This book describes itself as being suitable for non-statisticians and researchers working with longitudinal data from epidemiological and clinical studies. However, despite my statistical background, I turned the first page with some trepidation.
I was pleasantly surprised as I found this book both well written and easy to understand. The layout is good and there are plenty of figures, tables, and actual computer output coming mainly from SPSS and Stata. Finally, and perhaps most importantly, equations are kept to a minimum and where they do occur, notation is kept simple.
The book commences with an overview of observational and experimental longitudinal studies and is followed by the relatively simple concepts of paired t-tests and their non-parametric equivalent that in turn are followed by MANOVA complete with a number of numerical examples.
In chapter 4, generalized estimating equations (GEE) and random coefficient models (also known as random effects or multilevel models) are explained followed by a comparison between the two accompanied by a number of numerical examples. These two types of models are carried through the majority of the remaining chapters with a thorough investigation into models with continuous, categorical and dichotomous outcomes. A brief pause comes in chapter 5 when alternative methods such as time-lag and autoregressive models receive a mention, and again in chapter 8 where the analysis of change takes centre stage with a look at ANCOVA and alternatives, once again for a variety of outcomes.
Chapter 10 is devoted to missing data in longitudinal studies. This came as a pleasant surprise as such issues are often swept under the carpet. The concepts of MCAR (missing completely at random), MAR (missing at random), and MNAR (missing not at random) are described followed by an example showing how both ignorable and non-ignorable missing data can affect regression coefficients. Ways of testing which kind of missing data is present lead into the implications of this depending on the type of analysis planned. A further example shows in detail how various missing data mechanisms can affect estimates for both GEE and random coefficients models with both continuous and dichotomous outcomes. This chapter ends with an overview of different imputation methods and further examples of how these too can affect results.
This book ends with a lengthy discussion of the software packages Stata, SAS, Splus, SPSS, and MlwiN. How the models in this book can be fitted using each package is described, complete with the computer output one can expect and potential differences in results that can occur depending on the package chosenall very useful information.
This book introduces a dataset in chapter 1 consisting of an outcome, two continuous variables (one time dependent, one not), and two dichotomous variables (also one time dependent, one not). By using this one dataset throughout the majority of the book the reader is permitted to assess how the increasingly complex models affect the conclusions. An obvious question to ask is why this dataset has not been made available either as a disk or downloadable from a website. Hopefully, this can be rectified in future editions.
In conclusion, I can thoroughly recommend this book to anyone with an interest in longitudinal modelling. Furthermore, the reading of chapter 10 should be made compulsory to anyone working with longitudinal epidemiology data.