the
Modelling Seasonality and Structural Breaks 69

First cycle Second cycle

4 4

2 2

0 0

−2 −2

−4 −4

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Month Month

Fig. 3 The two seasonal cycles used for the seasonal component in the illustrative simulation

data. If required, we are able to identify important changes in the seasonal pattern
separately. Since the trend and seasonal break points, tj and tj0 respectively,
are estimated independently, they are not constrained to occur concurrently.
Of course, this does not preclude (some) trend and seasonal break points
coinciding if appropriate. In all data analysis and simulations we have followed
the recommendations of Bai and Perron (2003) and Zeileis et al. (2003), concerning
the fraction of data needed between breaks. For monthly seasonal data, we used
3 full years (36 observations) as a minimum, corresponding to 12 % of a 25-year
data span. However, a further consequence of the iterative estimation of trend and
seasonal breaks is that while any two breaks of the same type must have a minimum
separation (3 years here), the distance between a trend break and a seasonal break
has no constraints. This feature is of practical importance, e.g. as shown for the
arrivals data in Sect. 4, and is another desirable feature of the new iterative approach.
The importance of this method is now illustrated using simulated data. Consider
a time series with piecewise linear trend given by
8
ˆ
<20 C 0:05t;
ˆ t D 1; : : : ; 78
Tt D 23:9; t D 79; : : : ; 234
ˆ
:̂23:9 C 0:05.t  234/; t D 235; : : : ; 312

and piecewise fixed seasonal component with a break point at t D 156. The two
seasonal cycles are shown in Fig. 3, and are identical except for the ordering of the
Jan/Feb, Mar/Apr, Jul/Aug and Sep/Oct values. The data are given by

Y t D T t C S t C It ; It  i.i.d. N.0; 1/

and thus have three break points: two associated solely with the trend, and one
associated only with the seasonal component.
We simulated 500 independent series as above, and for each of them estimated
the trend and seasonal components simultaneously using the Bai and Perron (1998,
2003) approach. In particular, we restricted the parameters in both the trend and
70 J. Haywood and J. Randal

Fitted break points in complete model

n = 343 n = 146 n = 11
250
Observation

150

50
0
1 break 2, smaller 2, larger 3, smaller 3, middle 3, larger

Fitted break points in iterated model

n = 494 n=6 n = 500
250
Observation

150

50
0

2, smaller 2, larger 3, smaller 3, middle 3, larger seasonal

Fig. 4 The estimated break points for 500 simulated series. True break dates are at the horizontal
grey lines. Results in the upper panel are for the Bai and Perron methodology, with the new iterated
methodology in the lower panel. The number of cases is shown above each boxplot, while the total
number of estimated breaks is given below. In the iterated panel, trend break points are shown on
the left and the (single) seasonal break point on the right

seasonal components to change simultaneously; i.e., a pure structural change model.

Between breaks, the constant term in the estimated trend was corrected so that the
seasonal component added to zero. We also applied the iterated methodology to the
same series, fitting the trend and seasonal components separately. Figure 4 shows the
estimated break points for the 500 series using the two competing methodologies.
In both panels are boxplots of the estimated break points for the series, and these
have been grouped depending on how many break points were estimated.
The estimated break points obtained fitting both components simultaneously
are in the upper panel of Fig. 4. In 343 series (68.6 %) only one break point was
estimated, and the sample distribution of these is summarised in the extreme left
boxplot. The estimates appear to be unbiased for the central (seasonal) break point.
In the middle of the upper panel are the sample distributions of the two break points,
as estimated in 146 series (29.2 %). These appear to be biased estimates of the true
trend break points, with each being closer to the seasonal break point than to the
nearest end of the series. On the right, the sample distributions of the three break
points are displayed, estimated in 11 of the series (2.2 %). These appear unbiased,
although the sample size is very small. In conclusion, fitting the two components
simultaneously has allowed us to correctly identify the true break points in only
2.2 % of the series. Despite correctly dating the breaks in these 11 series, this
approach does not attribute the change to any one component of the structural
Modelling Seasonality and Structural Breaks 71

30

25

20

15

0 50 100 150 200 250 300

Observation

Fig. 5 Simulated data (solid) with true piecewise linear trend (dashed). Estimated piecewise linear
trends using the Bai and Perron methodology (grey) and the iterated methodology (black)

model. Clearly though, each true change affects only one of the trend or seasonal
components.
In the lower panel of Fig. 4 are comparable results for the new iterated approach.
All series have at least two estimated trend break points and a single seasonal break.
Two trend breaks are estimated in 494 series (98.8 %) and are displayed on the left;
they appear to be unbiased estimates of the true trend break points. In the remaining
six series (1.2 %), a third trend break is estimated. The estimated seasonal break
points for all 500 series are shown on the right; these are clearly unbiased, and
estimated relatively precisely. In conclusion, fitting the two components iteratively
allowed us to correctly identify and attribute the trend and seasonal break points
in 98.8 % of the series; a dramatic improvement over the simultaneous approach.
The use of sequential F -tests instead of BIC to select the number of breaks in the
simultaneous approach does not substantially change the number of series with the
correct number of estimated breaks.
Figure 5 shows a single example series that results in typical estimation
behaviour, with the true trend and both estimated trends also plotted. For the
complete model, BIC selects only one break point at t D 168. Applying the
iterated methodology to the data, two trend break points are estimated at t D 83
and t D 249, and a single seasonal break point is estimated at t D 157.
Note that the complete model induces a “quadratic” trend in the residuals on
either side of the single break, which is expected after viewing the estimated and
true trends together as in Fig. 5. That (local) trending behaviour is reflected by
significant residual sample autocorrelations at low lags; clear evidence that the
model is misspecified. In contrast, the residuals from the iterative approach show no
significant autocorrelations at low lags, reflecting the more appropriate modelling
of the true trend component.
This simulation shows the undesirable consequences of fitting two components
simultaneously when a parameter-rich seasonal component breaks at times other
72 J. Haywood and J. Randal

than those of a relatively parsimonious trend component. Our new iterated approach
to fitting such components addresses this concern. Next we apply this iterated
technique to the arrivals data.

4 Modelling the Arrivals Using an Iterated Approach

The seasonal variation of the arrivals typically increases with the level of the series
(Fig. 1). Applying the new iterated approach directly to the untransformed data
would certainly require seasonal breaks to account for the changes in amplitude of
the seasonal component. This is clearly undesirable because such changes usually
evolve smoothly, so should not be modelled as abrupt changes. Consequently a
stabilising transformation is needed.
A log transformation is one obvious possibility, but this does not yield an
optimal stabilising transformation for all these series and instead we estimate a
power transformation, identified using the robust spread-vs-level plots described
in Hoaglin et al. (1983). For each individual series we calculate the median and
interquartile range (IQR) of the monthly arrivals for each of the 25 calendar years,
then regress log IQR on log median. The appropriate stabilising transformation is
x 1slope , and the transformed series are shown in Fig. 6, with the estimated powers.
Confidence intervals for the slopes in these spread-vs-level regressions support
the use of logs only in the case of the UK, USA and Total arrivals (i.e. a power of
zero, or a slope of one). In the case of Germany the estimated power is negative, so
x 1slope is used to preserve order in the transformed arrivals. All further analysis
is conducted on the transformed data.
In the case of the transformed arrivals data, each linear time trend requires two
parameters, and each dummy seasonal an additional s  1 D 11. Figure 6 indicates
that for most series a linear time trend would need breaks. Further, while the
seasonal patterns generally have constant variation over the length of the series due
to the power transformations, we do not wish to preclude seasonal changes during
the data period. As the simulation study demonstrated, the parameter-rich trend-
plus-seasonal (complete) model would severely limit our ability to appropriately fit
the data, since the large number of seasonal dummies would reduce the possible
number of breaks, especially when selected by BIC.
As with the simulated data, we use a minimum period between breaks of 36
observations for estimation of both trend and seasonal components. In fact, when
estimating the trend and seasonal components iteratively, there is scope to reduce
that minimum period for estimation of the trend component, since it only requires
two parameters between breaks. This possibility further increases the flexibility of
trends estimated using our new iterated approach. However, to simplify comparisons
we have not pursued this option here. For the iterative approach, three iterations
were sufficient to ensure convergence of the estimated break points in all cases but
Other and Total, which each required four.
Modelling Seasonality and Structural Breaks 73

35

5
30
Australia

China

4
25

3
20

−0.352
1.60

Germany
1.50

−0.45
UK

12 14 1.85 1.95 2.05 2.15 1.40

3.6 4.0 4.4 −0.55

Other
USA

1.40 3.2
Japan

Total

1.36
1
8

1.32
6
2.6

1980 1985 1990 1995 2000 2005

Korea

2.2
1.8
1.4

1980 1985 1990 1995 2000 2005

Fig. 6 Power transformed monthly short-term visitor arrivals to New Zealand, by origin, from
January 1980 to December 2004. The power transformations are: Australia 0.3, UK 0.05, USA
0.08, Japan 0.27, Korea 0.11, China 0.18, Germany 0.11, Other 0.13, and Total 0.03

The estimated trend break points are shown in Table 2 along with estimated 95 %
confidence intervals. The confidence intervals have been formed with heteroscedas-
ticity and autocorrelation consistent (HAC) estimates of the covariance matrix
(Andrews 1991). These confidence intervals are computed and displayed as standard
output using the R package sandwich; they make use of a quadratic spectral
kernel with vector autoregressive prewhitening, as recommended by Andrews and
Monahan (1992). Details of the R implementation are given in Zeileis (2004, 2006).
Figure 7 displays the estimated parametric trends and break points (with confidence
intervals), along with nonparametric trends estimated by STL. September 2001 is
included in only two confidence intervals, indicating the possibility that the terrorist
events of 9/11 may be linked to a structural break in the trend of arrivals for those
two origins: Australia and Other. Other is difficult to interpret given its composite
nature, although it is plausible that the 9/11 events did have an effect on tourist
behaviour in some of these countries. An alternative (or perhaps complementary)
explanation is discussed in Sect. 5.
In the case of Australia, a break is estimated in the month following 9/11,
which results in an increased trend slope but a decreased intercept. A relevant
confounding effect is the collapse of Ansett Australia, which occurred just 3 days
after the terrorist attacks of 9/11; hence it is impossible to separate these two
effects with monthly data. The termination of flights by Ansett Australia and
Ansett International on 14 September 2001 certainly affected capacity and timing
of arrivals to New Zealand. In addition, in the following week, strike action targeted
at Air New Zealand occurred at Melbourne and Perth airports (Air New Zealand
74 J. Haywood and J. Randal

Table 2 Estimated trend break points for the transformed monthly visitor arrivals to New Zealand,
by origin, from January 1980 to December 2004
Australia China
1984(5) 1985(1) 1985(2) 1984(7) 1984(8) 1985(1)
1989(3) 1989(4) 1990(5) 1988(10) 1989(7) 1989(9)
1997(10) 1997(12) 1998(1) 1997(1) 2000(11) 2000(12)
2001(1) 2001(10) 2001(11) Germany
UK 1986(6) 1986(7) 1986(11)
1985(10) 1986(1) 1986(4) 1994(5) 1994(6) 1994(7)
1990(7) 1990(8) 1996(5) 1999(6) 1999(8) 2000(11)
USA Other
1982(12) 1983(3) 1986(12) 1983(1) 1983(3) 1983(4)
1988(9) 1988(10) 1990(1) 1985(6) 1986(8) 1986(9)
1998(6) 1998(8) 2001(6) 1990(7) 1990(10) 1990(12)
Japan 1992(11) 1994(1) 1994(3)
1987(3) 1987(6) 1987(8) 1997(3) 1997(6) 1997(8)
1996(2) 1996(8) 1996(10) 2001(4) 2001(7) 2001(9)
Korea Total
1982(8) 1983(12) 1984(4) 1982(12) 1983(1) 1983(3)
1990(9) 1990(10) 1990(11) 1987(10) 1987(12) 1988(4)
1994(9) 1994(11) 1994(12) 1989(8) 1990(12) 1991(4)
1997(10) 1997(11) 1997(12) 1997(1) 1997(3) 1997(6)
2000(10) 2000(11) 2001(1)
The middle column gives the estimated break points, while the first and third columns give the
lower and upper 95 % confidence limits respectively, estimated using a HAC estimate of the
covariance matrix

had acquired control of Ansett Australia during the year preceding its collapse).
Those strikes required the cancellation of all Air New Zealand trans-Tasman flights
operating from Melbourne and Perth. These physical constraints on passenger
numbers are a plausible explanation for a decrease in intercept, while the subsequent
increase in the rate of arrivals from New Zealand’s nearest neighbour is unlikely to
have any causal links from the terrorist events of September 2001.
Focusing on Fig. 7 more generally, we note that it is often difficult to distinguish
between the two alternative trend estimates; i.e. those from our iterated approach
and STL. In particular, the iterative parametric method achieves similar flexibility in
its trend estimate to the nonparametric technique, with the latter essentially fitting
linear time trends at each point in the series using only a local window of obser-
vations to estimate parameters. The break point technology allows instantaneous
changes in the trend, however, unlike the STL technique. In effect, STL requires
an “innovational outlier” approach to any structural changes in the data, while
our parametric procedure models the changes directly and permits an “additive
outlier” approach. (In a series of papers, Perron and coauthors popularised the
use of these “outlier” terms, to describe an approach which is attributed to the
intervention analysis work of Box and Tiao 1975.) An obvious contrast between
Modelling Seasonality and Structural Breaks 75

Australia arrivals UK arrivals USA arrivals

2.10
20 22 24 26 28 30

1.55

2.05
1.50

2.00
1.45

1.95
1980 1985 1990 1995 2000 2005 1980 1985 1990 1995 2000 2005 1980 1985 1990 1995 2000 2005

Japan arrivals Korea arrivals China arrivals

2.0 2.5 3.0 3.5 4.0 4.5 5.0

1.4 1.6 1.8 2.0 2.2 2.4 2.6
10 11 12
9
8
7

1980 1985 1990 1995 2000 2005 1980 1985 1990 1995 2000 2005 1980 1985 1990 1995 2000 2005

Germany arrivals Other arrivals Total arrivals

4.2

1.39
−0.40

4.0

1.37
3.8
−0.44

1.35
3.6
−0.48

3.4

1.33
1980 1985 1990 1995 2000 2005 1980 1985 1990 1995 2000 2005 1980 1985 1990 1995 2000 2005

Fig. 7 Estimated trends and trend break points for the transformed monthly visitor arrivals to
New Zealand, by origin, from January 1980 to December 2004. The solid line is the piecewise
linear time trend, while the dotted line is the estimated STL trend. The vertical dashed lines
and grey regions respectively indicate the fitted break points and their 95 % confidence intervals,
estimated using a HAC estimate of the covariance matrix

the two approaches is seen in the Korean data at the time of the Asian financial
crisis. The parametric break point is dated at November 1997 (with a narrow 95 %
confidence interval of October to December), which corresponds exactly to the
month that the financial crisis first affected Korea (Kaminsky and Schmukler 1999).
However, STL spreads the downward impact of the crisis over a number of months,
in contrast to the observed behaviour.
Table 3 gives the estimated seasonal break points for the transformed arrivals,
with the estimated seasonal components shown in Fig. 8. Korea, China, Germany
and Other have no estimated seasonal break points. As the power transformations
have effectively stabilised the seasonal variation, any changes in the seasonal
patterns more likely reflect behavioural changes in the time of year when visitors
arrive. For example, in Australia’s seasonal pattern the “middle” peak has moved
and one extra peak has been added, reflecting a shift from a three-term school year
to a four-term year in New South Wales in 1987 (NSW Department of Education
1985). The placement of the seasonal break point coincides exactly with the final
month under the old three-term system, with the first holiday in the new sequence
occurring in July 1987. The UK data show a shift in arrivals from the second half of
the year to the first and a shift in the peak arrivals from December to February. The
USA and Japanese arrivals have had relatively complex changes, while the Total
76 J. Haywood and J. Randal

Table 3 Estimated seasonal Origin Point estimate and 95 % CI

break points for the
transformed monthly visitor Australia 1987(1) 1987(6) 1987(9)
arrivals to New Zealand, by UK 1985(11) 1986(6) 1987(9)
origin, from January 1980 to USA 1995(1) 1995(4) 1995(12)
December 2004 Japan 1987(10) 1988(6) 1988(12)
Total 1987(3) 1987(7) 1988(1)
The middle column gives the estimated break points, while
the first and third columns give the lower and upper 95 %
confidence limits respectively, estimated using a HAC esti-
mate of the covariance matrix. Korea, China, Germany and
Other have no estimated seasonal break points

Australia UK USA
4

0.05
0.04
3
2

0.00
0.00
1
0

−0.05
−2 −1

−0.04

Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov

Japan Korea China

−0.5 0.0 0.5 1.0

−0.10 −0.05 0.00 0.05 0.10

0.10
0.00
−0.10
−1.5

Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov

Germany Other Total

0.00 0.02 0.04

0.3

0.015
0.2
0.1

−0.015 −0.005 0.005

−0.2 −0.1 0.0
−0.04

Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov

Fig. 8 The estimated seasonal components for visitors to New Zealand by origin. The solid line
is the final estimated seasonal component; it is the only estimate in four of the nine cases, where
no seasonal breaks were detected. The five dashed lines are the seasonal components prior to the
seasonal break points listed in Table 3

series has seen most change in the winter months. Note here the practical relevance
of allowing breaks in the seasonal component of any given series to be independent
of those in the trend, with no minimum separation between them: four of these
five seasonal breaks (all except USA) are less than 3 years away from at least one
corresponding trend break (see Tables 2 and 3). However, none of the dates for trend
and seasonal breaks coincide in any given series, which reinforces the need to allow
the components to break separately for additional flexibility in the fitted model.
Modelling Seasonality and Structural Breaks 77

Korea arrivals Other arrivals

2.8 4.2

2.6

4.0

2.4
Trend of transformed arrivals

Trend of transformed arrivals

2.2 3.8

2.0

3.6
1.8

1.6
3.4

1.4

1980 1985 1990 1995 2000 2005 1980 1985 1990 1995 2000 2005

Date Date

Complete model Iterated model STL decomposition Complete model Iterated model STL decomposition
1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6
ACF

ACF

ACF

ACF

ACF

ACF
0.4 0.4 0.4
0.2 0.2
0.2

0.0 0.0 0.0

−0.2
−0.2 −0.2

0.0 1.0 2.0 0.0 1.0 2.0 0.0 1.0 2.0 0.0 1.0 2.0 0.0 1.0 2.0 0.0 1.0 2.0
Lag Lag Lag Lag Lag Lag

Fig. 9 Trend estimates for the Korean arrivals and those from Other origins. The trend estimates
are based on the complete model (grey), the new iterated approach (black) and STL (dashed). Also
shown are sample autocorrelation functions for the residuals from the three methods

To conclude this section, we compare the trend estimates obtained from our
new iterated approach to the trends obtained fitting a complete structural break
model (with 13 parameters between breaks), and using STL. In Fig. 9 we present
trends for the Korean arrivals and those from Other origins. We also show sample
autocorrelation functions for the three sets of residuals from each series. The trends
are all similar, but the agreement is closest for the iterated approach and STL.
Some differences are evident particularly at the end of the series though, which
would be important for prediction. For Other arrivals, the number of parameters
required for the complete model clearly restricts the estimated number of breaks,
leading to greater departures from the STL trend than achieved by iteration. The
irregular components also favour the iterated approach over STL and the complete
model, as the residuals for the latter are highly autocorrelated, especially at low
lags. In contrast, the residuals of the iterated method exhibit far less autocorrelation,
indicating a better overall decomposition (see Fig. 9).

5 Discussion

The growth in the number of visitor arrivals to New Zealand was lower than
expected in late 2001 (e.g., by the New Zealand Ministry of Tourism, as noted in
Haywood and Randal 2004), yet there is no conclusive evidence to attribute this
78 J. Haywood and J. Randal

forecast error solely to the terrorist events of 9/11. The termination of flights by
Ansett Australia on 14 September 2001 certainly affected capacity and timing of
arrivals from Australia to New Zealand, and that would have affected Total arrivals
in September 2001 somewhat as well. Indeed Australia is the only (individual)
country of origin with a structural change in trend identified close to 9/11. The
subsequent rate of Australian arrivals to New Zealand in fact shows an increase,
following an initial drop which is plausibly explained by the Ansett effect; see
Table 2 and Fig. 7.
A further plausible cause for the lower than forecast number of visitors is the US
recession dated March 2001 (Hall et al. 2001), along with the world-wide flow-on
effects from a slow down in the US economy. The recession predates 9/11 by 6
months but that is consistent with observed features of the data. In particular, March
2001 corresponds exactly to the minimum in the second difference of an STL trend
of Total monthly (log) arrivals, indicating a maximum decrease in the slope at that
time. It is possible that the slow down seen in the Other (composite) arrivals series,
dated July 2001, may be due in part to the flow-on effects from this US recession.
It seems quite clear that the events of 9/11 did not have much influence on the
longer term numbers of visitors to New Zealand, and especially not a negative
influence. In contrast our analysis identifies other events which have had marked
structural effects on the trends in these data, especially from certain countries of
origin. In particular, the stock market crash of October 1987 preceded a dramatic
decline in arrivals from the USA, followed by a sustained period of only moderate
growth. In turn, both the intercept and slope of Total arrivals decreased in December
1987. Similarly, the Asian financial crisis of 1997–1998 precipitated a massive drop
in arrivals from Korea, with the intercept and slope of Total arrivals again both
decreasing in 1997. The SARS epidemic affected arrivals from China in a different
way, with a very short-lived but large reduction, which we class as temporary and not
structural. The overall effects of 9/11 might also be seen as temporary and negative,
but of a smaller magnitude than those associated with SARS.
Estimation of structural breaks was facilitated by a new implementation of Bai
and Perron (1998, 2003) work that is recommended for seasonal data. Specifically,
use of an iterative approach to estimate the trend and seasonal components
separately enabled us to locate structural breaks in the data, and to attribute these
to either changes in the trend or the seasonal pattern. Estimating these components
simultaneously did not achieve the same flexibility in the estimated components, nor
in the location of the break points. The agreement between the estimated parametric
trends from the iterated approach and the nonparametric STL trends is especially
pleasing, as is the lack of residual structure around those parametric trends when
compared to other trend estimates.

Acknowledgements Statistics New Zealand kindly supplied the data. We thank those who
commented on presentations at Statistics New Zealand, the Reserve Bank of New Zealand, Victoria
Management School, the ASC/NZSA 2006 Conference, and the TSEFAR 2006 Conference. We
also thank Peter Thomson for some helpful suggestions that improved the paper.
Modelling Seasonality and Structural Breaks 79

References

Andrews, D. W. K. (1991). Heteroskedasticity and autocorrelation consistent covariance matrix

estimation. Econometrica, 59, 817–858.
Andrews, D. W. K., & Monahan, J. C. (1992). An improved heteroskedasticity and autocorrelation
consistent covariance matrix estimator. Econometrica, 60, 953–966.
Australian Bureau of Statistics. (2003). Information paper: A guide to interpreting time series –
Monitoring trends. ABS Cat. No. 1349.0. Canberra: Australian Bureau of Statistics.
Bai, J., & Perron, P. (1998). Estimating and testing linear models with multiple structural changes.
Econometrica, 66, 47–78.
Bai, J., & Perron, P. (2003). Computation and analysis of multiple structural change models.
Journal of Applied Econometrics, 18, 1–22.
Box, G. E. P., & Jenkins, G. M. (1976). Time series analysis, forecasting and control (2nd ed.).
Oakland: Holden-Day.
Box, G. E. P., & Tiao, G. C. (1975). Intervention analysis with applications to economic and
environmental problems. Journal of the American Statistical Association, 70, 70–79.
Busetti, F., & Harvey, A. (2008). Testing for trend. Econometric Theory, 24, 72–87.
Cleveland, R. B., Cleveland, W. S., McRae, J. E., & Terpenning, I. (1990). STL: A seasonal-trend
decomposition procedure based on loess. Journal of Official Statistics, 6, 3–73.
Dungey, M., Fry, R., & Martin, V. L. (2004). Currency market contagion in the Asia-Pacific region.
Australian Economic Papers, 43, 379–395.
Findley, D. F., Monsell, B. C., Bell, W. R., Otto, M. C., & Chen, B. C. (1998). New capabilities and
methods of the X-12-ARIMA seasonal-adjustment program. Journal of Business and Economic
Statistics, 16, 127–177.
Franses, P. H. (1998). Time series models for business and economic forecasting. Cambridge:
Cambridge University Press.
Hall, R., Feldstein, M., Bernanke, B., Frankel, J., Gordon, R., & Zarnowitz, V. (2001). The
business-cycle peak of March 2001. Technical report, Business Cycle Dating Committee,
National Bureau of Economic Research, USA. http://www.nber.org/cycles/november2001/.
Hamilton, J. D. (1994). Time series analysis. Princeton: Princeton University Press.
Harvey, A. C. (1989). Forecasting, structural time series models and the Kalman filter. Cambridge:
Cambridge University Press.
Haywood, J., & Randal, J. (2004). Stochastic seasonality, New Zealand visitor arrivals, and the
effects of 11 September 2001. Research Report 04-1, School of Mathematics, Statistics and
Computer Science, Victoria University of Wellington, New Zealand.
Hoaglin, D. C., Mosteller, F., & Tukey, J. W. (Eds.) (1983). Understanding robust and exploratory
data analysis. New York: Wiley.
Kaminsky, G. L., & Schmukler, S. L. (1999). What triggers market jitters? A chronicle of the Asian
crisis. Journal of International Money and Finance, 18, 537–560.
Macaulay, F. R. (1931). The smoothing of time series. New York: National Bureau of Economic
Research.
Makridakis, S., Wheelwright, S. C., & Hyndman, R. J. (1998). Forecasting: Methods and
applications (3rd ed.). New York: Wiley.
NSW Department of Education (1985, April). 1987: Perspectives: Looking at Education, 8(4), pp.
3.
Pearce, D. (2001). Tourism. Asia Pacific Viewpoint, 42, 75–84.
R Development Core Team (2007). R: A language and environment for statistical computing.
Vienna, Austria: R Foundation for Statistical Computing.
Schwarz, G. (1978). Estimating the dimension of a model. The Annals of Statistics, 6, 461–464.
Statistics New Zealand. (2005). Tourism satellite account 2004. Wellington: Statistics
New Zealand.
US Department of State. (2004). Three years after 9/11: Mixed reviews for war on terror. http://
www.globalsecurity.org/security/library/news/2004/09/wwwh40915.htm.
80 J. Haywood and J. Randal

Zeileis, A. (2004). Econometric computing with HC and HAC covariance matrix estimators.
Journal of Statistical Software, 11(10), 1–17.
Zeileis, A. (2006). Object-oriented computation of sandwich estimators. Journal of Statistical
Software, 16(9), 1–16.
Zeileis, A., Kleiber, C., Krämer, W., & Hornik, K. (2003). Testing and dating of structural changes
in practice. Computational Statistics and Data Analysis, 44, 109–123.
Zeileis, A., Leisch, F., Hornik, K., & Kleiber, C. (2002). Strucchange: An R package for testing
for structural change in linear regression models. Journal of Statistical Software, 7(2), 1–38.
Forecasting the Risk of Insolvency Among
Customers of an Automotive Financial Service

Rita Allais and Marco Bosco

Abstract The aim of our study is to present an application of Generalized Linear

Models to the prevention of the risk of insolvency of an automotive financial service.
In order to forecast the payments (by installment) of the customers, we use a logit
multivariate regression model. Before fitting the model, we resort to sample logits,
Generalized Additive Models and univariate logistic regression in order to identify
the subset of best predictors and verify the assumptions underpinning the statistical
model. For the estimated model, we use the Wald statistics to assess the significance
of the coefficients, the Likelihood Ratio to test the goodness of fit of the model,
and the Odds Ratio to interpret the meaning of the estimated coefficients. In order
to verify the goodness of fit of the model, we utilize classifications tables and the
Receiver Operating Characteristic curve. Finally, we externally validate the fitted
model by means of a predictive-test on a training set.

Keywords Generalized additive model • Likelihood ratio • Logistic regression •

Receiver operating characteristic curve • Wald statistics

1 Introduction

Our paper concerns the policies of control and prevention of the risks of insolvency
of an automotive financial service. In order to forecast the performance of the
customers’ repayments (by instalment) we utilize a Generalized Linear Model.
Typically, at the expiry of the contract, the finance company classifies a customer as
being “good” or “bad” according to the correctness of his/her monthly re-payments.

R. Allais ()
Department of Social-Economic, Mathematical & Statistical Sciences, University of Turin,
Corso Unione Sovietica, 218 bis, 10134 Turin, Italy
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 81

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__7,
© Springer International Publishing Switzerland 2014
82 R. Allais and M. Bosco

The company uses a performance indicator which is based on the regularity and type
of the monthly re-payments.
Our goal is to establish the existence of a link between the performance indicator
and some features of the customer and of his or her contract, so as to be able to
forecast during the re-payment period, whether the customer will turn out to be a
good one or a bad one.
Since the response is a binary random variable, we use a logit regression model
to analyse our data. In order to identify, the subset of the best predictors for the
response (i.e. those variables which are strongly correlated with the response) and
to verify the assumptions of the statistical model (e.g., non-collinearity among
the predictors, linearity between each predictor and the logit of the response) we
adopt a step-by-step approach based on sample proportions and sample logits. For
continuous predictors we need to resort to Generalized Additive Models (GAMs)
in order to estimate the sample proportion. Furthermore, we check the results of
univariate logistic regression.
After fitting the multivariate model, we verify the goodness of fit of the model by
means of Odds Ratios, Wald statistics and the Likelihood Ratio (G-index). In order
to verify the goodness of fit of the model, we analyse the behaviour of the Sensitivity
and Specificity functions and construct the Receiver Operating Characteristic (ROC)
curve. We also externally validate the fitted model by means of a predictive test on
a training set.

2 The Response Variable and the Model

The financing house computes for each customer the value of the performance
index I , defined on the real interval Œ1; 1, taking into account some feature of
the customer related to the regularity of the re-payments, such as, the number of
paid instalments, the number of unpaid instalments, and the number of instalments
paid late. Furthermore, they classify a customer as bad if I  0:86. Therefore we
define the binary response random variable as

0 1  ;
Y D
1 ;

where  D P.I  0:86/.

For a binary response variable Y and a multiple explanatory variable X D
.X1 ; : : : ; Xp /, let

.x/ D EŒY jX D x D P.Y D 1jX D x/:

To investigate the relationship between the response probability .x/ and the
covariate vector x it is convenient (see McCullagh and Nelder 1989) to construct
Forecasting the risk of insolvency 83

a formal model thought capable of describing the effect on .x/ of changes in x.

Furthermore we suppose that the dependence of a transformation g./ on x occurs
through the linear combination

X
p
g./ D ˇi xi :
i D1

We select as link function the logit function

.x/ X p
g.x/ D logitŒ.x/ D log D ˇ0 C ˇi xi
1  .x/ i D1

and so we resort to the logistic regression model

Pp
exp.ˇ0 C i D1 ˇi xi /
.x/ D Pp :
1 C exp.ˇ0 C i D1 ˇi xi /

If some of the independent variables are discrete, nominal scaled variables, then
is inappropriate to include them in the model. This is because the numbers used to
represent the various levels are merely identifiers, and have no numeric significance.
In this situation the method of choice is to use a collection of dummy variables
(Anderson 1984). Therefore, if Xj is a categorical predictor with M categories,
defining, for each h 2 .1; : : : ; M /, the dummy variable

j 0 Xj ¤ h;
Xh D
1 Xj D h

and hence the logit function of the model becomes

X X
M 1
j j
g.x/ D ˇ0 C ˇi xi C ˇh xh : (1)
i ¤j hD1

We need a strategy to select the variables in order to identify the subset of

best predictors for the logit regression model (Hosmer and Lemeshow 2013). The
selection process should begin with a careful univariate analysis of each variable
and subsequently, following the fit of the multivariate model, the importance of each
variable included in the model should be verified. The univariate analysis involves
fitting a univariate logistic regression model and studying the effect on Y of each
predictor by itself and the interaction among variables using, for instance, plots of
sample logits.
If X j is a categorical predictor, with M categories, we define, 8 h 2 .1; : : : ; M /
the sample logits
 
ph
Q
g.x j
/ D log ;
1  ph
84 R. Allais and M. Bosco

where (Agresti 2010) the sample proportion ph is defined as

j
Nufyi D 1jxi D hgi D1;:::;n
ph D j
:
Nufxi D hgi D1;:::;n

For continuous covariates scatter diagrams are not very informative. Plotting
a fitted smooth function for a continuous predictor may reveal a general trend
without assuming a particular functional relationship. Hastie and Tibshirani (1990)
introduced the use of a GAM for the analysis of binary data. GAM replaces the
linear predictors of Generalized Linear Models by smooth function of the predictors:

X
p
g./ D sj .xj /;
j D1

where sj .:/ is an unspecified smooth function of predictor X j .

3 The Database and the Variables

Our database consists of 151,443 customers, whose application has already been
successfully scored, observed over a period of 4 years, and it includes 33 variables
arising from merged different data sources. The percentage of bad customers in the
database is 5.75 %. In order to construct the logit regression model, we consider
101;577 records letting the remainder (49,866) constitute the training-set, i.e. a
randomly selected control group, hold out of sample, to test the final model.
Aiming to identify good predictors for the model, we decide not to resort to
automatic methods, such as stepwise selection, which can lead to a biased set of
predictors, but instead to use a step-by-step approach based on sample proportions
and sample logits.
To this end, we shall distinguish among continuous (e.g., the price, the amount of
the advance, the amount of the financing, the monthly income, the age, the annual
interest rate, the net present value of the return on financing, etc.), ordinal (i.e., the
number of monthly instalments, the number of incorrect payments during the first 3
months, the loyalty of the customer, classes of cubic capacity of the car, the family
size of the customer, etc.) and dichotomous variables (i.e., the sex of the customer,
real-estate ownership, the final destination of the car, method of payment, etc.). We
shall furthermore choose those which are strongly correlated with the response but
not correlated among themselves, showing no interaction in order to avoid problems
of multicollinearity, and such that the assumption of linearity between each of them
and the sample logit of the response holds.
We verify the assumption of linearity between each predictor and the logit
function of the univariate form of model (1) by means of a plot of sample logit
against each predictor. The assumption of no-interaction effects among the variables
Forecasting the risk of insolvency 85

Table 1 Correlation matrix for all the continuous predictors

Price Advance Financing Income Age Int.rate

Price 1:00 – – – – –
Advance 0:75 1:00 – – – –
Financing 0:73 0:09 1:00 – – –
Income 0:15 0:13 0:09 1:00 – –
Age 0:11 0:01 0:14 0:12 1:00 –
Int.Rate 0:12 0:03 0:15 0:09 0:09 1:00

−2.0

−2.5
−2.5
Logit

Logit
−3.0
−3.5
−3.5

−4.0 −4.5
0 10000 20000 30000 40000 10000 20000 30000 40000 50000
Advance − Panel a Financing − Panel b

−2.0

−2.5
−2.5
Logit
Logit

−3.0
−3.5
−3.5

−4.5 −4.0
0 2000 4000 6000 8000 10000 12000 20 40 60 80 100
Monthly Income − Panel c Age − Panel d

Fig. 1 Checking the linearity of the sample logit of the response versus the continuous predictors.
The GAM approach

is checked by the plot of sample logit of one predictor conditioned by each level of
another one.
Finally, we shall examine the results of the logit regression models obtained
considering separately each of the selected predictors (Hosmer and Lemeshow
2013).
— Continuous variables
As continuous variables we consider the price of the car (PRICE), the amount
of the advance (ADV), the amount of the financing (FIN), the monthly income
of the customer (INC), the age of the customer (AGE), the annual interest rate
(IRATE).
Table 1 shows the correlation matrix of these continuous variables. Since it is
clear that ADV, FIN and PRICE are highly correlated among them, we decide to
drop from our study the price of the car.
For these continuous variables, instead of reclassifying them into classes, we
prefer to check their linearity with the sample logit of the response, simply
inspecting a plot of the smoothed prediction of the mean provided by a
GAM (Hastie and Tibshirani 1990). As displayed in Fig. 1, the assumption of
86 R. Allais and M. Bosco

−2.4 0.25

−2.6
0.20
−2.8

0.15
−3.0

Density
Logit

−3.2
0.10

−3.4
0.05
−3.6

−3.8 0.00

0 5 10 15 20 0 5 10 15 20

Annual interest rate (%) − Panel a Annual interest rate (%) − Panel b

Fig. 2 The annual interest rate. (Panel a) Checking its linearity with the sample logit of the
response via the GAM approach. (Panel b) Histogram and kernel density estimate of the annual
interest rate

linearity in the sample logit of the response is verified for the four continuous
predictors Advance, Financing, Income and Age. It must be pointed out that
those continuous predictors have strongly skewed distributions; for this reason
the results of GAM procedures are somewhat biased in the right tail. Our last
continuous variable, the annual interest rate, is not correlated with the previous
four predictors and hence it is a candidate to be a valid predictor itself. But we
find that the relation between the sample logit of the response and the annual
interest rate is not linear (see Fig. 2, panel a) and that its distribution is clearly
bimodal (see Fig. 2, panel b), suggesting it comes from a mixture of two densities.
Since we do not want to lose the prized information on customers’ interest rate,
we decide to transform this variable into a dichotomous one defined as

0 if interest rate  0:05;
IRATE D
1 otherwise:

— Discrete and ordinal variables

As discrete and ordinal variables, we consider the number of monthly instalments
(INST), the number of incorrect payments during the first 3 months (HIST),
i.e., the history of the customer during his early 3 months and the loyalty of
the customer (LOYAL). These variables are strongly correlated with the response
and for them we have the following sample proportions of bad customers listed
in Table 2.
These allow us to state that the assumption of linearity between the sample
logit of the response and each variable is verified. We therefore consider all these
three variables as good predictors.
— Dichotomous variables
As dichotomous variables, we consider the sex of the customer (SEX), the
occupation state of the customer (OCCU), real-estate ownership (OWNER), the
Forecasting the risk of insolvency 87

Table 2 Sample proportions of bad customers—I

INST HIST LOYAL
xh ph % xh ph % xh ph %
24 2:4 0 4:9 Low 8:1
36 4:7 1 4:0 Medium 3:8
48 8:8 2 5:5 High 3:4
60 13:8 3 7:6

Table 3 Sample proportions of bad customers—II

SEX OCCU OWNER DEST SEGM
xh ph % xh ph % xh ph % xh ph % xh ph %
Male 5:2 Empl. 3:1 No 8:4 Priv. 5:7 Low 5:3
Female 6:8 Self-empl. 7:1 Ye 4:8 Comm. 6:5 High 7:4
.p1  p0 /% C1:6 C4:0 3:6 C0:8 C2:1

0.10 0.15
Sample proportion of bad customers

Female
0.08 Male
Sample proportion

0.10
0.06

0.04
0.05

0.02 Female
Male

0.00 0.00

0 3000 6000 9000 12000 24 36 48 60

Monthly Income − Panel a # of Monthly Instalments − Panel b

Fig. 3 (Panel a) Gam of sample proportion vs. INC by levels of SEX. (Panel b) Sample proportion
vs. INST by levels of SEX

final destination of the car (DEST) and the segment of the car (SEGM). Among
all the variables in the database, these are the most strongly correlated with
the response and for them we have the following sample proportions of bad
customers, see Table 3.
Since the increment of the DEST is negligible, we are likely to drop this
variable as a predictor for the response. If we do so, we consider as predictors the
dichotomous variables SEX, OCCU, OWNER, SEGM and IRATE.
In our study there is no evidence of interaction between predictors. The analysis
is conducted, ignoring technicalities, simply by looking to see if the sample
proportions of predictor X conditioned on each level of another predictor, say X  ,
show the same behaviour. We do the same in the case of continuous predictors but
we resort to GAMs in order to estimate the sample proportions. Figure 3 shows the
approach we applied in the case of income (INC) versus sex and number of monthly
instalments (INST) versus sex.
The next and last step consists in examining the results of univariate logit
regression models, i.e., the models we obtain considering separately each predictor.
88 R. Allais and M. Bosco

Table 4 Results of univariate logit model for the predictors

Beta OR G-Index
ADV 0:000167 0:9199 1050.06
FIN 0:000130 1:0672 1278.00
INC 0:000274 0:8720 305.33
AGE 0:030860 0:9696 911.61
INST 0:051684 1:0530 2027.26
HIST 2:065065 7:8858 2757.48
LOYAL(Medium) 0:793644 0:4522 854.85
LOYAL(High) 0:899862 0:4070 854.84
SEX(Male) 0:122465 1:1303 180.84
OCCU(Self.) 0:895209 2:4478 786.78
OWNER(Yes) 0:588468 0:5551 422.04
DEST(Comm.) 0:139410 1:1490 9.99
SEGM(High) 0:335890 1:3991 112.25
IRATE(High) 1:117891 3:0584 1234.84
The starred ORs are proposed in terms of increments of 500 euros; e.g. OR.ADV / D exp.ˇADV 
500/

We use the Wald statistics to assess the significance of the coefficients and the
Likelihood Ratio (G-index) to test the goodness of fit of the model (Harrell 2001;
Balakrishnan 2013), being aware that these are not informative when, as here, the
sample size is very large whence most if not all of the coefficients exhibit statistical
significance. However, in order to interpret an estimated coefficient, we resort to the
Odds Ratio, which can be computed, for any i D 1; : : : ; p, as ORi D exp.ˇi /.
Table 4 shows the results of each univariate logit model in terms of the estimated
coefficients, the Odds-Ratios, eventually corrected by the unit of measure of the
predictor, and the corresponding G-index for the model. The p-values associated to
the Wald’s test are all of order approximately < 2e 16 , except for the variable DEST
for which is 0.013 %. This result together with the corresponding G-index (9:99)
which is very low, justifies our previous idea that the contribution of the variable
DEST to the response can be regarded as negligible and hence dropped from the
final model.
This preliminary data analysis leads us finally to the identification of the best
predictors to be considered in the logistic model.

