1. 2610 BOOK REVIEWS
3. META-ANALYSIS IN MEDICINE AND HEALTH POLICY. practical problem of heterogeneity of reporting is
Dalene K. Stangl and Donald A. Berry (eds), Mar- addressed for change over baseline outcomes in
cel Dekker, New York, 2000. No. of pages: 398. Chapter 2 and mixtures of continuous and binary
Price: $155.00. ISBN 0-8247-9030-8 outcomes in Chapter 5. This latter chapter con-
siders a latent-variables approach but the example
The use of meta-analysis to combine informa- presented also involves the indirect comparison
tion across sources in medicine continues to of treatments. This problem of indirect compar-
increase. The editors recognize the important isons is also considered in Chapter 7 where a
‘paradigm shift’ occurring over the last decade ÿxed-e ect model is ÿtted to some hypothetical
as the focus has moved from ÿtting a com- data. In Chapter 12 the authors apply a method to
mon estimate of e ect to estimating the extent account for potential publication bias allowing for
and investigating sources of variability among such bias to depend on the quality of the study.
studies. Substantial emphasis is placed on this Chapter 14 describes the approach taken by the
issue of heterogeneity throughout much of the Institute of Medicine of the National Academy of
book. Sciences in its review of two ‘politically charged’
The book is a collection of 15 individual papers policy issues, illustrating the di culties of inter-
by di erent authors. The alphabetical ordering of preting evidence. Chapter 15 provides an infor-
the chapters in the book is not followed in this mative review of available software, including
content review which aims to draw together papers identiÿcation of unique features of the packages
addressing similar issues. Chapter 13 reviews the examined.
purposes of meta-analysis and usefully summa- Almost all of the analyses presented are (fully)
rizes the important methodologic features to bear Bayesian, allowing uncertainty in several model
in mind when undertaking such analyses. Chapter 1 parameters to be incorporated. Most authors em-
gives a clear introduction to ÿxed- and random- ployed di use prior distributions although data
e ect models and to the frequentist and Bayesian priors were used in three examples. Several
perspectives, and provides an excellent source of authors assessed sensitivity across a range of
references for Bayesian meta-analyses. Chapters 3, prior distributions. In three chapters, the authors
4, 6, 8, 9, 10 and 11 discuss the issue of heterogene- also presented the results from the classical ap-
ity in the context of model uncertainty and selection proach for comparison. Two chapters describe
in di erent examples. Chapter 3 examines the ef- results from a classical rather than Bayesian
ÿcacy of mammography screening by comparing perspective.
three models with increasing generalization re- Generally the papers are well written. A strength
garding heterogeneity. Chapter 4 presents an inter- of this book is that nine chapters provide the data
esting example of heterogeneity in a meta-analysis from the real examples used to illustrate the meth-
of three mega-trials. Useful exploratory graphi- ods. The examples cover a range of study de-
cal analyses for model selection are described in signs, namely randomized controlled trials, epi-
Chapter 6 and applied to a meta-analysis of stud- demiological and pharmacokinetic studies, and in-
ies assessing the e ect of nicotine-replacement clude binary, continuous and time-to-event out-
therapy on smoking cessation. Chapter 8 consid- comes. One criticism is that in some chapters the
ers heterogeneity across epidemiological studies clinical interpretation is minimal. A greater empha-
in the e ect of duration of oestrogen exposure sis on publication bias might also have been ex-
on the incidence of endometrial cancer. Chapter pected given its threat to the validity of any meta-
9 illustrates the importance of examining re- analysis.
sults for a range of plausible prior distributions As the editors suggest, this book will be useful
for the between-study variance and proposes a for applied statisticians. The reader is assumed to
default half-normal prior for this parameter. Di- have taken an ‘undergraduate course in statistical
agnostics for examining the normality assumption theory and methods’, which is implicitly assumed
of study e ects are also presented. A hierarchi- to have included Bayesian inference. General fa-
cal two-compartment model is ÿtted to individ- miliarity with methods for parameter estimation in
ual patient data from pharmacokinetic studies a Bayesian analysis, in particular Gibbs sampling,
in Chapter 10 to incorporate within-individual, is assumed, although the emphasis of the book is
between-individual and between-study variabil- on application and interpretation and less so on
ity. Chapter 11 presents several Cox proportional implementation.
