IJE vol.33 no.5 © International Epidemiological Association 2004; all rights reserved.
Book Review |
Applied Longitudinal Data Analysis: Modelling Change and Event Occurrence. Judith D Singer, John B Willett. New York: Oxford University Press, 2003, pp. 644, £39.50 (HB). ISBN: 0-19-515296-4.
This volume deals with statistical methods for the analysis of longitudinal data, covering both the case when time is a predictor for an outcome of interest and when it is itself the outcome of interest. So multilevel models for individual change over time, and survival models for studying event rates, are given equal prominence. They are introduced in two distinct parts, each of which is about 300 pages and sufficiently detailed to be self-standing. The book's main aim however is to link these two approaches so that applied researchers would gain an overview and ‘be able to apply both methods synergistically to different research questions in the same study’.
The style is very didactical and, to my taste, too lengthy. The quality and depth of the material is nevertheless excellent, as is that of the advice regarding issues of model building and interpretation. Examples taken from existing studies, mostly arising in the behavioural and social sciences, are used throughout for illustration. Data and typical analyses carried out in several software packages (Mplus, MLwinN, HLM, SAS, Stata and SPSS) are accessible via a companion website (www.ats.ucla.edu/stat/examples/alda) which is very much worth a visit.
The first part of the book introduces multilevel models, starting from simple subject-specific estimation of individual trajectories over time and stressing the need for a more comprehensive approach. At first a linear model for time-structured data is introduced, followed by step-by-step generalizations which are supported by plots and tables that summarize concepts and notations. Among the generalizations, the most interesting ones concern piecewise and discontinuous specifications of the effect of time and the inclusion of time-varying exposures. The distinction between defined, ancillary, contextual, and internal variables guides the discussion of the issues that arise from the latter generalization and is probably one of the more important features that are shared with the second part of the book.
The material presented is also very educational from a practical point of view. It shows the advantages of centring explanatory variables, including time itself, and includes a comprehensive discussion of the range of possible estimation methods that is linked to the choice, and uses, of software.
As new concepts are always supported by examples and discussion, the material is suitable for newcomers to multilevel models. For those already familiar with the subject there are excellent practical hints scattered throughout and, more importantly, new insights into broader generalizations. These are touched in the last two chapters of this first part of the book where multilevel models are directly linked to covariance structure analysis and to its application to longitudinal data, latent growth models. This link enables researchers to relax some of the less-realistic modelling assumptions and to develop models for more complex, multivariate, dependencies over time.
The second part of the book deals with methods for investigating the distribution of time up to an event of interest. Unlike most textbooks on survival analysis, a far greater focus is given here to methods for discrete time data, i.e. of data collected at fixed points in time so that only the intervals during which events occurred are known. The proportional odds model and the complementary log-log model are introduced within this framework and then compared. Practical advice on how data have to be organized for the models to be successfully fitted is given, and some graphical representations of the estimated hazard and survival functions are recommended. These methods are then extended to the analysis of continuous survival time data. The presentation and development of Cox's proportional hazards model is sharp and includes an excellent overview of the main features of this model and of the most useful tools for assessing its assumptions, leading to alternative generalizations. The discussion of the available regression diagnostic tools is extremely useful, even for those already familiar with this material. Indeed its summary table is something I will copy and stick on my wall!
If I have a criticism, beside that regarding the lack of conciseness, it concerns the topics which are not covered. In the first part I would have wanted more obvious references to models for non-continuous outcomes and, in the second part, to non-proportional parametric models. But has this ambitious volume achieved its aim? Are the two separate parts better off because they are presented together? Overall the answer is yes: there are several cross-references that make the reader gain valuable insights because issues are discussed from different angles. Is the book of interest to epidemiologists? Again the answer is yes, although most of the examples are not directly relevant and the style is unfamiliarly lengthy. An epidemiologist who has no or only superficial knowledge of these two topics would certainly gain from reading this book; anyone else would find clarity and wise counsel from consulting its parts.
The two authors are internationally renowned social statisticians whose collaboration has been spanning for nearly 20 years with, interestingly, first authorship always determined by randomization (even in this case). They are driven by the wish to improve the quality of published empirical research and—to achieve this—most of their work is available on the web. This book is a great addition to their efforts and will certainly have a substantial impact on the analyses of longitudinal data carried out in many fields.
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||