4 The Estimated Logit Regression Model

At this point we consider 4 continuous, 2 discrete, 1 ordinal and 5 dichotomous

predictors, so that the full 12 predictors logit regression model can be written as

g.x/ Dˇ0 C ˇ1 ADV C ˇ2 FIN C ˇ3 INC C ˇ4 AGE C ˇ5 INSTC

ˇ6 HIST C ˇ7 LOYAL(Medium) C ˇ8 LOYAL(High)C
Forecasting the risk of insolvency 89

Table 5 Estimates of the coefficients of the full model ordered by OR

Beta SE (beta) OR
Const. 3:3400 0:10500
HIST 0:0194 0:04038 6:862
OCCU (self) 0:4289 0:04136 1:536
IRATE (high) 0:4224 0:04162 1:526
SEGM (high) 0:2033 0:04464 1:225
SEX (male) 0:1021 0:03181 1:108
FIN 0:0001 0:00001 1:036
INST 1:9260 0:00150 1:020
AGE 0:0129 0:00131 0:987
ADV 0:0002 0:00001 0:933
INC 0:0002 0:00002 0:932
OWNER (yes) 0:2606 0:03082 0:771
LOYAL (medium) 0:4911 0:03152 0:612
LOYAL (high) 0:6670 0:09473 0:513

ˇ9 SEX(Male) C ˇ10 OCCU(Self) C ˇ11 OWNER(Yes) C

ˇ12 SEGM(High) C ˇ13 IRATE(High):

Resorting to R software, defined the logit regression model via glm (Venables
and Ripley 2002), we obtain estimates of the coefficients shown in Table 5 along
with the correspondent standard errors.
We remark that all the p-values associated to the Wald’s test applied to each
coefficient are of order < 2e 16 , but for the predictor SEX for which it reduces to
0.133 %.
The G-index for the full model is 6,813.15 (on 13 d.f.) and due to the large
sample size its p-value is clearly zero; so it does not tell us much about the goodness
of fit of the model itself.
The elements of the matrix of the estimated covariances between the parameter
estimates in the linear predictor of the model are all close to zero, suggesting
non-collinearity among the predictors, as expected.
In order to have an idea about the contribution of each predictor to the response,
we can consider (see Table 5) the Odds Ratios of each estimated coefficient ordered
by their magnitude. Here, the ORs for ADV, FIN and INC are presented in terms of
increments of 500 euros.
According to this classification, the best predictors are HIST, OCCU and IRATE.
The very high value of the OR for HIST, and hence its heavy contribution to the
predicted response, is clearly obvious if we recall the definition of HIST itself. Due
to its low OR, we could drop LOYAL from the model. However, the society’s staff
argued that this predictor makes sense from the business point of view.
90 R. Allais and M. Bosco

Table 6 Classification—I
Observed Y
Predicted YO 0 1
0 Nu.yOi D 0jyi D 0/ Nu.yOi D 0jyi D 1/
1 Nu.yOi D 1jyi D 0/ Nu.yOi D 1jyi D 1/

Table 7 Classification—II
0 D 0:50 Observed Y
Predicted YO Good Bad Good Bad
Good 95;471 5;347 0:997 0:915
Bad 269 491 0:003 0:085

4.1 Classification Tables and the ROC Curve

Fixing a cut-off value 0 , we transform the n continuous fitted values O i of the final
model into n binary predicted yOi values according to

0 O i  0 ;
yOi D
1 O i > 0

and in this way, we are able to construct the Classification Table, see Table 6.
If we now set the cut-off level 0 D 0:50, we obtain the following Classification
Table, see Table 7.
Hence, among the bad customers (5.75 % of the entire population), our model
will correctly classify only 8.5 % of them as bad, while, among the good customers,
it classifies 99.7 % of them correctly.
We can improve the predictive behaviour of the model looking for a different
value for the cut-off. Varying  2 Œ0; 1 and defining the functions

Sensitivity./ D Nu.yOi D 1jyi D 1/=Nu.yi D 1/;

Specificity./ D Nu.yOi D 0jyi D 0/=Nu.yi D 0/:

We can construct the ROC curve (Agresti 2010) plotting Sensitivity./ as a

function of .1  Specificity.//. The area under ROC curve is the Concordance-
Index. The C-Index gives a measure of the concordance of predicted and observed
values of the response and hence of the predictive power of the model. For our
model we observe a value of the C-Index equal to 0:81 showing a quite good power
of discrimination (see Fig. 4, panel b).
Analyzing the shape of the functions Sensitivity and Specificity, we can choose
an optimal operating value for the cut-off. If we look at the plot of the two functions
in Fig. 4, panel a, we may argue that a good cut-off point should lie approximately
between 6 and 50 %. Thus we can re-define the optimal operating cut-off value as
Forecasting the risk of insolvency 91

1.0 1.0
Specificity

0.8 0.8

Sensitivity
0.6 0.6

0.4 0.4

0.2 0.2
π optimal Sensitivity C−Index = 0.81
0.0 0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

π − Panel a 1 − Specificity − Panel b

Fig. 4 Panel a: sensitivity and specificity functions and optimal cut-off point. Panel b: the
Receiver Operating Characteristic curve and the C-Index

Table 8 Classification—III
0 D 0:27 Observed Y
Predicted YO Good Bad Good Bad
Good 94;857 4;952 0:991 0:848
Bad 882 886 0:009 0:152

Table 9 Classification—IV
0 D 0:27 Good Bad Good Bad
Good 46;468 2;511 0:991 0:847
Bad 432 455 0:009 0:153

optimal such that Specificity./ D 0:99. In our case, we obtain optimal D 0:27,
corresponding to a value of Sensitivity.optimal / D 0:15. The new Classification
Table with 0 D optimal is as Table 8.
Now, among the bad customers, our model will classify 15.2 % correctly as bad,
while it correctly classifies 99.1 % of the good customers.
We remark that, comparing the two previous Classification Tables, the rate of bad
customer classified as good decreases from 91.5 to 84.8 %.
Since our goal is to use the fitted logistic model to predict the behaviour of
future customers, it seems natural to validate the fitted model on the training set.
The percentage of bad customer in the training set is 5.95 % (D 2966=49866). The
predictive power of the model on the training set is still good (C-index D 0:80) and
for 0 D 0:27 we have the following Classification Table, see Table 9.
We obtain (approximately) the same values for Sensitivity, Specificity and hence
for C-Index: these results assess the goodness of fit of the proposed model.

5 Final Remarks

Despite a very detailed and thorough analysis of the available data, the identification
of bad risk customers remains difficult. The current prediction model may be viewed
in the same way as an Epidemiological screening test. In that context, such tests
92 R. Allais and M. Bosco

are acceptable only when the sensitivity and specificity exceed 85 %. Accordingly,
the specificity result satisfies the Epidemiological criterion but the result for the
sensitivity does not. Thus, while the current model forecasts the reliability of the
customer during the repayment period in a fair way, it seems that there are other
important risk factors which have yet to be identified.

Acknowledgements Authors are indebted to the coordinating editors and to the anonymous
referees for carefully reading the manuscript and for their suggestions.

References

Agresti, A. (2010). Analysis of ordinal categorical data (2nd ed.). New York: Wiley.
Anderson, J. A. (1984). Regression and ordered categorical variables. Journal for the Royal
Statistical Society, B, 46, 1–30.
Balakrishnan, N. (2013). Handbook of the logistic distribution. New York: CRC Press.
Harrell, F. E. (2001). Regression modeling strategies. New York: Springer.
Hastie, T., & Tibshirani, R. (1990). Generalized additive models. London: Chapman and Hall.
Hosmer, D. W., & Lemeshow, S. (2013). Applied logistic regression (3rd ed.). New York: Wiley.
McCullagh, P., & Nelder, J. A. (1989). Generalized linear models. London: Chapman and Hall.
Venables, W. N., & Ripley, B. D. (2002). Modern applied statistics with S. New York: Springer.
On Joint Modelling of Constrained Mean
and Covariance Structures in Longitudinal Data

Jing Xu and Gilbert MacKenzie

Abstract A data-driven method for modelling the intra-subject covariance matrix

is developed in the context of constrained marginal models arising in longitudinal
data. A constrained iteratively re-weighted least squares estimation algorithm is
applied. Some key asymptotic properties of the constrained ML estimates are
given. We analyze a real data set in order to compare data-driven covariance
modelling methods with classical menu-selection-based modelling techniques under
a constrained mean model, extending the usual regression model for estimating
generalized autoregressive parameters. Finally, we demonstrate, via a simulation
study, that a correct choice of covariance matrix is required in order to minimise
not only the bias, but also the variance, when estimating the constrained mean
component.

Keywords Cholesky decomposition • Covariance modelling • Inequality

constraints • Longitudinal data • Marginal models

1 Introduction

In longitudinal studies, constrained problems are often of interest in biomedicine

and clinical trials. For example, in xenograft experiments, tumors in immunosup-
pressed mice usually continue to grow unchecked during the study period, leading
to a larger mean size in the control group thereby generating a natural inequality
constraint (Tan et al. 2005). Moreover, treatment effects arising in different groups in
comparative clinical studies may form a natural ordering (Crowder and Hand 1990).
Constrained problems of this type may be investigated using regression models with

J. Xu ()
Department of Mathematics, Statistics & Economics, Birkbeck College, London, UK
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 93

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__8,
© Springer International Publishing Switzerland 2014
94 J. Xu and G. MacKenzie

inequality constraints: see Xu and Wang (2008a,b), Fang et al. (2006), Tan et al.
(2005), Shi et al. (2005) for constrained estimation and Pilla et al. (2006), Cysneiros
and Paula (2004), Park et al. (1998), and Shin et al. (1996) for inequality-constrained
hypothesis testing. The aim of this paper is to develop estimation methods for the
parameters in constrained-mean covariance models.
Historically, intra-subject correlations have been modelled by menu selection
techniques or by working correlation structures (Liang and Zeger 1986). The
conventional approach is to select a particular covariance model from a menu of
potential candidate structures, e.g., compound symmetry, AR(1), ARMA or unstruc-
tured covariance (Diggle et al. 2002). However, such procedures may not work well
in practice. For example, misspecification of the working covariance structure may
lead to a large loss of efficiency of the estimators of the mean parameters (Wang
and Carey 2003). Although the unstructured covariance is assumed to approximate
the true covariance structure, the number of nuisance parameters, say j k , in the
resulting unstructured correlation matrix may be excessive and cause convergence
problems in the iterative estimation process (Dobson 2002). We note, too, that when
the true covariance structure is not contained in the menu set, this approach may
fail to identify the optimum covariance structure suggested by the data (Pan and
MacKenzie 2007). Therefore, in this paper, we adopted the data-driven method
proposed by Pourahmadi (1999) to fit the covariance matrix in constrained models.
Accordingly, in Sect. 2 we formulate the constrained model and outline the data
driven covariance modelling approach. In Sect. 3 we discuss constrained Maximum
Likelihood Estimation, and outline some key elements of the relevant asymptotic
theory for constrained estimators. In Sect. 4 we analyze data from a small diabetic
study, consider two simulation studies in Sect. 5 and discuss the findings briefly in
Sect. 6.

2 Model Formulation

2.1 Constrained Mean Covariance Model

Consider a balanced longitudinal study. Let m denote the common number of

measurement times. The response vector for subject i is denoted by yi . We assume
that the yi arise from the constrained marginal model

yi D Xi ˇ C "i for i D 1;    ; n (1)

s:t: Aˇ > b;

where s:t: means “subject to”; Xi is a known m  p design matrix for the i th
individual; ˇ is a p  1 vector of unknown coefficients to be estimated; A is a
k  p matrix and b D .b1 ;    ; bk /0 is a k  1 vector; "i D ."i1 ;    ; "im /0 are
Joint Constrained Mean and Covariance Modelling 95

independently distributed as N.0; †/ for i D 1;    ; n. The inequality constraint set

fˇ W Aˇ > bg is quite general, containing order restrictions as a special case.

2.2 Covariance Model

Since the subject-specific covariance matrix † is positive definite, there exists a

unique lower triangular matrix, T , with 1’s as main diagonal entries and a unique
diagonal matrix, D, with positive diagonal entries such that T †T 0 D D. The
below-diagonal entries of T are the negatives of the autoregressive coefficients, jk ,
j 1
in yOij D ij C †kD1 jk .yik  ik /, the linear least squares predictor of yij based
on its predecessors yi.j 1/ ;    ; yi1 . Here the ij s are the expectations of the yij s.
The diagonal entries of D are the innovation variances j2 D var.yij  yOij /; where
1  j  m and 1  i  n (Pourahmadi 1999). The parameters jk and &j logj2
are modelled as jk D z0jk
and &j D h0j . Here
and are vectors of order q C 1
and d C 1, respectively.

2.3 Joint Model

Thus, the constrained mean covariance model may be represented by an augmented

regression model:

ij D xij0 ˇ s:t: Aˇ > b;

jk D z0jk
; (2)
&j D h0j ;

where ˇ,
and are the three regression parameters of scientific interest and only
ˇ is constrained. Typically, the design matrices are time- or lag-dependent i.e, xij0 D
x.t/0ij ; z0jk D z.lag/0jk and h0j D h.t/0j . The latter two equations are said to define
the class, C  , of covariance structures when z0jk is a polynomial in lag and h0j is a
polynomial in time, respectively (MacKenzie 2006). Now, our focus is on the effect
of the constraints on simultaneous estimation of the parameters of interest.
The log-likelihood function for the parameters is given by

1 X 0 0 1
n
nm n 1
`.ˇ;
; / D  log.2/  logjT 1 DT 0 j  r T D Tr i ; (3)
2 2 2 i D1 i

where rij D yij  xij0 ˇ is the j th element of ri D yi  Xi ˇ, the vector of residual,

and the matrix Xi has row vectors xij0 .j D 1; 2;    ; m/.
96 J. Xu and G. MacKenzie

3 Constrained Maximum Likelihood Estimation

3.1 Estimation of Parameters

Denote the feasible solution set of the constrained model (2) by S D fˇ W Aˇ > bg
and the function (3) by `.ˇ;
; /. Then the constrained ML estimation problem is

max `.ˇ;
; /: (4)
ˇ2S

Based on the iteratively re-weighted least squares algorithm given by Pan and
MacKenzie (2003), we present the following procedure for solving the constrained
estimation problem (4).
Given
and , the constrained regression parameters are determined by

X
n
ˇ D arg min ri0 †1 ri : (5)
ˇ2S
i D1

Finding the optimal numerical solution for ˇO in (5) is equivalent to solving a

quadratic programming problem. Several methods are available in mathematical
programming such as: the active set method, dual method and interior point method,
(Fletcher 1971; Goldfarb and Ininani 1983). In practice, however, the solve.QP
O
function in the quadprog package in the R language, can be used directly to obtain ˇ.
Secondly, given ˇ and , the first order estimating equation for
is

X
n
0
U2 .
/ D Zi D 1 .ri  Zi
/ D 0; (6)
i D1

0 Pj 1
where the matrix Zi , of order m  .q C 1/, has typical row zij D kD1 rik z0jk . The
estimates of
can be obtained easily from Eq. (6). And finally, given ˇ and
, the
estimating equation for is

1 X 0 1
n
U3 . / D H .D ei  1m / D 0; (7)
2 i D1

where H D .h01 ; h02 ;    ; h0m /0 , ei D .ei1 ; ei 2 ;    ; eim /0 with eij D .rij  rOij /2 and
Pj 1
rOij D kD1 jk rik , are the m  .d C 1/ matrix of covariates and the m  1 vector of
squared fitted residuals, respectively, and 1m is the m  1 vector of 1’s. Equation (7)
may be solved iteratively by the Newton–Raphson algorithm to obtain . O
By initializing at † D Im , the iterative procedure proceeds within (5)–(7) until
convergence. We refer to it as a constrained iteratively re-weighted least squares
(CIRWLS) algorithm.
Joint Constrained Mean and Covariance Modelling 97

3.2 Asymptotic Properties

In this section some key asymptotic properties of the constrained estimates are only
stated as they can be derived directly in the spirit of Xu and Wang (2008b). Those
differences in computing certain moments, which are mostly due to the different but
smoother reparameterisation of †, can also be found in Pourahmadi (2000).
All of the asymptotic results (as n ! 1) take m; p; q and d to be fixed. Let
parameter spaces B,  and ƒ, where ˇ 2 B,
2  and 2 ƒ, be compact
subspaces of Rp , Rq and Rd . Let ˇ0 be the true parameter of ˇ lying in B and
˛0 D .
00 ; 00 /0 be the true parameter of ˛ D .
0 ; 0 /0 lying in   ƒ. Furthermore,
we denote all of the unknown parameters by  D .ˇ 0 ;
0 ; 0 /0 and the unknown
true values by 0 D .ˇ00 ;
00 ; 00 /0 . Let aj0 ; j D 1;    ; k, be the rows of the matrix A.
Define a .pCqCd C3/-dimensional vector Aj D .aj0 ; 0;    ; 0/0 ; for j D 1;    ; k.
Then the constrained ML estimation problem (4) becomes

max `./ W s:t: A0j  > bj ; j D 1;    ; k: (8)

Here Aj ; j D 1;    ; k are assumed to be linearly independent. The optimization

O Moreover, for
solution of the constrained estimators of problem (8) is denoted by .
the remainder of the article, it is assumed that
Condition I. For all ˛ ¤ ˛0 in   ƒ,

†.˛/ ¤ †.˛0 /;

where †.˛/ D T .
/1 D. /T 0 .
/1 .
Condition I is needed to guarantee that the density function f .yi I ˇ; ˛/ of Yi is
identifiable. Noting that the .1; 1/-th element in † is equal to the .1; 1/-th element
in D, Condition I can be always satisfied given linear regression models of . The
consistency of the maximum likelihood estimate is presented in Theorem 1.
Theorem 1. Suppose that the design matrices Xi for i D 1; 2;    , are bounded
uniformly in the sense that there exists a real number P c such that j.Xi /ls j  c where
.Xi /ls is the .l; s/th element of Xi ; the limit of n1 niD1 Xi0 †1 .˛/Xi exists for ˛ 2
  ƒ. Then the constrained ML estimators O D .ˇO 0 ;
O 0 ; O 0 /0 are strongly consistent
for the true value 0 D .ˇ00 ;
00 ; 00 /0 ; that is, O D .ˇO 0 ;
O 0 ; O 0 /0 ! 0 D .ˇ00 ;
00 ; 00 /0
almost surely as n ! 1.
Theorem 1 implies that O lies in a neighborhood of 0 when n is sufficiently large.
Hence it is sufficient to discuss the properties of O in a small neighborhood of 0 .
Next we give a brief description about the possible approximate representations
of the constrained estimates O when n is sufficiently large. More details are available
in Xu and Wang (2008b).
98 J. Xu and G. MacKenzie

First denote the feasible solution set of the model (8) by S D f W A0j  >
bj ; j D 1;    ; kg which has the following subsets

S ı D f W A0j  > bj ; j D 1;    ; kgI

Sj D f W A0j  D bj ; A0i  > bi ; i D 1;    ; k; i ¤ j gI
Sj1 ; ;jt D \tlD1 Sjl :

So the parameter 0 may be located in S ı , or the relative interior parts of Sj or their

intersections, Sj1 ; ;jt .
Denoting the relative interior of a set, say K, by ri.K/ and recalling the
consistency of O , the following results obtain. If 0 is in S ı , then O must be located
in S ı with a probability approaching to one for sufficiently large n. If 0 is in
ri.S1 / D f W A01  D b1 ; A0i  > bi ; i D 2;    ; kg, then O may be in S ı
or ri.S1 / with a probability approaching to one for sufficiently large n. If 0 is in
ri.S12 / D . W A01  D b1 ; A02  D b2 ; A0j  > bj ; j D 3;    ; k/, then most of O may
be in S ı ; ri.S1 /; ri.S2 / and ri.S12 /. Other cases can be analyzed similarly. Then,
if O is in, for instance, ri.S12 /, the inequality-constrained problem (8) is reduced
to an equality-constrained problem with a constraint set fA01  D b1 ; A02  D b2 g.
Thus, the approximate representation of O may be obtained by equality constrained
optimization method.
In conclusion, for a given location of 0 except S ı , the estimators n 2 .O 0 / will
1

have different approximate representations, hence different asymptotic distributions

(see Theorem 2), according to different locations of . O
Theorem 2 explains that the constrained ML estimator O has a piecewise
asymptotic normal distribution.
Theorem 2. Suppose that
ˇ
1X @ log f .Yi I ˇ; ˛/ @ log f .Yi I ˇ; ˛/0 ˇˇ
n
Vn11 .ˇ0 ; ˛0 / D E0 Œ ˇ ! V11 .ˇ0 ; ˛0 /I
n i D1 @ˇ @ˇ ˇ D ˇ0
˛ D ˛0

ˇ
1X @ log f .Yi I ˇ; ˛/ @ log f .Yi I ˇ; ˛/0 ˇˇ
n
Vn .ˇ0 ; ˛0 / D
22
E0 Œ ˇ ! V22 .ˇ0 ; ˛0 /;
n i D1 @˛ @˛ ˇ D ˇ0
˛ D ˛0

where f .yi I ˇ; ˛/ is the density function of Yi ; E0 denotes the expectation oper-

ator with ˇ D ˇ0 and ˛ D ˛0 . All the matrices Vn11 .ˇ0 ; ˛0 /, Vn22 .ˇ0 ; ˛0 /
and V11 .ˇ0 ; ˛0 /, V22 .ˇ0 ; ˛0 / are assumed to be positive definite. Further, let
V D diag.V11 .ˇ0 ; ˛0 /; V22 .ˇ0 ; ˛0 // and Vn D diag.Vn11 .ˇ0 ; ˛0 /; Vn22 .ˇ0 ; ˛0 //. The
computation of Vn at the true value 0 is given in the appendix.
Then, under the same assumptions in Theorem 1, when O is in S ı , the asymptotic
distribution of O follows

n 2 .O  0 / ! N.0; V 1 /0 I
1 L
Joint Constrained Mean and Covariance Modelling 99

when O is not in S ı , the asymptotic distribution of O follows

n 2 .O  0 / ! N.0; C V 1 /;
1 L

where C D I  V 1 G.G 0 V 1 G/1 /G 0 with G D .Aj1 ;    ; Ajt / with

fj1 ;    ; jt g D fi W A0i O D bi ; i D 1;    ; kg and I being a .p C q C d C 3/-order
L
identity matrix. Here “!” stands for convergence in distribution.

4 Re-analysis of Diabetic Patient Data

4.1 The Data

We reanalyze data on a comparative study among diabetic groups given by Crowder

and Hand (1990). These data have been analyzed by Shin et al. (1996) and Cysneiros
and Paula (2004) in the context of inequality hypothesis testing. Originally, there
were four patient groups, but following Shin et al. (1996) and Cysneiros and Paula
(2004), we only consider three groups: the control group .n1 D 8/, the diabetic
group without complications .n2 D 6/ and the diabetic group with hypertension
.n3 D 7/. For each patient the response, a physical task, was measured at
1; 2; 3; 4; 5; 6; 8 and 10 min. Additional responses were measured at 12 and 15 min,
but these were dropped because a high proportion were missing. One additional
missing response, at minute 8, was imputed by the mean response at that time.
Accordingly, the data selected for analysis are balanced, but irregularly spaced. They
comprise a challenging set because of the relatively small sample size. Data analysis
and simulations in this paper were conducted in the R software package (Version
2.8.1). The function solve.QP adopted in programmes for constrained optimization
is from the R package quadprog contributed by Berwin A. Turlach and Andreas
Weingessel.

4.2 Joint Constrained Mean Covariance Regression Models

Scatterplots of the response against time suggest that three constant means may
be appropriate for these data. Measurement times are rescaled to j D 1; 2;    ; 8:
Let yilj be the observed physical task for the i th patient of the lth group at the
time j . Here l D 1; 2; 3 is merely a group indicator. We assume model (2), namely:
yil D l C "i , where l D l  1 and l is a scalar representing the common mean
value in group l, where 1 is the .m  1/ unit vector. In particular, we have 1 >
2 > 3 . Moreover, we assume "i  N.0; †/ across subjects. Shin et al. (1996)
and Cysneiros and Paula (2004) adopted similar conventions in their analysis.
100 J. Xu and G. MacKenzie

Table 1 Diabetic patient data

j 1 2 3 4 5 6 7 8
1 10.1 0.973 0.969 0.978 0.932 0.931 0.887 0.879
2 0.970 10.1 0.966 0.954 0.895 0.901 0.812 0.808
3 0.533 0.404 9.2 0.975 0.925 0.930 0.851 0.844
4 0.633 0.174 0.545 9.9 0.950 0.951 0.875 0.871
5 0.372 0.346 0.106 0.904 11.9 0.966 0.915 0.905
6 0.050 0.002 0.086 0.331 0.627 11.3 0.888 0.879
7 1.902 1.157 0.187 0.698 0.836 0.082 18.5 0.982
8 0.171 0.101 0.139 0.282 0.031 0.021 0.997 19.4
O 2 10.1 0.5 0.5 0.3 1.1 0.6 2.0 0.7
Sample variances (along the main diagonal), correlations (above the main diagonal), generalized
autoregressive parameters (below the main diagonal) and innovation variances (last row)

Table 1 shows the sample variances (main diagonal) and the sample correlation
matrix (upper triangle) computed from the data. Inspection suggests that a model
with a stationary covariance structure and strong correlation would be a reasonable
model choice. Accordingly, classic covariance structures such as compound symme-
try CS, AR.1/ and ARMA.1; 1/ are strong competitors for a data-driven covariance
modelling approach.
We fitted model (2), using the group constraints described above and C  , and
estimated the generalized autoregressive parameters, jk ; k < j; j D 2;    ; 8, and
innovation variances, j2 ; j D 1    ; 8. The estimated generalized autoregressive
parameters (Table 1, lower triangle) showed strong dependence on both indices j
and k and we fitted the extended model

jk D
0 C
1 j C  C
q1 j q1 C
1 k C  C
q2 k q2 C
1 .j k/C  C
q
3
.j k/q3 ;

in order to capture this structure. Here j D 2;    ; 8 and k D 1;    ; 7 and

D .
0 ;
1 ;    ;
q1 ;
1 ;    ;
q2 ;
1 ;    ;
q
3
/0

with dim.
/ D q D q1 Cq2 Cq3 . This model extends the standard regression model
in which
 and
 are absent. We call the class of covariance structures indexed by
the extended model, C  . In the context of these data, it is a rather ambitious model,
but we sought a low order polynomial model in the indices and their interaction
satisfying the usual marginality constraints.

4.3 Model Selection

In the sequel we label the standard covariance model CM1 and the extended model
CM2. We fitted two versions of each: CM1a with q D 3 (generalized auto-
regressive parameters) and d D 3 (innovation variances) and CM1b with q D 5
Joint Constrained Mean and Covariance Modelling 101

Table 2 Data-driven, regression-based, covariance modelling for † compared with several menu-
selection methods
Structure of † No. of parameters lOmax AIC BIC
CS 2 310.26 29.74 29.84
AR(1) 2 269.46 25.85 25.95
ARMA(1,1) 3 267.77 25.79 25.94
CM1a 8 262.28 25.74 26.14
CM2a 11 258.26 25.64 26.19
CM1b 14 250.11 25.15 25.85
CM2b 15 247.11 24.96 25.71

Fig. 1 Fitted regressograms for the diabetic patient data, model CM2b . Left: fitted generalized
autoregressive parameters and surface; Right: fitted log-innovation variances

and d D 7; CM2a with q1 D 2; q2 D 2; q3 D 2 and d D 3 and CM2b with

q1 D 5; q2 D 4; q3 D 1 and d D 3. In addition, we fitted three stationary models:
CS, AR.1/ and ARMA.1; 1/, making seven models in total.
We used the BIC D .2=n/ Òmax C .q C d C 2/ log n=n as our model selection
criterion and include the AIC D .2=n/ Òmax C 2.q C d C 2/=n for comparison. The
smaller the AIC or BIC value, the better the model. We systematically searched C 
for the optimum standard model (Pan and MacKenzie 2003) and found CM1b above.
Then we searched the extended covariance class C  and found CM2b . The other
two covariance models are included for the purposes of comparison. The stationary
models were fitted directly using a specially written Fisher-scoring algorithm for
constrained mean optimization.
The results are shown in Table 2. The model with the minimum BIC is CM2b and
details of the fit are given in Fig. 1. However, the classical AR.1/ and ARMA.1; 1/
models are very close competitors. The extended covariance models, CM2, perform
best in terms of BIC, but the AR(1) model, which fits two parameters, is preferred
on the grounds of simplicity of interpretation. Overall, Table 2 allows us to exclude
the CS model and adopt the AR.1/ model on the grounds that its BIC is close to that
of the optimum model in C  (MacKenzie 2006).
Finally, we remark that all of this model fitting is exploratory rather than
confirmatory.
102 J. Xu and G. MacKenzie

5 Simulation Studies

Two different simulation studies are conducted. The first study aims to investigate
how mis-specification of the covariance structure affects the estimates of the mean
(˙s:e) in the constrained model. We also present the results for unconstrained
models. The emphasis is on non-stationary covariance models. The objective of
the second study is to investigate the utility (if any) of the extended model CM2.
For simplicity, however, we only consider balanced and regular designs in these
simulation studies.

5.1 Simulation I

The simulated data were generated from the following model

yil .tj / D l C Wi .tj / C "i .tj /

for: l D 1; 2; 3I i D 1;    ; nI j D 1;    ; m; and where l is the mean of group l.

The Wi .tj /, generating the within subject serial correlation, are sampled from n
independent copies of a non-stationary Gaussian process. The "i .tj /, measurement
error, are a set of N D m  n mutually independent Gaussian random variables and
Wi .tj / and "i .tj / are statistically independent.
We present representative results from our simulation study for the particular
case n D 60 (nl D 20; l D 1; 2; 3), time points tj W j D 1;    ; m D 8, with
a diabetic data-directed mean: 1 D 7:322, 2 D 6:860 and 3 D 4:191. We
sampled the Wi .tj / from an ARIMA.1; 1; 0/ process. Denoting Wi .tj /  Wi .t.j 1/ /
by 4Wi .tj /, the 4Wi .tj / can be represented as AR(1) model, i.e, 4Wi .tj / D
4Wi .t.j 1/ / C Zi .tj /. We set  D exp.0:5/. The Zi .tj / white noise random
variables were generated with E.Zi .tj // D 0 and Var.Zi .tj // D  2 .12 /, to give
Var.4Wi .tj // D  2 and we set  D 3. The random variables "i .tj / are mutually
independent with E."i .tj // D 0 and Var."i .tj // D 1.
In contrast to the stationary covariance structure in the diabetic patient data, our
simulated data have a non-stationary covariance structure as the variance of yi .tj /
increases with time tj and correlations between yi .tj / and yi .tj Ck / also increases
with time tj for a given lag k. For the constrained case we have 1 > 2 > 3 : The
number of replications was 500.
The following models were fitted: CM1 with q D 3 and d D 3, CS, AR.1/ and
ARMA.1; 1/. The results are shown in Table 3. The data-driven covariance model,
CM1, performs much better than the other three models in the constrained and
unconstrained cases. In particular, the smaller biases and reduced standard errors
should be noted, particularly in the constrained case. In the unconstrained case, the
bias is negligible for all models, but the standard errors in CM1 are 33–40 % of
those produced by the other models.
Joint Constrained Mean and Covariance Modelling 103

Table 3 Simulation I. Average maximum likelihood estimates and standard errors: for the true
model and four completing models in constrained and unconstrained cases
Constrained
Est. True CM1 CS AR1 ARMA
O1 7.322 7.433(0.732) 7.843(1.853) 7.827(1.820) 7.830(1.826)
O2 6.860 6.731(0.723) 6.505(1.619) 6.512(1.603) 6.513(1.613)
O3 4.191 4.150(0.843) 3.928(2.011) 3.938(1.996) 3.934(2.006)
lO 1,120.048 1,134.989 1,686.788 1,332.732 1,280.458
AIC 37.435 38.100 56.293 44.491 42.782
BIC 37.540 38.379 56.363 44.561 42.887

Unconstrained
Est. True CM1 CS AR1 ARMA
O1 7.322 7.306(0.832) 7.312(2.268) 7.318(2.215) 7.324(2.219)
O2 6.860 6.855(0.824) 6.861(2.034) 6.851(2.008) 6.847(2.019)
O3 4.191 4.153(0.848) 4.102(2.222) 4.107(2.191) 4.106(2.201)
lO 1,120.048 1,134.784 1,686.543 1,332.593 1,280.318
AIC 37.435 38.100 56.285 44.486 42.777
BIC 37.540 38.379 56.355 44.556 42.882

It is generally well known that constrained models produce a smaller mean square
error for the estimates of mean, but that this may conceal a larger bias (Liew 1976;
Xu and Wang 2008a). So the bias here must arise in part from the constraints
imposed. However, it is interesting to note that the data-driven model produces
much smaller biases compared with the menu-selection models in the constrained
case. Accordingly, we may conclude that covariance mis-specification exacerbates
the bias in the mean caused by the imposed constraints.

5.2 Simulation II

We simulated the data for the three groups data from the model

yil .tj / D l C ui C "i .tj /

for: l D 1; 2; 3I i D 1;    ; nI j D 1;    ; m. The random variables ui are mutually

independent with E.ui / D 0 and Var.ui / D  2 . The random variables "i .tj / are
mutually independent with E."i .tj // D 0 and Var."i .tj // D  2 . These two types
of random variables are independent. The remaining assumptions are the same as
in Simulation I. Denote by I and J the m  m identity matrix and the m  m unit
matrix. Let  D  2 =. 2 C  2 / and  2 D  2 C  2 . Then, † D  2 f.1  /I C J g has
a Compound Symmetric structure. The modified Cholesky decomposition of this
matrix † is T †T 0 D D where
104 J. Xu and G. MacKenzie

Table 4 Simulation II. Maximised log-likelihood estimates and AIC and BIC: comparison of CM2
and CM1 under a Compound Symmetry model
var.ui / D 22 , var."i / D 1 var.ui / D 32 , var."i / D 1
Est. True CM2 CM1 True CM2 CM1
lO 785.809 786.336 804.336 809.615 819.084 843.160
AIC 26.260 26.478 27.082 27.054 27.569 28.372
BIC 26.330 26.757 27.361 27.124 27.849 28.651

0 1
1 0 0 ::: 0
B   1 0 ::: 0C
B 1C C
B   C
T DB  1C2  1C2 1 ::: 0C
B C
B :: :: :: :: C
@ : : : :A
  
 1C.m2/  1C.m2/  1C.m2/ ::: 1

and

2 2 .m  1/2
D D diagf 2 ;  2 .1  /;  2 .1  /;    ;  2 .1  /g:
1C 1 C 2 1 C .m  1/

Notice that in T the values in every sub-diagonal decrease when the indices j and
k increase. Accordingly, this suggests modelling both indices simultaneously using
CM2, rather than CM1 which uses just one value to fit every sub-diagonal. Imposing
the inequality restriction on the model, we set q1 D 1, q2 D 1 and q3 D 1 and
d D 3 for CM2 and q D 3 and d D 3 for CM1. Furthermore, we consider two
different cases: var.ui / D 22 and var.ui / D 32 .
The results are in Table 4 and show, as expected, that model CM2 is closer to the
true model than model CM1 in sense of the likelihood and the AIC and BIC criteria,
although its superiority is not particularly compelling in the scenarios studied.

6 Discussion

This is our first foray into joint mean-covariance modelling, where the mean is
subject to inequality constraints. The asymptotic theory for constrained estimators
is much more complicated than the standard theory. Accordingly, we generalized
the work of Xu and Wang (2008a,b) on fixed effect estimation with inequality
constraints by adding a data-driven marginal covariance model via the modified
Cholesky decomposition and applied many of the covariance modelling techniques
developed in Pan and MacKenzie (2003) to analyse the diabetic data. For some
covariance structures, the generalized autoregressive coefficients, jk , vary with j
and k simultaneously and we extended the usual model to capture this variation,
Joint Constrained Mean and Covariance Modelling 105

with some limited success. In all of the analyses undertaken, the covariance models
performed well and the technique is clearly feasible even when dealing with small
sample sizes. Its real strength lies, perhaps, in modelling non-stationary features
where variances increase over time, and measurements equidistant in time are not
equi-correlated, as has been demonstrated in the analysis of Kenward (1987)’s
cattle data (Pourahmadi 2000; Pan and MacKenzie 2006). In the constrained mean
context we point to the smaller standard errors obtained and to the reduction in
bias as obvious advantages of the method. We have clearly demonstrated that mis-
specification of the covariance structure can impact adversely on the estimation of
the mean component (MacKenzie 2006).
This first paper has only addressed the constrained mean marginal model and
naturally it has not been possible to tackle other covariance modelling issues which
will form the subject of future work.

Acknowledgements This paper is one of several emanating from the BIO-SI project (07/MI/012)
supported by Science Foundation Ireland (SFI), see www3.ul.ie/bio-si. Dr. Xu is a SFI postdoctoral
research fellow and Professor MacKenzie is the principal investigator. The authors gratefully
acknowledge Professor Jianxin Pan, Manchester University, UK, for helpful discussions and thank
Professor Martin Crowder, Imperial College, London,UK, for supplying the interesting data set.

Appendix

Computation of Vn

Since the log-likelihood function L.Y1 ;    ; Yn I / is regular with respect to its first
and second derivatives, i.e.,

E@L=@ D 0 and  E@2 L=@@ 0 D E@L=@@L=@ 0 ;

the covariance matrix Vn can be obtained by computing the expectation of the

second order derivatives of L.Y1 ;    ; Yn I / at the true value  D 0 .
It is easy to verify that

ˇ X
1
ˇ 0ˇ 1
n
v11
n D n E0 @ L=@ˇˇ ˇ
2
Dn Xi0 †1
i .
0 ; 0 /Xi ;
 D0 i D1

where E0 denotes the expectation operator at the point  D 0 .

Setting ri Œj 1 D .ri1 ;    ; ri.j 1//0 and ZiŒj 1 D .zij1 ;    ; zij.j1/ /0 for j D
1;    ; mi , then we have
106 J. Xu and G. MacKenzie

ˇ
1
ˇ 0ˇ
v22
n D n E0 @ L=@

ˇ 2
 D0

X ˇ
n
ˇ
Zi Di1 Zi ˇˇ
0
D n1 E0
i D1  D0
ˇ
X
n X
mi
ˇ
ij2 zij zij ˇˇ
1 0
D n E0
i D1 j D2  D0

ˇ
X
n X
mi
ˇ
ij2 ZiŒj 1 ri0Œj 1 ri Œj 1 ZiŒj 1 ˇˇ
0
D n1 E0
i D1 j D2  D0

X
n X
mi
0
D n1 exp.h0ij 0 /ZiŒj 1 †i Œj 1 .
0 ; 0 /ZiŒj 1 ;
i D1 j D2

where †i Œj 1 D E0 ."0i Œj 1 "i Œj 1 / with "i Œj 1 D ."i1 ;    ; "i.j 1/ /0 .
Noticing that eij D .rij  rOij /2 D Tij0 ri ri0 Tij with Tij0 is the j -th row of Ti and
defining Ri by Ri D diagfei1 ;    ; eimi g, then we get
ˇ
1
ˇ 0ˇ
v33
n D n E0 @ L=@ ˇ
2
 D0

X ˇ
n
ˇ
D .2n/1 E0 Hi0 Di1 Ri Hi ˇˇ
i D1  D0

X
n
D .2n/1 Hi0 Di1 . 0 /d i agfTi10 †i .
0 ; 0 /Ti1 ;    ; Tim
0
i
†i .
0 ; 0 /Timi gHi :
i D1

After some matrix algebra, we can see that the factors containing random error in
the elements of the matrices @2 L=@ˇ@
0 , @2 L=@ˇ@ 0 and @2 L=@
@ 0 are the residual
vectors ri ; i D 1;    ; n, the expectations of which are zero at the true value  D 0 .
Thus

n D 0 for k ¤ l
vkl with k; l D 1; 2; 3:

References

Crowder, M. J., & Hand, D. J. (1990). Analysis of repeated measures. London: Chapman and Hall.
Cysneiros, F. J. A., & Paula, G. A. (2004). One sided tests in linear models with multivariate
t-distribution. Communications in Statistics: Simulation and Computation, 33, 747–771.
Diggle, P. J., Heagerty, P., Liang, K. Y., & Zeger, S. L. (2002). Analysis of longitudinal data (2nd
ed.). London: Oxford.
Joint Constrained Mean and Covariance Modelling 107

Dobson, A. J. (2002). An introduction to generalized linear models (2nd ed.). London: Chapman
and Hall.
Fang, H., Tian, G., Xiong, X., & Tan, M. (2006). A multivariate random-effects model with
restricted parameters: Application to assessing radiation therapy for brain tumours. Statistics
in Medicine, 25, 1948–1959.
Fletcher, R. (1971). A general quadratic programming algorithm. Journal of the Institute of
Mathematics and its Applications, 7, 76–91.
Goldfarb, D., & Ininani, A. (1983). A numerical stable dual method for solving strictly convex
quadratic programs. Mathematical Programming, 27, 1–33.
Kenward, M. G. (1987) A method for comparing profiles of repeated measurements. Applied
Statistics, 36, 296–308.
Liang, K. Y., & Zeger, G. (1986). Longitudinal data analysis using generalized linear models.
Biometrika, 73, 13–22.
Liew, C. K. (1976). Inequality constrained least squares estimations. Journal of the American
Statistical Association, 71, 13–22.
MacKenzie, G. (2006). In discussion of double hierarchical generalized linear models. Journal of
the Royal Statistical Society. Series: C, 55, 139–185.
Pan, J., & MacKenzie, G. (2003). On modelling mean-covariance structure in longitudinal studies.
Biometrika, 90, 239–244.
Pan, J., & MacKenzie, G. (2006). Regression models for covariance structures in longitudinal
studies. Statistical Modelling, 7, 49–71.
Pan, J., & MacKenzie, G. (2007). Modelling conditional covariance in the linear mixed model.
Statistical Modelling, 7, 49–71.
Park, T., Shin, D. W., & Park, C. G. (1998). A generalized estimating equations approach for testing
ordered group effects with repeated measurements. Biometrics, 54, 1645–1653.
Pilla, R. S., Qu, A., & Loader, C. (2006). Testing for order-restricted hypotheses in longitudinal
data. Journal of the Royal Statistical Society. Series: B, 68, 437–455.
Pourahmadi, M. (1999). Joint mean-covariance models with applications to longitudinal data:
Unconstrained parameterisation. Biometrika, 86, 677–690.
Pourahmadi, M. (2000). Maximum likelihood estimation of generalised linear models for multi-
variate normal covariance matrix. Biometrika, 87, 425–435.
Shi, N. Z., Zheng, S. R., & Guo, J. (2005). The restricted em algorithm under inequality restrictions
on the parameters. Journal of Multivariate Analysis, 92, 53–76.
Shin, D. W., Park, C. G., & Park, T. (1996). Testing for ordered group effects with repeated
measurements. Biometrika, 83, 688–694.
Tan, M., Fang, H., Tian, G., & Houghton, P. J. (2005). Repeated-measures models with constrained
parameters for incomplete data in tumour xenograft experients. Statistics in Medicine, 24, 109–
119.
Wang, Y. G., & Carey, V. (2003). Working correlation structure misspecification, estimation and
covariate design: Implications for generalized estimating equations performance. Biometrika,
90, 29–41.
Xu, J., & Wang, J. (2008b). Maximum likelihood estimation of linear models for longitudinal data
with inequality constraints. Communication in Statistics-Theory and Methods, 37, 931–946.
Xu, J., & Wang, J. (2008a). Two-stage estimation of inequality-constrained marginal linear models
with longitudinal data. Journal of Statistical Planning and Inference, 138, 1905–1918.
 Part III
Statistical Model Development
Hierarchical Generalized Nonlinear Models

Roger W. Payne

Abstract Hierarchical generalized linear models allow non-Normal data to be

modelled in situations when there are several sources of error variation. They extend
the familiar generalized linear models to include additional random terms in the
linear predictor. However, they do not constrain these terms to follow a Normal
distribution nor to have an identity link, as is the case in the more usual generalized
linear mixed model. They thus provide a much richer set of models, that may seem
more intuitively appealing. Another extension to generalized linear models allows
nonlinear parameters to be included in the linear predictor. The fitting algorithm for
these generalized nonlinear models operates by performing a nested optimization,
in which a generalized linear model is fitted for each evaluation in an optimization
over the nonlinear parameters. The optimization search thus operates only over
the (usually relatively few) nonlinear parameters, and this should be much more
efficient than a global optimization over the whole parameter space. This paper
reviews the generalized nonlinear model algorithm, and explains how similar
principles can be used to include nonlinear fixed parameters in the mean model
of a hierarchical generalized linear model, thus defining a hierarchical generalized
nonlinear model.