hazards models that allow for random-e ect meta- The knowledge assumed by the book limits
analyses of individual patient survival data. The its accessibility but several chapters provide ex-
Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:2607–2612
2. BOOK REVIEWS 2611
planations to help the statistician put across the PAULA WILLIAMSON
ideas to those without a statistical background. Division of Statistics and Operational Research
The book will also be a useful reference for Department of Mathematical Sciences
postgraduate students and provides a number of University of Liverpool
examples which could form the basis for a student Liverpool L69 3BX, U.K.
project. (DOI: 10.1002/sim.1071)
4. FUNDAMENTALS OF MODERN STATISTICAL METH- (normality, homoscedasticity) are violated. In
ODS. Rand R. Wilcox, Springer-Verlag, New York, chapter 6 the percentile bootstrap method is pre-
2001. No. of pages: xiii+258. Price: $ 49.95. ISBN sented and its practical value when making infer-
0-387-95157-1 ences about the slope of a regression line is shown.
Chapter 7 shows the devastating implications of
This book provides a comprehensive view of mod- the so-called ‘contaminated normal distributions’
ern basic statistical methods aimed at overcom- in terms of dramatic reduction of power in conven-
ing practical problems arising when assumptions tional inferential methods about means, regression
underlying conventional methods, foremost nor- slope and Pearson’s correlation coe cient.
mality, are violated. These ‘robust’ techniques are In chapter 8, two robust estimators of loca-
typically not dealt with in standard courses or tion are introduced: a trimmed mean and an M-
books on basic statistics and their use is still very estimator. The application of these estimators in
limited. The goal of the book is to bridge the testing hypothesis and computing conÿdence in-
gap between state-of-the-art in the development tervals is dealt with in chapter 9 and their relative
of these techniques and application in applied re- merits in terms of bias, control over type I error and
search. To this purpose, part I (chapters 2–7) of the power compared to the Student’s T -test are dis-
book highlights how and why standard methods cussed. Comparisons between groups are restricted
may be misleading and provides a framework for to the two-sample case and bibliographic notes on
intuitively understanding the practical advantages more complex experimental designs are given at
of modern techniques, while part II (chapters 8 – the end of the book. Chapter 10 illustrates the Win-
12) describes most basic methods and explains the sorized correlation, Spearman’s rho and Kendall’s
strategies they use. tau approaches to reduce the e ect of outliers in de-
Chapter 1 gives a brief historical overview of tecting associations between two variables. Global
basic theoretical foundations of statistics as they measures of associations for the detection of out-
were developed during the last three centuries. liers are also proposed. Robust methods for linear
Chapter 2 introduces the reader to practical prob- regression are dealt with in chapter 11. Finally,
lems that might arise in basic statistics (mea- chapter 12 focuses on contrasting robust methods
sures of location, dispersion and linear regres- for comparing two groups with ranked-based non-
sion) when standard techniques are used. Chap- parametric methods.
ter 3 deals with issues and methods related to out- The book is clearly written and main theoret-
liers detection and with the practical implications ical concepts are made easily accessible thanks
of the central limit theorem. In chapter 4 the ac- to a great didactic e ort. Technical details and
curacy of the sample mean and median in the esti- mathematics are kept at minimum, while graphical
mation of the population mean are compared and explanations and examples are frequently given.
a non-technical description of the Gauss–Markov A useful summary of key points is provided at the
theorem and of Laplace’s strategy for comput- end of each chapter. Classical inferential methods
ing a conÿdence interval is given. The reader is are refreshed, placing them within the historical
also introduced to practical implications, in terms development of statistical theory, thus contributing
of serious loss of accuracy, of violating the ho- to making reading more interesting and pleasant.
moscedasticity assumption in the estimation of All these features make the book well-tailored for
the linear regression slope. Chapter 5 focuses on the targeted user, namely the applied researcher
the factors that a ect power in hypothesis test- having a standard training in statistics.
ing and on practical problems arising with the The major merit of the book is to open the
Student’s T -test when the underlying assumptions mind of the reader about potential problems, in
Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:2607–2612