Keywords Hierarchical generalized linear models • Hierarchical generalized non-

linear models

R.W. Payne ()

Department of Computational and Systems Biology, Rothamsted Research, Harpenden, Herts
AL5 2JQ, UK
VSN International, 5 The Waterhouse, Waterhouse Street, Hemel Hempstead, Herts HP1
1ES, UK
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 111

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__9,
© Springer International Publishing Switzerland 2014
112 R.W. Payne

1 Introduction

The methodology of hierarchical generalized linear models (HGLMs) was devised

by Lee and Nelder (1996) to provide a flexible and efficient framework for
modelling non-Normal data in situations when there are several sources of error
variation. They extend the familiar generalized linear models (GLMs) to include
additional random terms in the linear predictor.
In an ordinary regression the model to be fitted is

y D  C ";

where  is the mean predicted by a model

 D Xˇ;

(e.g. aCbx), and " is the residual with Normal distribution N.0;  2 /. Equivalently,
we can write that y has Normal distribution N.;  2 /.
In a GLM the expected value of y is still

E.y/ D ;

but model now defines the linear predictor

 D Xˇ

with  and  related by the link function g

 D g./

and y has a distribution with mean  from the exponential family; see McCullagh
and Nelder (1989).
In a HGLM, the expected value E.y/ is still , and this is still related to the
linear predictor by a link function g./. The vector y still has a distribution from
the exponential family, but this is currently limited to binomial, gamma, Normal or
Poisson. The linear predictor now contains additional random terms
X
 D Xˇ C Zi vi
i

which have their own link functions

vi D v.ui /

and the vectors of random effects ui have beta, Normal, gamma or inverse gamma
distributions (these being the distributions that are conjugate to the distributions
available for y). For details see Lee et al. (2006).
HGNLMs 113

The analysis involves fitting an augmented GLM, known as the augmented mean
model, to describe the mean vector . This has units corresponding to the original
data units, together with an additional unit for each effect of a random term. The
augmented mean model is fitted as a GLM, but there may be different link functions
and distributions operating on the additional units from those on the original units.
The link function is the function v./, while the distribution is the one to which the
distribution of ui is conjugate; see Chap. 6 of Lee et al. (2006) for details. The data
values for the extra units contain the inverse-link transformations of the expected
values of the random distributions. The HGLM algorithm also involves further
GLMs, with gamma distributions and usually with logarithmic links, to model the
dispersion for each random term (including the residual dispersion parameter ).
The models are connected, in that the y-variates for the dispersion models are
deviance contributions from the augmented mean model divided by one minus their
leverages, while the reciprocals of the fitted values from the dispersion models act
as weights for the augmented mean model. So the models are fitted alternately until
convergence, as shown in Table 7.3 of Lee et al. (2006).
The methodology has been implemented in GenStat, as a suite of procedures (i.e.
sub-programs written in the GenStat language); see Payne et al. (2006b). From its
9th Edition GenStat also includes data files and programs to run many of the worked
examples from Lee et al. (2006).

1.1 Generalized Nonlinear Models

Another extension to GLMs allows for the inclusion of nonlinear parameters in the
linear predictor; see Lane (1996), or Sect. 3.5.8 of Payne et al. (2006a). The linear
predictor is still
X
 D Xˇ D xi ˇi ;

but some of the xi ’s may now be nonlinear functions of other explanatory variables
and parameters that must be estimated separately. For example, an exponential curve
can be written as

a C b  x;

where x D  r for parameter  and explanatory vector r. The methodology also

allows for nonlinear parameters in the link function (for example to model natural
mortality or immunity in probit analysis) or in the offset variate (for example for
survival models).
The nonlinear parameters are fitted by standard optimization methods such
as Gauss–Newton or Newton–Raphson. So, deviances are calculated on grids
of trial values of nonlinear parameters, as the algorithm finds its way to the
114 R.W. Payne

optimum. Each point on the grid involves fitting a GLM, but this is a relatively
quick and straightforward operation. So the process is much more efficient than a
global optimization over the whole parameter space of both nonlinear and linear
parameters. Furthermore, because the model essentially remains a GLM, all the
familiar inference techniques can be used, for example to assess the fit, or to
determine which parameters are needed in the model.

1.2 Hierarchical Generalized Nonlinear Models

The same idea can be used to define a hierarchical generalized nonlinear model
(HGNLM). The linear predictor is
X
 D Xˇ C Zi vi
i

as before, but some columns of X may be derived as nonlinear functions of other

explanatory variables and parameters. Again the nonlinear parameters are fitted by
standard optimization methods, but now it is the augmented mean model that is
fitted at each point of the grids of trial nonlinear parameter values.

2 Example

As an example we use the Cake data from Cochran and Cox (1957), also analysed
in Lee et al. (2006). The data concern an experiment carried out at Iowa State
College in which three recipes were used to prepare the cake mixtures. Fifteen
replicates were performed. In each of these a batch of mixture was produced using
each recipe, and then subdivided into enough for six different cakes which were each
baked at a different temperature: 175; 185; 190; 205; 215 or 225 ıC. The analysis
variate is the angle at which the cakes break. The experiment thus has the structure
of a split-plot design with batch as the whole-plot factor, corresponding to recipe
as the whole-plot treatment factor, and temperature as the sub-plot treatment factor.
Cochran and Cox (1957) analyse the data in this way, using a standard linear mixed
model, but Lee et al. (2006) suggest that the breaking angles are non-Normal and
fit a gamma generalized linear mixed model. Currently HGNLMs can be fitted only
with conjugate HGLMs, i.e., those in which the distribution of the random effects
is the conjugate of the distribution of y. So instead, below, we fit a gamma/inverse-
gamma HGLM (which may, in any case, be an intuitively more appealing model).
The first command, SPLOAD in line 2, loads the data from a GenStat spread-
sheet. HGRANDOMMODEL defines the random terms using the model for-
mula Replicate/Batch, giving random terms Replicate (replicates) and Replicate.
HGNLMs 115

Batch (batches within replicates). HGFIXEDMODEL defines the fixed terms using
the model formula Recipe*Temperature, giving terms Recipe and Temperature
representing the main effects of recipe and temperature, and Recipe. Temperature
representing their interaction (this is an interaction as both main effects occur earlier
in the model). HGANALYSE then fits the model.

Output from GenStat

1 " Cake data (see Lee, Nelder & Pawitan (2006) Sections 5.5 and 6.4.1) "
2 SPLOAD ’Cake.gsh’

Loading Spreadsheet File

Catalogue of file Cake.gsh

Data imported from GenStat Server on: 21-Jun-2005 12:27:25

Sheet Type: vector

Index Type Nval Name

1 factor 270 Replicate
2 factor 270 Batch
3 factor 270 Recipe
5 factor 270 Temperature
6 variate 270 Angle
Note: Missing indices are used by unnamed or system structures. These store
ancillary information, for example factor labels.
3 " fit a gamma inverse-gamma HGLM "
4 HGRANDOMMODEL [DISTRIBUTION=inversegamma; LINK=log] Replicate/Batch
5 HGFIXEDMODEL [DISTRIBUTION=gamma; LINK=log] Recipe*Temperature
6 HGANALYSE [PRINT=model,fixed,dispersion,likelihood] Angle

Hierarchical Generalized Linear Model

Response variate: Angle

Mean Model

Fixed terms: Recipe

Temperature
Distribution: gamma
116 R.W. Payne

Link: logarithm
Random terms: Replicate/Batch
Distribution: inversegamma
Link: logarithm
Dispersion: free

Dispersion Model

Distribution: gamma
Link: logarithm

Estimates from the Mean Model

Covariate estimate s.e. t(252)

constant 3.3737 0.05884 57.341
Recipe 2 0.0783 0.05617 1.395
Recipe 3 0.0541 0.05617 0.963
Temperature 185 0.0933 0.05053 1.846
Temperature 195 0.0614 0.05053 1.216
Temperature 205 0.1491 0.05053 2.951
Temperature 215 0.2874 0.05053 5.688
Temperature 225 0.1928 0.05053 3.816
Recipe 2 .Temperature 185 0.0094 0.07146 0.131
Recipe 2 .Temperature 195 0.0978 0.07146 1.368
Recipe 2 .Temperature 205 0.0286 0.07146 0.400
Recipe 2 .Temperature 215 0.0495 0.07146 0.693
Recipe 2 .Temperature 225 0.0818 0.07146 1.145
Recipe 3 .Temperature 185 0.0512 0.07146 0.716
Recipe 3 .Temperature 195 0.0745 0.07146 1.043
Recipe 3 .Temperature 205 0.0388 0.07146 0.543
Recipe 3 .Temperature 215 0.0648 0.07146 0.907
Recipe 3 .Temperature 225 0.0576 0.07146 0.806
HGNLMs 117

Estimates from the Dispersion Model

Estimates of Parameters

antilog of
Parameter estimate s.e. t(
) estimate
phi 3.9584 0.0947 41.79 0.01909
lambda Replicate 3.541 0.394 8.99 0.02900
lambda Replicate.Batch 5.401 0.342 15.80 0.004514
Message: s.e.s are based on dispersion parameter with value 1.

Likelihood Statistics
2 h.yjv/ 1,507.637
2 h 1,351.700
2 Pv .h/ 1,614.954
2 Pˇ;v .h/ 1,696.710
2 EQD.yjv/ 1,506.777
2 EQD 1,350.735
2 Pv .EQD/ 1,613.989
2 Pˇ;v .EQD/ 1,695.744

Initially temperature has been fitted as a factor, with no account taken of the actual
values represented by its levels. If we form predictions for the temperatures and
plot these against the temperature values, the relationship appears to be fairly linear,
as can be seen in Fig. 1. So the next analysis instead fits regression coefficients
of temperature (including interactions with recipe, i.e., to investigate whether the
regression coefficient differs according to the recipe).

Output from GenStat

7 " form and plot predictions at the various temperatures "

8 HGPREDICT [PRINT=prediction,se; PREDICTION=Mean] Temperature
118 R.W. Payne

Fig. 1 Predicted breaking angle by temperature ı C

Predictions from Regression Model

The standard errors are appropriate for interpretation of the predictions as sum-
maries of the data rather than as forecasts of new observations.
Response variate: Angle
Mean se
Temperature
175 3.330 0.04909
185 3.403 0.04909
195 3.448 0.04909
205 3.475 0.04909
215 3.579 0.04909
225 3.569 0.04909
Message: s.e’s, variances and lsd’s are approximate, since the model is not linear.
Message: s.e’s are based on the residual deviance.
9 PEN 1; SYMBOL=2; CSYMBOL=1; CFILL=1
10 FRAME 1,2; BOX=include; BOXKEY=bounded;\
11 XLOWER=0,0.5; XUPPER=1,0.95; YLOWER=0,0.1; YUPPER=1,0.2
12 XAXIS 1; TITLE=’Temperature~^{o}C’
13 YAXIS 1; TITLE=’Predicted breaking angle’
14 GETATTRIBUTE [ATTRIBUTE=levels] Temperature; SAVE=TempAtt
15 VARIATE DegreesC; VALUES=TempAtt[’levels’]
16 DGRAPH Mean; DegreesC
HGNLMs 119

17 " fit a linear relationship with temperature "

18 CALCULATE TempC = Temperature
19 HGRANDOMMODEL [DISTRIBUTION=inversegamma; LINK=log] Replicate/Batch
20 HGFIXEDMODEL [DISTRIBUTION=gamma; LINK=log] Recipe*TempC
21 HGANALYSE [PRINT=model,fixed,dispersion,likelihood] Angle

Hierarchical Generalized Linear Model

Response variate: Angle

Mean Model

Fixed terms: Recipe

TempC
Distribution: gamma
Link: logarithm
Random terms:
Replicate/Batch
Distribution: inversegamma
Link: logarithm
Dispersion: free

Dispersion Model

Distribution: gamma
Link: logarithm

Estimates from the Mean Model

estimate s.e. t(264)

constant 2.5621 0.1788 14.329
Recipe 2 0.1697 0.2453 0.692
Recipe 3 0.1263 0.2453 0.515
TempC 0.0047 0.0009 5.486
TempC.Recipe 2 0.0006 0.0012 0.474
TempC.Recipe 3 0.0003 0.0012 0.278
120 R.W. Payne

Estimates from the Dispersion Model

Estimates of Parameters

antilog of
Parameter estimate s.e. t(
) estimate
phi 3.9451 0.0922 42.78 0.01935
lambda Replicate 3.540 0.394 8.99 0.02902
lambda Replicate.Batch 5.410 0.344 15.75 0.004473
Message: s.e.s are based on dispersion parameter with value 1.

Likelihood Statistics

2 h.yjv/ 1,523.414
2 h 1,366.924
2 Pv .h/ 1,629.995
2 Pˇ;v .h/ 1,681.544
2 EQD.yjv/ 1,522.544
2 EQD 1,365.947
2 Pv .EQD/ 1,629.019
2 Pˇ;v .EQD/ 1,680.567
22 HGPREDICT [PRINT=*; PREDICTION=LinMean] TempC; LEVELS=DegreesC

The nonlinearity with temperature can be assessed using the change in the likelihood
statistic 2  Pv .h/; see Sect. 6.5 of Lee et al. (2006). The difference in the
statistic between the two models is 1629:995  1614:954 D 15:041 corresponding
to a change of 12 in the number of fixed parameters. So there is scant evidence
of nonlinearity. However, it may still be worth trying a nonlinear relationship,
for interest and to illustrate the HGNLM methodology. So in the next section of
the example, we use the HGNONLINEAR procedure (added in the 10th Edition
of GenStat for Windows) to define a Box-Cox transformation of temperature:
BoxCoxTemp is the derived column to include in the design matrix X , and
BoxCox is the nonlinear parameter. Notice that, to avoid potential overflows in
the calculation, the temperatures are first divided by 100. We then refit the model
with BoxCoxTemp fixed (taking the estimated value of BoxCox) so that we can
form predictions to compare in a plot with the predictions made originally for each
individual temperature and those made assuming a linear model. (GenStat does not
currently allow predictions to be formed from nonlinear GLMs.)
HGNLMs 121

Output from GenStat

23 " try a Box-Cox transformation of temperature

-24 (scaled to avoid overflow) "
25 CALCULATE xtemp = TempC / 100
26 EXPRESSION [VALUE=BoxCoxTemp = (BoxCox==0) * LOG(xtemp) +\
27 (BoxCox/=0) * (xtemp**BoxCox-1)/(BoxCox+(BoxCox==0))]\
28 BoxCoxCalc
29 SCALAR BoxCox; VALUE=-1
30 CALCULATE #BoxCoxCalc
31 HGFIXEDMODEL [DISTRIBUTION=gamma; LINK=log] Recipe*BoxCoxTemp
32 HGNONLINEAR [CALCULATION=BoxCoxCalc; VECTORS=BoxCoxTemp,xtemp]\
33 BoxCox; INITIAL=-1; STEP=0.01
34 HGANALYSE [PRINT=model,fixed,dispersion,likelihood] Angle

Hierarchical Generalized Linear Model

Response variate: Angle

Mean Model

Fixed terms: Recipe

BoxCoxTemp
Distribution: gamma
Link: logarithm
Random terms:
Replicate/Batch
Distribution: inversegamma
Link: logarithm
Dispersion: free

Dispersion Model

Distribution: gamma
Link: logarithm
122 R.W. Payne

Estimates from the Mean Model

estimate s.e. t(263)

BoxCox 1.523 2.699 0.5643
constant 2.373 1.208 1.9641
Recipe 2 0.198 0.330 0.6013
Recipe 3 0.128 0.301 0.4248
BoxCoxTemp 2.665 4.884 0.5457
BoxCoxTemp.Recipe 2 0.338 0.924 0.3664
BoxCoxTemp.Recipe 3 0.163 0.746 0.2185

Estimates from the Dispersion Model

Estimates of Parameters

antilog of
Parameter estimate s.e. t(
) estimate
phi 3.9446 0.0924 42.68 0.01936
lambda Replicate 3.540 0.394 8.99 0.02902
lambda Replicate.Batch 5.412 0.344 15.75 0.004462
Message: s.e.s are based on dispersion parameter with value 1.

Likelihood Statistics

2 h.yjv/ 1,522.585
2 h 1,365.968
2 Pv .h/ 1,629.083
2 Pˇ;v .h/ 1,642.647
2 EQD.yjv/ 1,521.714
2 EQD 1,364.991
2 Pv .EQD/ 1,628.106
2 Pˇ;v .EQD/ 1,641.670
35 " reanalyse using the Box-Cox transformed temperatures
-36 (i.e. taking the Box-Cox parameter as fixed) to allow
-37 so that predictions can be calculated "
38 CALCULATE #BoxCoxCalc
39 HGRANDOMMODEL [DISTRIBUTION=inversegamma; LINK=log] Replicate/Batch
40 HGFIXEDMODEL [DISTRIBUTION=gamma; LINK=log] Recipe*BoxCoxTemp
HGNLMs 123

Fig. 2 Predicted breaking

angle by temperature ı C

41 HGANALYSE [PRINT=*] Angle

42 " make predictions "
43 CALCULATE BoxCoxLevs = LOG(DegreesC/100)*(BoxCox==0) \
44 + ((DegreesC/100)**BoxCox-1)/(BoxCox+(BoxCox==0))*(BoxCox/=0)
45 HGPREDICT [PRINT=*; PREDICTION=BCMean] BoxCoxTemp; LEVELS=BoxCoxLevs
46 " plot predictions "
47 PEN 2,3; SYMBOL=0; CLINE=1; METHOD=monotonic; LINESTYLE=1,2
48 XAXIS 1; TITLE=’Temperature~^{o}C’
49 DGRAPH Mean,LinMean,BCMean; DegreesC; PEN=1,2,3

The plot in Fig. 2 shows some nonlinearity in the Box-Cox model, but this is
only slight, and that is borne out by the fact that the likelihood statistic 2  Pv .h/
is virtually unchanged, at 1,629.083. Nevertheless, the example does illustrate the
potential usefulness of being able to assess nonlinearity of fixed terms in an HGLM.

3 Conclusion

HGNLMs provide a further enhancement to the HGLM methodology in the

10th Edition of GenStat for Windows. Additional information can be found at
http://genstat.co.uk/.
124 R.W. Payne

References

Cochran, W., & Cox, G. (1957). Experimental designs (2nd ed.). New York: Wiley.
Lane, P. (1996). Generalized nonlinear models. In COMPSTAT 1996 Proceedings in Computa-
tional Statistics (pp. 331–336).
Lee, Y., & Nelder, J. (1996). Hierarchical generalized linear models (with discussion). Journal
of the Royal Statistical Society Series B, 58, 619–678.
Lee, Y., Nelder, J., & Pawitan, Y. (2006). Generalized linear models with random effects: Unified
analysis via H-likelihood. Boca Raton: CRC Press.
McCullagh, P., & Nelder, J. (1989). Generalized linear models (2nd ed.). London: Chapman and
Hall.
Payne, R., Harding, S., Murray, D., Soutar, D., Baird, D., Welham, S., et al. (2006a) The guide to
GenStat release 9, part 2 statistics. Hemel Hempstead, UK: VSN International.
Payne, R., Lee, Y., Nelder, J., & Noh, M. (2006b). GenStat release 10 reference manual, part
3, procedure library PL18 (pp. 372–389). Hemel Hempstead, UK: VSN International [proce-
dures HGANALYSE, HGDISPLAY, HGDRANDOMMODEL, HGFIXEDMODEL, HGKEEP,
HGNONLINEAR, HGPLOT, HGPREDICT and HGRANDOMMODEL].
Comparing Robust Regression Estimators
to Detect Data Clusters: A Case Study

Alessandra Durio and Ennio Isaia

Abstract It is well known that in all situations involving the study of large data
sets where a substantial number of outliers or clustered data are present, regression
models based on M -estimators are likely to be unstable. Resorting to the inherent
properties of robustness of the estimates based on the Integrated Square Error
criterion we compare the results arising from L2 estimates with those obtained from
some common M -estimators. The discrepancy between the estimated regression
models is measured by means of a new concept of similarity between functions and
a system of statistical hypothesis. A Monte Carlo Significance test, is introduced to
test the similarity of the estimates. Whenever the hypothesis of similarity between
models is rejected, a careful investigation of the data structure is necessary to check
for the presence of clusters, which can lead to the consideration of a mixture of
regression models. Concerning this, we shall see how L2 criterion can be applied in
fitting a finite mixture of regression models. The requisite theory is outlined and the
whole procedure is applied to a case study concerning the evaluation of the risk of
fire and the risk of electric shocks of electronic transformers.

Keywords Minimum integrated square error • Mixture of regression models •

Robust regression • Similarity between functions

1 Introduction

Regression is one of the widespread tools used to establish the relationship

between a set of predictor variables and a response variable. However, in many
circumstances, careful data preparation may not be possible and hence data may

A. Durio ()
Department of Economics & Statistics “S. Cognetti de Martiis”, University of Turin, Lungo Dora
Siena 100/A, 10124 Turin, Italy
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 125

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__10,
© Springer International Publishing Switzerland 2014
126 A. Durio and E. Isaia

be heavily contaminated by a substantial number of outliers. In these situations,

the estimates of the parameters of the regression model obtained by the Maximum
Likelihood criterion are fairly unstable.
The development of robust methods is underlined by the appearance of a wide
number of papers and books on the topic including: Huber (1981), Rousseeuw and
Leroy (1987), Staudte and Sheather (1990), Davies (1993), Dodge and Jurečkova
(2000), Seber and Lee (2003), Rousseeuw et al. (2004), Jurečkova and Picek (2006),
Maronna et al. (2006) and Fujisawa and Eguchi (2006).
The approach based on minimizing the Integrated Square Error is particularly
helpful in those situations where, due to large sample size, careful data preparation
is not feasible and hence data may contain a substantial number of outliers (Scott
2001). In this sense the L2 E criterion can be viewed as an efficient diagnostic tool
in building useful models.
In this paper we suggest a procedure of regression analysis whose first step
consists in comparing the results arising from L2 estimates with those obtained from
some common M -estimators. Afterwards, if a particular test of hypothesis leads us
to reject the conjecture of similarity between the estimated regression models, we
investigate the data for the presence of clusters by analyzing the L2 minimizing
function. The third step of the procedure consists in fitting a mixture of regression
models via the L2 criterion.
Below, we introduce the Integrated Square Error minimizing criterion for regres-
sion models, define a new concept of similarity between functions and introduce
a Monte Carlo Significance (M.C.S.) test. We also illustrate the whole procedure
by means of some simulated examples involving simple linear regression models.
Finally, we present an analysis of a case study concerning the evaluation of the risk
of fire and the risk of electric shocks in electronic transformers.

2 Parametric Linear Regression Models and Robust

Estimators

Let f.xi1 ; : : : ; xip ; yi /gi D1;:::;n be the observed data set, where each observa-
tion stems from a random sample drawn from the p C 1 random variable
.X1 ; : : : ; Xp ; Y /. The regression model for the observed data set being studied
is yi D mˇ .xi / C "i , with i D 1; : : : ; n, where the object of our interest is the
regression mean

X
p
mˇ .xi / D EŒY jxi  D ˇ0 C ˇj xij (1)
j D1

and the errors f"i gi D1;:::;n are assumed to be independent random variables with zero
mean and unknown finite variances.
Comparing Robust Regression Estimators 127

2.1 Huber M -Estimator

The presence of outliers is a problem for regression techniques; these may occur for
many reasons. An extreme situation arises when the outliers are numerous and they
arise as a consequence of clustered data. For example, a large proportion of outliers
may be found, if there is an omitted unknown categorical variable (e.g. gender,
species, geographical location, etc.) where the data behave differently in each
category. In parametric estimation, the estimators with good robustness proprieties
relative to maximum likelihood are the M -estimators. The class of M -estimators of
the vector ˇ is defined as (e.g., Hampel et al. 2005)

X
n

Ǒ M D arg min  yi  mˇ .xi / ; (2)
ˇ i D1

where  W R ! R is absolutely continuous convex function with derivative .

If we assume that the r.v.s "i are independent and identically distributed as the r.v.
"  N .0; /, the least-squares estimator gives the Maximum Likelihood Estimate
(MLE) of the vector ˇ, i.e.:

X
n
 2
Ǒ MLE D arg min yi  mˇ .xi / :
ˇ i D1

Since in the presence of outliers MLEs are quite unstable, i.e., inefficient and
biased, for our purpose in the class of M -estimators we shall resort to the robust
Huber M-estimator (HME) for which
8
ˆ 1
< .yi  mˇ .xi //2 if jyi  mˇ .xi /j  k;
.yi  mˇ .xi // D 2 k
:̂k jyi  mˇ .xi /j .1  / if jyi  mˇ .xi /j > k;
2

where the tuning constant k is generally set to 1:345 .

2.2 L2 -Based Estimator

We investigate estimation methods in parametric linear regression models based on

the minimum Integrated Square Error and the minimum L2 metric. In the ˛-family
of estimators proposed by Basu et al. (1998), L2 estimator, briefly L2 E, is the more
robust to outliers, even if it is less efficient than MLE.
Given the r.v. X , with unknown density f .xj 0 /, for which we introduce the
model f .xj/, the estimate for  0 minimizing the L2 metric will be:
128 A. Durio and E. Isaia

Z
O L2 E D arg min Œf .xj/  f .xj 0 /2 dx D
 R
Z 
D arg min f 2 .xj/ dx  2 E Œf .xj 0 / D (3)
 R
"Z #
2 X
n
D arg min f .xj/ dx 
2
f .xi j/ ;
 R n i D1

where, the so-called expectedPheight of the density, E Œf .xj 0 / is replaced with its
estimate EO Œf .xj 0 / D n1 niD1 f .xi j/ and where (Basu et al. 1998),
Z Z
1X
n
f .xj/ dx D
2
f 2 .xi j/ dxi : (4)
R n i D1 R

We turn now our attention to illustrate how the estimates based on L2 criterion can
be applied to parametric regression models. Assuming that the random variables
Y jx are distributed as a N .mˇ0 .x/; 0 /, i.e. fY jx .yjˇ 0 ; 0 / D .yjmˇ0 .x/; 0 /, the
L2 estimates of the parameters in ˇ 0 and 0 are given by Eq. (3), which in this case
becomes
"Z #
2 X
n
Ǒ
. ; O /L2 E D arg min  .yjmˇ .x/; / dy 
2
.yi jmˇ .xi /; /
ˇ; R n i D1
" # (5)
2X
n
1
D arg min p  .yi jmˇ .xi /; / ;
ˇ; 2  n i D1

since from Eq. (4)

Z Z
1X
n
1
 2 .yjmˇ .x/; / dy D  2 .yi jmˇ .xi /; / dyi D p :
R n i D1 R 2 

Clearly Eq. (5) is a feasible computationally closed-form expression so that L2

criteria can be performed by any standard non-linear optimization procedure, for
example, the nlm routine in the R library. However, it is important to recall that,
whatever the algorithm, convergence to the global optimum can depend strongly on
the starting values.

3 The Similarity Index and the M.C.S Test

To compare the L2 E performance with respect to some other common estimators

we resort to an index of similarity between regression models introduced in Durio
and Isaia (2010). In order to measure the discrepancy between the two estimated
regression models, the index of similarity takes into account the space region
Comparing Robust Regression Estimators 129

Fig. 1 Data points and two estimated regression models m O T 0 .x/ and m
O T 1 .x/. In panel (b) the
domains D pC1 and C pC1 upon which the si m.T0 ; T1 / statistic is computed

between m O T0 .x/ and m

O T1 .x/ with respect to the space region where the whole of
the data points lie. Let T0 and T1 be two regression estimators and Ǒ T0 , Ǒ T1 the
corresponding vectors of the estimated parameters. Introducing the sets:
 
I p D min.xi1 /I max.xi1 /  : : :  Œmin.xip /I max.xip / ;
I D Œmin.yi /I max.yi / D ŒaI b ;

we define the similarity index as

R
def pC1 d t
sim.T0 ; T1 / D RD
C pC1 d t

C pC1 D I p  I (6)
˚ 
D pC1 D .x; y/ 2 RpC1 W .x/  y  .x/; x 2 I p \ C pC1

with .x/ D min .m O T0 .x/; mO T1 .x// and .x/ D max .m O T0 .x/; m

O T1 .x//.
Figure 1 shows how the similarity index given by Eq. (6) can be computed in
the simple case where p D 1. In panel (a) we have the cloud of data points and
the Ǒ Ǒ
R two estimated models T0Rand T1 . The shaded area of panel (b) corresponds to
D pC1 d t, while the integral C pC1 d t is given by the area of the dotted rectangle,
in which data points lay.
In order to compute the integrals of Eq. (6), we employ the fast and accurate
algorithm proposed by Durio and Isaia (2010).
If the vectors Ǒ T0 and Ǒ T1 are close to each other, then sim.T0 ; T1 / will be close
to zero. On the other hand, if the estimated regression models m O T0 .x/ and m
O T1 .x/ are
130 A. Durio and E. Isaia

dissimilar we are likely to observe a value of sim.T0 ; T1 / far from zero. We therefore
propose to use the sim.T0 ; T1 / statistic to verify the following system of hypothesis
(
H0 W ˇ 0 D Ǒ T0
(7)
H1 W ˇ 0 ¤ Ǒ T0

Since it is not reasonable to look for an exact form of the si m.T0 ; T1 / distribution,
in order to check the above system of hypothesis we utilise a simplified M.C.S. test
originally suggested by Barnard (1963) and later proposed by Hope (1968).
Let simT0 T1 denote the value of the sim.T0 ; T1 / statistic computed on the observed
data. The simplified M.C.S. test consists of rejecting H0 if simT0 T1 is the m ˛-th
most extreme statistic relative to the corresponding quantities based on the random
samples of the reference set, where the reference set consists of m  1 random
samples, of size n each, generated under the null hypothesis, i.e., drawn at random
from the model m O T0 .x/ with  D O T0 . In other words we generate m  1 random
samples under H0 and for each of them we compute simT0 T1 and we shall reject
the null hypothesis, at the ˛ significance level, if and only if the value of the test
statistic simT0 T1 is greater than all the m  1 values of simT0 T1 . We remark that if we
set m ˛ D 1 and fix ˛ D 0:01, we have m  1 D 99 (while fixing ˛ D 0:05 would
yield m  1 D 19).

4 Simple Linear Regression and Examples

Since for our case study we shall consider the simple linear regression model
yi D ˇ0 C ˇ1 xi C "i , the L2 criterion according to Eq. (5) reduces to the following
computationally closed-form expression
" #
2X
n
1
. Ǒ ; O /L2 E D arg min p  .yi jˇ0 C ˇ1 xi ; / : (8)
ˇ; 2  n i D1

In the following we introduce two simulated examples in order to demonstrate

the behaviour of the L2 criterion in the presence of outliers and in the presence of
clustered data. To evaluate its performance, we shall use the Maximum Likelihood
estimator and the robust Huber M estimator. Given T1 D L2 E, we shall perform
the M.C.S. test two times: the first one, fixing T0 D MLE, for sim.MLE; L2 E/, the
second one fixing T0 D HME, for sim.HME; L2 E/. We remark that, as p D 1, in
both situations we have I p D Œmin.xi /I max.xi / and that clearly the integrals of
Eq. (6) are defined on bi-dimensional domains.
Example I. Let us consider a simulated dataset of n D 200 points generated
according to the model Y D X C ", where X  U.0; 10/ and "  N .0; 0:8/.
We then introduce m D 10.30/ points according to the model Y D 3 C X C ",
Comparing Robust Regression Estimators 131

Table 1 Results of simulated Example I

m D 10 m D 30
MLE HME L2 E MLE HME L2 E
ˇO0 0.3078 0.1616 0.0353 0.2884 0.2081 0.0139
ˇO1 0.9054 0.9509 0.9886 0.8635 0.8944 0.9975
O 0.9889 0.9972 0.7926 1.2352 1.2389 0.9712

10 Data points 10 Data points

MLE MLE
RME RME
L2E L2E
8 8

Response y
Response y

6 6

4 4

2
2

0
0

0 2 4 6 8 10 0 2 4 6 8 10
x − Panel (a) x − Panel (b)

O ML .x/, m
Fig. 2 Data points of Example I and estimated models m O HM .x/ and m
O L2 .x/. In panel
(a) we set m D 10 outliers while in panel (b) m D 30

where X  U.8; 10/ and "  N .0; 0:4/, so that they can be considered as outliers.
Resorting to the estimators ML, HM and L2 we obtain the following estimates of
the parameters ˇ0 , ˇ1 and  listed in Table 1 (also see Fig. 2).
Applying the M.C.S. test, with ˛ D 0:01, to the estimated models m O ML .x/
and m O L2 .x/, we reject the null hypothesis of system (7) as we have simML;L2 D
0:0203 > max.simML;L2 / D 0:0128. Turning our attention to models m O HM .x/ and
mO L2 .x/, the M.C.S. test leads us to accept the null hypothesis since simHM;L2 D
0:0091 < max.simHM;L2 / D 0:0123.
In the case we add m D 30 outliers to the sample data, the results of the M.C.S.
tests lead us to different conclusions. In both situations we reject the null hypothesis
of system (7) as we have

simML;L2 D 0:0364 > max.simML;L2 / D 0:0159

simHM;L2 D 0:0289 > max.simHM;L2 / D 0:0103

When the outliers are few, the estimated regression model m O HM .x/ and m O L2 .x/
do not differ significantly. This is not the case when the number of outliers increases;
in this sense it seems that L2 estimator can be helpful in cluster detection.
132 A. Durio and E. Isaia

Table 2 Results of simulated MLE HME L2 E

Example II
ˇO0 2.6755 2.4956 1.7340
ˇO1 0.4607 0.5086 0.6856
O 1.4021 1.4074 1.1633

10 1.5
Data points
MLE
8 RME
L2E 1.0
Response y

Slope
0.5

0.0
2

−0.5

2 4 6 8 10 0 1 2 3 4 5 6
x − Panel (a) Intercept − Panel (b)

Fig. 3 (Panel a) Data points of Example II and estimated models m O ML .x/, m

O HM .x/ and m
O L2 .x/.
(Panel b) Contour plot of function g.ˇj  / of Eq. (9) evaluated at   D 0:5 O L2 E

Example II. Let us consider a dataset of n D 300 points, 200 of which arise from
model Y D 1 C 0:8 X C "1 while the remaining from model Y D 5  0:2 X C "2 ,
where "1  N .0; 1/, "2  N .0; 0:5/ and X  U.1; 10/. Again, resorting to the
estimators ML, HM and L2 we obtain the following estimates of the parameters ˇ0 ,
ˇ1 and  listed in Table 2 (also see Fig. 3, panel a). Considering the models m O ML .x/
and m O L2 .x/ the M.C.S. test, with ˛ D 0:01, indicates that they can be considered
dissimilar, as we observe simML;L2 D 0:0582 > max.simML;L2 / D 0:0210. This is
still true if we consider the estimated models m O HM .x/ and m
O L2 .x/, in fact from the
M.C.S. test we have simHM;L2 D 0:0451 > max.simMH;L2 / D 0:0156. Also in this
situation the L2 estimator seems to be helpful in detecting clusters of data when
compared with the Maximum Likelihood and the Huber M estimators.

5 Mixture of Regression Models via L2

It seems to the authors that the properties of robustness of L2 estimates, as outlined

above, can be helpful in pointing out the presence of clusters in the data, e.g. Durio
and Isaia (2007).
This in the sense that whenever sample data belong to two (or more) clusters,
mO L2 .x/ will always tend to fit the cluster with the heaviest number of data points and
hence big discrepancies between m O ML .x/ and mO L2 .x/ will be likely to be observed,
as illustrated by the previous examples. Investigating more accurately function (5)
Comparing Robust Regression Estimators 133

for a fixed value of  it can be seen that in all situations where sample data are
clustered it can show more than one local minimum. A simple way forward is to
investigate the behaviour of the function

2X
n
1
g.ˇj  / D p  .yi jmˇ .xi /;   / (9)
2  n i D1

for different values of   on its parameter space, for instance, the interval
0; 2  O L2 E . In fact, whenever sample data are clustered, function g.ˇj  / given
by Eq. (9) shows one absolute and one or more local points of minimum.
Whenever the presence of clusters of data is detected by L2 criterion, we can
use L2 estimator assuming that the model that best fits the data is a mixture of
K  2 regression models. Assuming that each data point .xi ; yi / comes from
the k-th regression model yi D mˇk .xi / C "i k with probability pk , we suppose
that the random variables Y jx are distributed as a mixture of K Gaussian random
variables, i.e.,

X
K
fY jx .yj 0 / D pk0 .yjmˇ0 .x/; k0 /: (10)
k
kD1

We are now able to derive the following closed-form expression for the estimates of
 0 D Œp 0 ; ˇ 0 ;  0 ; in fact, according to Eq. (9) and recalling Eq. (4), we have
2
1 XXX
n K K
O L2 E Darg min 4 pj ph .0jmˇj .xi /  mˇh .xi /; j2 C h2 /
p;ˇ; n i D1 j D1
hD1
#
2 XX
n K
 pk .yi jmˇk .xi /; k / :
2
(11)
n i D1
kD1

Solving Eq. (11) we obtain the estimates of the vector of the weights, i.e. pO D
Œp1 ; : : : ; pK T , the vector of the parameters, i.e. Ǒ D Œˇ01 ; : : : ; ˇd1 ; : : : ; ˇ0K ; : : : ;
ˇdK T and the vector of the standard deviations of the error of each component of
the mixture, i.e. O D Œ1 ; : : : ; K T .
Example II (continued). Referring to the situation of Example II, for which O L2 D
1:1633, the contour plot of function g.ˇj  / of Eq. (9) and displayed in Fig. 3,
panel b, evaluated at   D 0:5 O L2 E , shows the existence of one absolute minimum
corresponding to the estimates of the parameters of the model Y D 1 C 0:8 X C "1
and one local minimum close to the values of the parameters of the model Y D
5  0:2 X C "2 . We therefore consider a mixture of K D 2 simple linear regression
models. Since in this situation Eq. (10) becomes

fY jx .yj 0 / D p10 .yjˇ001 C ˇ101 x; 10 / C p20 .yjˇ002 C ˇ102 x; 20 /;

134 A. Durio and E. Isaia

10 10 1 1
1
1 L2E_Model_1 1
Data points 1 11 111
L2E_Model_1 2 L2E_Model_2 1 1 11 1 1
u Unknown 1 1 11 1 1
L2E_Model_2 1 1
8 8 o Outlier o 1 1
111 1 1
1 1 1
1 1 11 1
1 1 1

Response y
Response y

1
111 1 11 11 1 11
1
11 1 1 11
o 1 1 1 11
1 11
2 u u u u 111
11u
u 1
6 6 2 u
uuu u u uu u u
u u u uuu u u
uu u
u u u u uuu u uu u u
22 22 u u u u u u u 2
uu u u u uu uuuu u u uuu u uu u 2 2
22 uu u u u u
u uu u u u u uu 22
2 u uu uu u uuu uu uu u u
u u 222 2 2 2
u u u uuu u uu uu uuu u 22 2
4 4 u uu u uu u uuu
u
u
u u 2 2 2 22 22
u uu u uu u u 22 2
uu u u u u u u 2 2 2 2
u u u uu u u
1 u
1 1 11 1
1 1 1 1 11
2 2 1 11 1
11 1
1 1
1
111

2 4 6 8 10 2 4 6 8 10
x − Panel (a) x − Panel (b)

Fig. 4 (Panel a) Data points and estimated components of the mixture of two simple regression
models via L2 . (Panel b) Data points assignment according to the “quick classification rule” with

D3

the L2 estimates of the vector  0 , according to Eq. (11), will be given by solving
"
2X
n
p12 2 C p22 1
O L2 E D arg min p C p1 p2 .0jˇ01 C ˇ11 xi  ˇ02  ˇ12 xi ; 12 C 22 /
p;ˇ; 2 
1 2  n
i D1
#
2 X

n
 p1 .yi jˇ01 C ˇ11 xi ; 12 / C p2 .yi jˇ02 C ˇ12 xi ; 22 / : (12)
n
i D1

From numerical minimization of Eq. (12), we obtain (see Fig. 4, panel a) the
following estimates of the eight parameters of the mixture

L2 E Model_1: pO1 D 0:646 ˇO01 D 1:0281 ˇO11 D 0:8109 O 1 D 0:8411

O O
L2 E Model_2: pO2 D 0:354 ˇ02 D 4:8267 ˇ12 D 0:0576 O 2 D 0:5854

which are quite close to the true values of the parameters.

From a practical point of view, it would be interesting to be able to highlight
which data points belong to each component of the mixture; to this end we resort
to a quick classification rule based on the assumption that the density of the errors
follows a Normal distribution, i.e. 8 i D 1; : : : ; n

if jO"i1 j 5
O 1 ^ jO"i2 j >
O 2 ! .xi ; yi / 2 Model L2 E  I
if jO"i1 j >
O 1 ^ jO"i2 j 5
O 2 ! .xi ; yi / 2 Model L2 E  II
if jO"i1 j 5
O 1 ^ jO"i2 j 5
O 2 ! .xi ; yi / 2 Unknown model
if jO"i1 j >
O 1 ^ jO"i2 j >
O 2 ! .xi ; yi / 2 Outlier; (13)

where
is an appropriate quantile of a N .0; 1/.
Comparing Robust Regression Estimators 135

Table 3 Classification I L2 estimates of pO Quick rule

L2 E Model_1 64:6 % 34:6 % .103/
L2 E Model_2 35:4 % 12:7 % .38/
Unknown model – 52:7 % .157/

Fixing
D 3, if we apply the quick rule and drop two points that are classified as
outliers we obtain (see Fig. 4, panel b) the following classification table, see Table 3.
Clearly, the high percentage of not assigned points (52:7%) is due to the specific
structure of the two clusters which are quite confused.

6 The Case Study

A firm operating in the field of diagnosis and decontamination of electronic

transformers fluids assesses the risks of fluid degradation, electric shocks, fire or
explosion, PCB contamination, decomposition of cellulosic insulation, etc. With the
aid of well-known models and relying on the results of chemical analysis, the firm’s
staff estimate the value of the risk on continuous scales.
In order to determine if their methods of assigning risk values are independent
of specific characteristics of the transformers (age, voltage, fluid mass, etc.) we
conducted an analysis based on a database of 1;215 records of diagnosis containing
oil chemical analysis, technical characteristics and risk values.
Taking into account the risk of fire (Y ) and the risk of electric shocks (X ),
it was natural to suppose a linear dependence between the two variables, i.e., we
considered the simple regression model with mˇ .xi / D ˇ0 C ˇ1 xi .
Resorting to the estimators ML, HM and L2 we obtained the following estimates
of the parameters ˇ0 , ˇ1 and  listed in Table 4.
Although the estimates of the vector of the parameters ˇ are quite close, the
corresponding three estimated models differ in some way, e.g., Fig. 5, panel a.
Computing the values of the sim./ statistics, the M.C.S. test led us to the
conclusion that the L2 estimated model can be considered dissimilar from both
mO ML .x/ and mO HM .x/ models, as

simML;L2 D 0:0220 > max.simML;L2 / D 0:0051

simHM;L2 D 0:0203 > max.simHM;L2 / D 0:0031

Probing more deeply, we found that function g.ˇj  / of Eq. (9) presents two
points of minimum for   D 0:5 O L2 E D 0:0755, as shown in Fig. 5, panel b.
Therefore we decided to model our data by means of a mixture of two simple
regression models. Considering the L2 criterion and solving Eq. (12), we found that
about 57 %.D pO1 %/ of the data points follow the model

O ˇ1 .x/ D 0:4042 C 1:7705 x

m ! L2 E Model_1
136 A. Durio and E. Isaia

Table 4 Estimates of the MLE HME L2 E

parameters after resorting
ˇO0 0.4321 0.4423 0.5330
ˇO1 1.7110 1.7199 1.8115
O 0.1472 0.1471 0.1509

5 2.0

Data points
4 MLE
RME
L2E 1.8
*
3
Risk of fire

Slope
1.6

2 *

1.4
1

0 1.2

0.5 1.0 1.5 2.0 2.5 3.0 −0.7 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0.0
Risk of electric shocks − Panel (a) Intercept − Panel (b)

Fig. 5 Case study. (Panel a) Data points and estimated models m O ML .x/, m
O HM .x/ and m O L2 .x/.
(Panel b) Contour plot of function g.ˇj  / of Eq. (9) evaluated at   D 0:5 O L2 E , with O L2 E D
0:151

for which O 1 D 0:0547, while the remaining 43 %.D pO2 %/ of the data points follow
the model

O ˇ2 .x/ D 0:3955 C 1:5847 x

m ! L2 E Model_2

for which O 2 D 0:0775. Panel a of Fig. 6 shows the two estimates models.
Applying the quick rule we were able to classify the data according to whether
they followed the first or the second regression model. From the L2 estimates of pO
and the quick rule (dropping two points that were classified as outliers) we obtained
the following classification table, see Table 5.
In order to classify the 266 (D 22:0 %) points belonging, according to the
quick rule, to the Unknown Model, we had to investigate more deeply the specific
characteristics of the transformers themselves.
Examining our database, we found that 40 % of the transformers has a fluid mass
5500 kg and the L2 criterion gave us an estimate of 43 % for the weight of points
belonging to L2 E Model_1 while our quick rule assigned the 36:9 % of data points
to L2 E Model_2.
Furthermore, our quick classification rule assigns 419 out of the 448 points
(93:5 %) to L2 E Model_2 and these have a fluid mass less (or equal) than 500 kg,
while all the 499 transformers imputed to L2 E Model_1 have a fluid mass greater
than 500 kg, see Table 6.
Comparing Robust Regression Estimators 137

5 5

Data points Fluid <= 500 kg

4 L2E_Model_1 4 Fluid > 500 kg
L2E_Model_2

3 3
Risk of fire

Risk of fire
2 2

1 1

0 0

0.5 1.0 1.5 2.0 2.5 3.0 0.5 1.0 1.5 2.0 2.5 3.0
Risk of electric shocks − Panel (a) Risk of electric shocks − Panel (b)

Fig. 6 Case study. (Panel a) Data points and estimated models m O ˇ1 .x/ and mO ˇ2 .x/. (Panel b)
Final data points assignment according to the fluid mass of the electrical transformers

Table 5 Classification II L2 estimates of pO Quick rule

L2 E Model_1 57:0 % 41:1 % .499/
L2 E Model_2 43:0 % 36:9 % .448/
Unknown model – 22:0 % .266/

Table 6 Fluid mass of the Fluid mass 5 500 kg Fluid mass > 500 kg
model
L2 E Model_1 0 (0:0 %) 499 (100 %)
L2 E Model_2 419 (93:5 %) 29 (6:5 %)
Unknown model 65 (24:4 %) 201 (75:6 %)

From the above, we decided to use the fluid mass as clustering variable and so
we assigned the transformers with a fluid mass equal or less than 500 kg to Model
L2 E Model_2 while the transformers with a fluid mass greater than 500 kg were
assigned to the L2 E Model_1 regression line. The final assignment is shown in
Fig. 6, panel b.
These results allowed us to state that, at fixed level of risk of electric shocks,
the risk of fire was evaluated in a different way for the two groups of transformers,
i.e., the relationship between the two variables depended on the fluid mass of the
transformers.
However, the chemical staff of the firm could not find any scientific reason to
explain the different risks of fire in the two types of transformers, so they decided to
change the model used by assigning different weights to the hydrocarbon variable
in order to better reflect the differential risks of fire.

Acknowledgements The authors are indebted to the coordinating editors and to the anonymous
referees for carefully reading the manuscript and for their many important remarks and suggestions.
138 A. Durio and E. Isaia

References

Barnard, G. A. (1963). Contribution to the discussion of paper by m.s. bartlett. Journal of the Royal
Statistical Society, B, 25, 294.
Basu, A., Harris, I. R., Hjort, N., & Jones, M. (1998). Robust and efficient estimation by
minimizing a density power divergence. Biometrika, 85, 549–559.
Davies, P. L. (1993). Aspects of robust linear regression. Annals of Statistics, 21, 1843–1899.
Dodge, Y., & Jurečkova, J. (2000). Adaptive regression. New York: Springer.
Durio, A., & Isaia, E. D. (2007). A quick procedure for model selection in the case of mixture of
normal densities. Journal of Computational Statistics & Data Analysis, 51(12), 5635–5643.
Durio, A., & Isaia, E. D. (2010). Clusters detection in regression problems: A similarity test
between estimate. Communications in Statistics Theory and Methods, 39, 508–516.
Fujisawa, H., & Eguchi, F. (2006). Robust estimation in the normal mixture model. Journal
of Statistical Planning Inference, 136, 3989–4011.
Hampel, F. R., Ronchetti, E. M., Rousseeuw, R.,J., & Stahel, W. A. (2005). Robust regression and
outlier detection. New York: Wiley.
Hope, A. C. (1968). A simplified monte carlo significance test procedure. Journal of the Royal
Statistical Society, B, 30, 582–598.
Huber, P. J. (1981). Robust statistics. New York: Wiley.
Jurečkova, J., & Picek, J. (2006). Robust statistical methods with R. Boca Raton: Chapman & Hall.
Maronna, R., Martin, D., & Yohai, V. (2006). Robust statistics: Theory and methods. New York:
Wiley.
Rousseeuw, P. J., & Leroy, A. M. (1987). Robust regression and outlier detection. New York:
Wiley.
Rousseeuw, P. J., Van Alest, S., Van Driessen, K., & Agulló, J. (2004). Robust multivariate
regression. Technometric, 46, 293–305.
Scott, D. W. (2001). Parametric statistical modeling by minimum integrated square error. Techno-
metrics, 43, 274–285.
Seber, G. A. F., & Lee, A. J. (2003). Linear regression analysis (2nd ed.). New York: Wiley.
Staudte, R. G., & Sheather, S. J. (1990). Robust estimation and testing. New York: Wiley.
Finite Mixture Model Clustering of SNP Data

Norma Bargary, J. Hinde, and A. Augusto F. Garcia

Abstract Finite mixture models have been used extensively in clustering applica-
tions, where each component of the mixture distribution is assumed to represent
an individual cluster. The simplest example describes each cluster in terms of a
multivariate Gaussian density with various covariance structures. However, using
finite mixture models as a clustering tool is highly flexible and allows for the
specification of a wide range of statistical models to describe the data within
each cluster. These include modelling each cluster using linear regression models,
mixed effects models, generalized linear models, etc. This paper investigates using
mixtures of orthogonal regression models to cluster biological data arising from a
study of the sugarcane plant.

Keywords Clustering • Finite-mixtures models • Orthogonal regression • SNP

data

1 Introduction

Sugarcane is the highest tonnage crop among cultivated plants. Almost 70 % of the
world’s sugar supply is derived from sugarcane while the remaining 30 % comes
from sugar beet. Sugarcane is cultivated in approximately 110 different countries
and is an important industrial crop in many regions such as Brazil and South
Asia, where approximately 50 % of all sugarcane production occurs. According to
Palhares et al. (2012), sugarcane is a cost-effective renewable resource produced
for use as sugar, in animal feeds, alcohols and fertilizers. In addition, there is an
increased emphasis in its use as a bio-fuel which requires a crop with high yield

N. Bargary ()
Department of Mathematic & Statistics, University of Limerick, Limerick, Ireland
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 139

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__11,
© Springer International Publishing Switzerland 2014
140 N. Bargary et al.

and fibre content. Thus there is an interest in developing breeds with high sugar
yield, increased resistance to drought, high fibre content, etc. Such work requires
identifying gene(s) that contribute to these complex traits that could possibly
provide DNA markers for use in marker-assisted breeding. One of the primary issues
in examining the genetic make-up of the sugarcane genome is its polyploid structure.
Polyploidy implies that the sugarcane plant has multiple sets of chromosomes rather
than two as for diploid organisms (such as humans). For example, the S. spontaneous
breed has between 5 and 16 sets of chromosomes, while the species S. officinarum
has 8 sets of chromosomes. Each chromosome carries a particular form of each
gene, i.e., an allele for that gene. For example, the gene for eye colour has a number
of alleles; blue, brown, green, grey. The particular combination of these alleles
results in different genotypes, and consequently phenotypes in each individual.
Since sugarcane is polyploid, individual alleles can appear in varying numbers
and combinations, resulting in much genetic variation in sugarcane. It is therefore
difficult to obtain information about its complex genetic makeup. A primary aim is
to identify the many different alleles and associated genotypes and phenotypes for
the various sugarcane crossbreeds. One way of doing this is through the analysis of
single nucleotide polymorphisms (SNPs).
SNPs occur during cell division, when an existing cell divides in two by first
copying its DNA so the new cells will each have a complete set of genetic
instructions. Cells sometimes make mistakes (called SNPs) during the copying
process and these lead to variations in the DNA sequence at particular locations.
An SNP is a single base pair mutation at a specific locus and typically consists of
two alleles. Much research has centered on the identification of new SNPs since
they act as markers to identify genomic regions controlling traits of interest and can
be used for genotyping (Olivier 2005). Figure 1 gives an example for a sample of
diploid individuals. Here the intensities for two alleles (C and G for example) of a
particular SNP are measured for a number of individuals and plotted on the x and
y-axes respectively. These observations can be clustered into three clear groups; one
group along the y-axis, one along the x-axis and one along the line corresponding to
=4. Since diploid individuals have two sets of chromosomes and each chromosome
has one of the two alleles (either C or G), there are three possible genotypes for this
SNP. If the alleles on both chromosomes are the same, the individual is said to be
homozygous with possible genotypes G/G or C/C. If an individual has different
alleles on each chromosome they are said to be heterozygous with genotype G/C.
The group along the y-axis consists of individuals with genotype G/G (since the
contribution from the C allele is essentially 0), the group along the x-axis consists
of individuals with genotype C/C, and the group along the line =4 consists of
individuals with genotype G/C. Therefore the clusters produced contain information
about the genotype of each individual. In addition, the proportion of individuals
within each cluster are related to expected segregation ratios and to the ploidy
level (the number of chromosomes) of the individuals measured, while the angles
between the clusters are also informative for this reason. In this simple example,
the expected segregation ratio is 1:2:1 for the G/G:G/C:C/C genotypes and the
angle between the clusters is =4 (since the intensity of the G allele and C allele
Clustering SNP Data 141

Fig. 1 Measured intensities 60

for a single SNP with two
alleles; G (y-axis) and C 50
(x-axis)
40

30

20

10

0 10 20 30 40 50 60

in this group will be equal). Thus, by comparing the proportions of individuals in

each cluster with the expected segregation ratio and examining the angles between
the clusters, it would be possible to infer that these individuals are diploid if that
information was not already available.
Such an approach would also prove useful in the analysis of sugarcane since
SNPs occur frequently in the sugarcane genome, approximately 1 in every 50 bases.
However, its polyploid structure results in a much more complex problem. The
ploidy level, on which segregation ratios etc. are based, is generally unknown and
thus it is difficult to determine the number of unique genotypes that exist among
individuals. This information could have implications for sugarcane breeding since
high yield potential may be due to the presence of a specific allele, or a particular
number of copies of a specific allele(s) present at a gene locus, or possibly a
combination of both. Cordeiro et al. (2006) state that the frequency of a SNP base
(A, T, C, G) at a locus is determined by
• the number of chromosomes carrying the gene;
• the number of different alleles for a gene;
• the frequency of each allele possessing each SNP base.
In sugarcane, the proportional frequency of each SNP base varies depending on
the number of alleles containing the SNP locus. The frequency of a SNP base at
a gene locus will be determined by both the number of chromosomes carrying the
gene, the number of different alleles and the frequency of each allele possessing
each SNP base. This in turn provides information about the ploidy level and genetic
make-up of the sugarcane plant.
Figure 2 displays the data collected for two of the SNPs analyzed in this paper.
Each point represents the intensity of two SNP bases, h.L is the intensity of the C
allele and h.H is the intensity of the G allele. The data on the LHS in Fig. 2 can
clearly be clustered into two groups—the group along the y-axis and the group
along the line with a particular (unknown) angle. These groups correspond to two
genotypes and thus clustering is essential for genotyping. In the example shown on
142 N. Bargary et al.

80
50

60 40

30
h.H

h.H
40

20

20
10

0 0

0 20 40 60 80 0 10 20 30 40 50
h.L h.L

Fig. 2 LHS: Raw data for SNP Contig89b17; RHS: Raw data for SNP Contig2312b2

the RHS in Fig. 2, it is not clear how many clusters (i.e. genotypes) are present and
therefore the aim of this research is to develop a technique that can:
(1) provide a probabilistic clustering to identify individuals that have high proba-
bility of belonging to a particular cluster (i.e. individuals that have a particular
genotype) and those that are regarded as an unclear genotype;
(2) determine the number of clusters present;
(3) calculate the proportion of individuals in each cluster;
(4) determine the angles between clusters.
The above criteria suggest that using an appropriate clustering algorithm may prove
useful for these data. Clustering methods such as k-means clustering (Hartigan and
Wong 1978), hierarchical clustering (Eisen et al. 1998; Spellman et al. 1998), clus-
tering on self-organizing maps (Kohonen 1997; Tamayo et al. 1999), model-based
clustering (Fraley and Raftery 2002; McLachlan et al. 2002, 2003, 2006), fuzzy
c-means clustering (Futschik and Carlisle 2005) and tight clustering (Tseng and
Wong 2005), have been used extensively in many applications to group observations
such that individuals within a cluster are more alike than individuals in different
clusters. Here we use model-based clustering, a parametric clustering technique that
assumes a particular statistical model for the data within each cluster. In the simplest
case, the data in each cluster are assumed to have a normal distribution, however,
more elaborate models can also be used. These include mixtures of linear regression
models, mixtures of mixed effects models (Celeux et al. 2005; Ng et al. 2006),
mixtures of generalized linear models (Leisch 2004; Grün and Leisch 2008), etc.
In contrast, this paper uses finite mixtures of orthogonal regression lines to cluster
SNP data arising from the analysis of the genetic traits of sugarcane.
The remainder of the paper is outlined as follows. Section 2 describes finite
mixture models and discusses the use of orthogonal regression lines to describe
the data within each cluster. Section 3 presents the results of applying the proposed
method to SNP data arising from sugarcane. Section 4 provides a brief discussion.
Clustering SNP Data 143

2 Methodology

2.1 Finite Mixture Models

Model-based clustering assumes that the data y D .y1 ; : : : ; yn / arise from a mixture
of G components

X
G
f .yi I ‚/ D g fg .yi I  g /;
gD1

where f .yi I ‚/ is the density of the data, fg .yi I  g / is the density of the data in
the gth component, which if assumed to be a normal distribution has parameters
 g D .g ; † g /, and g are mixing proportions such that

X
g D 1:
g

To fit the latter model, it is necessary to obtain estimates of . 1 ; : : : ;  G ; 1 ; : : : ;

G /, which is typically achieved using the Expectation–Maximization (EM) algo-
rithm of Dempster et al. (1977).
The EM algorithm determines maximum likelihood estimates of the parameters
in a statistical model by maximizing the log-likelihood in the presence of “missing”
data. The missing data in a clustering context are the vectors of cluster membership
probabilities zi D .zi1 ; : : : ; ziG / such that

1 if gene i belongs to cluster g;
zig D
0 otherwise:

The EM algorithm maximizes the “complete data” log-likelihood

X
n X
G
`C . g ; g ; zig jy/ D zig Œlog g fg .yi j g /;
i D1 gD1

by iterating between the E-step, where the zig values are replaced with
their expected values conditional on the current model parameter estimates
zOig D EŒzig jyi ;  1 ; : : : ;  G , and the M-step, where the model parameters
. 1 ; : : : ;  G ; 1 ; : : : ; G / are updated given the current zOig values.
144 N. Bargary et al.

2.2 Finite Mixtures of Least Squares Regression Lines

For these data, fitting a regression line to the data in each cluster would facilitate
determining the number of clusters and the angles between clusters (as the angles
between the regression lines describing the data in each cluster), which contains
information about the number of different alleles, the corresponding genotypes and
phenotypes and the ploidy level as required. Here, the data for each individual
consists of a p-length vector yi D .h.H; h.L/, where p D 2. To cluster the data
using a mixture of regression lines, it was initially assumed that one of the measured
variables, h.L or h.H, was the response variable y and the other was the explanatory
variable x. Information from the geneticists indicated that the regression lines
should pass through the origin and thus it was assumed that a linear regression line
through the origin

yi D ˇ1 xi C "i

could be fitted to the data in each cluster. The cluster densities were then univariate
normal such that

fg .yi jˇ1g xi ; g2 / D N.ˇ1g xi ; g2 /;

where ˇ1g is the slope in the gth component and g2 is the variance. The process was
repeated by reversing the roles of h.H and h.L, i.e., if h.H was the assumed response
variable in the first analysis, it was treated as the explanatory variable in the second
analysis. The results of both analyses were then compared.
Using h.L as the response variable y and h.H as the explanatory variable x
yielded more interpretable results than using y = h.H and x = h.L. This was
primarily due to the presence of many h.L values that had the same corresponding
h.H value, resulting in a regression line parallel to the y-axis when y = h.H was
used. Such lines have infinite slope and thus cannot be estimated by standard least
squares. In addition, for these data it is not clear which variable is the true response
variable and which is the true explanatory variable and both values contain some
measurement error/noise. As a result an alternative approach using finite mixtures
of orthogonal regression lines is proposed.

2.3 Finite Mixtures of Orthogonal Regression Lines

Total Least Squares (also known as errors in variables or rigorous least squares) is
a regression method where observational errors on both the dependent variable and
p independent variables are taken into account during model fitting. Computation
of the total least squares regression line is achieved using the singular value
decomposition (SVD) as described in the general case as follows. Let
Clustering SNP Data 145

y D Xˇ

be a system of equations that we wish to solve for ˇ, where X is n  p and y is n  1.

Therefore, we wish to find ˇ that minimizes the matrix of errors E for X and vector
of errors r for y. That is,

argminE;r kŒE rkF ; .X C E/ˇ D y C r;

where ŒE r is the augmented matrix and k  kF is the Frobenius norm.

Golub and Van Loan (1980) show that this can be achieved using a SVD of ŒX y
such that
  T
  DX 0 VXX VXy
ŒX y D UX Uy
0 Dy VyX Vyy

and setting some of the singular values to zero;

  T
  DX 0 VXX VXy
Œ.X C E/ .y C r/ D UX Uy :
0 011 VyX Vyy

This implies
  T
  011 0 VXX VXy
ŒE r D  UX Uy
0 DX VyX Vyy
  T
  VXy VXy
D Xy
Vyy Vyy

and
 
VXy
Œ.X C E/ .y C r/ D 0:
Vyy

If Vyy is nonsingular, right multiply both sides by V1

yy to get:

" #  
VXy V1 ˇ
Œ.X C E/ .y C r/ yy
D Œ.X C E/ .y C r/ D0
Vyy V1
yy I11

and thus

ˇO D VXy V1
yy :

Since the sugarcane data is two-dimensional, Deming regression (a special case of

total least squares where there is one dependent variable y and one independent
146 N. Bargary et al.

variable x), is appropriate. Then both x and y are assumed to be measured with
error such that

xi D xi C i ;
yi D yi C i ;

where i and i are independent and the ratio of their variances

2
ıD
2

is assumed to be known. For our purposes we use orthogonal regression (through

the origin), which assumes ı D 1. Applying the estimation approach outlined above
implies that ˇO1 D VXy =Vyy and VXy /Vyy are scalars rather than matrices.
Once the slope of the orthogonal regression line is determined using the above
method, it is necessary to determine the fitted values. This requires finding the
equation of the line that passes through the original data point .xi ; yi / but orthogonal
to the fitted regression line with slope given by ˇO1 . The point at which both lines
intersect gives the fitted point. Let yOi D xO i ˇO1 denote the line with slope ˇO1
that passes through the origin (i.e. the fitted orthogonal regression line). Then the
equation of the line orthogonal to this will have slope 1=ˇO1 and corresponding
equation

xO i
yOi D  C b:
ˇO1

This line passes through our original data point .xi ; yi / implying that
xi
b D yi C
ˇO1

and

xO i xi
yOi D  C yi C :
ˇO1 ˇO1

The point at which the two lines intersect is given by:

xO i xi
xO i ˇO1 D  C yi C ;
ˇO1 ˇO1
xO i ˇO12 D xi C yi ˇO1 C xi ;
Clustering SNP Data 147

yi ˇO1 C xi
xO i D ;
1 C ˇO 2 1

yOi D xO i ˇO1 :

Therefore the fitted point is

!
yi ˇO1 C xi
xO i D ; yOi D xO i ˇO1 :
1 C ˇO 2 1

The corresponding residuals are

rix D xi  xO i ;
riy D yi  yOi ;

and since orthogonal regression assumes the variances 2 and 2 are equal (but
independent), an overall estimate of  2 is given by,

P
n P
n
.xi  xO i /2 C .yi  yOi /2
i D1 i D1
O D2
:
2.n  1/

Implementing this in a clustering context requires maximizing the “complete-data”

log-likelihood

X
n X
G
`C . g ; g ; zig jxi ; yi / D zig Œlog g fg .xi ; yi j g /; (1)
i D1 gD1

where fg .xi ; yi j g / is now a bivariate normal distribution such that

fg .xi ; yi j g / D N.g ; † g /;

and
 ! !
2
Xg g 0 g2 0
g D ; †g D 2 D ;
Yg 0 g 0 g2

2
since g D g
2
. Estimates of Xg D xO ig and Yg D yOig are given by

yi ˇO1g C xi
xO ig D ; (2)
1 C ˇO 2 1g

yOig D xO ig ˇO1g : (3)

148 N. Bargary et al.

The estimated regression coefficients ˇO1g , g D 1; : : : ; G are found for each

component via SVD wherepthe data values .xi ; yi / are multiplied by their estimated
weights in component g, zOig , such that
  T
  DX 0 VXX VXy
ŒXwgt ywgt  D UX Uy ; (4)
0 Dy VyX Vyy
p p p p p
where Xwgt D . zOig x1 ; zOig x2 ; : : : ; zOig xn /T , ywgt D . zOig y1 ; zOig y2 ; : : : ;
p
zOig yn /T .
Then

ˇO1g D VXy =Vyy

and g2 is estimated using

P
n P
n
zOig .xi  xO ig /2 C zOig .yi  yOig /2
i D1 i D1
O g2 D : (5)
P
n
2 zOig
i D1

The following outlines the EM algorithm for the orthogonal regression problem.
Steps in EM algorithm for orthogonal regression:
(1) Initialize zOig , e.g. by randomly allocating each individual to a particular cluster.
(2) M-step:
• For each cluster g, calculate ˇO1g using (4), O g2 using (5) and

P
n
zOig
i D1
O g D :
n
• Calculate the fitted values for the i th individual in each component using (2)
and (3).
(3) E-step:
• For each component, calculate fg .xi ; yi j O g / (a bivariate normal distri-
O g; †
2 2
bution). Since g and g are assumed to be independent and equal, this
implies that

fg .xi ; yi j O g / D fg .xi jxO ig ; O 2 / fg .yi jyOig ; O 2 /;

O g; † g g

i.e., the product of two univariate normal distributions.

• Update the individual weights using
Clustering SNP Data 149

O g fg .xi ; yi j O g/
O g; †
zOig D :
P
G
O h fh .xi ; yi j O g/
O g; †
hD1

(4) Iterate between Steps 2 and 3 until convergence.

2.4 Choosing the Number of Clusters

Cluster analysis requires choosing the number of clusters G and model-based

clustering facilitates using model-selection criteria such as the Akaike Information
Criterion (AIC) or Bayesian Information Criterion (BIC) to determine an optimal
value for G. Typically in cluster analysis the BIC is used and is written as

BIC D 2 log L C d log nI

where log L is given in (1), d = number of parameters to be estimated in the

model and n is the sample size. Here d consists of G slope parameters ˇO1g ; g D
1; : : : ; G, G variance parameters O g2 ; g D 1; : : : ; G and G  1 mixing proportions
1 ; : : : ; G1 , implying that d D 3G  1. The model fitting process is repeated
for varying values of G and the value for G corresponding to the solution with
minimum BIC is typically chosen. However, in many instances in this work, a drop
in BIC could be attributed to a solution having many empty clusters. As a result, the
statistical “elbow” in the BIC plot was identified and the value for G at which this
“elbow” occurred was then chosen as the number of clusters.

3 Results

The proposed method was applied to five different SNPs: Contig89b17, Con-
tig168b1, Contig628b21, Contig875b2 and Contig2312b2. The aim was to deter-
mine the number of clusters, the estimated orthogonal regression lines modelling
the data in each cluster and the angle between the resulting clusters. In each case the
algorithm was run using either h.H or h.L as the response variable, to determine if
the clustering results were consistent when the roles of the variables were reversed.
The results presented in this section are for y = h.H and x = h.L; however, it should
be noted that the results obtained from reversing the roles of h.H and h.L were
identical (as expected). The number of clusters was determined using the BIC and
outright assignment of data points into clusters was achieved using the map function
in the mclust package. This function assigns point i to cluster g such that
150 N. Bargary et al.

8
<1 if g D arg max zOih ;
zOig D h
:0 otherwise:

To ensure the EM algorithm did not converge to a local optimum, the algorithm
was run from five initial starting points obtained by randomly allocating individuals
to clusters. The solution with maximum log-likelihood from these five starting
positions was retained. The angles between adjacent clusters were calculated using

m2  m1 180
 D arctan  ;
1 C m1 m2 

where m1 and m2 are the slopes of the orthogonal regression lines describing the
data in these two clusters.

3.1 Contig89b17

The top of Fig. 3 displays the raw data and BIC plot for Contig89b17. The BIC plot
indicated that there were three clusters in these data (the reduction in BIC at later
stages is due to a number of empty clusters for each solution). The corresponding
cluster assignments and estimated orthogonal regression lines for each cluster are
given in the bottom of Fig. 3. It is clear that the proposed method can fit a regression
line to the group parallel to the y-axis, which was not possible using mixtures of
least squares regression lines. The estimated slopes of the orthogonal regression
lines (from right to left) were ˇO11 D 4:294, ˇO12 D 5:496 and ˇO13 D 6; 377:499, and
thus the angle between Cluster 1 and Cluster 2 (C1;C 2 ) was calculated using
!
ˇO12  ˇO11
C1;C 2 D arctan
1 C ˇO11 ˇO12

5:496  4:294 180
D arctan 
1 C 4:294 5:496 
D 2ı 800 :

The angle between Cluster 2 and Cluster 3 was


6377:499  5:496 180
C 2;C 3 D arctan  D 10ı 300 :
1 C .5:496/.6377:499/ 
Clustering SNP Data 151

80

1000

60

800

BIC
h.H

40

600

20

400

0 20 40 60 80 2 3 4 5 6 7 8
h.L G

80

60
h.H

40

20

0 20 40 60 80
h.L

Fig. 3 Top LHS: Raw data for SNP Contig89b17. Top RHS: BIC plot for SNP Contig89b17.
Bottom: Cluster allocations and fitted orthogonal regression line for SNP Contig89b17

3.2 Contig168b1

The analysis was repeated for SNP Contig168b1. In this case, the BIC plot indicated
that there were five clusters (top of Fig. 4) in this dataset. The corresponding cluster
assignments and estimated orthogonal regression lines for each cluster are shown in
the bottom of Fig. 4. The estimated slopes of the orthogonal regression lines (from
right to left) were ˇO11 D 2:050, ˇO12 D 2:840, ˇO13 D 3:813, ˇO14 D 5:654 and
ˇO15 D 33:992. The corresponding angles between adjacent clusters (from right to
left) were

C1;C 2 D 6ı 610 C 3;C 4 D 4ı 670

C 2;C 3 D 4ı 700 C 4;C 5 D 8ı 340
152 N. Bargary et al.

2200

80

2000

60

1800

BIC
h.H

40

1600

20

1400

0 20 40 60 80 2 4 6 8 10
h.L G

80

60
h.H

40

20

0 20 40 60 80
h.L

Fig. 4 Top LHS: Raw data for SNP Contig168b1. Top RHS: BIC plot for SNP Contig168b1.
Bottom: Cluster allocations and fitted orthogonal regression line for SNP Contig168b1

3.3 Contig628b21

For this SNP, there was an obvious group of individuals parallel to the y-axis. Again
it can be seen that the proposed methodology could cluster the data in this group
without difficulty. The BIC plot in Fig. 5 indicated that there were three clusters in
this dataset, with the cluster assignments and estimated orthogonal regression lines
for each cluster displayed in the bottom of Fig. 5. From right to left, the estimated
slopes of the orthogonal regression lines were ˇO11 D 4:075, ˇO12 D 8:491 and
ˇO13 D 1; 087:018. The angles between adjacent clusters (from right to left) were

C1;C 2 D 7ı 70
C 2;C 3 D 6ı 660
Clustering SNP Data 153

2000
100

1800
80
1600

60

BIC
h.H

1400

40
1200

20 1000

0 800

0 20 40 60 80 100 2 4 6 8 10 12 14
h.L G

100

80

60
h.H

40

20

0 20 40 60 80 100
h.L

Fig. 5 Top LHS: Raw data for SNP Contig628b21. Top RHS: BIC plot for SNP Contig628b21.
Bottom: Cluster allocations and fitted orthogonal regression line for SNP Contig628b21

3.4 Contig875b2

The data for this SNP did not exhibit a particularly obvious grouping structure;
however, there were a number of points close to the y-axis which were quite
different from the remaining data. The proposed method clustered all of these
observations into the same cluster, described by an orthogonal regression line that
was virtually parallel to the y-axis. The BIC plot for these data indicated that
choosing G D 6 clusters was appropriate, with estimated slopes (from right to
left) given by ˇO11 D 1:532, ˇO12 D 1:874, ˇO13 D 2:394, ˇO14 D 2:858, ˇO15 D 4:906
and ˇO16 D 114:623. The drop in BIC from G D 10 onwards was attributable to
a number of empty clusters arising in the clustering solution (Fig. 6). The angles
between adjacent clusters were calculated as
154 N. Bargary et al.

35

1400
30

25
1200
20

BIC
h.H

15
1000

10

5 800

0 5 10 15 20 25 30 35 2 4 6 8 10 12 14
h.L G

35

30

25

20
h.H

15

10

0 5 10 15 20 25 30 35
h.L

Fig. 6 Top LHS: Raw data for SNP Contig875b2. Top RHS: BIC plot for SNP Contig875b2.
Bottom: Cluster allocations and fitted orthogonal regression line for SNP Contig875b2

C1;C 2 D 5ı 050 C 4;C 5 D 7ı 770

C 2;C 3 D 5ı 420 C 5;C 6 D 11ı 020
C 3;C 4 D 3ı 380

3.5 Contig2312b2

Finally, the algorithm was applied to SNP Contig2312b2. This was the most
complex dataset, since there was no clear grouping structure evident a priori.
Examining the BIC plot, G D 8 clusters were chosen with estimated regression lines
and cluster memberships displayed in the bottom of Fig. 7. The slope coefficients
(from right to left) for the eight clusters were ˇO11 D 0:582, ˇO12 D 1:186,
Clustering SNP Data 155

1900
50

1800
40
1700

30

BIC
h.H

1600

20
1500

10 1400

0 1300

0 10 20 30 40 50 2 4 6 8 10 12 14
h.L G

50

40

30
h.H

20

10

0 10 20 30 40 50

h.L

Fig. 7 Top LHS: Raw data for SNP Contig2312b2. Top RHS: BIC plot for SNP Contig2312b2.
Bottom: Cluster allocations and fitted orthogonal regression line for SNP Contig2312b2

ˇO13 D 1:645, ˇO14 D 2:283, ˇO15 D 2:890, ˇO16 D 4:214, ˇO17 D 7:116 and
ˇO18 D 19:111, with corresponding angles between adjacent clusters given by

C1;C 2 D 19ı 660 C 4;C 5 D 4ı 570 C 7;C 8 D 5ı 000

C 2;C 3 D 8ı 860 C 5;C 6 D 5ı 740
C 3;C 4 D 7ı 630 C 6;C 7 D 5ı 350

4 Discussion

This paper has presented a simple but effective way of clustering bivariate SNP data
using mixtures of orthogonal regression lines, where both variables are assumed to
be measured with error and there is no natural distinction between the response
156 N. Bargary et al.

and explanatory variable. The approach can handle groups of observations that
form clusters parallel to the y-axis, which could not be fitted using mixtures of
simple linear regression lines and provides estimates of the number of clusters, the
number of individuals within each cluster and the angles between clusters. This
information can then be used to make inferences about the (unknown) ploidy level,
the segregation ratios and the different genotypes/phenotypes associated with a
particular sugarcane genome. It should be noted that orthogonal regression is closely
related to calculating the principal components for a particular dataset. However, in
this instance it is assumed that the regression line passes through the origin, and thus
this is not equivalent to calculating the principal components (Joliffe 2002). In other
applications, calculation of the principal components may be equivalent to the total
least squares estimate of the orthogonal regression lines.
An alternative approach to consider is that of Fujisawa et al. (2004), who
transformed SNP data to polar co-ordinates before clustering using normal mixture
model clustering with a penalized likelihood. Their approach examined data from
a diploid organism and used prior knowledge (based on the known ploidy level)
about the angles between clusters to improve the clustering results. We explored
transforming these data to angular co-ordinates before using standard model-based
clustering; however, in our case this did not produce particularly interpretable results
since the ploidy level was unknown and no prior knowledge about the angles
between clusters was available.
A future extension of this work involves constraining the angles between clusters
to be the same (or based on multiple(s) of a baseline angle). In this instance, each
cluster would be represented by an orthogonal regression line, but some clusters
may now be empty. This could imply that more clusters are required, but might
more accurately reflect the polyploid structure and perhaps explain the behavior of
the BIC encountered in the analysis presented.

Acknowledgements Norma Bargary (nee Coffey) carried out this work at the National University
of Ireland, Galway while supported by Science Foundation Ireland under Grant No. 07/MI/012
(BIO-SI project). The authors would also like to thank Dr. Anete P. Souza who carried out the
laboratory work and is the co-ordinator of the project that produced the data analyzed.

References

Celeux, G., Martin, O., & Lavergne, C. (2005). Mixture of linear mixed models for clustering gene
expression profiles from repeated microarray experiments. Statistical Modelling, 5, 243–267.
Cordeiro, G., Eliott, F., McIntyre, C., Casu, R. E., & Henry, R. J. (2006). Characterization of single
nucleotide polymorphisms in sugarcane ESTs. Theoretical and Applied Genetics, 113, 331–343.
Dempster, A., Laird, N., & Rubin, D. (1977). Maximum likelihood from incomplete data via the
EM algorithm. Journal of the Royal Statistical Society, Series B, 39, 1–38.
Eisen, M., Spellman, P., Brown, P., & Botstein, D. (1998) Cluster analysis and display of genome-
wide expression patterns. Proceeding of the National Academy of Sciences of the USA, 95,
14863–14868.
Clustering SNP Data 157

Fraley, C., & Raftery, A. (2002). Model-based clustering, discriminant analysis, and density
estimation. Journal of the American Statistical Association, 97, 611–631.
Fujisawa, H., Eguchi, S., Ushijima, M., Miyata, S., Miki, Y., Muto, T., et al. (2004). Genotyping
of single nucleotide polymorphism using model-based clustering. Bioinformatics, 20, 718–726.
Futschik, M., & Carlisle, B. (2005). Noise-robust soft clustering of gene expression time-course
data. Journal of Bioinformatics and Computational Biology, 3, 965–988.
Golub, G., & Van Loan, C. (1980). An analysis of the total least squares problem. SIAM Journal
of Numerical Analysis, 17, 883–893.
Grün, B., & Leisch, F. (2008). FlexMix Version 2: Finite mixtures with concomitant variables and
varying and constant parameters. Journal of Statistical Software, 28, 1–35.
Hartigan, J., & Wong, M. (1978). A k-means clustering algorithm. Applied Statistics, 28, 100–108.
Joliffe, I. (2002). Principal component analysis. New York: Springer.
Kohonen, T. (1997). Self-organizing maps. New York: Springer.
Leisch, F. (2004). FlexMix: A general framework for finite mixture models and latent class
regression in R. Journal of Statistical Software, 11, 1–18.
McLachlan, G., Bean, R., & Peel, D. (2002). A mixture model-based approach to the clustering of
microarray expression data. Bioinformatics, 18, 413–422.
McLachlan, G., Ng, S., & Bean, R. (2006). Robust cluster analysis via mixture models. Austrian
Journal of Statistics, 35, 157–174.
McLachlan, G., Peel, D., & Bean, R. (2003). Modelling high-dimensional data by mixtures of
factor analyzers. Computational Statistics and Data Analysis, 41, 379–388.
Ng, S., McLachlan, G., Wang, K., Ben-Tovim Jones, L., & Ng, S. (2006). A mixture model with
random-effects components for clustering correlated gene-expression profiles. Bioinformatics,
22, 1745–1752.
Olivier, M. (2005). The invader assay for SNP genotyping. Mutational Research, 573, 103–110.
Palhares, A., Rodrigues-Morais, T., Van Sluys, M. A., Domingues, D., Maccheroni, W., Jordão, H.,
et al. (2012). A novel linkage map of sugarcane with evidence for clustering of retrotransposon-
based markers. BMC Genetics, 13(51), 1–16.
Spellman, P., Sherlock, G., Zhang, M., Iyer, V., Anders, K., Eisen, M., et al. (1998). Comprehensive
identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray
hybridization. Molecular Biology of the Cell, 9, 3273–3297.
Tamayo, P., Slonim, D., Mesirov, J., Zhu, Q., Kitareewan, S., Dmitrovsky, E., et al. (1999).
Interpreting patterns of gene expression with self-organizing maps: Methods and application
to hematopoietic differentiation. Proceedings of the National Academy of Science USA, 96,
2907–2912.
Tseng, G., & Wong, W. (2005). Tight clustering: A resampling-based approach for identifying
stable and tight patterns in data. Biometrics, 61, 10–16.
Discrepancy and Choice of Reference Subclass
in Categorical Regression Models

Defen Peng and Gilbert MacKenzie

Abstract For categorical regression models we derive the optimal design alloca-
tion of observations to subclasses and provide a statistic, based on generalized
variance and its distribution for measuring the discrepancy between the optimal
allocation and the observed allocations occurring in observational studies in the
general linear model and extend our methods to generalized linear models. The
focus is on techniques which maximize the precision of the resulting estimators.
We explore the general form of optimal design matrix for general linear models with
categorical regressors, and propose an algorithm to find the optimal design matrix
for generalized linear models when the design matrix is of high dimension. We also
find that the proposed statistic can be used to show whether secondary criteria for
the choice of reference subclasses is required in parametric categorical regression
models. The development and use of the techniques and tools are illustrated by
means of simulation studies and the analysis of a set of lung cancer survival data.

Keywords Categorical variables • Design matrix parametrization • D-optimality

• GLMs • Reference subclass choice • Regression models

1 Introduction

An ongoing question concerns the choice of reference subclass when encoding

categorical variables in regression models. Other questions are focussed on the
categorical encoding of continuous variables (Pocock et al. 2004; Altman and
Royston 2006). The literature on the former issue is rather conflicting. For example,
William (2005) claimed that the choice of reference category may be made on

D. Peng ()
The Centre for Biostatistics, University of Limerick, Limerick, Ireland
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 159

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__12,
© Springer International Publishing Switzerland 2014
160 D. Peng and G. MacKenzie

the basis of: (a) subject matter considerations which may suggest the choice of a
particular category or, (b) the largest category should be chosen as it yields the
smallest standard errors. On the other hand, Berk (2008), argued that there is no
statistical justification for choosing one category over another. However, Frøslie
et al. (2010), in the hyperglycaemia and adverse pregnancy outcome (HAPO)
study, showed empirically that an unfortunate choice of reference category led to
less precise estimators and noted that the largest reference category gave narrower
confidence intervals.
The focus of this paper is on techniques which maximize the precision of the
resulting estimators in a D-optimal sense. We found that, when the sample allocation
is close to its ideal allocation in a D-optimal sense, the choice of a reference category
can be regarded as arbitrary—there being no need to use a secondary criterion. In
the case of a sample allocation being distant from its ideal allocation, the choice of
a reference category affects the precision of estimators. Accordingly, the use of a
secondary criterion is advantageous. For general linear models, we confirmed the
finding of Frøslie et al. (2010). However, in the other cases the results are more
complicated, see Peng and MacKenzie (2014).
The paper is organized as follows. In Sect. 2 we formulate the problem and
derive an optimum allocation strategy. In Sect. 3 we develop an index measuring
the “distance” any particular allocation is from the optimum. In Sect. 4 we extend
the methodology to GLMs. In Sect. 5, we collect up the simulations results on which
some of the earlier findings are based. In Sect. 6 we re-analyze population data on
lung cancer survival in Northern Ireland before concluding with a short discussion.

2 Model and Formulation

2.1 General Linear Model

To make matters concrete we consider the general linear model

Y D Xˇ C ;

where: Y is a continuous response variable, X is an np design matrix, ˇ is a p 1

column vector of regression parameters. We will also assume that i  N.0;  2 /
when required, for i D 1;    ; n, then E./ D 0 and E. 0 / D  2 In , where, In is a
n  n identity matrix. It follows immediately that

ˇO D .X 0 X /1 X 0 Y

and that

O D  2 .X 0 X /1
V .ˇ/
Discrepancy and Choice of Reference Subclass 161

which implies, under the Gaussian assumption, that the observed information
matrix is

I.ˇ/ D .X 0 X /= 2

when we assume  2 is known. It is also possible to consider other forms of the

response variable Y , for example a binary response, leading to binary regression
(Feldstein 1966).
If the design matrix X encodes a single categorical variable with p subclasses,
the resulting cross-product matrix, X 0 X , may take one of several, essentially
equivalent, patterned forms. Two main cases are, either,

X 0 X D diag.n1 ; n2 ;    ; np / (1)

or, on letting k D p  1

0 1
n n2    nk
n1
Bn1 0  0 C
n1
B C
B n2    0 C
X 0 X D Bn2 0 C: (2)
B: ::
:: :: C
@ :: :: :A
nk 0 0    nk

In (1) we have included exactly p D k C 1 binary indicator variables and in (2) we

have included an intercept term and exactly k binary indicator variables. Version (1)
does not involve a reference category, and omits the intercept term, whilst version
(2) involves treating one of the subclasses as a reference category. For example, if
x0i D 1 8 i is the intercept term, then for i 2 reference category W x1i D 0 \    \
xki D 0. Version (2) occurs most frequently and is the subject on this note.
With these arrangements the coefficients in the linear regression model have the
following interpretation

ˇ0 D E.Y jx0 D 1; x1 D 0;    ; xk D 0/;

ˇj D E.Y jx0 D 1; x1 D 0;    ; xj D 1; : : : ; xk D 0/
 E.Y jx0 D 1; x1 D 0;    ; xk D 0/ (3)

for j D 1;    ; k. In (3), the intercept, ˇ0 is the conditional expectation of Y in

the reference subclass and the usual regression coefficients, the ˇj s, of primary
scientific interest, are differences between two conditional expectations.
162 D. Peng and G. MacKenzie

2.2 Optimum Allocation

Now we consider the idea of optimal allocation from the design perspective of
D-optimality. Given n observations and p D k C 1 subclasses how should
the observations be allocated to subclasses in order to minimize the generalized
variance, GV.ˇ/, of the parameter ˇ? The generalized variance is defined as

GV .ˇ/ D detŒI 1 .ˇ/ D detŒ 2 .X 0 X /1 ; (4)

where, e.g., detŒA stands for the determinant of the matrix A. Minimizing (4) is
equivalent to maximizing detŒI.ˇ/ so that the D-optimal design may be found as
˚ 
Dp .n0 / D argmax detŒ.X 0 X /= 2 (5)
nj s

subject to (†j D1 nj D n/, where X 0 X is given by (2) and Dp .n0 / is the design
p

comprising a vector of optimal subclass numbers n0 D .n1 ; n2 ;    ; np /.

It is then easy to show that the optimal allocation is uniform. It follows that the
D-optimal design matrix is

0 1
n n1
n2    nk
Bn1 n10  0 C
B C
B  n2    0 C
.X 0 X  / D Bn2 0 C;
B: :::: :: C
@ :: : : :A
nk 0 0    nk

where nj D n=p, j D 1;    ; p are corresponding to the D-optimal design matrix

X  . Moreover, the minimum generalized variance is then from (4),

Y
p
GV  .ˇ/ D  2 = nj : (6)
j D1

O
It will be recalled that the D-optimal solution does not guarantee that the se.ˇ/
is uniformly minimal, rather it is equivalent to minimizing the volume of Gaussian
theory confidence regions for ˇ (Isham 1991).
This is a useful result when we can control the allocation, but in observational
studies this is not possible. Then, of course, we might choose to view this result
simply as a counterfactual experiment representing an ideal allocation and ponder
how distant any sample allocation is from this ideal.
Discrepancy and Choice of Reference Subclass 163

3 Measures of Discrepancy

As a measure of discrepancy from GV  .ˇ/ we consider the general index

 
GD .ˇ/ D loge GV .ˇ/=GV  .ˇ/
h i
D loge detŒI  .ˇ/=detŒI.ˇ/ ; (7)

where I  .ˇ/ is the observed information matrix, corresponding to the D-optimal

design matrix X  . The index GD ./ is a random variable with support on Œ0 ; 1/
and GD ./ D 0 implies that the sample allocation is optimal. We retain the unknown
ˇ in the notation to accommodate the GLM cases discussed later.
In order to explore how this index varies with the sample data we shall have
to vary .n1 ; n2 ;    ; np / systematically away from .n1 ; n2 ;    ; np / using positive
compositions. Whilst (7) is natural in context, a potential draw-back is that the
distribution of the index, over the finite set of positive compositions, is unknown,
although it can always be obtained by direct enumeration (below). For a single
categorical variable with p D k C 1 subclasses in the model, from (4) and (6),
 0 
GD .ˇ/ D loge  2 det.X  X  /= 2 det.X 0 X /
X
p
D loge .nj =nj /; (8)
j D1

a very simple form.

However, there are other, contending indices, one of which is the generalized
Chi-squared test statistic X 2 . For a single categorical variable with p D k C 1
subclasses in the model, the X 2 is the ordinary Chi-squared goodness-of-fit test of
H0 W n0 D n0 , versus the alternative hypothesis of discrepancy, H1 W n0 ¤ n0 ,

X
p
X D2
Œnj  nj 2 =nj
j D1

from which the null distribution of discrepancy is readily available, although the
adequacy of this approximation needs to be checked in the current context. When
H0 is rejected, we may choose to regard 2 , where  D p  1 D k, as a measure of
standardized squared distance that n0 is from n0 .

3.1 Compositions

First we consider some computational issues involving integers associated with this
problem. For p subclasses, the set of integer numbers, .n1 ; n2 ;    ; np /, falling in
164 D. Peng and G. MacKenzie

Table 1 Distributions of Index Null n p Non-Null

measures of discrepancy
GD ./
.˛0 ; 0 )
 500 D2
(0.5, 1)
>2
(˛1 ; 1 /
> 500 Œ2  9
(˛1 ; 1 /
X2 2 2;
p D k C 1,  D k and, .˛0 ; 0 / and .˛1 ; 1 / can be found
by simulation,
. ; / is the Gamma distribution and 2; is the
non-central Chi-squared distribution

the subclasses forms a positive composition (nj > 0; 8 j ) of the total sample
size n. For example, for n D 3 and p D 3 there are ten compositions;
.3; 0; 0/; .0; 3; 0/;    ; .1; 1; 1/, nine of these involve zero cells and one, the last,
the
 only positive
composition, does not. n and
 For general p there are exactly
nCp1 nCp1 n1
compositions of which  involve zeros and
n n p1
are rejected by algorithms generating positive compositions. Such algorithms are
required in order to study the variation in the indices proposed above. Moreover,
the set of positive compositions, when they exist, is symmetric around n0 , so
that in practice only half need be generated. Nijenhuis and Wilf (1978) give
Fortran algorithms which we translate into R software scripts (Appendix 1) and
use to generate positive compositions by rejecting all compositions involving
zeros. The second script generates random positive compositions and is useful in
simulation studies. As we need to consider the distribution of the indices on the
Null hypothesis of optimal allocation, i.e., their Testing Distributions, we generate
positive compositions using the multinominal function in R, which enables us to
control the cell probabilities.

3.2 Indices and Their Distributions for a Single Category

The results of a comprehensive simulation study (later) showed that the Null
distribution of the index, GD ./, follows a Gamma.˛0 ; 0 / distribution, where
.˛0 ; 0 / depend on n and p in a way described in the simulation section. In the
non-Null case, the distribution of the index is generally Gamma.˛1 ; 1 /, where the
parameter ˛1 and 1 are fractional and depends on n and p in a way described in the
simulation section. When n  500 and p D 2 we found that 2GD ./  2 , where
 D 1. See Table 1 for more details and the simulation section below.
A key finding from thePsimulation is that the Null distribution of the ordinary Chi-
squared statistic, X 2 D j D1 Œnj  nj 2 =nj is well described by a 2 distribution
p

where, as expected,  D p  1 D k degrees of freedom. And the non-Null

distribution is the corresponding non-central Chi-squared with parameters (p1; ),
where the non-centrality parameter, , is defined in the simulation section.
Discrepancy and Choice of Reference Subclass 165

The two indices are implicitly functionally related, viz

X
p
nj X
p
 1 @X 2
GD ./ D loge . /D loge 1 C ; (9)
j D1
nj j D1
2 @nj

but are not easily separable. However, this equation shows that the two indices mea-
sure the same underlying discrepancy. We may also conclude that the distribution
of the index GD ./ is inherently more complicated than its Chi-squared competitor.
The values of the ˛1 and 1 parameters, which depend on n and p, must be known
in advance in order to use the index; for example, as given in this paper (later).
Secondly, the extension to multiple categorical variables requires us to formulate
and compute the ideal allocation, i.e., to find n0 , for the general case.

3.3 General Expression for the Ideal Allocation

0
In order to pursue the comparison of the indices, we must find .X  X  / representing
the ideal allocation when there is more than one categorical variable in the model.
Accordingly, suppose there are m categorical variables in the model X1 ;    ; Xm ,
such that each variable has p` subclasses, ` D 1; 2;    ; m. Then, it can be shown
that the optimal design matrix takes the general form
!
0 1 v
.X X  / D n 0 p
; (10)
vp M

P
where vp D . p11  10p1 ; p12  10p2 ;    ; p1m  10pm /, a vector with dimension . ` p`  m/,
and 1p` are .p`  1/ vectors with element 1, ` D 1; 2;    ; m. M is a symmetric
matrix with diagonal elements p1` I` , I` is .p`  1/  .p`  1/ identity matrix, ` D
1; 2;    ; m, while the elements in the upper triangle are pi1pj 1ij , where 1ij is .pi 
1/  .pj  1/ matrix with element 1, i D 1; 2;    ; m  1; j > i .
This general pattern was derived by considering lower order cases and using
multiple Lagrange multipliers to deal with the defining constraints and recognizing
that, whatever the number of categorical variables included, finally, only one
substitution is permitted. This finding was also confirmed for higher order cases
by simulation which showed that, as the sample allocation departs from (10), the
value of det[ .X 0 X /] becomes smaller.
We present two simple examples of the pattern: .X1 ; X2 / with .2; 2/ subclasses
and three categorical variables .X1 ; X2 ; X3 / with .2; 3; 4/ subclasses, respectively
to illustrate the construction. The ideal allocation design matrices are then
166 D. Peng and G. MacKenzie

0 1 0 1 0 1
n n1 n2 n n
2
n
2
1 1
2
1
2
0 @ A @ A D n@ 1 A;
.X X  /.2;2/   
D n1 n1 n12 D n2 n
2
n
4 2
1
2
1
4
n2 n12 n2 n
2
n
4
n
2
1
2
1
4
1
2

where n1 is associated with X1 , n2 is associated with X2 and n12 is the cross term
between X1 and X2 , †x1i
 
x2i D n=.p1 p2 / D n=4 (Appendix 2) and similarly
0 1
1 1=2 1=3 1=3 1=4 1=4 1=4
B 1=2 1=2 1=6 1=6 1=8 1=8 1=8 C
B C
B 1=3 1=6 1=3 0 1=12 1=12 1=12 C
B C
0 B C
.X  X  /.2;3;4/ D n B 1=3 1=6 0 1=3 1=12 1=12 1=12 C
B C
B 1=4 1=8 1=12 1=12 1=4 0 0 C
B C
@ 1=4 1=8 1=12 1=12 0 1=4 0 A
1=4 1=8 1=12 1=12 0 0 1=4

from which we note that within each categorical variable: (a) the ideal allocation is
inversely proportional to the corresponding number of subclasses and (b) the cross
terms are zero. Finally, the cross terms between any two categorical variables are
inversely proportional to the product of the numbers of subclasses involved. Thus,
these findings allow us to generalize the index, GD .ˇ/ to models with multiple
categorical variables.
The availability of a general expression for the ideal allocation for the index
allows us to compute the value of the corresponding X 2 statistic, compare the Null
distributions and compute the power functions for the two indices in the simulation
section. Meanwhile, we turn now to investigate reference subclass choice.

4 Generalized Linear Models

We consider the canonical generalized linear model (Cox andPSnell 1989) for
independent responses Yi with E.Yi / D i D g.i /, where i D kuD0 xui ˇu is the
linear predictor, g./ is the link function, for i D 1;    ; n independent observations
and u D 0; P   ; k, representing the p unknown regression parameters. The log
likelihood is i .si i  K.i //, where si is a function of Yi . Then the observed
information matrix for ˇ is

I.ˇ/ D .Oˇ  T /.O O K/.Oˇ  T /T D .Oˇ T /.O O K/1 .Oˇ T /T : (11)

Discrepancy and Choice of Reference Subclass 167

Table 2 Canonical link functions and the form of wi

Distribution Density(mass) function Link function wi
.y/2
Normal f .yI ;  / D p2
1
e 2 2 xˇ D  D   2
Exponential f .yI / D e  y
xˇ D  1
D .xi0 ˇ/2
 .y/2
.xi0 ˇ/3=2
1
IG f .yI ; / D . 2y

3/ e
2 22 y xˇ D 2 D 
4
Poisson y 
f .yI / D yŠ e xˇ D log./ D  exp.xi0 ˇ/

ni ni exp.xi0 ˇ/
p y .1  p/ni y

Binomial f .yI ni ; p/ D xˇ D log. .1 / D  .1Cexp.xi0 ˇ//2
y

Geometric f .yI p/ D .1  p/y1 p xˇ D log. .1 /D 1
1Cexp.xi0 ˇ/

Cox and Snell (1989) commented that this completes the close formal and
conceptual connection between maximum likelihood estimation and generalized or
weighted least squares estimation (WLS).
When the ˇ0 is the intercept, (11) can be expressed as the .p  p/ matrix
 P P 0
w x w
I.ˇ0 ; ˇc / D .X 0 W X / D P i i P i ci 0i ; (12)
i xci wi i xci xci wi

where: p D k C 1 and we have partitioned xi0 D .x0i D 1; xci 0

/, where xci0
D
0
.x1i ;    ; xki / represents the k binary indicator variables and ˇc D .ˇ1 ;    ; ˇk /
their effects. Moreover, W D O O K is a diagonal matrix with the elements
wi D h.xi0 ˇ/. In Table 2 we list some special cases from this family with their
corresponding link functions and structural weights, wi , derived in the next section
below. These quantities play a key role in the sequel.

4.1 Covariance Matrix: One Categorical Variable

Suppose we have only one categorical covariate with exactly p D k C 1 categories

with p parameters .ˇ0 ; ˇ1 ;    ; ˇk /, the observed information matrix (12) becomes
0P P P P 1
i wi iŒ1 wi iŒ2 wi    iŒk wi
BP P C
B iŒ1 wi iŒ1 wi 0  0 C
BP P C
B 0 C
I.ˇ0 ; ˇc / D B iŒ2 wi 0 iŒ2 wi    C;
B : :: :: :: C
B : C
@P : : :
P
: A
iŒk wi 0 0    iŒk wi
168 D. Peng and G. MacKenzie

taking the general form

0 P 1
nr h.ˇ0 /C j nj h.ˇ0 C ˇj / n1 h.ˇ0 C ˇ1 / n2 h.ˇ0 C ˇ2 /    nk h.ˇ0 Cˇk /
B C
B n1 h.ˇ0 C ˇ1 / n1 h.ˇ0 Cˇ1 / 0  0 C
B C
DB
B
n2 h.ˇ0 C ˇ2 / 0 n2 h.ˇ0 Cˇ2 /  0 C
C
B :: :: :: :: C
@ : : : : A
nk h.ˇ0 C ˇk / 0 0    nk h.ˇ0 C ˇk /
(13)

and the inverse of (13) is

0 1
1 1 1  1
B1 1 C .q1  nr
/ 1  1 C
B n1 C
1 B C
I 1 .ˇ0 ; ˇ/ D B1 1 1 C .q2  n2   
nr
/ 1 C;
nr h.ˇ0 / B
B :: :: :: ::
C
C
@ : : : : A
1 1 1    1 C .qk  nr
nk /
(14)

where i Œj  means subject i 2 j th category, whence xij D 1 for i 2 j th category,

and qj D h.ˇ0 /= h.ˇ0 C ˇj /, nr and nj are allocated numbers in the reference
category and the other categories respectively, j D 1; 2;    ; k. This matrix is an
obvious generalization of the inverse arising in the general linear model.

4.2 Optimal Allocation and GD ./

Now, in the GLM family, with one categorical variable and using multiple Lagrange
constraints we can show that the optimal allocation is still uniform, namely: n1 D
n2 D    D nkC1 D n=.k C 1/, and does not depend on ˇ0 or on any of the other
ˇs. Moreover, for GLMs involving a single categorical variable, from (7) and (13)
the general index is
 
GD .ˇ0 ; ˇc / D loge GV .ˇ0 ; ˇc /=GV  .ˇ0 ; ˇc /
 
D loge det.I  .ˇ0 ; ˇc //=det.I.ˇ0 ; ˇc //
X
p
D loge .nj =nj /
j D1

which is identical to (8). Under this condition, the distributional properties of GD ./
for all the GLMs listed in Table 2 are exactly those presented in Table 1.
Discrepancy and Choice of Reference Subclass 169

4.3 Two or More Categorical Covariates

When we have more than one categorical covariate in the GLMs, the ideal allocation
is no longer uniform. In this case the ideal allocations depend on the unknown
parameters and the problem of their construction is therefore necessarily more
complicated than that for classical linear models or for GLMs with one covariate.
The construction of optimal designs in GLMs is an active area of current research.
Our approach is to fix ˇ D ˇO and use numerical methods to obtain the optimum
allocation of the nj s. Thus, similar to (5), we obtain the optimal design as
˚ 
Dp .n0 / D argmax detŒ.X 0 W X / ;
nj s

P
where p  D p` ; ` D 1; 2;    ; m. Having found the optimal allocation we may
proceed to compute the general index, GD ./, given in (7).
For the canonical GLMs listed in Table 2 with some fixed parameters, we found
that both the testing distributions and the non-Null distributions of GD ./ are still
well described by Gamma.˛; /. See more details in the simulation section.

4.4 Indices and Choice of Reference Subclass

From the study above we know that, the distribution of the GD ./ or X 2 corre-
sponding to a particular null and alternative hypothesis can be explicitly determined
by numerical methods, then they can directly be used to form decision regions (to
accept/reject the null hypothesis). When we have a sample allocation, if the test
results on the indices show that this sample allocation is far away from its ideal
allocation significantly, then a poor choice of reference category may lead to a loss
of efficiency of the regression parameter, otherwise, choice of reference category
may be arbitrary, in a D-optimal sense. We present more details in the simulation
section.

5 Simulation Studies

5.1 Distribution of the Index

We investigated the distribution of the index by conducting a detailed simula-

tion study covering the following scenario space. For one categorical variable
in the linear regression model, we consider the number of subclasses: p D
2; 3; 4; 5; 6; 7; 8; 9, and n D 50; 100; 200; 500; 1;000. In practice, greater values
of p are unusual. When p and n are small, we use the exact distribution which
170 D. Peng and G. MacKenzie

1.2
p=2 p=3

0.0 0.2 0.4 0.6

Density

Density
0.8
0.4
0.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 0 1 2 3 4 5 6

G D(β ) G D (β)

p=4 p=5

0.30
0.4
Density

Density

0.15
0.2

0.00
0.0

0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

G D(β ) G D (β)

p=6 p=7
0.00 0.10 0.20 0.30

0.30
Density

Density

0.15
0.00

0 2 4 6 8 0 2 4 6 8 10

G D(β ) G D (β)

p=8 p=9
0.20

0.20
Density

Density
0.10

0.10
0.00

0.00

0 2 4 6 8 0 2 4 6 8 10

G D(β ) G D (β)

Fig. 1 Histogram of GD .ˇ/ index with different numbers of subclasses p for (n=100); Gamma
distribution shown as smooth curve

is enumerable and take the number of replications, to be the exact number

of positive compositions obtained (Appendix 1). For larger p and n, we use
1,000 replicates in the simulation algorithm to generate the positive compositions
randomly (Appendix 1).
Figure 1 shows just eight scenarios from our simulation with a single categorical
regressor in the model confirming that the histograms of the index GD ./ in the
non-Null case change with different p. Here n D 100, which we consider to
be a relatively small sample size. For larger sample sizes the fit improves. The
smooth curve in the histograms is the density curve of Gamma(˛; O /
O (density
x ˛1 x=
f .x/ D .˛/˛ e for x  0 and ˛;  > 0), where O D Var.GD .//=mean.GD .//
and ˛O D mean.GD .//=O are the estimators from the Gamma distribution density
with the simulated GD ./. The results in Fig. 1 show that the distribution of GD ./ in
the non-Null case is well described by a Gamma(˛1 ; 1 / when p > 2, and 21 when
Discrepancy and Choice of Reference Subclass 171

p D 2. The other two histograms (not shown) indicate that the distribution of GD ./
in the Null case is well described by a Gamma(˛0 ; 0 / and the distribution of X 2 in
the Null case is well described by a 2.p1/ random variable.

5.2 Testing Goodness of Fit

We carried out some statistical hypothesis tests for each scenario: the null and
alternative hypotheses were: H0 : the data follow the specified Gamma or Chi-
squared distribution, and H1 : the negation of H0 . We adopted the Chi-squared
goodness-of-fit test, Anderson–Darling test of fit and Kolmogorov–Smirnov test. For
the Chi-squared goodness-of-fit test, we created categories such that the minimum
expected cell frequency was attained. For Anderson–Darling test, we computed
the p-value from the simulated data because there are no tabulated critical values
available for the Gamma distribution for this test. Moreover, since the parameters
are estimated from the sample data, we tested whether the cumulative distribution
function follows the standard uniform distribution (Shapiro 1980).

5.3 Results: Single Categorical Variable

All three tests, the Chi-squared goodness-of-fit test, Anderson–Darling test and the
Kolmogorov–Smirnov goodness of fit test we have conducted are in agreement
on our finding shown in Sect. 5.1. It appears that the Gamma(˛1 ; 1 ) distribution
provides a reasonable fit to the proposed index GD ./ in the non-Null case. And
when p D 2; n  500, it is distributed as 21 .
To find the testing distribution (Null distribution) of GD ./ and X 2 , the proce-
dures are: (a) generating the sample compositions from multi-nominal distribution
by setting the cell probabilities equal to the proportions of the subclass obtained
from the ideal allocation (here, uniform), (b) calculating the corresponding index
GD ./ and X 2 by using the formulae given above, (c) using the test methods
described in Sect. 5.2 to test whether GD ./ and X 2 follow some distributions.
The three tests also showed that both testing distributions, i.e., the proposed index
GD ./ and X 2 in the Null case are well described by a Gamma(˛0 ; 0 ) and 2.p1/
respectively. We have summarized our findings in Table 1.
In order to estimate .˛1 ; 1 / and .˛0 ; 0 /, we carried out a multiple regression
analysis with (Y D ˛; O or; Y D ; O X1 D n; and X2 D p) based on three blocks
of the simulated data, 120 values for Null distribution, while 108 values by deleting
p D 2; n D 50; 100; 200; 500 for non-Null distribution. The results are presented in
Table 3.
Thus, for fixed p and n, we can find or estimate ˛ and  in the corresponding
Gamma distribution with Tables 1 and 3.
172 D. Peng and G. MacKenzie

Table 3 Formulae for estimating parameters on the distributions of GD ./

YO Formulae R2
Non-NULL:
˛O 1 D 0:9076p  0:1569 log.n/  0:0372p log.n/ 0.9966
O 1 = 0:0237p C 0:1752 log.n/ 0.9947
NULL:
˛O 0 D 0:2741p
  0:0565 log.n/ C 0:0294p log.n/ 0.9929
O 0 = exp 1:1572 log.p/  1:0262 log.n/ 0.9978

R2 in Table 3 is the multiple correlation coefficient. We noted that ˛O 0 and O 0

are very close to those estimates obtained in the simulation with the corresponding
p and n, and using them to check the significant level 5 %, the testing results are
similar to those obtained by using 2.p1/ . Further, we conducted the power of
analysis by using the distributions of GD ./ and 2.p1/ respectively. On the Null
hypothesis, in the Multi-nominal with p categories, H0 W 1 D 2 D    D
p D o , while on H1 : at least i ¤ j , for some i ¤ j; 1  i; j  p where
P p
i D1 i D 1.
For the X 2 index we calculated the exact power at 100 different values of
the effect size, 4, by using the non-central 2.p1; / distribution, where D
Pp p
i D1 .ni  n0 / =n0 (Cohen 1988) and accordingly, 4 D
2
=n (cran.r project
2009). For the analysis involving GD ./, we estimated the nominal significance
level (4 D 0) and power by simulation, repeating m =1000 statistical tests of
the null hypothesis at each value of the effect size in H1 . Recall that under H0 ,
GD ./  Gamma.˛0 ; 0 ) where .˛0 ; 0 / are estimated by using the formulae listed
in Table 3. Figure 2 shows the power functions for GD ./ and X 2 , against effect size.
They have very similar behaviour, showing that the two indices are equivalent.
Thus, for a single categorical variable in the LM, one can use the formulae listed
in Table 3 to obtain the parameters for testing the Gamma distribution and then test
the discrepancy.

5.4 Results: Two or More Categorical Variables

When we have m > 1 categorical variables, each one has p` ; ` D 1; 2;    ; m

subclasses, to be regressed. In our simulation scenarios we considered p` ; ` D
1; 2;    ; m running from 2 to 9, n D .50; 100; 200; 500; 1;000/, and replicating
1,000 times. To find the distribution of the index GD ./, we used the general
findings on the ideal allocation (10) in Sect. 3.3 and sampled from the Multinominal
distribution with the cell probabilities which guaranteed the exact cross terms in
.X 0 X /. We set the cell probabilities equal to the proportions of subclasses obtained
from the sample allocation (Appendix 1) to find the non-Null distribution of GD ./,
and equal to the proportions of subclasses obtained from the ideal allocation to
Discrepancy and Choice of Reference Subclass 173

1.0 1.0

0.8 0.8

0.6 0.6
power

power
0.4 0.4

0.2 0.2

0.1 0.2 0.3 0.4 0.5 0.6 0.1 0.2 0.3 0.4 0.5 0.6
esize esize

Fig. 2 Power as a function of effect size; n D 100; p D 3, esize=4. Left panel: power of GD ./;
right panel: power of X 2

find the Null distribution of GD ./ respectively. The procedures are similar to those
described in Sect. 5.3.
When m D 2, the Kolmogorov–Smirnov goodness of fit test showed that the
distribution of the index GD ./ in the non-Null case was well described by a
Gamma(˛; ) distribution except for the cases in n D 50; p1  4; p2  4.
Similarly, (˛; ) are the function of n; p1 ; p2 approximately, and for some special
cases,  D 1. The distribution of the index GD ./ in the null case was well described
by a Gamma(˛; ) distribution, again, (˛; ) are functions of n; p1 ; p2 approxi-
mately (not shown). The simulation procedure and the corresponding findings can
be extended to m P 3 categorical variables.
p Cp
That X 2 D j 1D1 2 Œnj  nj 2 =nj  2 , with  D Œ.p1  1/ C .p2  1/ in
the null case with m D 2 is not surprising. Since m D 2, the nj s are the marginal
frequencies from the corresponding two-dimensional contingency table. The finding
holds when varying the second categorical variable in the range Œ2; 9. We extended
the procedure to m D 3, and found that X 2  2 , with  D Œ.p1  1/ C .p2  1/ C
.p3  1/, when we varied the second and the third categorical variable in the range
Œ2; 9. Accordingly, suppose there are m categorical variables in model X1 ;    ; Xm ,
such that each variable has p` subclasses, ` D 1; 2;    ; m, then, for the generalized
174 D. Peng and G. MacKenzie

Table 4 Estimated testing distribution of GD ./ 

.˛; / for two different scenarios in various
GLMs
Scenario 1a Scenario 2b
GLM Dist. GD ./ GLM Dist. GD ./
Poisson
.1:7247; 0:0315/ Poisson
.5:3968; 0:0332/
Exponential
.4:5642; 0:0347/ Exponential
.23:892; 0:0349/
Binomial
.1:4750; 0:0206/ Binomial
.2:5490; 0:0239/
Geometric
.1:5251; 0:0255/ Geometric
.2:6363; 0:0327/
IG( =1)
.2:6408; 0:0348/ IG( =1)
.12:655; 0:0350/
a
Scenario 1=(ˇ0 D 0:5; ˇ1 D 0:3; ˇ2 D 0:3; n D 100)
b
Scenario 2=(ˇ0 D 0:5; ˇ1 D 0:3; ˇ2 D 0:3; ˇ3 D 0:5; n D 100)

P Pp`  2 
X2 D m j D1 .nj `  nj ` / =nj ` we conjecture is distributed as  , where  D
2
Pm `D1
. `D1 p`  m/. For example, when m D 4, and .p1 ; p2 ; p3 ; p4 / D .2; 3; 4; 5/, then
 D 10, etc.
Thus, for two or more categorical variables in LM, one can find the testing
distribution by using the general ideal allocation form (10) and similar procedures
described in Sect. 5.3, or alternatively, one can use the Chi-squared distribution
found above to test the discrepancy.

5.5 Results: Two or More Categorical Variables in GLMs

As mentioned earlier the determinant of the observed information matrix .X 0 W X /

depends on the unknown parameters. Accordingly, in order to find the distribution of
GD ./ we conducted a small simulation study, fixing the values of ˇ in the scenario
space.
Suppose we have two binary covariates X1 and X2 in the GLM models listed in
Table 2 and let the scenario space be: ˇ1 2 Œ0:5; 0:5, ˇ2 2 Œ0:5; 0:5, ˇ0 D
0:5; n D 100, and replicating 1,000 times. We used R function “optim” to find the
ideal allocation within each scenario.
Table 4 lists the results corresponding to two scenarios (complete results not
shown). The main findings are that the approximate testing distribution of GD ./ is
well described by a Gamma.˛; / distribution. In Scenario 2 the first categorical
variate has three subclasses. The approximate distribution of GD ./ under non-Null
is also well described by a Gamma.˛; /, but the details are omitted.
We just showed that for two or more categorical variables in a GLM for lower
dimensional cases of .X 0 W X / one can use the R function “optim” to find the ideal
allocation and a similar procedure to that described in Sect. 5.3 to find the testing
distribution. However, for higher dimensional cases of .X 0 W X /, for example when
p > 8, we used the following Monte Carlo algorithm.
Discrepancy and Choice of Reference Subclass 175

We suppose there are m categorical variables in model X1 ;    ; Xm , such that

each variable has p` subclasses, ` D 1; 2;    ; m, our algorithm to find an ideal
allocation is,
1. Initialize allocation nQ D .nQ 1 ;    ; nQ ` ;    ; nQ m /; nQ ` D .n`;1 ;    ; n`;p` /, ` D
1; 2;    ; m.
2. Use the proportions rQn` D .n`;1 =n;    ; n`;p` =n/ as cell probabilities for the
multinominal distribution, to generate sets of binary variables X` , for ` D
1; 2;    ; m., thus building the design matrix X .
3. For the fixed (given) ˇ, compute detŒI.ˇ/ D det.X 0 W X /.
4. Find n` D .n`;1 ;    ; n`;p` / maximizing detŒI.ˇ/ D det.X 0 W X /.
5. Set nQ ` D n` , and rQn` D .n`;1 =n;    ; n`;p` =n/.
6. Repeat steps (2)–(5) until convergence.
The initial allocation can be sample compositions generated by using the code in
Sect. 7 in Appendix 1 for each categorical variable, however, by taking the uniform
allocation as the starting allocation, the search is faster. We cross-checked the
results of the algorithm with the R function “optim” for lower dimensional case
of .X 0 W X /, and found that the methods always agreed for GLMs fitted.
In addition, the distribution of X 2 , utilizing the ideal allocation numbers as
the expected values, still followed a 2 distribution as in Sect. 5.4. However, we
note that in these cases the central Chi-squared distribution approach for testing
discrepancy is not suitable, see the illustrative example in Sect. 5.6 below for the
explanation.

5.6 Results: Illustrative Example—Drawback of X 2

in Complicated GLMs

In this example, we choose three compositions from the simulated compositions

(using algorithm in Sect. 7 in Appendix 1). The example illustrates, inter alia, the
use of the GD ./ and X 2 and the testing results based on Poisson regression model
with two binary covariates with n D 100 and two settings, ˇ D .0:5; 0:3; 0:3/ and
ˇ D .0:5; 0:3; 1:5/ assumed, respectively. The results are presented in Table 5.
The approximate ideal allocations and the corresponding testing distributions are
obtained by using the similar procedure described in Sect. 5.5.
In the Poisson GLM setting with two binary covariates the simple link between
X 2 and GD ./ as (9) is lost. For example, X 2 rejects the first composition chosen,
but GD ./ does not. Also in the lower panel of Table 5 where the value of ˇ2 has been
changed from 0.3 to 1.5, increasing the weight of the second binary variable, we note
that the values of X 2 and GD ./ no longer rank the second and third compositions
similarly. For example, the index ranks the second composition closer to the ideal
allocation, while X 2 ranks the third composition closer to the ideal allocation. Recall
that the ˇ information is incorporated in the determinants defining the index, while
176 D. Peng and G. MacKenzie

Table 5 An example showing the GD ./, and X 2 statistics, based on simulated Poisson GLM data
with two binary covariatesa
Critical values
Model-Composition.X1 I X2 I X12 / GD ./ X2 GD ./ X2
ˇ D .0:5; 0:3; 0:3/ - 0.1354 5.9914
Ideal-(45.544, 54.456; 54.456, 45.544; 25.984)
(68, 32; 56, 44; 15) 0.1115 20.189
(30, 70; 23, 77; 14) 0.9426 49.057
(4, 96; 48, 52; 35) 1.9317 73.786
ˇ D .0:5; 0:3; 1:5/ - 0.2733 5.9914
Ideal-(38.479, 61.521; 64.076, 35.924; 32.364)
(68, 32; 56, 44; 15) 0.2402 39.799
(30, 70; 23, 77; 14) 1.5213 79.476
(4, 96; 48, 52; 35) 2.1485 64.943
a
With ˇ D .0:5; 0:3; 0:3/ and under H0 , GD ./  Gamma.1:7247; 0:0315/ from Table 4, and
with ˇ D .0:5; 0:3; 1:5/, GD ./  Gamma.5:3663; 0:0283/ from simulation, X 2  2 ./;  D
.2  1/ C .2  1/ D 2

X 2 uses the ideal allocation numbers generated by maximizing the determinant of

the information matrix which appears in the index.
We conjecture that this lack of concordance in ranking persists for two more
categorical variables. Accordingly, we prefer to rely on the index GD ./ as a measure
of discrepancy for the complicated GLMs.

5.7 Results: Illustrative Example—GD ./ and the Choice of

Reference Subclass

As an illustrative example, we choose three compositions and the corresponding

values of GD ./ from the simulated results when n D 100 and one single categorical
variable with p D 4. The example illustrates, inter alia, the use of the index GD ./
and the choice of reference subclass in the linear and Poisson regression model. The
results are presented in Table 6.
From the foregoing analytical work, that the index, GD ./, takes the same values
on the compositions in the LM and GLM Poisson regression model, and X 2  23
can be used too. With n D 100; p D 4, using the formulae listed in Table 3,
we have ˛O 0 D 1:3778 and O 0 D 0:0441, i.e., the testing distribution of GD ./ is
Gamma.1:3778; 0:0441/ approximately. Comparing with the critical values 0:1629
and 7:8147 at 5 % significant level obtained from the two testing distributions, i.e.,
Gamma.1:3778; 0:0441/ and 23 , with n D 100; p D 4, respectively, only the first
allocation, which is close to ideal allocation, is not rejected.
To show the relationship between the indices and the choice of reference
subclass, we used Vr , called total variance of the estimators as a measure of precision
Discrepancy and Choice of Reference Subclass 177

Table 6 An example showing the GD ./ statistic and the choice of reference category, based on
simulated Poisson GLM data with two binary covariates
GD ./ Switch Linear model GLM(Poisson)
Composition (X 2 / rmax a ! r Vr Vr
(18, 33, 26, 23) 0.0950 2!2 0.2587 0.1482
(4.5446) 2!3 0.2832 0.1920
2!4 0.2982 0.1582
2!1 0.3345 0.1891
(22, 48, 5, 25) 1.0849 2!2 0.3688 0.2896
(36.941) 2!4 0.4263 0.3112
2!1 0.4427 0.3202
2!3 0.9063 0.8535
(2, 3, 34, 61) 3.4465 4!4 0.9283 0.4254
(95.396) 4!3 0.9674 0.4546
4!2 1.8791 0.7760
4!1 2.3791 1.1510
a
rmax represents that the reference is the most numerous group

P O
of the estimators, where Vr D j ŒVar.ˇj  and the subscript r represents the
corresponding reference. In LM, we calculated Vr by setting h.ˇ0 / D 1 and
h.ˇj / D 1 in form (14) where k D 3 and ignore  2 . In the Poisson GLM case,
we assumed ˇ0 D 0:5; ˇ1 D 0:1; ˇ2 D 0:2; ˇ3 D 0:3, and simulated response
variable y from Poisson distribution with rate exp.Xˇ/, where X is the design
matrix which is determined by the three chosen sample compositions. With y and X
we fitted the Poisson GLM under different choice of reference subclass, accordingly,
the estimated Vr were obtained, see Table 6. The results show that, when the
allocation is close to the ideal, the Vr are very close with each other for each possible
reference chosen. In this case, a choice of reference subclass can be arbitrary for
both LM and Poisson GLM considered. However, as the allocation becomes further
away from its ideal allocation, the corresponding total variance, Vr changes when
switching from the largest subclass to the smallest subclass, and hence the penalty
on the precision of the estimators increases. In these two cases, a poor choice of
reference category may lead to a loss of efficiency of the regression parameter, and
choosing the reference category with the largest number of observations for each
categorical variable is the optimal strategy.

6 Application

A prospective epidemiological study designed to measure the annual incidence

of lung cancer in Northern Ireland was carried out between October 1st, 1991,
and September 30th, 1992 (Wilkinson 1995; MacKenzie 1996). During this 1-year
period, 900 incident cases were diagnosed and followed up for nearly 2 years.
Comprehensive clinical including information on treatment and prognosis were
178 D. Peng and G. MacKenzie

abstracted from the hospital and the general practitioner’s records. Despite extensive
enquires, the outcome could not be determined in 25 cases, and in another 20 cases
the diagnosis was made at the post mortem. In total there were 855 (95 %) cases on
whom survival information was complete and these cases are analyzed here.
The following nine factors were selected for study in order to determine their
influence, if any, upon survival: (a) Patient Characteristics; age, sex, smoking
status, (b) Disease Status/Markers; WHO performance status, cell type, metastases,
sodium and albumen levels and (c) Treatment Details; surgery, radiotherapy,
chemotherapy, or palliative treatment. Further details of the categorical factors
studied may be found in Wilkinson (1995) and in Table 7.
We analyze survival time T , the time from diagnosis to death from lung cancer
or censoring, using Cox’s proportional hazards regression model (Cox 1972), in
order to identify the independent effect of factors studied simultaneously. The 5 %
level of statistical significance was used in performing statistical tests and in the
construction of confidence intervals (CIs).
To estimate the model we first choose the natural reference category for each
factor, construct the corresponding design matrix and apply a stepwise backward
algorithm to the nine factors studied in order to estimate the reduced model. The
natural reference subclass for a categorical is usually the least or most hazardous.
Later, we switch the reference subclasses to the most numerous subclasses and re-
estimate the model parameters, their standard errors and corresponding p values.
Of the nine factors studied, only seven were found to independently influence
survival. In particular, survival did not depend on the gender of the patient, nor,
rather surprisingly, on the age of the patient. The results for the reduced model,
using the natural subclasses are shown in the left panel of Table 7. In the right hand
side panel we present the effect of switching the reference to the most numerous
subclass. The numbers of patients in each subclass are shown in brackets.
For these data, .p1 ; p2 ; p3 ; p4 ; p5 ; p6 ; p7 / D .5; 5; 4; 3; 3; 3; 4/, and choosing the
ˇO based on the model with most numerous reference subclasses, we used the Monte
Carlo algorithm given in Sect. 5.5 with 10,000 replications to maximize det[I.ˇ/]
associated with Cox’s proportional hazard model. We found that the approximate
ideal allocation (cross terms omitted) to be:

n D .194; 183; 138; 207; 133I 189; 198; 155; 77; 236I 180; 219; 230; 226I
269; 260; 326I 294; 248; 313I 325; 247; 283I 233; 208; 220; 194/:

Under H0 , using the similar procedure described in Sect. 5.5, the finding testing
distribution GD ./  Gamma(8.3699, 0.0697) approximately, which has a critical
value 0.9503 associated with a 5 % significant level. The calculated GD ./ for
these data corresponding to the ideal allocation is 11:2443 >> 0:9503. Hence,
the allocation of these data is significantly far away from the ideal allocation. In
this case, the choice of reference category is not arbitrary and choosing the most
numerous subclasses as references, reveals the ˇO has the smallest variances (see
Table 7).
Discrepancy and Choice of Reference Subclass 179

Table 7 Results of PH regression analysis: backwards solution involving seven of nine factors:
natural and most numerous reference subclasses
Natural reference Largest subclass as reference
Model (n) ˇO O
se.ˇ/ p-value Model ˇO O
se.ˇ/ p-value
1. Treatment
Pall. (441) – – – Pall. (441) – – –
Surg. (79) 1.209 0.249 < 0:01 Surg. (79) 1.209 0.249 < 0:01
Chemo. (45) 0.487 0.209 0.020 Chemo. (45) 0.487 0.209 0.020
Radio. (256) 0.296 0.100 0.003 Radio (256) 0.296 0.100 0.003
C + R. (34) 0.908 0.237 < 0:01 C + R. (34) 0.908 0.237 < 0:01
2. WHO
Normal (78) – – – No Work (286) – – –
Light work (278) 0.091 0.184 0.619 Light work (278) 0.443 0.102 < 0:01
No work (286) 0.534 0.185 0.004 Normal (78) 0.534 0.185 0.004
Walking (191) 1.014 0.197 < 0:01 Walking (191) 0.480 0.104 < 0:01
Bed/Chair (22) 1.680 0.285 < 0:01 Bed/Chair (22) 1.146 0.232 < 0:01
3. Cell
Squamous (247) – – Other (379) – – –
Small (121) 0.729 0.154 < 0:01 Small (121) 0.518 0.141 < 0:01
Adeno ca (108) 0.313 0.141 0.026 Adeno ca (108) 0.102 0.130 0.436
Other (379) 0.211 0.102 0.039 Squamous (247) 0.211 0.102 0.039
4. Sod. mmol/l
 136 (505) – – –  136 (505) – – –
< 136 (310) 0.327 0.085 < 0:01 < 136 (310) 0.327 0.085 < 0:01
Missing (40) 0.093 0.219 0.672 Missing (40) 0.093 0.219 0.672
5. Alb. g/l
35 (458) – – –  35 (458) – – –
< 35 (315) 0.421 0.091 < 0:01 < 35 (315) 0.421 0.091 < 0:01
Missing (82) 0.460 0.165 0.005 Missing (82) 0.460 0.165 0.005
6. Metastases
No Met. (188) – – – Met (428) – – –
Met (428) 0.771 0.120 < 0:01 No met (188) 0.771 0.120 < 0:01
Missing (239) 0.351 0.132 0.008 Missing (239) 0.421 0.096 < 0:01
7. Smoking
Non-smoker (88) – – – Current (416) – – –
Current (416) 0.384 0.139 0.006 Non-smoker (88) 0.384 0.139 0.006
Ex-smoker (330) 0.252 0.142 0.076 Ex-smoker (330) 0.132 0.085 0.119
Missing (21) 0.298 0.273 0.275 Missing (21) 0.086 0.253 0.734

We also noted that switching the reference subclass in the WHO performance
status variable affected the survival time significantly, however, a similar switch in
Cell Type did not. Overall, switching the reference subclasses, increases precision
and hence yields narrower confidence intervals compared with retaining natural
reference subclasses. Our findings, show clearly the magnitude of the penalty in
precision associated with choosing the so-called “natural” reference subclasses.
180 D. Peng and G. MacKenzie

Clearly, in such an observational study one has some freedom to choose the ref-
erence subclasses and one general approach may be to select the largest subclasses,
hence minimizing the variance of the estimators.

7 Discussion

In this paper we have tried to collate existing knowledge and develop some new
tools from a mathematical statistical perspective with a view to clarifying decision
making in this arena.
We were led first to create GD ./, an index which measures the discrepancy
between observed categorical distributions and the corresponding D-optimal ideal
allocation. We have shown that the index and its distribution is invariant across
GLMs in the case of a single categorical covariate. When we have more than
one covariate in the model the properties of the index are more complicated,
but we can still obtain its null distribution as a Gamma.˛; / distribution, or,
alternatively, make use of the testing distribution of the X 2 statistic which follows
a 2 distribution in the Linear Model case based on the general form of the ideal
allocation (10) found. However, for GLMs with two or more categorical variables,
we must obtain the ideal allocation by numerical methods and GD ./ is preferred.
In relation to the choice of subclass we have shown that the strategy of choosing
the largest subclasses as references is optimal both in the Linear Model case and
in other GLMs when a test on GD ./ showed that the sample allocation is far away
from its ideal allocation significantly. Otherwise, the choice of reference subclass
could be arbitrary. It may be argued that we have ignored the covariance terms when
defining Vr , the total variance, but, overall the results are similar when they are
included.
There is a relationship between proportional hazards models and Poisson
regression models. For example, McCullagh and Nelder (1989) has a chapter on
converting PH models to GLM models. We have also found an observed information
matrix with a similar structure to (12) when working on the interval censored data
with an Exponential regression model (MacKenzie and Peng 2013). This implies
that all of the findings from the GLMs hold for PH regression models. Yet another
extension of GLMs are generalized additive models (GAMs) and we conjecture that
the findings may also hold for some regression models in this class.
In further work we have shown that in the linear model switching to the most
numerous subclasses also has the beneficial effect of reducing multi-collinearity
among the columns of the design matrix as measured by the condition number of
.X 0 X /.
Finally, we hope our findings will clarify some of the issues surrounding
reference subclass choice and impact positively on practice in regression analysis.

Acknowledgements The work in this paper was supported by two Science Foundation Ireland
(SFI, www.sfi.ie) project grants. Professor MacKenzie was supported under the Mathematics
Discrepancy and Choice of Reference Subclass 181

Initiative, II, via the BIO-SI (www.ul.ie/bio-si) research programme in the Centre of Biostatistics,
University of Limerick, Ireland: grant number 07/MI/012. Dr. Peng was supported via a Research
Frontiers Programme award, grant number 05/RF/MAT 026 and latterly via the BIO-SI project.

Appendix 1: R Script to Generate Compositions

Generating Compositions

#
# Algorithm nexcom - generates compositions one at a time
#
nexcom<-function(n,k,r,t,h,Qtest)
{
if (Qtest == TRUE)
{ if(t > 1) {h <-0}
{ h <- h+1
t<-r[h]
r[h] <- 0
r[1] <- t-1
r[h+1] <- r[h+1]+1
Qtest <- (r[k] != n)
return(list(Comp=r,Ind= Qtest, Tee=t, Hee=h))}
}
else if (Qtest==FALSE)
{
#set up 1st pass
r[1]<-n
t<-n
h<-0
if (k ==1){ Qtest <- (r[k] != n)
return(list(Comp=r,Ind= Qtest, Tee=t, Hee=h))}
{r[2:k]<-0
Qtest <- (r[k] != n)
return(list(Comp=r,Ind= Qtest, Tee=t, Hee=h))}
}

}
#
# Example
#
#choose n<-6 ;k<-3 - pouring 6 balls into 3 urns
n<-6 ;k<-3
#compute myrow =number of compositions
myrow<-choose(n+k-1,n)
#initialise generation parameters
r<-rep(NA,k);Qtest<-FALSE; t<-NA; h<-NA
#
#set up matrix to hold compositions
compos<-matrix(NA,nrow=myrow,ncol=k)
#
# generate all possible compositions
#
for (i in 1:myrow)
{
test<- nexcom(n,k,r,t,h,Qtest)
182 D. Peng and G. MacKenzie

r<-test$Comp
compos[i,]<-r
Qtest<-test$Ind
t<-test$Tee
h<-test$Hee
}
compos<-data.frame(compos)
#
# Create a binary indicator for positive compositions
#
z<-rep(NA, myrow)
for (i in 1:myrow)
{
tst<-all(compos[i,1:k]>0)
z[i]<-ifelse(tst, 1,0)
}
#
# Reject non-positive compositions
#
compos<-cbind(compos,z)
attach(compos)
poscompos<-subset(compos,z==1)
poscompos
npos<-dim(poscompos)[1]
npos

Generate Random Positive Compositions

#
# Algorithm simcom - generates random positive compositions one at a time
#
simcom <- function(n,p)
{
x <- seq(1, n+p-1)
freq <- c(rep(0, p))
while(any(freq == 0))
{
rand <- sample(x, size=p-1)
rand <- sort(rand)
freq[1] <- rand[1]-1
if(p>2)
{
for(j in 2:(p-1))
{freq[j] <- rand[j]-rand[j-1]-1}
freq[p] <- n+p-1-rand[p-1]
}
else
{freq[p] <- n+p-1-rand[p-1]}
}
return(list(Iter=n,Comp=freq))
}
test<-simcom(6,3)
numsim<-test$Iter
numsim
resmat<-test$Comp
resmat
Discrepancy and Choice of Reference Subclass 183

Appendix 2: Derivation of Optimal Allocation with Two

Binary Covariates

Proof. To find the optimal allocation when we have .X1 ; X2 / with .2; 2/ subclasses,
we need to maximize det.X 0 X / D nn1 n2 C 2n1 n2 n12  n1 n22  n21 n2  nn212 subject
to two equivalent constraints n1 C n10 D n or n2 C n20 D n, where n12 is the cross
term of .X 0 X /, n10 and n20 are the numbers at X1 D 0 and X2 D 0 respectively.
Using the method of Lagrange multipliers, we have

ƒ.n1 ; n10 ; n2 ; n12 ; / D n10 n1 n2 C2n1 n2 n12 n1 n22 n1 n212 n10 n212  .n1 Cn10 n/:

Setting On1 ;n10 ;n2 ;n12 ; ƒ.n1 ; n10 ; n2 ; n12 ; / D 0, we have

n10 n2 C 2n2 n12  n22  n212  D 0; (15)

n1 n2  n212  D 0; (16)
n10 n1 C 2n1 n12  2n1 n2 D 0; (17)
n1 n2  n1 n12  n10 n12 D 0; (18)
n1 C n10  n D 0: (19)

From (15) and (16), we have

n10  n1 C 2n12  n2 D 0; (20)

and from (17) and (20), we have n1 D n2 . From (18) and (19) and the finding
n1 D n2 , after some algebra, we have n12 D n21 =n. Finally, from (17), (19) and the
findings n1 D n2 and n12 D n21 =n, after some algebra, we have n1 D n=2. Thus, for
the optimal allocation for this case we found that n1 D n2 D n=2; n12 D n=4.

References

Altman, D. G., & Royston, P. (2006). Statistics notes - the cost of dichotomising continuous
variables. British Medical Journal, 332, 1080.
Berk, R. (2008). Statistical learning from regression perspective. New York: Springer.
Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Hillsdale:
Lawrence Erlbaum.
Cox, D. R. (1972). Regression models and life-tables (with discussion). Journal of the Royal
Statistical Society, 34, 187–220.
Cox, D. R., & Snell, E. J. (1989). The analysis of binary data (2nd ed.). London: Chapman & Hall.
cran.r project. (2009). R project. Retrieved 2010, R Package “pwr”. http://cran.r-project.org/web/
packages/pwr/pwr.pdf.
184 D. Peng and G. MacKenzie

Feldstein, M. S. (1966). A binary variable multiple regression method of analysing factors affecting
peri-natal mortality and other outcomes of pregnancy. Journal of the Royal Statistical Society,
Series A, 129, 61–73.
Frøslie, K. F., Røislien, J., Laake, P., Henriksen, T., Qvigstad, E., & Veierød, M. B. (2010).
Categorisation of continuous exposure variables revisited. A response to the hyperglycaemia
and adverse pregnancy outcome (hapo) study. BMC Medical Research Methodology, 10, 1471–
2288.
Isham, V. (1991). Statistical theory and modelling by edited by Hinkley, D. V., Reid, N. and Snell,
E. J.. London: Chapman & Hall.
MacKenzie, G. (1996). Regression models for survival data: The generalised time dependent
logistic family. Journal of the Royal Statistical Society, 45, 21–34.
MacKenzie, G., & Peng, D. (2013). Interval-censored parametric regression survival models and
the analysis of longitudinal trials. Statistics in Medicine, 32, 2804–2822.
McCullagh, P., & Nelder, J. A. (1989). Generalized linear models (2nd ed.). London: Chapman &
Hall.
Nijenhuis, A., & Wilf, H. S. (1978). Combinatorial algorithms for computers and calculators (2nd
ed.). London: Academic.
Peng, D., & MacKenzie, G. (2014, forthcoming). Choice of reference subclass in parametric
regression models with categorical variables.
Pocock, S. J., Collier, T. J., Dandreo, K. J., De Stavola, B. L., Goldman, M. B., Kalish, L. A., et al.
(2004). Issues in the reporting of epidemiological studies: A survey of recent practice. British
Medical Journal, 329, 883–887.
Shapiro, S. S. (1980). How to test normality and other distributional assumptions, Statistical
Techniques, 3, 1–78.
Wilkinson, P. (1995). Lung cancer in northern Ireland (MD thesis). Queen’s University, Belfast.
William, G. J. (2005). Regression III: Advanced Methods. Lecture notes. East Lansing, MI:
Department of Political Science Michigan State University. Accessed 28 Dec 2010 http://polisci.
msu.edu/jacoby/icpsr/regress3.
 Part IV
Applied Statistical Modelling
Statistical Methods for Detecting Selective
Sweeps

David Ramsey

Abstract The emigration of humankind from Africa and the adoption of agriculture
have meant that the selective pressures on humankind have changed in recent
evolutionary times. A selective sweep occurs when a positive mutation spreads
through a population. For example, a mutation that enables adults to digest lactase
has spread through the Northern European population, although it is very rare in the
African population. Since neutral alleles that are strongly linked to such a positive
mutation also tend to spread through the population, these sweeps leave a signature,
a valley of low genetic variation.
This article reviews the development of statistical tests for the detection of
selective sweeps using genomic data, particularly in the light of recent advances
in genome mapping. It also points out directions for future research.

Keywords Coalescent theory • Computational statistics • Genome mapping •

Selective sweeps

1 Introduction

The genetic code of a sample of individuals gives us information about the

present population. Since these codes result from the evolutionary and demographic
processes within a population, they also contain information regarding these
processes. Selection pressure on humankind has changed in the recent past due to
the emigration from Africa and later adoption of agriculture. Previously neutral or
deleterious alleles may become positive in such a new environment and thus spread
through a population. A strong selective sweep occurs when a positive mutation that

D. Ramsey ()
Department of Computer Science and Management, Wrocaw University of Technology, Wrocaw,
Poland
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 187

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__13,
© Springer International Publishing Switzerland 2014
188 D. Ramsey

was previously not observed in a population spreads. A weak selective sweep occurs
when a previously neutral variant that is already present in a population becomes
positively selected for. The lower the initial frequency of such a variant the more
such a sweep will resemble a strong selective sweep.
When a sweep occurs, neutral variants that are tightly linked to such a positive
variant will also spread through the population. Thus immediately after such a
sweep, there will be a “valley” of low genetic variation around the site affected by
selection. During such a sweep, the population will be split into those that have the
positive mutation and those that do not. As above, those individuals with the positive
mutation will tend to have very similar genetic codes around the site affected by
selection. The remaining individuals will tend to have much more varied genetic
codes around this site. Hence, there will be an association between the variant at the
site affected by selection and the variants at surrounding sites. This phenomenon is
known as linkage disequilibrium.
The layout of the article is as follows. Section 2 briefly describes the data used,
nucleotide sequences, and commonly used descriptive statistics based on this data.
Other forms of data can be used, but are not considered here. Section 3 considers
the Wright–Fisher model, which models the evolution of a population under the
assumptions of fixed population size, random mating and no selection. Section 4
considers coalescent processes. A coalescent tree is a way of modelling evolution in
reverse from a sample of individuals back to a common ancestor. All the information
relevant to the evolutionary processes involved in defining the genetic makeup of the
sample is contained in such a coalescent tree. The standard coalescent is based on
the Wright–Fisher model, but this model has been generalised to allow for selection,
population structure and changes in population size. At the end of this section the
effects of demographic changes and selective sweeps on the genetic makeup of a
population are briefly considered. Section 5 considers classical tests for selective
sweeps within a population, which were developed before the era of genome
scanning. The intuition behind these tests is explained using coalescent theory.
Section 6 briefly considers a classical test for differential selection in different
subpopulations. Section 7 considers models that have been developed in recent years
that use data from genome scans. Section 8 outlines some of the challenges and
possible areas of development in the near future.

2 The Genetic Data Used

The data used consist of strings of nucleotides, which are the building blocks of
the genome. There are four nucleotides: adenine, thymine, cytosine and guanine,
denoted A, T, C, and G, respectively. Recently, Consortium (2003) and Project
(2004) have provided extensive maps of the genome of samples of humans
from different subpopulations. At a large majority of the sites observed, all the
individuals have the same nucleotide. Sites at which variation is observed are
called segregating sites. At virtually all such sites just two nucleotides (variants) are
Detecting Selective Sweeps 189

observed. The variation observed at a segregating site is termed a single nucleotide

polymorphism (SNP). Suppose the genomes of n individuals are scanned. The
following are commonly used as summary statistics:
1. The non-polarised frequency spectrum. This is made up of the frequencies of the
least common (minor) variant at each of the k segregating sites, together with the
position of each segregating site. The non-polarised frequency at a segregating
site is between 1 and b n2 c, where bxc is the integer part of x.
2. The polarised frequency spectrum. This is made up of the frequencies of the wild
type variant (assumed to have been the prevalent variant in a recent ancestral
population) at each of the k segregating sites, together with the position of each
segregating site. The polarised frequency at a segregating site is between 1 and
n  1.
3. Measures of linkage disequilibrium between segregating sites.
Let p1 and q1 be the frequencies of chosen variants at sites 1 and 2, respectively.
Let p11 be the frequency with which these variants are observed together. Define
D D p11  p1 q1 . The standardised coefficient of linkage disequilibrium is DS D
jDj
Dmax , where Dmax D minfp1 .1q1 /; .1p1 /q1 g. This measure does not depend on
which variants are chosen. Clearly, if the variants at these two sites are independent,
then the standardised coefficient of linkage disequilibrium is equal to zero.
Two individuals are said to have the same extended haplotype over an interval of
the genome, if they have the same sequence of nucleotides on that interval.

3 The Wright–Fisher Model

The Wright–Fisher model assumes:

1. The population is of constant size 2N alleles (i.e., N diploid individuals).
2. Generations do not overlap and reproduction is asexual.
3. There are no mutations, recombination or selection (i.e., alleles may be under-
stood to be very short nucleotide sequences).
4. Each allele in generation j is the parent of a randomly chosen allele in generation
j C 1 with probability 2N1
(i.e., reproduction in one generation is independent of
what happens in other generations).
It follows from these assumptions that the number of offspring of allele i in
1
generation j , Xi;j , has a Bin.2N; 2N / distribution. For reasonably large N this
distribution can be assumed to be Poisson(1).
It should be noted that since the population size is fixed, for a given j the
random variables Xi;j are not independent. The correlation between the numbers
of offspring of different alleles is negative; given that one allele has a large number
of offspring, the expected number of offspring of the remaining alleles is less than
1. However, for large N these correlations will be very small.
190 D. Ramsey

Fig. 1 A realization of the • • • • • •

Wright–Fisher Process
• • • • • •

• • • • • •

• • • • • •

• • • • • •

• • • • • •

An example of a realization of the Wright–Fisher model is given in Fig. 1. Lines

indicate parentage.
In this example, all the alleles in the final generation are descendants of the allele
on the top left. It should be noted that Cannings (1974) generalized these models.
Suppose there are two alleles, denoted A and a (which may be understood
to contain all the alleles which are not A). Define the frequency of A alleles in
generation j to be Pj . According to the Wright–Fisher model, the number of A
alleles in generation j C 1 has a binomial distribution and

pj .1  pj /
E.Pj C1 jPj D pj / D pj I Var.Pj C1 jPj D pj / D :
n
The variance of the frequency of the A allele is a measure of genetic drift. The
effective size of a population is defined to be the size of a population which
reproduces according to the Wright–Fisher model and exhibits the same level of
genetic drift as observed in the population of interest. That is to say, for a fixed actual
population size the higher the variance in the number of offspring, the lower the
effective population. In practice, the effective population is estimated by assessing
the time to the most recent common ancestor (MRCA) of a sample of individuals
by observing mitochondrial DNA (which passes down the matrilineal line) or Y-
chromosomes (which pass down the patrilineal line). In both cases no recombination
occurs. The relation between the time to the MRCA and the effective population size
will be considered in Sect. 4.

4 Coalescent Trees

For a more detailed account of applications of coalescent processes in statistical

genetics, see Hein et al. (2004).
Detecting Selective Sweeps 191

Fig. 2 A realisation of the • MRCA

Common Ancestor Process
• •

• •

• • • Coalescence of ancestral lines

• • • • •

• • • • • •

4.1 The Standard Coalescent

Consider a sample of alleles. Looking back far enough in history, these alleles
have a MRCA. The birth–death process starting from this individual contains all
the information regarding this sample. This process is called the common ancestor
process (CAP). The realisation of this process corresponding to the realisation of
the Wright–Fisher model given in Fig. 1 is illustrated in Fig. 2.
These lines contain all the information regarding the genetic composition of the
present population.
The standard coalescent process models the CAP in reverse time. The present
will be denoted as time 0 and time i will denote i generations ago. Suppose
the evolution of a population follows the Wright–Fisher model. In reverse time
these lines coalesce when two individuals have a common ancestor in the previous
generation. Let Gk be the number of generations for which k distinct ancestral lines
exist. We now consider the distribution of Gk when the population size is large.
First we consider G2 , the time until two ancestral lines coalesce. The probability
that two randomly chosen alleles in the population have different “parents” (i.e., one
individual has a different parent from the other) is 1  2N1
. Since reproduction in a
given generation is independent of reproduction in other generations, it follows that
the probability that two alleles have distinct ancestors i generations in the past is
.1  2N1 i
/ . Hence,

1 i
P .G2 > i / D .1  /:
2N
Now suppose the unit of time t is defined to be 2N generations (i.e., t D 1
corresponds to 2N generations, thus i D 2tN ). Let T2 be the time to coalescence
of these two lines measured in these units. Then,

1 2tN
lim P .T2 > t/ D lim .1  / D e t :
N !1 N !1 2N

It follows from this that T2 has an exponential distribution with mean 1 (i.e., 2N
generations). Hence, the coalescence of two lines occurs as a Poisson process with
192 D. Ramsey

MRCA

t2 M1 •
{1 2 3 4} {5 6}

t3
{1 2} M2 •
t4
{3 4}
t5 • M
3
t6 M4 •
{1} {2} { 3} {4} {5} {6}

Fig. 3 A realisation of the standard coalescent

rate 1. It can be shown in a similar way that the probability of three individuals
having a common ancestor in the previous generation is O. N12 /.
Now consider Tk , this is the time (considering the reverse process) until the
first coalescence of two of k ancestral lines. Each pair of lines coalesce at rate 1.
Since there are k.k1/
2 pairs of lines, it follows that Tk  ExpΠk.k1/ 2 . It should
be noted that the Tk are independent. Since the probability of three lines coalescing
simultaneously is O. N12 /, when a coalescence occurs in the limiting process, with
probability one it occurs only between two lines. Hence, after such a coalescence
there will be k  1 ancestral lines.
Figure 3 illustrates a realization of a coalescent process starting with six
individuals at the present (time 0). This process can be thought of as a tree. Each
vertical line represents an ancestral line. A horizontal line represents the coalescence
of two lines. Each set denotes the individuals from the sample corresponding to a
line.

4.2 Mutations Within the Coalescent Tree and Polarity

In the absence of mutations and recombination, the descendant alleles would be

identical to the most common ancestor. Mutations can be introduced in a very
simple way. The simplest model is the infinite sites model, which assumes that the
number of sites in an allele is very large and the mutation rate is small. Hence, every
mutation that occurs in a genetic tree occurs at a different site. This site may be
chosen by generating a number at random from the uniform distribution on Œ0; 1.
Let u be the probability that a mutation occurs. It is assumed that u is of order N1 . Set
 D 4N u. In the limiting process (as N ! 1 and 1 unit of time is 2N generations),
mutations occur in the coalescent process as a Poisson process of rate 2 along each
ancestral line. This is illustrated in Fig. 3.
Detecting Selective Sweeps 193

Mutation M1 is common to individuals 1–4, mutation M2 is common to

individuals 3 and 4. Mutations M3 and M4 each only appear in one individual (1
and 5, respectively). Such mutations as M3 and M4 are called singletons and occur
in the external line of a tree. Under the infinite sites model, the individuals at the
end of each line only show variation at the sites of these mutations. Such sites are
called segregating sites. The number of segregating sites is thus equal to the number
of mutations in the coalescent tree.
A more advanced model for generating mutations is the finite sites model. The
simplest such model assumes that mutations are equally likely to occur at each site.
Improvements to this model have been made by assuming, e.g., that the mutation
rate is greater in non-coding regions than in coding regions. Since a mutation can
occur at a site more than once, the number of mutations is at least the number of
segregating sites. However, if the mutation rate is reasonably small and the number
of sites large, the infinite sites model will be a reasonable approximation.
Since it is assumed that only two variants can appear at a site, the genetic code
of the descendants may be described by a binary code whose length is equal to the
number of segregating sites. Two types of coding systems are used:
Non-polarised The most common variant at a segregating site is coded using 0.
Polarised The code for the most recent common ancestor is assumed to be made
up of zeroes. Mutant variants are represented by ones. It should be noted that
when dealing with data, it is often assumed that the ancestral code is the most
common code in the most similar ancestral species or an ancestral population
(i.e., nucleotide sequences for the African human population can be used to
define the ancestral code for the European human population).
The non-polarised frequency at a site is the frequency of the minor variant. The
polarised frequency is the frequency of the derived (mutant) variant.
Suppose the order of the sites of the mutation along the nucleotide sequence is
M1 ; M2 ; M3 and M4 . Then the polarised codes of the six descendants are:
Individual 1—1010, Individual 2—1000, Individuals 3 and 4—1100, Individual
5—0001, Individual 6—0000.
Since mutation M1 is the only mutation seen in the majority of individuals, the
non-polarised codes are obtained by swapping 0 with 1 in position 1. The polarised
frequency at site 1 is 4, while the non-polarised frequency is 2.
One simple test of the infinite sites model without recombination is based on
the pattern of possible sequences under this model. Suppose mutation M1 occurred
before M2 . There are two possibilities: (a) mutation M2 occurs on a lower branch
that emanates from the branch in which M1 occurred, (b) mutation M2 occurs on a
branch that does not emanate from the branch in which mutation M1 occurred. In the
first case, there is no individual who does not have mutation M1 , but has mutation
M2 . In the second case, there is no individual who has both mutations. Hence, only
three of the four possible codes for these two positions (00, 10, 01, 11) can occur. If
all four possibilities occur at any pair of sites, then the infinite sites model without
recombination cannot hold.
194 D. Ramsey

4.3 Population Size Variation over Time

The assumption of a population of fixed size may be relaxed by assuming that

the population size t units of time ago was 2N.t/ alleles, where one unit of time
represents 2N.0/ generations. The rate of convergence of ancestral lines at time t
relative to the rate of convergence in the standard coalescent is N.0/=N.t/ (since the
probability of the coalescence of lines is inversely proportional to population size).
The coalescent tree starting with n lines can be generated using the fact that the first
coalescence of any two of k lines occurs as a non-homogeneous Poisson process of
rate k.k1/N.0/
2N.t /
. In this case, the times between coalescent points are not independent.
For example, suppose a population is expanding (i.e., N.t/ is decreasing in t, as
we are considering the reverse process). If the time to the first coalescent point is
relatively long, then the expected time to the next coalescence is relatively short,
since going back in time the coalescence rate is increasing for fixed k.
The coalescent tree for an expanding population will tend to have relatively long
external branches and short upper branches. Hence, for expanding populations we
expect a larger number of singletons and a smaller number of mutations that occur
in a large number of individuals.

4.4 The Introduction of Recombination into the Coalescent

Tree

It is more difficult to introduce recombination into a coalescent tree. Looking back in

time, when recombination occurs an allele will have two parents. One of the parents
contains the information to the left of the recombination point and the other contains
the information to the right of the recombination point. Hence, recombination is
represented by the splitting of one ancestral line into two. Suppose the probability
of recombination along an allele, r, is of order N1 . Then setting  D 2rN , each
ancestral line is subject to recombination according to a Poisson process of rate .
The recombination point is chosen at random. Since the rate of coalescence is of
the order of the square of the number of lines, with probability one the number
of ancestral lines will fall to one in finite time. Using the coalescent tree, we
can trace which information was passed on. Suppose m recombinations occur in
such a coalescent tree and the order of the recombination points along the tree
is R1 ; R2 ; : : : ; Rm . Let R0 and RmC1 denote the end points of the allele. Since
no recombination occurs between Ri and Ri C1 , it follows that the ancestral lines
corresponding to this section can be described by a standard coalescent tree. Hence,
the coalescent process with recombination can be described by a set of m C 1
standard coalescent trees. Figure 4 gives a simple example starting with just two
individuals. The genetic information contained in A is in light grey. The genetic
information in B is denoted in dark grey. Black denotes material that is passed
down to both individuals. White denotes material that is of unknown source when
Detecting Selective Sweeps 195

Fig. 4 The coalescent tree

with recombination

the initial tree is being generated. The source of this material becomes apparent
at coalescent points. For example, at t3 the two central lines converge and so
must contain the same genetic information. From the graph the central section is
transferred to both A and B, hence the material is common to both.
Suppose the recombination points are at 0.3 and 0.7, in the standard coalescent
tree corresponding to the left hand section the two lines coalesce at time t4 . In the
case of the central and right hand sections, the coalescent times are t3 and t5 ,
respectively. Mutations can be added to these standard coalescent trees as before.
It should be noted that these models have been extended to account for population
structure (see Slatkin 2000) and different forms of selection (see Neuhauser and
Krone 1997). Various computer programs are available to simulate samples of
nucleotide sequences under various assumptions regarding population structure and
selection. Hudson (2002) made available a program that carries out such simulations
under various demographic scenarios when selection does not act. Spencer and
Coop (2004) extended this to models that included selection.

4.5 Selective Sweeps and the Coalescent Tree

A selective sweep occurs when a mutation that was previously not present in the
population spreads throughout the population. Variants at sites that are closely
linked to such a positive mutation will tend to also spread. This process is known
as hitchhiking. Suppose such a selective sweep has just finished. Consider the
coalescent tree for the nucleotide sequence immediately surrounding the positive
mutation. Due to a lack of recombination between neighbouring sites, the genetic
information in the present population will have been inherited from the individuals
in the population that carried the positive mutation. These individuals form a
subpopulation which grows rapidly in size. Hence, the coalescent tree will tend to
196 D. Ramsey

resemble the coalescent tree for a nucleotide sequence in an expanding population.

However, due to recombination the genetic information at sites not linked to the
positive mutation will tend to be inherited from the population as a whole. The
general conclusion is that the demographic history affects the genome uniformly,
while the effect of a selective sweep is just local. Coalescent theory has been applied
to predict how a selective sweep affects genetic variation. It is normally assumed
that the frequency of a positive mutant changes randomly according to a birth/death
process, until it reaches some frequency and then spreads deterministically through
the population (see Durrett and Schweinsberg 2004).

5 Classical Tests for Selective Sweeps Within a Population

In this section we will describe a classical test for a selective sweeps: Tajima’s
test (Tajima 1989) and mention some others. In order to do this, we first consider
some statistical properties of the standard coalescent tree and two related measures
of diversity, which can be used to estimate the mutation rate under the standard
coalescent model.

5.1 The Height and Length of a Standard Coalescent Tree

The height of a standard coalescent tree starting from n lines, Hn , is the time
required for all the lines to converge. Since Tk is the time to the first coalescence of
any pair of lines from k lines and the Tk are independent, it follows that

X
n X
n
2 2
EŒHn  D EŒTk  D D2 ;
k.k  1/ k
kD2 kD2

X
n X
n
4
VarŒHn  D VarŒTk  D :
k 2 .k  1/2
kD2 kD2

Hence, under the standard coalescent the expected time to the most common ances-
tor is always less than two time units (4N generations). For large n, VarŒHn  
4 2
3
 12.
Another important measure is the length of a coalescent tree, Ln . This is the sum
of the lengths of the branches. It can be seen from Fig. 3 that k lines exist for time
Tk . It follows that
Detecting Selective Sweeps 197

X
n X
n
2 X1
n1
EŒLn  D kEŒTk  D D2 :
k1 k
kD2 kD2 kD1

X
n
4
VarŒLn  D :
.k  1/2
kD2

For large n, we may use the approximations EŒLn   ln.n/C0:577 and VarŒLn  
2 2
3
.
Since mutations occur along each line as a Poisson process with rate 2 , it follows
that the expected number of mutations, E.S /, is given by E.S / D EŒL n
2 .
It should be noted that under the infinite sites model without recombination, the
number of mutations is the number of segregating sites. Under this model, we can
estimate the mutation rate from a sample of n nucleotide sequences using

2S S
OL D D Pn1 :
EŒLn  kD1 1=k

5.2 Nucleotide Diversity

The Hamming distance between sequence i and sequence j , Di;j is the number of
positions at which the sequences differ. The nucleotide diversity, …, of a set of n
nucleotide sequences is defined to be the average Hamming distance between pairs
of sequences, i.e.

2 X
…D Di;j : (1)
n.n  1/ i <j

According to the standard coalescent tree, the expected Hamming distance between
two nucleotide sequences is the expected number of mutations on a coalescent tree
starting with two lines. Since the expected length of such a tree is one, the nucleotide
diversity is an estimator of the mutation rate i.e. O… D ….

5.3 Tajima’s Test for Selection

It should be noted that these estimators are only valid when the Wright–Fisher
model holds, i.e., there is no selection and the population size does not vary in
time. As argued above, after a selective sweep the number of singletons is expected
198 D. Ramsey

to be relatively high and the number of mutations that are common to many
lines relatively low compared to the standard coalescent. Each of these mutations
corresponds to a segregating site, but affects the measure of nucleotide diversity in
different ways. A singleton mutation will increase the Hamming distance between
the individual in which it occurs and the remaining n  1 individuals by one (i.e.,
increase the sum in Eq. (1) by n1). Consider the following extreme case in which a
mutation occurs when there are only two lines and both of these lines correspond to
n n2
2 individuals. This difference will be counted in 4 comparisons (and hence increase
2
the sum in Eq. (1) by n4 ). It can thus be seen that when a selective sweep occurs,
the estimator of the mutation rate based on the number of segregating sites, OL , is
expected to be larger than the estimator based on the nucleotide diversity, O… .
Tajima noted that under the Wright–Fisher model for reasonably large samples
there is almost no correlation between OL and O… . He introduced the standardised
D statistic to test for the effect of selection, where

O…  OL
DDq :
O .O…  OL /
Var

Under the null hypothesis that the standard coalescent model holds:

O .O…  OL /  e1 S C e2 S.S  1/;

Var
nC1 1
e1 D  ;
3an .n  1/ an2
 
1 2.n2 C n C 3/ n C 2 bn
e2 D 2  C 2 ;
an C bn 9n.n  1/ nan an

P Pn1 1
where an D in1 D1 i and bn D
1
i D1 i 2 .
Large negative realisations of the test statistic are assumed to be evidence that a
selective sweep has occurred. Tajima observed that assuming that D has a standard
normal distribution leads to a conservative test (i.e., when the standard coalescent
model holds, the probability of rejecting the null hypothesis is less than the nominal
significance level). He observes that a scaled beta distribution gives a better fit to
the distribution of D, but the resulting test is still conservative.
However, Tajima’s test does not distinguish between effects of demography and
selection, since we also expect negative realisations of the test statistic D given that
a population is rapidly expanding. The effect of such demographic changes should
be uniform over the genome, while the effect of a selective sweep will be local.
However, before the age of gene mapping it was difficult to practically make use of
this idea.
Detecting Selective Sweeps 199

The assumption that a selective sweep is complete is also important. Macpherson

et al. (2008) note that partially completed sweeps may be associated with positive
values of Tajima’s D statistic, as at intermediate frequencies a positive mutation
is likely to be linked with other alleles at intermediate frequency and thus the
nucleotide diversity is likely to be relatively large. Also, varying recombination rates
may produce a signal, since areas of low recombination tend to be characterised by
low genetic variation.

5.4 Other Classical Tests for a Selective Sweep Within a

Population

Fu and Li (1992) proposed a test based on estimating the mutation rate in two ways
using the number of non-singleton and singleton mutations,  and  D n  ,
respectively. Under the Wright–Fisher model, the coefficient of correlation between
these two estimators is smaller than the coefficient of correlation between O… and
OL . They suggested that the standardised difference between these two estimators
should be used to test the standard coalescent model. As before, if there has been a
recent selective sweep there will be a relatively large number of singletons. Hence,
we would expect large negative realisations of the test statistic. However, this is also
true if the population is expanding. Also, Achaz (2008) notes that the number of
singletons is very sensitive to the probability of wrongly reading a nucleotide and
so such a test will not be reliable.
Fu and Li note that other estimates of the mutation rate can be used to test the
standard coalescent model in the same way. Zeng et al. (2007) base tests on the joint
distribution of such statistics. The joint distribution is estimated using simulations
based on the Wright–Fisher model.
However, all such tests suffer from not being able to separate demographic effects
from the effects of a selective sweep, although the tests of Zeng et al. (2007) are
somewhat more robust to changes in population size. Möller et al. (2007) also note
that such tests are sensitive to the choice of population. For example, species wide
measures of D give lower average values of D than measures based on carefully
defined subpopulations. Including several subpopulations in one sample tends to
increase the number of SNPs, as many will be specific to just one population. Hence,
the number of segregating sites will be large in relation to the nucleotide diversity.
Another problem is that different forms of selection have different effects on the
pattern of variation in a population. A soft selective sweep would not have such
a clear signature. This signature is also affected by whether the favoured allele
is dominant, co-dominant or recessive. The appropriate signal to use depends on
the development of the sweep. When a sweep is in progress the level of linkage
disequilibrium will be a clearer signal than the frequency spectrum, which is more
useful at the end of or after a sweep.
200 D. Ramsey

6 A Classical Test for Differential Selection in

Subpopulations

Consider the case in which an ancestral population has split into n subpopulations
each occupying some specified environment. Suppose the frequency in subpopula-
tion i of a variant at a site is pi . Due to genetic drift, p1 ¤ p2 . Mutations counteract
the effect of genetic drift (if the frequency of A increases, then there will be more
mutations from A to the other variant). Hence, if there is no selection, the frequency
of the variants in each subpopulation will be similar. Lewontin and Krakauer (1973)
introduced a test for differential selection using a measure, FST , of the difference
s2
between the subpopulations using FST D p.1p/ p
, where sp2 is the variance in the
frequency of A between subpopulations and p is the mean frequency of A in the
population as a whole. In order to carry out such a test, it is necessary to estimate
the average value of FST at sites where two variants are relatively common (by
assumption most of these sites are unaffected by selection). Denote this estimate by
F . They noted that under the Wright–Fisher model the test statistic T D .n1/F F
ST

2
has approximately a n1 distribution. Large realisations of T indicate differential
selection, i.e., A is relatively more favourable in one of the environments. However,
there are some problems associated with this approach:
1. The test does not work well when the (effective) sizes of the subpopulations are
very different.
2. The test is sensitive to changes in the size of subpopulations.
3. If A is selected for (or against) in both environments to a similar extent, then this
test is unlikely to detect selection.

7 Tests of Selection in the Era of Genome Mapping

McVean et al. (2005) outline many possibilities arising from the existence of the
HapMap data. One of the obvious advantages is that there is enough data to be
able to differentiate between the local effect of selection and the global effect of
the demographic history of the population. New methods of detecting selecting
sweeps can be broadly separated into two main classes: (1) empirical methods and
(2) composite likelihood methods.

7.1 Empirical Methods

Using such an approach a score is calculated for all the sites along a chromosome.
For example, the score for a site might be Tajima’s D statistic calculated using data
from all the sites in a specified neighbourhood. If the population is growing, then
Detecting Selective Sweeps 201

we would expect negative scores. However, if an area is affected by selection, then

all the scores in a region are expected to be particularly low. Thus the intuition
behind the empirical approach is to select a small proportion of the genome (say
1–5 %) as areas that seem to have been affected by a selective sweep and use then
use biological methods to further investigate these regions.
Carlson et al. (2005) use a score based on the approach described above using
a 100 kb window around a site. They investigated human populations of African,
European and Chinese origin and selected the 1 % of genes (taken over the three
groups as a whole) for which the lowest scores were observed. They concluded that
recent selection seems to have played a lesser role in the African population and has
acted in clearly distinct ways in the three populations. According to their criterion,
they selected 7 genes thought to be subject to selection in the African population,
23 in the European population and 29 in the Chinese population. Only four genes
were thought to be subject to selection in at least two populations. As in the case of
Tajima’s test, this method is not robust to varying recombination rates.
Voight et al. (2006) use a different procedure based on the concept of extended
haplotype homozygosity (EHH) introduced by Sabeti et al. (2002). The EHH for
a given interval of the genome is defined to be the probability that two randomly
chosen individuals from a population have the same nucleotide sequence on that
interval. This measure will fall to 0 as the length of this interval grows. Their
method is designed to detect an incomplete strong sweep, hence the site affected
will be an SNP. Denote by D the (derived) variant that is suspected to be favoured
by selection at an SNP. The other (ancestral) variant is denoted by A. If D is in
reality favoured by selection, then the individuals which have D will show less
variation around that site than those who have A. Define iHHA to be the sum of
the estimates of EHH over all the intervals centered around an SNP for which EHH
> 0:05 in the subpopulation which has variant A. The measure iHHD is defined
analogously based on the subpopulation that has the variant D. Their standardised
score, calculated for each SNP at which the frequency of the least common variant
is >5 %, is given by

 h
i
ln iHH A
iHH D Ep ln iHHiHH A

h
i
D
iHS D ;
iHH A
SDp ln iHH D
h
i h
i
iHH A iHH A
where Ep ln iHH and SD p ln are the estimates of the expected value
D

 D
iHH
iHH A
and standard deviation of ln iHH D
based on the empirical observations. Large
negative values of this score indicate that individuals with the derived variant show
less variation around that site. This is an indication that the site is affected by
selection. They note that the empirical distribution is close to normal and there may
be errors in determining which of the variants is the derived variant. Hence, they
consider any absolute score of above 2.5 (around 1 % of the scores) as being an
indication of selection. When selection occurs, it is expected that these high scores
202 D. Ramsey

will be clustered so they also use windows of length 100 kb to illustrate the variation
of scores over a whole chromosome.
They note that since the score is based on a ratio of homozygosity measures,
then this method is less sensitive to local changes in the recombination rate than
the method described above. This is due to the fact that low recombination rates
will lead to low measures in both subpopulations, which does not greatly affect
the score. They considered three human subpopulations: from East Asia, Nigeria
and Europe. Unlike Carlson et al. (2005), they found that more sites in the African
population were selected as candidates for sites affected by selection than in the
other two populations (note this test is more sensitive to incomplete sweeps).
Akey et al. (2002) adapt the method of Lewontin and Krakauer (1973). They
estimated FST for each of 25 549 SNPs based on three human populations: of
African, European and East Asian origin. The average measure of FST was 0.123.
Compared to simulations based on a model with migration but no selection, there
was an excess of low and high FST scores, which is in agreement with the action
of natural selection. Also, extreme values of the estimates of FS T based on the data
tended to be highly clustered.
The authors chose sites where the FST score was greater than the 97.5 % quartile
of the empirical distribution to be candidates for sites affected by a selective sweep.
This led to 174 candidate genes, among those were a gene associated with cystic
fibrosis and a gene associated with diabetes.
Kimura et al. (2007) adapt the concept of EHH to the problem of detecting the
site of a selective sweep that affects a particular subpopulation. They use sequences
from two populations, where one is assumed to be the ancestral population. They
define the most frequent haplotype homozygosity (MHH) to be the probability that
two individuals in a subpopulation both have the most frequent haplotype over a
given interval. For the ancestral population and derived population, this measure is
denoted as MHH A and MHH D , respectively. Blocks of low variation in the derived
population are defined to be intervals for which MHH D  0:9 that include at least
two sites where variation is seen in the two populations as a whole. They define
HTMH D to be the probability that two individuals in the derived population have the
same haplotype as the most common haplotype in the ancestral population. From
this they calculate the relative most frequent haplotype homozygosity (rMMH)
for the two populations as rMHH D HTMH D
MHH A . A very small value of rMHH for
a block indicates that the most common haplotype in the ancestral population
is rare in the derived population. Hence, such a block may have been affected
by selection. Similarly, the relative extended haplotype homozygosity (rEHH) is
defined as rEHH D EHH EHH A
D
; where EHH A and EHH D are the EHH over a chosen
block in the ancestral and derived population, respectively. If selection has only
occurred in the derived population, then this measure will also be small. Using
results from simulations, the authors defined candidate blocks as those for which
rMHH < 0:05 and rEHH < 0:3. They found that such a test compared favourably
with using a test based on estimating FST . Unlike the test of Voight et al. (2006)
this test is aimed at detecting complete or almost complete sweeps, since the chosen
blocks must display very little variation in the derived population. It should be noted
Detecting Selective Sweeps 203

also that the test is designed to detect a sweep specific to the derived population. If
there was a similar sweep at a site in both populations, the observed value of rEHH
would be relatively large, since EHH A would be large.

7.2 Advantages and Disadvantages of Empirical Methods

One of the obvious advantages of an empirical approach is simplicity. However,

there are serious disadvantages. Teshima et al. (2006) reviewed the reliability
of empirical methods based on simulations which considered different levels of
dominance and demographic histories, as well as both hard and soft selective
sweeps. One problem regarding the theory based on coalescent processes is that
it is normally assumed that there is no dominance. By selecting appropriate sites
from these simulations, the authors obtained data sets in which a known proportion
of sites were closely linked to the site of a selective sweep. They consider scores
based on nucleotide diversity , Tajima’s D and EHH (see Yamasaki et al. 2005;
Carlson et al. 2005; Voight et al. 2006, respectively). Candidate sites were chosen so
that the proportion of candidates was equal to the proportion of sites closely linked
to a selective sweep.
They conclude that the error rate is higher for methods based on D than for
tests based on  or EHH. However, the authors admit that methods based on 
and EHH are more sensitive to changes in the mutation rate and they assumed in
their simulations that mutation rates were constant over the whole genome. Also,
the methods are sensitive to the demographic history of a population. For example,
the test is more sensitive if a sweep occurred after the expansion of a population
rather than before.
As argued above, the stage of development of the sweep is important in
determining the efficacy of such methods. Methods based on EHH are effective
if a positive variant in a strong selective sweep has increased in frequency to around
40 %. However, this method is not effective for complete/almost complete sweeps.
On the other hand, methods based on  and D are effective only when a sweep
is almost complete. Also, since a positive recessive mutation will, in general, take
longer to invade a population than an undominated mutation, the power to detect
sweeps in which a recessive variant invaded a population is relatively low as a higher
level of variation is maintained.
Moreover, the nature of the sweep is very important. They simulated soft sweeps
in which a previously neutral variant starts to be selected for when its frequency
is 5 %. The methods based on  and D detect recently completed hard selective
sweeps well. When 5 % of the sites are affected by selection, the “false discovery
rate” (the proportion of chosen candidates which are not strongly linked to a site
affected by a selective sweep) was 8.3 % using  and 20.3 % using Tajima’s D.
However, these methods were poor at detecting recently completed soft selective
sweeps. The corresponding “false discovery rates” were 73.8 and 82.2 %.
Innan and Kim (2004) also observe that such tests are sensitive to the type of
sweep (hard or soft). In a later paper (Innan and Kim 2008), they note that tests
204 D. Ramsey

using samples from both ancestral and derived populations are more sensitive to
soft selective sweeps than tests based on samples from just one population. A soft
selective sweep has a weak signature, since local variation is not reduced by as
much as under a strong selective sweep. However, if a soft selective sweep occurs in
a derived population, then the frequencies of the variants in the two subpopulations
will be clearly different. This fact is felt to be important especially in the case
of agricultural crops, where artificial selection generally occurred from standing
variation. On the other hand, when a new population breaks away from its ancestral
population, a population bottleneck is a very likely scenario and this might explain
such differences in the frequency spectra of the two resulting populations.
Finally, there are also some technical problems associated with empirical
methods. Firstly, they cannot be used to estimate the strength of a sweep. Secondly,
we are unsure of the proportion of the genome that is affected by selection. Hence,
choosing the threshold above which we choose a site to be a candidate is, to a large
degree, guesswork. If we choose too low a threshold, we will obtain a large number
of “false positives” (candidates that are not linked to sites affected by selection).
By choosing too high a threshold, we will miss many sites that are affected by a
selective sweep.

7.3 Composite Likelihood Methods

Unlike empirical methods, composite likelihood methods test the hypothesis that
there is no selective sweep affecting a section of the genome. The procedure used is
similar to a standard likelihood ratio test in which the alternative is that a selective
sweep has occurred. These tests do not, in general, consider the correlation between
variant frequencies at sites and for simplicity calculate the composite likelihood.
This is done by multiplying the likelihoods of the variant frequencies observed at
each segregating site under the appropriate hypothesis.
Kim and Stepham (2002) present such a test, in which the null hypothesis is
that the population follows the Wright–Fisher model (i.e., constant population size
and no selection). The distribution of the variant frequencies under this alternative
was derived by Kimura (1971). The composite likelihood of the data can be
calculated simply by multiplying the individual probabilities of the observed variant
frequencies. Under the alternative, it is assumed that a strong selective sweep has
just finished. The distribution of the variant frequencies under the alternative, which
depends on the strength of selection and the rate of recombination between a site
and the location of the mutation, was given by Maynard Smith and Haigh (1974).
It is possible to find maximum likelihood estimates of the location and strength of
a putative selective sweep using a composite likelihood function. The test statistic
is based on the ratio between the composite likelihoods obtained under the two
hypotheses.
Jensen et al. (2005) note that the test of Kim and Stephan is sensitive to the
demographic history of a population. This is not surprising as the null hypothesis
Detecting Selective Sweeps 205

is that the Wright–Fisher model holds. For example, the authors carried out
simulations in which a population had a recent, strong bottleneck, but there was
no selection. Kim and Stephan’s test rejected the null hypothesis around 90 % of the
time. They proposed an improved version of the original test as follows: (1) Carry
out Kim and Stephan’s test. If the null hypothesis of no selection is rejected, then (2)
they carry out a goodness of fit test in which the null hypothesis is that a selective
sweep has occurred with strength and location given by the maximum likelihood
estimators calculated at stage 1. If the null hypothesis is rejected in stage 2, then
they conclude that demographic factors caused the changes from the predictions
made under the Wright–Fisher model. Of course there are problems associated with
such an approach, such as the possibility of multiple sweeps.
Nielsen et al. (2005) use an intuitive approach to deal with the problem of
assumptions regarding the demographic history of a population. Since demographic
history affects genetic variation at the level of the whole genome, they assume
that when there is no selection the distribution of the variant frequencies is some
background distribution. Any local divergence from this background distribution
may be attributed to the action of selection. Under the assumption of no selection,
this distribution can be estimated from the data. Consider the case in which we
have non-polarised frequencies at k SNPs for n individuals. Since the non-polarised
frequency is defined to be between 1 and imax D b n2 c, the background distribution
can be estimated by estimating the probability of a non-polarised frequency of i , pi ,
using pOi D kki , where ki is the number of times a frequency of i is observed. The
Q
composite likelihood of the data under the null hypothesis is thus iimax ki
D1 pOi .
The alternative is that a selective sweep has occurred. The distribution of the non-
polarised frequencies under such a hypothesis can be approximated using coalescent
theory. This is based on an estimate of the proportion of the population that, at
the end of a selective sweep, have a neutral variant that initially appeared with the
original positive mutation. This probability is given by e ˛d . Here ˛ D r ln.2N s
/
,
where d is the distance between a site and the positive mutation, r the recombination
rate, N the population size and s is the selection coefficient (see Durrett and
Schweinsberg 2004). Under this assumption, one can calculate the composite
likelihood under the hypothesis that there is a selective sweep of strength s at a given
locus. The maximum likelihood estimator of the strength and location of a sweep
on a chromosome is found by maximising over all the loci along a chromosome.
This is used to define a likelihood ratio test. It should be noted that intrinsically
this test involves multiple testing (the possibility of a selective sweep at each site
is considered). Hence, the authors find appropriate critical values for the test by
simulating the distribution of the test statistic under the hypothesis that the Wright–
Fisher model holds.
They show that the test is not sensitive to demographic history using a wide
range of simulations. Also, by analysing chromosome 2 for the European population
(using data from the HapMap project), they show that the test identifies a selective
sweep in the neighbourhood of the lactase gene. Although the alternative hypothesis
206 D. Ramsey

is that one selective sweep has occurred, the likelihood scores do indicate that
multiple sweeps have occurred on this chromosome.
In addition, they correct for the so-called ascertainment bias (see also Nielsen
et al. 2004). Often, to reduce costs a small subsample of individuals is fully
sequenced (the ascertainment process). The remaining individuals in the sample
are only sequenced at the sites where variation was observed in the subsample. This
means that there may be more SNPs in the sample than we actually observe and
high frequency variants are over-represented. This correction is a relatively simple
process based on Bayes’ laws.
Some work has been carried out on the use of linkage disequilibrium measures.
Kim and Nielsen (2004) use information from linkage disequilibrium measures in
addition to data on the frequency spectrum of SNPs used in Kim and Stepham
(2002), in order to detect a recently completed sweep. However, they conclude
that such an approach hardly increases the power of the test. Jensen et al. (2007)
consider the same problem and show that various demographic histories, such as a
population bottleneck, can lead to a similar pattern of linkage disequilibrium as a
selective sweep. They conclude that linkage disequilibrium measures can be useful
in detecting selective sweeps, but they note that linkage disequilibrium disappears
relatively rapidly after a sweep finishes. One problem with estimating linkage
disequilibrium lies in the fact that such estimation tends to be inaccurate when
variants are rare (as expected around the site of a selective sweep, Jensen et al.
(2007) did not use singletons in their analysis). Pfaffelhuber et al. (2008) analyse the
evolution of linkage disequilibrium under a selective sweep using coalescent theory.
They conclude that the pattern of linkage disequilibrium observed around the site of
a complete selective sweep is similar to the pattern of linkage disequilibrium around
a recombination hotspot, but a selective sweep will also reduce variation.

7.4 Advantages and Disadvantages of Composite Likelihood

Methods

One of the obvious advantages of such methods is that maximum likelihood

estimators can be found for the strength and location of a sweep. However, one
should remember that the standard assumption of such tests is that a strong sweep
has just ended. Using simulations, given this assumption Nielsen et al. (2005)
showed that these estimators are accurate. On the other hand, if a sweep finished
some time in the past or a weak sweep occurred, the strength of the sweep will tend
to be underestimated and the variance of the estimator of the location will increase.
If the sweep is not close to completion, it would be better to use methods based on
linkage disequilibrium or comparing two subpopulations.
More accuracy could be obtained by using the real likelihood function (which
takes into account linkage disequilibrium) rather than the composite likelihood
function. Although a relatively large amount of work has been done on what
Detecting Selective Sweeps 207

frequency spectrum should be expected after a selective sweep, relatively little has
been done on the form of the linkage disequilibrium. Once such work has been
carried out more fully, it may be possible to use ideas from covariance modelling
(see Pan and MacKenzie 2003).
Another problem is that it is assumed that there was only one selective sweep on
the section of the genome considered. Nielsen et al. (2005) show that their method
indicates the occurrence of many selective sweeps. Also, considering the role of
genes in the region of these sites, it is reasonable to assume that many of these
signals are true. The effect of multiple sweeps is unknown. It may be the case that
such sweeps are sufficiently rare that their effects on the frequency spectrum can
be assumed to be independent. However, if two positive mutations occur at sites
which are close together, the resulting effect is difficult to predict (it will depend
on the order in which the mutations occurred and the time between them). This is
particularly important when we consider agricultural crops, where it is likely that
artificial selection caused multiple soft sweeps occurring almost simultaneously.

8 Conclusion

Jensen et al. (2008) considered the problem of detecting selective sweeps using
firstly, data from whole chromosomes and secondly, using partial maps. He con-
cluded that low density maps lower the power of tests to detect selective sweeps.
Even using an initial map as a guide to decide upon areas of the genome to sequence
more fully is unreliable. Hence, the availability of genome maps is undoubtedly a
huge resource for the analysis of genetic data.
It has been shown that the frequency spectrum is dependent on both selection
and the demographic history of the population. It is also very important whether
the selective sweep originated from a positive mutation (a strong sweep) or from
standing variation (a weak sweep). In the second case, the signal will be weaker.
The time at which the sweep occurred is also important, as tests based on the variant
frequencies are most powerful at detecting sweeps that have just completed. If a
sweep is in progress, methods based on comparing the local variation of individuals
with one variant at a site with the remaining individuals (or considering the level of
linkage disequilibrium) are more powerful.
However, as Hamblin and Di Rienzo (2000) show, there is evidence that the
frequency spectrum of a section affected by a completed selective sweep may well
not resemble the one expected. They investigated a 1.9 kb section of the genome
around the Duffy blood group locus. The wild type allele is thought to be FY B,
as this is observed in apes. This allele is observed in the European population.
All 24 individuals from 5 sub-Saharan subpopulations were FY O homozygous.
It has been noted that such individuals are immune to vivax malaria and there is
no evidence to indicate that the FY O allele is selected against in the European
population. Hence, it has been hypothesised that a selective sweep occurred after
the European population split from the African one. In the section investigated there
208 D. Ramsey

were five SNPs in the African population and at two of these sites both variants were
frequent (the frequency of the minor variant was 0.23), which is very unexpected
under the theory of selective sweeps. These two sites were in complete linkage
disequilibrium. It is possible that the population structure plays an important role
here, as the least common variant was only observed in the populations towards the
southeast.
The method of Nielsen et al. (2005) seems to be relatively robust to demographic
factors. However, the test is based on the assumption that one strong selective sweep
has occurred, hence there are still questions to be answered. Barnes (2006) notes that
inbred lines of mice exhibit linkage disequilibrium even for sites lying on different
chromosomes. This indicates that selection acts on combinations of alleles. The
implicit assumption of the theory presented here assumes that the effects of selection
are additive and extension of the theory to include such interactions would be overly
complex.
However, there seem to be areas in which advances could be made. A couple of
very recent papers indicate the problem of errors in sequencing (see Achaz 2008;
Johnson and Slatkin 2006). The probability of wrongly reading a nucleotide is small,
however the number of nucleotides that need to be read is very large. Since most
sites are not SNPs, most of these errors result in singletons, i.e., low frequency
variants are over-represented due to this phenomenon. This may lead to false signals
of selective sweeps. Correcting for such a bias is not as simple as correcting for
ascertainment bias, as the probability of making an error needs to be estimated very
accurately and more practical work needs to be done in order to do this (even a
very small error rate can have a very large effect on the frequency spectrum). It is
quite possible that ascertainment will not be used in the long run and so errors in
reading will become the sole source of errors in the data. The expected number of
false singletons increases linearly in n, the number of individuals sequenced, but the
expected number of segregating sites is proportional to ln.n/. This may explain to
some degree the observation that the test of Nielsen et al. (2005) seems to work best
for intermediate sample sizes.
The methods described above treat each site equivalently, whereas it is known
that:
1. Some sections of the genome are non-coding.
2. Triples of nucleotides code for proteins, so the position of a nucleotide is
significant and many mutations (especially of the central nucleotide in a triple) do
not result in a change in the protein produced, these are the so-called synonymous
mutations.
Williamson et al. (2005) use such an approach to simultaneously infer the effect of
demographic history and selection on patterns of variation in the human genome.
Since mutations in non-coding regions can be assumed to be neutral, the sequences
in such sections are used to make simple inferences regarding the demographic
history of the population based on coalescent theory. This seems another possible
way of estimating the background frequency spectrum. Information regarding the
Detecting Selective Sweeps 209

sequences surrounding non-synonymous mutations is then used to estimate the

action of selection.
The use of linkage disequilibrium measures to detect partial selective sweeps
is definitely a powerful approach and linkage disequilibrium measures do seem to
give a slight improvement in the power of tests based on the frequency spectrum
when a sweep has recently finished. More theoretical work on the form of linkage
disequilibrium at various stages of selective sweeps would be welcome and as
suggested earlier this might give us the insight to adopt methods adapted from
covariate modelling. Another possible approach of making use of the spatial form
of the data is to adapt the theory of Markov models (see Husmeier and McGuire
2002) or change points (see Avery and Henderson 1999) to such problems.
Finally, nearly all the ideas presented here were based on coalescent theory.
Simulating data using such reverse time processes is efficient, but results in a loss
of accuracy. Recently, algorithms have been formulated to simulate evolutionary
processes in forward time (see Hoggart et al. 2007; Padhukasahasram et al. 2008).
Development of this theory will allow us to carry out more accurate simulations and
develop new theory and practice.

Acknowledgements This work was carried out in the Department of Mathematics & Statistics at
the University of Limerick, Ireland, and was supported by Science Foundation Ireland, Grant No.
07/MI/012 (BIO-SI project, www3.ul.ie/bio-si).

References

Achaz, G. (2008). Testing for neutrality in samples with sequencing errors. Genetics, 179, 1409–
1424.
Akey, J. M., Zhang, G., Zhang, K., Jin, L., & Shriver, M. D. (2002). Interrogating a high-density
snp map for signatures of natural selection. Genome Research, 12, 1805–1814.
Avery, P. J., & Henderson, D. A. (1999). Detecting a changed segment in dna sequences. Applied
Statistics, 48, 489–503.
Barnes, M. R. (2006). Navigating the hapmap. Briefings in Bioinformatics, 7, 211–224.
Cannings, C. (1974). The latent roots of certain markov chains arising in genetics: A new approach,
i. haploid models. Advances in Applied Probability, 6, 260–290.
Carlson, C. S., Thomas, D. J., Eberle, M. A., Swanson, J. E., Livingston, R. J., Rieder, M. J.,
et al. (2005). Genomic regions exhibiting positive selection identified from dense genotype
data. Genome Research, 15, 1553–1565.
Consortium, T. I. H. (2003). The international hapmap project. Nature, 426, 789–796.
Durrett, R., & Schweinsberg, J. (2004). Approximating selective sweeps. Theoretical Population
Biology, 66, 129–138.
Fu, Y. X., & Li, W. H. (1992). Statistical test of neutrality of mutations. Genetics, 133, 693–709.
Hamblin, M. T., & Di Rienzo, A. (2000). Detection of the signature of natural selection in humans:
Evidence from the duffy blood group locus. American Journal of Human Genetics, 66, 1669–
1679.
Hein, J., Schierup, M., & Wiuf, C. (2004). Gene genealogies, variation and evolution: A primer in
coalescent theory. Oxford, UK: Oxford University Press.
210 D. Ramsey

Hoggart, C. J., Chadeau-Hyam, M., Clark, T. G., Lampariello, R., Whittaker, J. C., De Iorio, M.,
et al. (2007). Sequence-level population simulations over large genomic regions. Genetics, 177,
1725–1731.
Hudson, R. R. (2002). Generating samples under a wright-fisher neutral model of genetic variation.
Bioinformatics, 18, 337–338.
Husmeier, D., & McGuire, G. (2002). Detecting recombination with mcmc. Bioinformatics, 18,
345–353.
Innan, H., & Kim, Y. (2004). Pattern of polymorphism after strong artificial selection in a
domestication event. Proceedings of the National Academy of Sciences, 101, 10667–10672.
Innan, H., & Kim, Y. (2008). Detection of local adaptation using the joint sampling of polymor-
phism data in the parental and derived population. Genetics, 179, 1713–1720.
Jensen, J. D., Kim, Y., DuMont, V. B., Aquadro, C. F., & Bustamante, C. D. (2005). Distinguishing
between selective sweeps and demography using dna polymorphism data. Genetics, 170, 1401–
1410.
Jensen, J. D., Thornton, K. R., & Aquadro, C. F. (2008). Inferring selection in partially sequenced
regions. Molecular and Biological Evolution, 25, 438–446.
Jensen, J. D., Thornton, K. R., Bustamante, C. D., & Aquadro, C. F. (2007). On the utility of linkage
disequilibrium as a statistic for identifying targets of positive selection in nonequilibrium
populations. Genetics, 176, 2371–2379.
Johnson, P. L. F., & Slatkin, M. (2006). Inference of population genetic parameters in metage-
nomics: A clean look at messy data. Genome Research, 16, 1320–1327.
Kim, Y., & Nielsen, R. (2004). Linkage disequilibrium as a signature of selective sweeps. Genetics,
167, 1513–1524.
Kim, Y., & Stepham, W. (2002). Detecting a local signature of genetic hitchhiking along a
recombining chromosome. Genetics, 160, 765–777.
Kimura, M. (1971). Theoretical foundation of population genetics at the molecular level. Theoret-
ical Population Biology, 2, 174–208.
Kimura, R., Fujimoto, A., Tokunaga, K., & Ohashi, J. (2007). A practical genome scan for
population-specific strong selective sweeps that have reached fixation. PLoS One, 2, e286.
Lewontin, R. C., & Krakauer, J. (1973). Distribution of gene frequency as a test of the theory of
the selective neutrality of polymorphisms. Genetics, 74, 175–195.
Macpherson, J. M., González, J., Witten, D. M., Davis, J. C., Rosenberg, N. A., Hirsh, A. E.,
et al. (2008). Nonadaptive explanations for signatures of partial selective sweeps in drosophila.
Molecular and Biological Evolution, 25, 1025–1042.
Maynard Smith, J., & Haigh, J. (1974). The hitchhiking effect of a favourable gene. Genetics, 23,
23–35.
McVean, G., Spencer, C. C. A., & Chaix, R. (2005). Perspectives on human genetic variation from
hapmap project. PLoS Genetics, 1, e54.
Möller, D. A., Tenaillon, M. I., & Tiffin, P. (2007). Population structure and its effects on patterns
of nucleotide polymorphism in teosinte. Genetics, 176, 1799–1809.
Neuhauser, C., & Krone, S. M. (1997). The genealogy of samples in models with selection.
Genetics, 145, 519–534.
Nielsen, R., Hubisz, M. J., & Clark, A. G. (2004). Reconstituting the frequency spectrum of
ascertained snp data. Genetics, 168, 2373–2382.
Nielsen, R., Williamson, S., Kim, Y., Hubisz, M. J., Clark, A. G., & Bustamante, C. (2005).
Genomic scans for selective sweeps using snp data. Genome Research, 15, 1566–1575.
Padhukasahasram, B., Marjoram, P., Wall, J. D., Bustamante, C. D., & Nordborg, M. (2008).
Exploring population genetic models with recombination using efficient forward-time simu-
lations. Genetics, 178, 2417–2427.
Pan, J., & MacKenzie, G. (2003). On modelling mean-covariance structures in longitudinal studies.
Biometrika, 90, 239–244.
Pfaffelhuber, P., Lehnert, A., & Stephan, W. (2008). Linkage disequilibrium under genetic
hitchhiking in finite populations. Genetics, 179, 527–537.
Detecting Selective Sweeps 211

Project, S. S. (2004). SeattleSNPs. NHLBI Program for Genomic Applications. Seattle, WA:
SeattleSNPs. http://pga.gs.washington.edu
Sabeti, P. C., Reich, D. E., Higgins, J. M., Levine, H. Z., Richter, D. J., Schaffner, S. F., et al. (2002).
Detecting recent positive selection in the human genome from haplotype structure. Nature, 419,
832–837.
Slatkin, M. (2000). A coalescent view of population structure. In R. S. Singh & C. B. Krimbas
(Eds.), In evolutionary genetics. New York: Cambridge University Press.
Spencer, C. A., & Coop, G. (2004). Selsim: A program to simulate population genetic data under
coalescent models. Bioinformatics, 20, 3673–3675.
Tajima, F. (1989). Statistical method for testing the neutral mutation hypothesis by dna polymor-
phism. Genetics, 123, 585–595.
Teshima, K. M., Coop, G., & Przeworski, M. (2006). How reliable are empirical genomic scans
for selective sweeps? Genome Research, 16, 702–712.
Voight, B. F., Kudaravalli, S., Wen, X., & Pritchard, J. K. (2006). A map of recent positive selection
in the human genome using empirical scores. PLoS Biology, 4, e72.
Williamson, S. H., Hernandez, R., Fledel-Alon, A., Zhu, L., Nielsen, R., & Bustamante, C. D.
(2005). Simultaneous inference of selection and population growth from patterns of variation in
the human genome. Proceedings of the National Academy of Sciences, 102, 7882–7887.
Yamasaki, M., Tenaillon, M. I., Bi, I. V., Schroeder, S. G., Sanchez-Villeda, H., Doebley, J. F., et al.
(2005). A large-scale screen for artificial selection in maize identifies candidate agronomic loci
for domestication and crop improvement. Plant Cell, 17, 2859–2872.
Zeng, K., Shi, S., & Wu, C. I. (2007). Compound tests for the detection of hitchhiking under
positive selection. Molecular and Biological Evolution, 24, 1898–1908.
A Mixture Model and Bootstrap Analysis
to Assess Reproductive Allocation in Plants

Caroline Brophy, D. Gibson, P.W. Wayne, and J. Connolly

Abstract In this paper we discuss issues that arise in predicting from complex
models for the analysis of reproductive allocation (RA) in plants. Presenting
models of RA requires prediction on the original scale of the data and this can
present challenges if transformations are used in modelling. It is also necessary
to estimate without bias the mean level of RA as this may reflect a plant’s ability
to contribute in the next generation. Several issues can arise in modelling RA
including the occurrence of zero values and the clustering of plants in stands which
can lead to the need for complex modelling. We present a two-component finite
mixture model framework for the analysis of RA data with the first component a
censored regression model on the logarithmic scale and the second component a
logistic regression model. Both components contain random error terms to allow for
potential correlation between grouped plants. We implement the framework using
data from an experiment carried out to assess environmental factors on reproductive
allocation. We detail the issues that arose in predicting from the model and present
a bootstrap analysis to generate standard errors for the predictions from and to test
for comparisons among predictions.

Keywords Bias in back-transformation • Bootstrap analysis • Censored regres-

sion • Mixture model • Reproductive allocation

C. Brophy ()
Department of Mathematics & Statistics, National University of Ireland Maynooth, Maynooth,
Co Kildare, Ireland (Current address)
UCD School of Mathematical Sciences, Environmental & Ecological Modelling Group,
University College Dublin, Belfield, Dublin 4, Ireland
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 213

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__14,
© Springer International Publishing Switzerland 2014
214 C. Brophy, et al.

1 Introduction

Several statistical issues can arise in the analysis of reproductive allocation (RA) in
plants and these include the occurrence of zero values and the clustering of plants
in pots. A suitable framework for the analysis of RA data that deals with these
issues is presented in Brophy et al. (2007). An application of this model to data
from an experiment assessing environmental factors on RA is presented in Brophy
et al. (2008). In this paper we present the issues that arose when predicting from the
model fitted there. The model is a two-group finite mixture model (McLachlann and
Peel 2000) in which the sub-model for the first group is a censored regression model
(Schmid et al. 1994) on the logarithmic scale and the sub-model for the second group
is a logistic regression model (Collett 1993). When predicting from this model it
is desirable to predict on the scale of the original data, i.e. on the RA scale, but
this can lead to bias in predictions from back-transforming model components. It is
also desirable to predict at the mean RA without bias as this may reflect a plant’s
ability to contribute in the next generation and biased predictions here could provide
misleading model inference. We present the predictions and a bootstrap analysis to
calculate standard errors for the predictions and to test for comparisons among them.

2 Methods

Log-log linear allometric regression has often been used to describe the relationship
between reproductive allocation (RA, defined here to be the ratio of reproductive
biomass to aboveground biomass) and aboveground biomass (Harper 1977). How-
ever, frequently many plants in experiments do not reproduce. This has previously
been dealt with using censored regression (Schmid et al. 1994). Censored regression
is a suitable method when the sole cause of non-reproduction is small plant size;
however, this may not always be the case. We propose using a modelling approach
that allows for two possible groups within the RA responses (Brophy et al. 2007).
The first group contains all reproducing and some non-reproducing plants, but all
plants are assumed to have the ability to reproduce. The second group contains the
remaining non-reproducing plants which are assumed to be unable to reproduce
under their experimental conditions. This framework was applied to data from an
experiment examining reproductive allocation in plants where approximately 40 %
of plants did not reproduce (Fig. 1) (Brophy et al. 2008).
The experiment was carried out as follows. Seeds of Sinapis arvensis were
directly sown into 5.5 L, 25 cm diameter, round pots. Seeds were sown at six
densities: 1, 2, 4, 8, 16 and 32 plants pot1 and were grown under two CO2
concentrations, 350 or 700 L L1 . There were 6 replicate pots per CO2 by density
combination, except for combinations at the lowest density which were replicated 12
times, giving 84 pots in total. Individual plants were harvested and the aboveground
biomass (DM) and the reproductive biomass (flowers and fruits) were recorded for
A Mixture Model and Bootstrap Analysis 215

Fig. 1 Log(reproductive biomass) versus log(aboveground biomass) for individual plants of

Sinapis arvensis. The log of non-reproducing plants is represented by 5.299 and ca. 40 %
of individuals in the experiment did not reproduce. There appears to be a strong allometric
relationship between log(reproductive biomass) and log(aboveground biomass); however, the
circled values suggest a second group of values that do not follow this relationship

each of the 704 available plants. Further details are in Wayne et al. (1999). A prefix
of L indicates the natural log scale, e.g. LDM D log.DM/.
Figure 1 suggested two groups within the experimental data indicating that the
framework of Brophy et al. (2007) would be suitable for its analysis. Using this
framework we developed a mixture model (McLachlann and Peel 2000) in which
the proportions of the two groups were (1-p) and p respectively. For the first group
we used a variant of censored regression to model the relationship between the RA
and aboveground biomass (DM) and CO2 . In the second group we modelled the
proportion p as a logit function of DM, CO2 and Ratio (size of an individual plant
relative to the average size of plants in their pot). The model for each group included
a random pot effect to allow for correlation between plants within a pot and we
tested these terms using a likelihood ratio test. The likelihood function was of the
form:
Y Y
Likelihood D ..1  p/f .LRAij // ..1  p/F .LRAij / C p/; (1)
RAij >0 RAij D0

where

1 .LRAij  ˇ0  ˇ1 LDMij  ˇ2 CO2  uj /2
f .LRAij / D q exp (2)
212 212
216 C. Brophy, et al.

and F .LRAij / D

Z ij
LDM 
1 .LRAij  ˇ0  ˇ1 LDM ij  ˇ2 CO2  uj /2
q exp dLRAij (3)
212 212
1

and is the log of the smallest value of reproductive biomass for reproducing
plants. Model parameters were estimated using maximum likelihood and relevant
interactions were tested. p can be included in the model as a constant or as a function
of experimental variables. We modelled p as a function of DM, CO2 and Ratio
giving (after testing relevant terms) the models:

LRAij D ˇ0 C ˇ1 LDM ij C ˇ2 CO2 C uj C ij ; (4)

p
log. /D˛0 C ˛1 LDM ij C ˛2 CO2 C ˛3 LRati oij C˛4 LDM ij
LRatioij C wj ;
1p
(5)

where ˇ0 , ˇ1 and ˇ2 are regression coefficients, uj is a pot-specific random effect

and ij is the residual term; ij and uj are assumed to be NID.0; 12 / and NID.0; 22 /
respectively. And where ˛0 , ˛1 , ˛2 , ˛3 , and ˛4 are regression coefficients; wj is a
pot-specific random effect assumed to be NID.0; 32 / and its covariance with uj is
.
To assess the model, RA was predicted at a range of values of the covariates,
CO2 , DM and Ratio. There were three steps involved in predicting from the model:
predicting RA conditional on being from Group 1 using the estimated Eq. (4),
predicting the probability of being in Group 2 from the estimated Eq. (5) and
predicting overall RA by combining these two predictions. Several issues arose
in predicting at the mean overall RA. These included the back-transformation of
random and fixed components and allowing for the censored distribution of the
relationship in Eq. (4). We predicted from Eq. (4) allowing for the censored nature of
the relationship using first moment equations for censored distributions as described
in Jawitz (2004). Using moment equations also allowed the bias caused by the
back-transformation of the fixed components from the log scale McCulloch and
Searle (2001) to be adjusted for. Predicting the probability of being in Group 2
was calculated by inserting the covariate values required into the estimated Eq. (5)
and applying the anti-logit function. However, this predicts the probability at the
median value of the random term wj on the anti-logit scale, not it’s mean. We carried
out a simulation study to estimate the bias caused by this and multiplied predicted
probabilities by the relevant bias adjustment.
Standard errors for predictions and relevant comparisons among predictions were
calculated using a bootstrap analysis (Efron and Tibshirani 1986, 1993). Since
we cannot assume independence of the individual plants due to the clustered (in
pots) nature of the data, in applying the bootstrap method we sampled the data
at two levels; first we sampled pots with replacement within CO2 by density
combinations and secondly we sampled plants with replacement within pots. The
A Mixture Model and Bootstrap Analysis 217

Fig. 2 Predicted RA (%) vs. plant biomass (g) for various combinations of Ratio and CO2 . The
vertical lines aid the comparisons at 3 and 6 g discussed in Sect. 3. Adapted from Fig. 2a in Brophy
et al. (2008) with kind permission of Oxford University Press. Thec Author 2008, published by
Oxford University Press on behalf of the Institute of Botany, Chinese Academy of Sciences and
the Botanical Society of China. All Rights Reserved

first level of sampling was done within each CO2 by density combination to retain
the structure of the original experiment design. We repeated the sampling process
1,000 times. We fitted the final model to each bootstrapped dataset and computed
the prediction or comparison of interest for each fitted model. The standard error
over the bootstrapped predictions estimated the standard error for the prediction
from the model. Comparisons of predictions were tested for significance using BCa
confidence intervals (Efron and Tibshirani 1993).

3 Results

Predicted RA, assessed jointly from the two components of the mixture model, is
presented in Brophy et al. (2008) and in Fig. 2 for a range of covariate values. The
effect of CO2 was always positive and often significant. At DM D 3 g and RatioD1
predicted RA was 3.7 % at ambient and 4.9 % at elevated (Fig. 2) and these two
values differed significantly .p < 0:05/. We found a strong positive effect of Ratio
(relative size within a pot): a plant with biomass 3 g grown at ambient CO2 allocated
2.5 % on average to reproduction if it was half its pot average size (Ratio D 0.5) but
this increased to 3.7 and 4.4 % if the plant was equal to (Ratio D 1) or double
(Ratio D 2) its pot average size respectively (each of these predictions differed;
p < 0:05). At plant biomass = 6 g and at ambient CO2 there was a significant
difference between predicted RA at Ratio = 1 and 2 .p < 0:05/. Biological
interpretations from these results are presented in Brophy et al. (2008).
218 C. Brophy, et al.

4 Discussion

Many issues arise when predicting and performing inference using predictions from
a complex model such as the one described here. These include predicting from a
censored distribution, combining predictions from multiple parts of a mixture model
and back-transforming predictions to an original scale when the model includes
fixed and random components. When assessing RA data, it is particularly necessary
to interpret models on the scale of the original data and to predict the mean RA
without bias since RA values represent a measure of a plant’s ability to continue
into the next generation.
Had the biases not been adjusted here the biological interpretations would have
differed significantly. In particular, the back-transformation of the fixed terms in
the censored regression model from the log scale would have resulted in the RA
values being underestimated by approximately 25 %. Failing to allow for the back-
transformation of the random term in the logistic regression model would have
also resulted in significant bias. An alternative to using the simulation study we
applied here to adjust for bias in the back-transformation of the random term in the
logistic regression model would be to model p using a probit analysis which has
an explicit formulation for the adjustment needed in back-transforming the random
term (McCulloch 1994; McCulloch and Searle 2001). While we could replace our
logistic regression model with a probit model, the empirical methodology we used
is a general approach for adjusting for bias in situations where explicit formulae are
not available.

Acknowledgements CB is the holder of an Environmental Protection Agency (Ireland) Doctoral

Scholarship, JC has received support from an Enterprise Ireland International Collaboration grant
for this work, and DG received support from the United States Department of Agriculture.

References

Brophy, C., Gibson, D., Wayne, P., & Connolly J. (2007). A modelling framework for analysing
the reproductive output of individual plants grown in monoculture. Ecological Modelling, 207,
99–108.
Brophy, C., Gibson, D. J., Wayne, P. M., & Connolly J. (2008). How reproductive allocation and
flowering probability of individuals in plant populations are affected by position in stand size
hierarchy, plant size and CO2 regime. Journal of Plant Ecology, 1, 207–215.
Collett, D. (1993). Modelling binary data. London: Chapman & Hall/CRC.
Efron, B., & Tibshirani, R. (1986). Bootstrap methods for standard errors, confidence intervals and
other measures of statistical accuracy. Statistical Science, 1, 54–77.
Efron B., & Tibshirani, R. (1993) An introduction to the bootstrap. New York: Chapman & Hall.
Harper, J. (1977). Population biology of plants. London: Academic.
Jawitz, J. (2004). Moments of truncated continuous univariate distributions. Advances in Water
Resources, 27, 269–281.
McCulloch, C. (1994). Maximum likelihood variance components estimation for binary data.
Journal of the American Statistical Association, 89, 330–335.
A Mixture Model and Bootstrap Analysis 219

McCulloch, C., & Searle, S. (2001). Generalised, linear, and mixed models. In W. Shewhart &
S. Wilks (Eds.), In probability and statistics. New York: Wiley.
McLachlann, P., & Peel, D. (2000). Finite mixture models. New York: Wiley.
Schmid, B., Polasek, W., Weiner, J., Krause, A., & Stoll, P. (1994). Modelling of discontinuous
relationships in biology with censored regression. American Naturalist, 143, 494–507.
Wayne, P., Carnelli, A., Connolly, J., & Bazzaz, F. (1999). The density dependence of plant
responses to elevated CO2 . Journal of Ecology, 87, 183–192.
On Model Selection Algorithms
in Multi-dimensional Contingency Tables

Susana Conde and Gilbert MacKenzie

Abstract We present a review focussed on model selection in log-linear models and

contingency tables. The concepts of sparsity and high-dimensionality have become
more important nowadays, for example, in the context of high-throughput genetic
data. In particular, we describe recently developed automatic search algorithms
for finding optimal hierarchical log-linear models (HLLMs) in sparse multi-
dimensional contingency tables in R and some LASSO-type penalized likelihood
model selection approaches. The methods rely, in part, on a new result which iden-
tifies and thus permits the rapid elimination of non-existent maximum likelihood
estimators in high-dimensional tables.

Keywords Comorbidity • Hierarchical log-linear model • LASSO penalized

likelihood • Smooth LASSO • Sparse high-dimensional contingency table •
Stepwise search algorithms

1 Introduction

Our interest in model selection in contingency tables stemmed originally from an

analysis of comorbidity data. A comorbidity is a disease that coexists with an index
disease under study (Feinstein 1970). Typically, comorbidity data comprise several
comorbidities. The scientific interest lies in the study of the dependence structure
between the binary comorbidities. Such data are typically high-dimensional and
sparse, which poses many challenges for the statistical analysis.
Traditionally, binary comorbidities have been measured using comorbidity
indices. These are linear combinations of the comorbidities with some specified

S. Conde ()
Department of Mathematics, Imperial College, London, UK
e-mail: [email protected]

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 221

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5__15,
© Springer International Publishing Switzerland 2014
222 S. Conde and G. MacKenzie

weights. The weights can be calculated statistically (e.g. Charlson Comorbidity

Index), or decided clinically by a doctor (e.g., Davies index or others (Davies
et al. 2002; Hall et al. 2004)).Thus in this modelling scheme, possible associations
between comorbidities are not taken into account (Hall et al. 2004; Conde and
MacKenzie 2007). When two or more diseases exist together, they can reflect a risk
of mortality greater than the sum of their individual weights. Of course interaction
terms could be included, but surprisingly this is rare, with no examples appearing
in the literature reviewed so far (Harnett, et al., 2006, A 5 year prospective study of
factors which influence selection for and survival on dialysis, unpublished; Charlson
et al. 1987; Davies et al. 2002; Hall et al. 2004).
Conde and MacKenzie (2007) propose a more comprehensive statistical
approach, which basically consists of constructing a multi-dimensional contingency
table and using a log-linear model (Agresti 2002). Here the dependence structure
is measured by interactions rather than by correlations. The latter are the natural
measures for multivariate Gaussian data, but inappropriate for binary data. The
log-linear modelling scheme allows a richer dependence structure between the
comorbidities and can be applied to any other type of categorical variables arising
in other settings. The main class of models employed is that of Hierarchical Log-
Linear (HLL) models (Birch 1963; Goodman 1971; Agresti 2002). Other families
of models may be useful: for example, Graphical Models (GM) (Darroch et al.
1980), which are a subset of HLL Models; Graphical Models are implemented in
the Mixed Interaction Modelling (MIM) program (Edwards 2000).
In this paper we present a review of statistical approaches for analyzing such
data. We focus on methods for model selection including stepwise search algorithms
and penalized likelihood approaches (Conde and MacKenzie 2008; Conde 2011).
First we formulate the log-linear model, describe flat tables and design matrices,
and present inference in Sect. 2. Then in Sect. 3 we comment on some log-linear
model classes: hierarchical, graphical, decomposable. We then describe sparseness
including a new definition in Sect. 4. We review some more recent tests of goodness
of fit and methods for analysing residuals in Sect. 5. Next, in Sect. 6, we describe
the stepwise search algorithms and penalized likelihood model selection methods.
Finally, we present some illustrations of the methods in Sect. 7 and end with a short
discussion in Sect. 8.

2 Model Formulation

2.1 Contingency Table

Consider a vector of p binary random variables (C1 ; : : : ; Cp ) where each Cr has

the event space XCr D f1; 2g for r D 1; : : : ; p and, for example, 1 = “absent”
and 2 = “present”. For the sth independent subject, s D 1; : : : ; n, let .c1s ; : : : ; cps /
p

be an element of the joint event space XC D rD1 XCr and where 
stands for
Model Selection for Binary Data 223

Table 1 Example of raw Comorbidities

data
MI CHF  PVD
Patient C1 C2  Cp
1 1 1  1
2 2 1  1
3 1 1  1
4 1 1  1
5 2 2  2
:: :: :: :: ::
: : : : :
n 2 2  1
MI myocardial infarction, CHF congestive heart fail-
ure, PVD peripheral vascular disease

the Cartesian product. The cardinality of XC is q D 2p . Suppose .c1s ; : : : ; cps / D

.1; : : : ; 1/, then the subject is not comorbid. PThe total number of subjects observed
sD1 ½f.c1s ; : : : ; cps / D .1; : : : ; 1/g,
n
to be free from comorbidity is then y1 D
where ½ is the indicator function such that ½fa D bg = 1. The quantity y1 is called
the observed frequency of cell .1; : : : ; 1/. Similarly, yq is the observed frequency of
the last cell .2; : : : ; 2/.
Table 1 shows an example of a raw data set which can be transformed into
a p-dimensional contingency table by counting. Table 2 P presents an example of
q
such a table formed with p D 3 binary variables where i yi D n is the total
number of observations and i is the true, but unknown, probability of a subject
belonging to cell i . The i , (i D 1; : : : ; q D 2p D 8) form a probability distribution
and the vector of random variables .Y1 ; : : : ; Yq / follows a Multinomial distribution,
MN.n; 1 ; : : : ; q / with probability function

 nŠ Y q
pr Y1 D y1 ; : : : ; Yq D yq D Qq .i /yi ;
nn i D1 yi Š i D1
P
where i D ni and i D n.
More generally, let gr be the number of categories of variable Cr with gr  2,
g
r D 1; : : : ; p so XCr WD f1; : : : ; gr g. The gr categories may be labelled lr1 ; : : : ; lr r ,
g
r D 1; : : : ; p. We can associate flr1 ; : : : ; lr r g WD f1; : : : ; gr g; r D 1; : : Q
: ; p and
p
consider the corresponding joint event space XC containing exactly q WD rD1 gr
cells. As above, we assume n independent realizations of the p variables and then,
mutatis mutandis, we again obtain a multinomial random variable.

2.2 Bijective Mapping

In Table 2 we have written the frequencies using one index (yi ) or with multiple
indices (yj1 j2 j3 ), where jr D 1; 2, r D 1; : : : ; 3. However, in general, for p
224 S. Conde and G. MacKenzie

Table 2 A three-dimensional contingency table

C3 .1/ C3 .2/
C2 .1/ C2 .2/ C2 .1/ C2 .2/
C1 .1/ y1 a .y111 / 1 y3 .y121 / 3 y5 .y112 / 5 y7 .y122 / 7
C1 .2/ y2 .y211 / 2 y4 .y221 / 4 y6 .y212 / 6 y8 .y222 / 8
a
Design matrix notation

categorical variables each with gr categories and r D 1; : : : ; p O’Flaherty and

MacKenzie (1982) showed that

X
p
f1; : : : ; qg ! Cr
rD1

i 7! .j1 ; : : : ; jp /

is a bijective map. The bijective map between the two index sets is not unique (for
example, leftmost, or rightmost subscript could vary fastest). When the leftmost
subscript varies fastest this is known as “Standard Order” (Montgomery 2001,
Chap. 6) and we adopt this convention hereafter.
Given one of the bijective mappings i 7! .j1 ;    ; jp /, we can define an observed
contingency table as y WD y1 ; : : : ; yq 2 Rq such that yi D #f.j1 ; : : : ; jp /; C1 D
j1 ; : : : ; Cp D jp g representing the count of the i th element in XC . The quantity
P
yi is fixed. Then we can represent the contingency table as the matrix

C1 C2    Cp y
0 1
1 1  1 y1
B 2 1  1 y2 C
B C
B   C
B C
B C
Cg1 ;:::;gp WD B g1 1  1 C:
B C
B 1 2  1 C
B C
@    A
g1 g2    gp yq

2.3 Flat Tables

It is convenient to represent multi-dimensional contingency tables as flat tables,

because the latter are more easily handled as design matrices (i.e., model matrices).
See Conde (2011, p. 11) for the flat table representation of the previous example.
The flat tables produced by R contain only the main effect columns of the
corresponding design matrix. Usually interaction effects are required and their
associated columns must be generated in the design matrix. In a model with p
Model Selection for Binary Data 225

Qp.C1 ; : : : ; Cp / and each with gr levels, r D 1; : : : ; p, there are exactly

factors
q D rD1 gr effects in total including the constant. A design matrix which contains
the columns representing all these effects is called saturated. Design matrices
containing interactions of any order may be generated recursively from flat tables by
an algorithm (MacKenzie and O’Flaherty 1982).

2.4 Sampling Schemes

In this paper we mainly focus on the multinomial sampling scheme. Other common
sampling schemes are the independent multinomial (Birch 1963) (when some sums
of the yi s are fixed,
Pq for example, say some row marginal totals), or the independent
Poisson (when i D1 yi is not fixed but random; then Y1 ; : : : ; Yq are independent
Poisson random variables with parameters 1 ; : : : ; q ). A clear example of each of
these schemes is in Agresti (2002, pp. 40–41).
Fisher (1922) showed that q independentP Poisson random variables with param-
eters i (i D 1; : : : ; q) conditioned on Yi D n, have a multinomial distribution
.n; i / where i D i =n. This is now a well-known result (Agresti 2002, pp. 8–9;
Christensen 1997, p. 20; Bishop et al. 1975, p. 441).

2.5 Log-Linear Modelling

Let p be the number of binary comorbidities or any other variables in consideration,

with p any fixed integer  2 (Conde and MacKenzie 2007). Consider the
p-dimensional contingency table with exactly q D 2p cells. The possible depen-
dencies between the p variables can be quantified using a log-linear model (Bishop
1969; Goodman 1971; MacKenzie and O’Flaherty 1982) in which the dependence
is specified by the presence of interactions. In this framework, dependence is more
naturally measured in terms of interactions rather than in terms of correlation.
If we define i WD E.Yi /, the expected value in the i th cell, we consider a log-
linear regression model with k parameters (with k  q) (MacKenzie and Conde
2014; Conde and MacKenzie 2014):

X
k
ln .i / D aij j
j D1

(Birch 1963; Agresti 2002; Bishop et al. 1975; Christensen 1997), where
0 1
a11    a1k
B : :: : C
B
A D @ :: : :: CA;
aq1    aqk
226 S. Conde and G. MacKenzie

is the .q  k/ design matrix, where k is the number of linearly independent

parameters and , is a vector of unknown parameters measuring the influence of the
constant, main effects and interactions on the response, and belongs to a parameter
space ‚. In general we consider models such that rank A D k. If the design matrix
has maximal order, namely q columns, it is called saturated.
There are mainly two design matrix coding schemes used in this area. The Yates’
coding scheme uses .C1; 1/ to encode the presence or absence of a comorbidity
(Montgomery 2001, pp. 222–223, 242) and the binary scheme uses .C1; 0/. There
are advantages and disadvantages to each scheme. Yates’ code is D-optimal and
the columns of the design matrix, A, are orthogonal which has implications for
increased stability, in terms of the condition number, in sparse tables. On the
other hand the binary coding scheme is useful for identifying inestimable effects
(MacKenzie and Conde 2014).
We note that a log-linear regression model is a generalized linear model
(McCullagh and Nelder 1997).

2.6 Inference

Given a sample y, the likelihood is

 nŠ Y q
L 1 ; : : : ; q j y D Qq .i /yi
nn i D1 yi Š i D1
P
such that i D n. If all the expected values are > 0 then the log-likelihood is:
!
nŠ Y q
` .1 ; : : : ; n j y/ D ln Qq .i / yi
nn i D1 yi Š i D1

X
q
DKC yi ln .i /
i D1
P
for a certain constant K and i D n. The log-likelihood with respect to  is
0 1
X
q
Xk
`. j y/ / yi @ aij j A ;
i D1 j D1

subject to
0 1
X
q
X
k
exp @ aij j A D n:
i D1 j D1
Model Selection for Binary Data 227

Birch (1963) showed that, if the maximum likelihood estimators Oi exist, they
coincide for
P multinomial
 sampling, for independent Poisson (provided that
Pq k O
i D1 exp a 
j D1 ij j D n), and for independent multinomial sampling
(provided that the parameter/s corresponding to the fixed marginal sum/s are
included in the model). Then, for estimation, the independent Poisson scheme
can always be used (Dobson 2002, p. 164).
The score function is
2 0 13
@` X q
X k
./ D air 4yi  exp @ aij j A5
@r i D1 j D1

and the .r; s/th element of the Fisher information matrix is:

@2
ir;s ./ D E  `./
@r @s
0 1
X
q
X
k
D air ais exp @ aij j A ;
i D1 j D1

where 1  r; s  k, say, where k D dim./  n with equality in the saturated case,

i.e., k D n. Accordingly, the asymptotic (k  k) variance-covariance matrix is then
†./ D I./1 where I./ has typical element ir;s ./.

3 Model Classes

It is convenient to introduce some additional notation for the components of the

parameter  in the case of models with p binary variables. Accordingly, we denote
 D .1 ; 2 ; : : : ; pC1 ; pC2 ; : : : / by (c0, c1, : : : , cp, c1c2, : : : ).
A HLL model (Birch 1963) is a log-linear model such that, if a parameter that
represents an interaction is zero, the interaction is said to be null. Moreover, all the
higher order interactions that include it are also null (Birch 1963; Agresti 2002).
Equivalently, if an interaction is non-null, all the lower order interactions included
in it (and their corresponding main effects) are non-null. For example, the model
M1 D fc0, c1, c2, c1c2g is hierarchical. A non-hierarchical model is a model that is
not hierarchical. The model M2 D fc0, c2, c1c2g is non-hierarchical because c1c2
is non-null and the main effect c1 is null.
A HLL model is characterised by its generating set (Krajewski and Siatkowski
1990). The generating set is the set of maximal effects, maximal in the sense of
the inclusion relation, and it is unique (Edwards and Havránek 1985). For example,
228 S. Conde and G. MacKenzie

Table 3 Generating sets of Classical set Rest of the hierarchical models

all the possible hierarchical
models with p D 3 f.null/g fc1, c2g
fc1g fc1, c3g
fc2g fc2, c3g
fc3g fc1, c2c3g
fc1c2g fc2, c1c3g
fc1c3g fc3, c1c2g
fc2c3g fc1c2, c1c3g
fc1c2c3g fc1c2, c2c3g
fc1c3, c1c3g
fc1, c2, c3g
fc1c2, c1c3, c2c3g

model M1 has generating set G D fc1c2g and the model with generating set G D
fc1, c2c3g is

M3 D fc0; c1; c2; c3; c2c3g:

Non-hierarchical models do not have a generating set. Furthermore, they are

“scientifically uninteresting” because the significance of effects in the model is not
invariant to the choice of design matrix (MacKenzie and Conde 2014).
Let us consider the set Ep of all the 2p effects (i.e., constant, main effects and
interactions) in the saturated log-linear model. For example, for p D 3,

E3 D fc0, c1, c2, c3, c1c2, c1c3, c2c3, c1c2c3g:

We define a kind of multiplication rule

in Ep by: for all E1 ; E2 2 Ep ; E1
E2 D
E1 E2 : where c0 is the identity element. For example, c1
c3 D c1c3; c1
c2c3
D c1c2c3, c0
c1=c1
c0=c1 from which we have that .Ep ;
/ is an Abelian group
(Lang 1992).
Table 3 displays all the possible hierarchical models for three variables. In the
left column, we can see the classical set, where the generating set is composed only
by one element. In the right column, the rest of hierarchical models.
Lemma. A model is decomposable if and only if it is graphical and its graph
contains no chordless cycles of length exceeding three.
The simplest nongraphical model is f.A; B/ ; .A; C /; .B; C /g (the saturated model
has the same graph of association and it is graphical). The simplest graphical
model with at least one chordless cycle of length exceeding three is the model
fAB; BC; CD; DAg. It follows that decomposable models are characterised by two
key properties: (a) that the ML estimating equations have explicit solutions and (b)
the models can be interpreted in terms of conditional independence, independence
and equiprobability (Darroch et al. 1980).
Model Selection for Binary Data 229

The classes of models are then:

fDecomposableg  fGraphicalg  fHierarchicalg  fLog-Linearg:

p
For p factors, there are exactly 22 1 possible models, e.g., 8:988466  10307
possible models when p D 10. See Darroch et al. (1980) for the total number of
hierarchical, graphical, or decomposable models for p  5. The most appropriate
model is found by searching this set.

4 Sparseness

Agresti (2002, p. 391) defined a sparse table as a table with small frequencies in
some cells. For more details, see Hu (1999). Suppose a sparse table has some zeros.
These can be either structural zeros, which are indeed cells for which “observations
are impossible” (i.e., j D 0), or sampling zeros (the observed value is 0 in that
sample, but j > 0 and then it may be non-zero in another sample).

4.1 Structural Zeros

Goodman (1968) and Mantel (1970) analyse a 22 contingency table with structural
zeros. They define an incomplete table as a table with structural zeros. Mantel
follows and completes the work from Goodman (1968) with respect to models of
quasi-independence in 2  2 tables which are incomplete. Fienberg (1972) used
quasi-log-linear models (like log-linear models but without considering the zero
cells). All these authors use the method of maximum likelihood to estimate the
parameters. Baker et al. (1985) note that, if the zeros are structural (in a sparse
table), this does not affect the existence of the MLEs because those cells are omitted
from the analysis.

4.2 Sampling Zeros

In a sparse table with sampling zeros, it may happen that the MLEs of the parameters
of a log-linear model may not exist. Haberman (1970, pp. 51–52) and more recently
Fienberg and Rinaldo (2006, pp. 15–17) show some examples of tables with some
zeros, where the MLEs do not exist. Glonek et al. (1988) summarised the basic
results about existence of MLEs in the class of HLL models. Also they determined
the exact conditions which must be satisfied for the MLEs to exist (see below). Baker
et al. (1985) said that the MLEs of the expected values always exist, if instead of
using the equation of a log-linear model, we define the model in a more general
230 S. Conde and G. MacKenzie

way (constraints definition). Fienberg and Rinaldo (2006, 2012) used algebraic
statistics and geometry to show the conditions of the existence of the MLEs in a
(sparse) multi-dimensional contingency table with independent Poisson and product
multinomial sampling schemes.
MacKenzie and Conde (2014) presents a new result concerning the nonexistence
of maximum likelihood estimators. This result is based on a simple calculation that
involves a binary design matrix. Hence, we can know exactly which effects have
non-existent MLEs. Moreover, the result uses the properties of hierarchical models
in relation to detecting non-existent MLEs.

4.3 Definition of Sparsity

If some effects cannot be estimated the saturated model cannot be fitted in the
usual way, but we may consider a model which is maximal, i.e., which fits all of
the available degrees of freedom. The question of how to identify the maximal
model then arises. MacKenzie and Conde (2014) prove a theorem to identify the
redundant columns of A and generate a design matrix for the remaining estimable
effects. It turns out that the resulting maximal model is always hierarchical. This
procedure is very fast for high-dimensional tables. Although it nearly always
identifies the maximal model, in some pathological cases further work involving
the condition number of †./ (Wissmann et al. 2007) is required. Backwards
elimination methods can be used when a maximal model has been identified. The
proposed scheme can also handle cases when the number of parameters being fitted
is greater than the number of observations.

5 Goodness of Fit and Residuals

As for the goodness of fit tests of a certain model, Hu (1999) highlights that in sparse
tables, the Pearson statistic (written here with the estimated expected values):

X q
.yi  O i /2
X2 D
i D1
O i

(Pearson 1900), or the likelihood ratio test statistic:

X
q 
yi
G2 D 2 yi ln
i D1
O i

(Wilks 1935, 1938), where yi and O i are respectively the observed, and maximum
likelihood estimate of the expected value in the i th cell; can be seriously flawed,
Model Selection for Binary Data 231

and proposes, instead of the asymptotic 2 distribution for these statistics (with  D
number of free parameters), a posterior predictive check (PPC) distribution (which is
asymptotically a gamma). Kim et al. (2008), for sparse tables without any structural
zero, compare X 2 , D 2 and Lr , where

X q
yi
D2 D X 2  ;
i D1
Oi

O i >0


(Zelterman 1987); and Lr is a new goodness of fit test statistic in log-linear models
with conditional independencies (Maydeu-Olivares and Joe 2005); they recommend
the use of D 2 or Lr . As the contingency table becomes sparser (i.e., 2p =n gets
larger), the distribution of D 2 becomes more skewed.
The adjusted residuals are

stdri
adjri D
1=2 ;
1  hO ii

where hO ii is the estimate of the i th diagonal element of the Hat matrix which is

H D W 1=2 A.AT WA/1 AT W 1=2

(Agresti 2002), where A has been defined previously and W is diagonal (q  q) with
the expected values in the diagonal (for example Agresti 2002, p. 339); and

yi  O i
stdri D ;
O .Yi /g1=2
fvar

P 
called the standardized residuals, where O i D exp k O
j D1 aij j , the estimated
expected values, i D 1; : : : ; q.
The adjusted residuals are preferable to the standardised ones, as they have
asymptotic standard normal distributions (Christensen 1997). If the random variable
is a multinomial, then

O .Yi / D nO i .1  O i /
var
O i
D O i .1  /:
n

If the random variables are independent Poisson, then varO .Yi / D O i ; thus
the standardised residuals are the square roots of the components of the Pearson
statistic.
232 S. Conde and G. MacKenzie

6 Model Selection Methods

We are now interested in finding a good log-linear model in terms of fit. Typically,
the saturated model space containing 2p terms must be searched to find a more
parsimonious model which is “best” supported by the data. This is particularly
onerous in high-dimensional problems. Thus, it can be useful to have an automatic
search algorithm at our disposal.
The SPSS software package has an automatic algorithm (the HILOGLINEAR
procedure) in order to find a best-fitting HLL model in a contingency table.
However, the maximum number of variables that can be included in the model
search, is limited to 10.
Conde and MacKenzie (2007, 2008) reimplemented the SPSS algorithm in R
and constructed other similar search algorithms. The algorithms are backwards
elimination, backwards elimination 2 (BE, BE2), forward selection (FS) with
terminology taken from Goodman (1971); and MacKenzie–Conde Backwards
Elimination (MCBE); all the algorithms can be used with any number of binary
variables (but in practice available computer memory limits the performance when
p becomes large).
Other approaches include penalized likelihood. For example, in regression (and
using the usual least squares function), Tibshirani (1996) defined the Least Absolute
Shrinkage and Selection Operator (LASSO) estimate of the parameters. The LASSO
estimate finds sparse models, and does variable selection and parameter estimation
simultaneously (Tibshirani 1996; Fan and Li 2001; Kou and Pan 2008). Dahinden
et al. (2007) provide an extension using the LASSO penalty in a likelihood from
a multinomial random variable in multi-dimensional contingency tables. Conde
and MacKenzie (2012, 2014) propose other LASSO-type related penalties in this
context.

6.1 Classical Stepwise Search Algorithms

These algorithms use the iterative proportional fitting algorithm (Deming and
Stephan 1940) and work in a stepwise fashion eliminating (or adding) one effect
at a time until they arrive at the final step. They compare two nested models at each
step using the likelihood ratio criterion. See Conde (2011, Chap. 5).
Ideally one wishes to find a best-fitting model and discover the dependence
structure between the categorical variables. However, in high-dimensional problems
these goals may be unattainable and in practice we may be forced to consider
a reduced problem involving the dependence structure of lower-order interactions
such as two-way or three-way interactions.
The algorithms perform (optionally for BE2 and FS) the tests of the hypothesis
that whether the m-way effects are zero, or the tests of all  m-way effects D 0. For
Model Selection for Binary Data 233

Table 4 Meaning of the tests of m-way effects are zero

m Null hypothesis Nested comparison
1 H0 : All main effects D0 Compare(M0ways , M1ways )
2 H0 : All 2-ways D0 Compare(M1ways , M2ways )
:: :: ::
: : :
p H0 : The p-way effect D0 Compare(M.p1/ways , Mpway )

p variables we can define: Mi ways WD the all i-ways model, where i D 0; : : : ; p. In

Table 4 we can see the tests of m-way effects are zero, with m, the null hypotheses,
and the test of comparison in the last column. In the tests of all  m-way effects
D 0, the comparisons are always conducted with the saturated model.
Backwards Elimination starts from the saturated model, it eliminates one effect
at a time, comparing the model, with the model from which the effect has been
dropped. Then, eliminating one effect at a time, the effects that are eliminated
always belong to the generating set of the current model in each step.
Moreover, at each step BE shows us the 10 results of the tests of comparisons of
the effects that can or cannot be eliminated, both ordered decreasingly with respect
to the p-values. For example, suppose that at one step, 5 effects can be eliminated,
and 13 can not. BE will print, for each of the eliminable effects: the effect, the LR
test statistic, degrees of freedom, and p-value of the test of the comparison, ordered
from the lowest to the highest significant effect. Next, it will print at most ten effects
which cannot be eliminated also ordered in the same way.
Backward Elimination 2 is akin to BE, working backwards, but starting with
a model with all the maxorder-way interactions. This means that, for example,
suppose a model with all main effects does not fit (i.e., the difference between
log-likelihoods between this model and the saturated model is significantly large);
and suppose a model with all the 2-way interactions does fit (this information
may be available from the tests of m- or  m-way effects are zero). In this
case maxorder D 2. BE2 will start its search with the model with all the 2-way
interactions until we arrive at the end (i.e., until no other effect can be eliminated).
Forward Selection (FS) starts with the null model and adds one effect at a time
until a model that fits the data is found. As opposed to the method in Edwards
(2012), these effects can be potentially of any order and not “only” based on 2-way
interactions. Another version of FS can start with the main effects model.
Conde and MacKenzie (2008) have another algorithm that calculates the tests
of partial associations (Christensen 1997, pp. 217–218): here, the saturated model
is always the basis for the comparisons. For each m-way effect that is present in
the HLL model, these tests compare the model with all the m-way effects against
a model in which the m-way effect in question is dropped out. Besides, the list is
ordered increasingly with respect to the p-values, i.e., from lowest to highest.
234 S. Conde and G. MacKenzie

More recently Conde and MacKenzie (2014) propose an enhanced version of the
backwards elimination algorithm: MCBE, which can remove inestimable effects
(that is, effects whose maximum likelihood estimator does not exist) and starts the
search backwards from the remaining parameter space.

6.2 Penalised Likelihood

The idea of Penalised Likelihood (Demidenko 2004) is to attach a penalty to the

usual likelihood function. Different penalties may be adopted to achieve various
desirable properties: e.g., sparsity or smoothness of solutions, etc. (Hastie et al.
2001; Friedman 2008). Here we are primarily interested in encouraging sparse
solutions in order to identify a more parsimonious model.
The LASSO estimate has an effect of shrinking proportionally the coefficients in
the model with respect to the Ordinary Least Squares estimator. Typically, some of
the coefficients go exactly to zero, depending on the regularization parameter. Thus
the LASSO estimate performs variable selection and parameter estimation at the
same time. Dahinden et al. (2007) use it in order to find a sparse log-linear model
that fits a table where the variables involved are genetic sites (splicings of genes
with presence or absence of exons).
Following the latter approach, the LASSO estimate is given by

O LASSO .t/ WD arg min f`mult ./g;

 2‚

P
where `mult is the log-likelihood of a multinomial, subject to kj D2 j j j  t for
some constant t, the tuning parameter. The definition is equivalent to ’
8 9
< X
k =
.OLASSO / WD arg min `mult ./ C jj j
 2‚ : ;
j D2

for some regularisation parameter  0, where a large means that all the
estimates have gone to 0; and D 0 means that the solution is .O LASSO /0 , O
O
the MLEs. We can name  LASSO as the maximum penalised likelihood estimators
(MPLEs), according to Green and Silverman (1994)’s methodology. The regu-
larisation parameter may be estimated, e.g., by cross-validation (Dahinden et al.
2007; Conde and MacKenzie 2012) or using the Bayes Information Criterion (BIC)
(Conde and MacKenzie 2012). In the literature, Yates’ coding scheme is used for
the design matrix, so the columns of A are orthogonal. The LASSO penalty is
non-differentiable and one needs a path following algorithm such as presented in
Dahinden et al. (2007). Moreover, the LASSO estimate does not automatically
impose hierarchical rules and the final models selected may be non-hierarchical.
Accordingly, this entire method is suspect in routine applications. Another criticism
Model Selection for Binary Data 235

is that when the set of effects which are to be treated as null is identified, these should
be eliminated from the model and the remaining effects re-estimated (non-LASSO).
This step is usually forgotten and interpretation is erroneously based on the LASSO
analysis.
Conde and MacKenzie (2012) consider and propose other LASSO-type related
penalties such as the LASSO defined only in the interactions (in order to keep the
hierarchical structure of the model), and the Smooth LASSO.

6.2.1 The Smooth LASSO

This is a new parametric, convex, analytic approximation to the LASSO. The

advantage of this approximation is that for computing the solutions of the equation
above, one can apply Newton–Raphson type methods thereby eliminating the need
for specialized algorithms such as the method of coordinate descent (Friedman et al.
2010). Moreover, the computation of the standard errors is very easy, such as in
Muggeo (2010). Conde and MacKenzie (2012) estimates using the method of
cross-validation (Conde 2011, p. 112).

7 Applications
7.1 Classical Stepwise Algorithms

For the first application we use data comprising 48,158 subjects where the cases are
patients with Chronic Obstructive Pulmonary Disease (COPD) and the age-matched
controls are COPD-free. The 15 binary variables analyzed indicate the presence or
absence of comorbidities. See other details of the data in Conde and MacKenzie
(2007). Table 5 shows the tests that the m-way effects are zero in these data.
Next, the BE was tested using sets up to and including ten comorbidities (Conde
and MacKenzie 2008) in this data set. The solutions given by BE in R were identical
to those obtained by SPSS and we limited the number of variables to 10 so that we
could effect direct comparison. The tables here are not sparse for small p, and they
become highly sparse for larger values of p.
Figure 1 displays the (ln) timings for each program on the vertical axis, whilst the
horizontal axis contain the number of comorbidities in the data set. The curves are
smoothing curves, i.e., loess’, formed with local quadratic polynomials (Cleveland
1979; Cleveland and Devlin 1988). From five variables upwards the times become
progressively longer in R compared with SPSS; R is an interpreted language while
SPSS uses compiled code in FORTRAN. Accordingly, for large values of p this
combinatorial algorithm becomes infeasible and other approaches will be required.
These timings refer to a standard desktop computer with 2 GHz processor and 4 GB
of main memory.
236 S. Conde and G. MacKenzie

Table 5 Tests of m-way m df LR P

effects are zero with p D 15
comorbidities 1 15 689,703.2000 0.0000
2 105 5,998.8920 0.0000
3 455 480.3214 0.1198
The best-fitting model may contain some 2-way inter-
action terms

16 ~46d
8 ~58m
Loess lBE_R 14 Loess lFS
7 Loess lBE_SPSS Loess lBE2 ~9d
BE_R 12
6 ~26h
~5.1m
5 10
ln(time/s)

ln(time/s)
4 8 BE2
3 6
FS
2 4
BE_SPSS
1 2
0 ~0.6s 0
−1 ~0.5s
−2
2 3 4 5 6 7 8 9 10 2 4 6 8 10 12
Number of comorbidities Number of comorbidities

Fig. 1 Comparison of times of the algorithms

Figure 1 displays analogous ln timings for FS and BE2 (right panel), and these
were obtained using a different set of (up to and including) 13 comorbidities. For
p > 4, BE2 takes longer than BE. Note that  p BE2 may start with a model that
contains a lot of effects in the generating set ( m for some 0 < m < p), as opposed
to the saturated model or the null model, which only contains one (or zero) elements
respectively. For large p, FS is the fastest algorithm. However, the FS algorithm
often selected models which were unsatisfactory in terms of fit, as judged by the
pattern of residuals. From this perspective, the backward elimination algorithms
were more satisfactory, but the final models were always more complex.
As an illustration, Figs. 2 and 3 display histograms (left panel) and Q–Q plots
(right panel) of the adjusted residuals for the BE and FS solutions respectively for
p D 8 comorbidities from the set of the left panel in Fig. 1. The final model found
by BE or BE2 is fc7c8, c5c6, c2c4c5, c1c4c5, c1c3c7, c1c3c4, c1c2c4c7, c1c2c5c7,
c1c2c3c5, c1c2c6c7g, with an LRD 48:9310, and 206 degrees of freedom, which
is an overparameterised model. The final model found by FS is fc6, c7, c8, c1c2,
c1c3, c2c4, c2c3, c4c5, c3c4g, with an LRD 276:4611, PearsonD 578:0098, and
241 degrees of freedom (i.e., a much more parsimonious model).
The graphs contain only the residuals for the cells with a non-zero frequency, and
non-zero estimated value, too (for the estimated values, it was considered non-zero
those >0.0001). Then the graphs contain 65 data, out of the 256. The blue curves are
the graphs of the probability density functions of a standard normal rescaled such
that the area under the curve coincides with that from the histogram.
Model Selection for Binary Data 237

30 22
19
25 A*N(0, 1)
16
20 13

Adj. res.
10
Count

mean= 0.36
15 7
sd= 1.528
4
10 1
5 −2
−5
0 −8
−3 −2 −1 0 1 2 3 4 5 6 7 −3 −2 −1 0 1 2 3
adjr Normal ordered statistics

Fig. 2 Histograms and Q–Q plots of adjusted residuals for p D 8, set 1, BE. The quantity A is
the area of the histogram

90
22
80
19
70 16
60 13
mean= 1.06
Adj. res.

10
Count

50
sd= 4.309
40 A*N(0, 1) 7
4
30
1
20
−2
10 −5
0 −8
−6 −3 0 3 6 9 12 15 18 21 −3 −2 −1 0 1 2 3
adjr Normal ordered statistics

Fig. 3 Histograms and Q–Q plots of residuals for p D 8, set 1, FS. The quantity A is the area of
the histogram. The quantity A is the area of the histogram

The distribution of adjusted residuals of Fig. 2 is a bit platykurtic compared

with the normal distribution, slightly right-skewed because of the presence of an
outlier corresponding to the cell in which comorbidities 2, 4, 7 and 8 were present
and the rest absent, and this cell has 1 observation. There is usually some outlier
(MacKenzie 2006). In the Q–Q plot the observations are approximately close to the
diagonal, apart from the outlier. The Kolmogorov–Smirnov test finds a p-value of
0.04786.
Both graphs in Fig. 3 indicate clearly that the distribution of adjusted residuals is
not normally distributed. The variance of the adjusted residuals in the FS solution
(12.982) is more than six times greater than the corresponding variance in the BE
solution (2.097). The Kolmogorov–Smirnov test has a p-value of 7:124  107 ,
clearly significant with the standard confidence levels.
In addition to testing with real data sets, the algorithms (BE2, FS) were tested in
a comprehensive simulation study (Conde 2011, Chap. 6). A core scenario involved
generating m D 1;000 random compositions (Nijenhuis and Wilf 1978) for p D
2; 3; 4; 5; n D 100; 500 and m D 150 compositions for p D 10; n=10,000. The
238 S. Conde and G. MacKenzie

Table 6 Final models found with the comorbidity table

LASSO
MCBE, BE2, FS CV BIC Smooth LASSOa
Comorb. [c1, c2, c3] [c1c2, c1c3, [c1c2c3] [c1, c2, c3]
data c2c3]
Variables mean c1: mild liver disease; c2: diabetes; c3: lung cancer
a
We removed  D 0 from the path as nlm did not converge

results of this simulation were: the final models found by the algorithms, the Wald
test statistics (Wald 1943) of their parameters, the likelihood, and other quantities.
For example, for p D 2, the algorithms always found the correct model except
for the saturated case, and in the latter case the percentages of success are very
high (Conde 2011, p. 98). Moreover, we also simulated tables corresponding to a
specific model, with p D 2; 3; 5; 7; 10; n D 200; 500; 2;000; 5;000; 10;000; 50;000;
m D 1;000 or 100 simulations; and all models (when p D 2), or all models
with up to and including the t-way interactions, t D 1; 2; 3. Overall, for small p,
the final model found by the algorithms (BE2 and FS) was the same or a model
in the neighbourhood of the correct model. For larger p, where the dimension
of the parameter space becomes very large, the algorithms found a model in the
neighbourhood of the correct model (Conde 2011, pp. 101–102).

7.2 MCBE and Penalized Likelihood

Conde and MacKenzie (2012) present an analysis of a three-dimensional con-

tingency table from the comorbidity data and the results of a simulation study.
The variables are: mild liver disease, diabetes and lung cancer. According to
MacKenzie’s theorem (MacKenzie and Conde 2014), the MLEs of the effects c1c2,
c1c3 and c1c2c3 are non-existent. Table 6 displays the results of the final models
found by MCBE and the diverse penalised likelihood approaches. Depending on the
method used, the results may be very different.
The three classical algorithms and the Smooth LASSO found the main effects
model, i.e., the presence of mild liver disease is not affected by diabetes and lung
cancer and vice versa with all the combinations of the three comorbidities. In
contrast, the LASSO method and when estimating by cross-validation, found a
model with all 2-way interactions, i.e., there is a complicated pattern of interaction
between the comorbidities. When estimating with the BIC, the final model is
the saturated model. Moreover, the penalized likelihood approaches contain diverse
effects whose MLEs do not exist in their final models.
The simulation study comprise 100 simulated random compositions with p D
2. Each random composition was given to MCBE, BE2, FS, LASSO using the
methods of cross-validation and BIC, and the Smooth LASSO. Table 7 displays the
Model Selection for Binary Data 239

Table 7 Percentages of final models found; p D 2, in 100 simulated tables

%
LASSO
p n model MCBE BE2 FS CV BIC Smooth LASSOa
2 50 null 4 8 5 0 1 23
{c1} 11 8 13 0 3 12
{c2} 6 5 6 3 2 9
{c1, c2} 20 20 20 15 11 10
sat. 55 55 52 78 79 40
No fitb 4
Total 100 96 96 96 96 94
2 10 null 18 27 25 7 13 69
{c1} 15 8 15 4 4 4
{c2} 15 13 16 5 5 4
{c1, c2} 13 13 13 14 11 0
sat. 22 22 14 53 50 4
No fitb 17
Total 100 83 83 83 83 81
CV: Five-fold cross-validation. BIC: BIC with a LASSO penalty. The Smooth LASSO approxima-
tion is used with ! D 1 and 5-fold cross-validation
a
We removed tables when nlm did not converge; (2, 2, respectively, in each scenario)
b
SSM does not fit

results. By using MacKenzie’s Theorem (MacKenzie and Conde 2014), to detect

inestimable effects, four contingency tables in the first scenario and twenty-two
tables in the second scenario had at least one inestimable effect. The acronym
SSM stands for the sparse saturated model, which is the maximal model that can
be fitted in a sparse contingency table after eliminating effects with non-existent
MLEs detected by MacKenzie’s theorem.
The classical algorithms find always hierarchical, sparser (i.e., more parsimo-
nius) models, and those which are free of inestimable effects in the case of MCBE.
On the other hand, none of the penalized likelihood approaches take into account the
hierarchical rules; furthermore, these (penalized likelihood) approaches generally
include effects with non-existent maximum likelihood estimators in their final
models.

8 Discussion

We have presented a short review of modern methods for analysing sparse contin-
gency tables. One of the most interesting developments concerns the elimination
of non-existent MLEs by pre-processing contingency tables prior to formal sta-
tistical analysis (MacKenzie and Conde 2014). This approach is potentially very
efficient and should allow the researcher to tackle higher dimensional sparse
240 S. Conde and G. MacKenzie

problems. Another key development is that of the Smooth LASSO which being
twice differentiable simply dispenses with the need for specialised optimization
algorithms such as the method of coordinate descent. An attractive feature is
that it casts the analysis of LASSO problems in a classical statistical mould. A
weakness with the regularization approach has been the failure of the technique
to produce hierarchical solutions. Non-hierarchical solutions are not design matrix
invariant and this can trap the unwary applied researcher. Another weakness of
current LASSO methods is that their solutions often contain effects which may
be shown not to exist in a classical analysis. This is a serious inconsistency and
implies that non-existent effects should be eliminated prior to any analysis. Overall,
our new approach, can in principle reconcile all of these problems by eliminating
non-existent effects and fitting the Smooth LASSO optimization into a hierarchical
algorithm. We expect that in time our methods will impact positively on practice.

Acknowledgements The work in this paper was conducted in the Centre of Biostatistics,
Limerick, Ireland, and supported by the Science Foundation Ireland (SFI, www.sfi.ie), project grant
number 07/MI/012 (BIO-SI project, www3.ul.ie/bio-si). The first author’s Ph.D. scholarship was
supported GlaxoSmithKline, England, UK.

References

Agresti, A. (2002). Categorical data analysis (2nd ed.). Hoboken, NJ: Wiley.
Baker, R. J., Clarke, M. R. B., & Lane, P. W. (1985). Zero entries in sparse contingency tables.
Computational Statistics and Data Analysis, 3, 33–45.
Birch, M. W. (1963). Maximum likelihood in three-way contingency tables. Journal of the Royal
Statistical Society. Series B (Methodological), 25(1), 220–233.
Bishop, Y. M., Fienberg, S. E., & Holland, P. W. (1975). Discrete multivariate analysis: Theory
and practice. Cambridge: MIT Press, The Massachusetts Institute of Technology.
Bishop, Y. M. M. (1969). Full contingency tables, logits, and split contingency tables. Biometrics,
25(2), 383–399.
Charlson, M. E., Pompei, P., Ales, K. L., & MacKenzie, C. R. (1987). A new method of classifying
prognostic comorbidity in longitudinal studies: Development and validation. Journal of Chronic
Diseases, 40(5), 373–383.
Christensen, R. (1997). Log-linear models and logistic regression (2nd ed.). New York: Springer.
Cleveland, W. S. (1979). Robust locally weighted regression and smoothing scatterplots. Journal
of the American Statistical Association, 74(368), 829–836.
Cleveland, W. S., & Devlin, S. J. (1988). Locally weighted regression: An approach to regression
analysis by local fitting. Journal of the American Statistical Association, 83(403), 596–610.
Conde, S. (2011). Interactions: Log-linear models in sparse contingency tables (Ph.D. thesis).
University of Limerick, Ireland.
Conde, S., & MacKenzie, G. (2007). Modelling high dimensional sets of binary co-morbidities. In
J. del Castillo, A. Espinal, & P. Puig (Eds.), Proceedings of the 22nd International Workshop
on Statistical Modelling, Barcelona (pp. 177–180).
Conde, S., & MacKenzie, G. (2008). Search algorithms for log-linear models in contingency tables.
Comorbidity data. In P. H. Eilers (Ed.), Proceedings of the 23rd International Workshop on
Statistical Modelling, Utrecht (pp. 184–187).
Model Selection for Binary Data 241

Conde, S., & MacKenzie, G. (2011). LASSO penalised likelihood in high-dimensional contingency
tables. In D. Conesa, A. Forte, A. López-Quílez, & F. Muñoz (Eds.), Proceedings of the 26th
International Workshop on Statistical Modelling, Valencia (pp. 127–132).
Conde, S., & MacKenzie, G. (2012). Model selection in sparse contingency tables: LASSO
penalties vs classical methods. In A. Komárek & S. Nagy (Eds.), Proceedings of the 27th
International Workshop on Statistical Modelling, Prague (pp. 81–86).
Conde, S., & MacKenzie, G. (2014). The smooth LASSO in sparse high-dimensional contingency
tables (in preparation).
Dahinden, C., Parmigiani, G., Emerick, M. C., & Bühlmann, P. (2007). Penalized likelihood for
sparse contingency tables with an application to full-length cDNA libraries. BMC Bioinformat-
ics, 8, 476.
Darroch, J. N., Lauritzen, S. L., & Speed, T. P. (1980). Markov fields and log-linear interaction
models for contingency tables. The Annals of Statistics, 8(3), 522–539.
Davies, S. J., Phillips, L., Naish, P. F., & Russell, G. I. (2002). Quantifying comorbidity in
peritoneal dialysis patients and its relationship to other predictors of survival. Nephrology
Dialysis Transplantation, 17(6), 1085–1092.
Demidenko, E. (2004). Mixed models. New York: Wiley.
Deming, W. E., & Stephan, F. F. (1940). On a least squares adjustment of a sampled frequency
table when the expected marginal totals are known. The Annals of Mathematical Statistics,
11(4), 427–444.
Dobson, A. J. (2002). An introduction to generalized linear models. New York: Chapman &
Hall/CRC.
Edwards, D. (2000). Introduction to graphical modelling (2nd ed.). New York: Springer.
Edwards, D. (2012). A note on adding and deleting edges in hierarchical log-linear models.
Computational Statistics, 27, 799–803.
Edwards, D., & Havránek, T. (1985). A fast procedure for model search in multidimensional
contingency tables. Biometrika, 72(2), 339–351.
Fan, J., & Li, R. (2001). Variable selection via nonconcave penalized likelihood and its oracle
properties. Journal of the American Statistical Association, 96(456), 1348–1360.
Feinstein, A. R. (1970). The pre-therapeutic classification of co-morbidity in chronic disease.
Journal of Chronic Diseases, 23(7), 455–468.
Fienberg, S. E. (1972). The analysis of incomplete multi-way contingency tables. Biometrics,
28(1), 177–202 [special Multivariate Issue].
Fienberg, S. E., & Rinaldo, A. (2006). Computing maximum likelihood estimates in log-linear
models. Manuscript extracted from Rinaldo’s Ph.D. thesis.
Fienberg, S. E., & Rinaldo, A. (2012). Maximum likelihood estimation in log-linear models. The
Annals of Statistics, 40(2), 996–1023.
Fisher, R. A. (1922). On the interpretation of 2 from contingency tables, and the calculation of P.
Journal of the Royal Statistical Society, 85(1), 87–94.
Friedman, J. H. (2008). Fast sparse regression and classification. In P. H. Eilers (Ed.), Proceedings
of the 23rd International Workshop on Statistical Modelling, Utrecht (pp. 27–57).
Friedman, J. H., Hastie, T., & Tibshirani, R. (2010). Regularization paths for generalized linear
models via coordinate descent. Journal of Statistical Software, 33(1), 1–22.
Glonek, G. F. V., Darroch, J. N., & Speed, T. P. (1988). On the existence of maximum likelihood
estimators for hierarchical loglinear models. Scandinavian Journal of Statistics, 15, 187–193.
Goodman, L. A. (1968). The analysis of cross-classified data: Independence, quasi-independence,
and interactions in contingency tables with or without missing entries. R. A. Fisher memorial
lecture. Journal of the American Statistical Association, 63(324), 1091–1131.
Goodman, L. A. (1971). The analysis of multidimensional contingency tables: Stepwise procedures
and direct estimation methods for building models for multiple classifications. Technometrics,
13(1), 33–61.
Green, P. J., & Silverman, B. W. (1994). Nonparametric regression and generalized linear models:
A roughness penalty approach (Vol. 58). Monographs on statistics and applied probability (1st
ed.). London: Chapman & Hall.
242 S. Conde and G. MacKenzie

Haberman, S. J. (1970). The general log-linear model (Ph.D. thesis). Department of Statistics,
University of Chicago, Chicago, IL.
Hall, W. H., Ramachandran, R., Narayan, S., Jani, A. B., & Vijayakumar, S. (2004). An electronic
application for rapidly calculating Charlson comorbidity score. BMC Cancer, 4, 94.
Hastie, T., Tibshirani, R., & Friedman, J. (2001). The elements of statistical learning. New York:
Springer.
Hu, M. Y. (1999). Model checking for incomplete high dimensional categorical data (Ph.D. thesis).
University of California, Los Angeles.
Kim, S. H., Choi, H., & Lee, S. (2008). Estimate-based goodness-of-fit test for large sparse
multinomial distributions. Computational Statistics and Data Analysis, 53(4), 1122–1131
Kou, C., & Pan, J. (2008). Variable selection in joint modelling of mean and covariance structures
for longitudinal sata. In P. H. Eilers (Ed.), Proceedings of the 23rd International Workshop on
Statistical Modelling, Utrecht (pp. 309–314)
Krajewski, P., & Siatkowski, I. (1990). Algorithm AS 252: Generating classes for log-linear
models. Journal of the Royal Statistical Society. Series C (Applied Statistics), 39(1), 143–176.
Lang, S. (1992). Algebra (3rd ed.). Delhi: Pearson Education.
MacKenzie, G. (2006). Screening multivariate comorbidities. Presentation. Assess, York, http://
www.staff.ul.ie/mackenzieg/Assess/assess.html.
MacKenzie, G., & Conde, S. (2014). Model selection in sparse contingency tables (in preparation).
MacKenzie, G., & O’Flaherty, M. (1982). Algorithm AS 173: Direct design matrix generation
for balanced factorial experiments. Journal of the Royal Statistical Society. Series C (Applied
Statistics), 31(1), 74–80.
Mantel, N. (1970). Incomplete contingency tables. Biometrics, 26(2), 291–304.
Maydeu-Olivares, A., & Joe, H. (2005). Limited- and full-information estimation and goodness-of-
fit testing in 2n contingency tables. Journal of the American Statistical Association, 100(471),
1009–1020.
McCullagh, P., & Nelder, J. A. (1997). Generalized linear models. London: Chapman & Hall.
Montgomery, D. C. (2001). Design and analysis of experiments (5th ed.). New York: Wiley.
Muggeo, V. M. R. (2010). LASSO regression via smooth L1 -norm approximation. In A. W.
Bowman (Ed.), Proceedings of the 25th International Workshop on Statistical Modelling,
Glasgow (pp. 391–396).
Nijenhuis, A., & Wilf, H. S. (1978). Combinatorial algorithms. New York: Academic.
O’Flaherty, M., & MacKenzie, G. (1982). Algorithm AS 172: Direct simulation of nested Fortran
DO-LOOPS. Journal of the Royal Statistical Society. Series C (Applied Statistics), 31(1), 71–74.
Pearson K. (1900) On the criterion that a given system of deviations from the probable in the case
of a correlated system of variables is such that it can be reasonably supposed to have arisen from
random sampling. Philosophy Magazine Series, 50(5), 157–174.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society, 58(1), 267–288.
Wald, A. (1943). Tests of statistical hypotheses concerning several parameters when the number of
observations is large. Transactions of the American Mathematical Society, 54(3), 426–482.
Wilks, S. S. (1935). The likelihood test of independence in contingency tables. The Annals of
Mathematical Statistics, 6(4), 190–196.
Wilks, S. S. (1938). The large-sample distribution of the likelihood ratio for testing composite
hypotheses. The Annals of Mathematical Statistics, 9(1), 60–62.
Wissmann, M., Toutenburg, H., & Shalabh (2007) Role of categorical variables in multicollinearity
in the linear regression model. Technical Report 008, Department of Statistics, University of
Munich.
Zelterman, D. (1987). Goodness-of-fit tests for large sparse multinomial distributions. Journal of
the American Statistical Association, 82(398), 624–629.
Obituary: Professor Ennio Isaia

In the course of preparing this book we were saddened by the unexpected news of
the death of Professor Ennio Isaia of the University of Torino, following a short
illness.
Ennio Isaia was a brilliant, passionate, serious statistician and a wonderful
person. Ennio died in February 2010 at the age of 54. He was a prolific researcher
at the Department of Statistics and Applied Mathematics “D. de Castro” in the
University of Torino, the establishment where he spent most of his academic
life. In his rather short life, Ennio produced many interesting research papers
(over 35) in both methodological and applied statistics. He was also the author of
several textbooks for the student body. His research interests were wide-ranging and
encompassed:
• Robust statistics, outliers and clusters detection.
• Model selection algorithms.
• Mixture models and clustering.
• Computational statistics and high-dimensional data analysis.
• Nonparametric and semi-parametric statistical methods.
• Statistical quality and process control.

G. MacKenzie and D. Peng (eds.), Statistical Modelling in Biostatistics 243

and Bioinformatics, Contributions to Statistics, DOI 10.1007/978-3-319-04579-5,
© Springer International Publishing Switzerland 2014
244 Obituary: Professor Ennio Isaia

He had presented the results of this research at many Workshops and Conferences
around the world. His last presentation was at the XIII International Conference,
Applied Stochastic Models and Data Analysis (ASMDA) held at Vilnius in June
2009, entitled, The minimum density power divergence approach in building robust
regression models.
Ennio was an active participant in International Workshops. He was also a mem-
ber of the Programme Committee which helped organise the WCDM Workshop
held in Torino in 2004. He was an enthusiastic supporter of the 2007 Workshop in
Limerick and produced the Latex templates for that event.
Accordingly, it is with great fondness that he is remembered by his friends and
colleagues.

Alessandra Durio & Gilbert MacKenzie

Autumn 2013