Bio[statistics]philia

Authors


Experimental Design and Data Analysis for Biologists. Quinn, G. P., and M. J. Keough. 2002. Cambridge University Press, Cambridge, United Kingdom. 537 pp. £75.00 (hardcover). ISBN 0–521–81128–7. £29.95 (paperback). ISBN 0–521–00976–6.

Most conservation biologists must draw on experimental design and data analysis in their roles as teachers, empiricists, authors, and reviewers, but few have an extensive background or aptitude in biometry. As a result, we typically experience mild discomfort when confronted with issues of statistical rigor and validity. Depending on the situation, our needs run the gamut from a refresher course in basic hypothesis testing to an overview of new methods that may be unfamiliar but are robust and innovative. In response, Gerry Quinn and Michael Keough have supplied us with a book that is exhaustive without being exhausting. Like the hip-pocket trail guides prized by backpackers, Experimental Design and Data Analysis for Biologists helps navigate the statistical path without making us feel foolish or clumsy. Ecologists themselves, Quinn and Keough frame their material using language and contextual examples that resonate with their audience. They frequently remind us that “for biologists, statistics is a tool that we use to illuminate and clarify biological problems” (p. xvi; emphasis theirs). Statisticians who evaluate this reference might find something to quibble about, but the overall assessment of most biologists likely will be quite positive.

Several features of the book render it especially friendly to biologists. First, the authors recognize that readers are likely to use statistical software to conduct most of the tests described, but they do not advocate or emphasize any particular package. Instead, they point out common quirks of statistical software that affect how tests are implemented. For example, they caution that there are two types of coding for turning categorical predictor variables into continuous dummy variables for ordinary least-squares regression and that interpretation of coefficients and odds ratios depends on which method is used. Second, descriptions of statistical techniques are accompanied by diverse examples from recent ecological publications. These examples illustrate real-world applications of various analyses and suggest that the authors keep abreast of current literature. To assist readers in understanding both statistical models and statistical software, the authors provide raw data files for examples included in the book on a public-access Web site. Third, most chapters conclude with “General Issues and Hints for Analysis,” which serves as both a summary of key points and a cheat sheet for data manipulation. The book is well-written and designed, and it maintains an accessible, conversational tone.

Chapter 1 presents a broad introduction to the scientific method, with several elements that are well-suited to applied ecology. For example, the chapter explains the differences between empirical and theoretical models and between research and statistical hypotheses. It also examines alternatives to Popperian falsification that derive from Kuhn (1970), Lakatos (1978), and other theorists. Quinn and Keough believe that manipulative experiments are ideal, but they recognize the challenges of drawing inferences from small-scale manipulative experiments to large scales of space and time. They emphasize that although observational data cannot demonstrate causality, well-designed sampling can be used effectively to refute a null hypothesis.

Chapter 2, “Estimation,” reviews common parameters and statistics and standard errors. It also explains two popular resampling methods, the bootstrap and the jacknife. Given the increasing emphasis on resampling in the ecological literature, I appreciated that these topics were introduced early in the book as opposed to being relegated to a back chapter. Chapter 3 covers a wide range of issues related to hypothesis testing, including Type I and Type II errors, parametric and nonparametric tests, and meta-analysis. An outline of statistical tests of hypotheses that explains some of the historical and philosophical underpinnings of those tests is followed by a critique of statistical hypothesis tests in general and significance tests in particular. The authors state their own opinion clearly, but by no means do they imply that readers who come to a different informed conclusion are misguided. Although they comment that “misuse of statistical testing does not mean that the process is flawed … it can provide a sensible and intelligent means of evaluating biological hypotheses” (p. 54), they also point out that “only by planning studies and experiments so they have a reasonable power to detect an effect of biological importance can we relate statistical and biological significance” (p. 54).

Sections in the first three chapters provide a lucid synopsis of Bayesian approaches, or “degrees of belief” (p. 8), compared with frequency interpretations of probability. The authors favor frequentist methods, but they are not dismissive of Bayesian techniques. Rather, they recommend that biologists become familiar with Bayesian philosophy and methods and understand how they can be used “as an alternative way of dealing with conditional probabilities” (p. 9). Chapter 3, for instance, includes an objective discussion of the theoretical and practical differences between frequentist and Bayesian frameworks in the context of hypothesis testing.

Chapter 4 examines graphical exploration of data—exploratory analysis, transformations, standardizations, and outliers. It stresses that, even in its earliest stages, data analysis need not and should not be a random walk but instead should proceed according to clear objectives and methods. Chapters 5 and 6 focus on how correlation and regression can be used to describe the relationship between two or more continuous variables to explain the variability in a dependent variable and, once relationships between variables are understood, to make predictions. Chapter 6, “Multiple and Complex Regression,” is an outstanding guide to determining which predictors are important; detecting and accounting for collinearity, interaction terms, and categorical variables; and alternative methods for finding the “best” model, including information criteria and stepwise procedures. In addition, it introduces hierarchical partitioning, regression trees, and path analysis. Anyone who has attempted to distill defensible inferences from a voluminous matrix of data—or confronted subsequent reviewers' comments on that work—will find tremendous value in these pages.

Discussions of experimental design and power analysis (chapter 7) and single-factor and multifactor analyses of variance (chapters 8 and 9) harbor few surprises. Nonetheless, they are cogent and thorough. Sections on replication, controls, randomization, and independence speak directly to the aspects of field studies that often lead to “messy” data, such as difficulty in replicating large-scale or controversial treatments. Several of the worked examples incorporate multiple statistical issues rather than treating each type of model or test in a vacuum. Unbalanced, nested (hierarchical), and factorial designs all are investigated at length.

Chapter 10 examines unreplicated two-factor designs, including randomized complete blocks, repeated measures, time as a blocking factor, Latin squares, and other complex designs. Chapter 11 then considers split-plot and repeated-measures designs. These chapters have considerable potential utility for deciding how to set up an experimental study and how to analyze the resulting data. Students in particular are likely to appreciate the attention to hypotheses associated with each design, their power and ease of interpretation, and main effects versus interactions.

A full chapter is allocated to analyses of covariance. A useful section explains how to deal with fitting a model for heterogeneous slopes, particularly with statistical software that might assume homogeneity of slopes by default. Chapter 13 deals with logistic regression, Poisson regression, and other techniques for modeling data that do not have a normal distribution. Such models are increasingly common in ecological studies, and understanding of their assumptions and most appropriate uses often seems to be taken for granted. These chapters are a helpful overview for biologists whose basic courses or textbooks on statistics did not emphasize generalized linear modeling. Analysis of frequencies (chapter 14) receives competent if unremarkable treatment.

Chapters 15 to 18 are an excellent exposition of techniques for detecting patterns and structure in complex multivariate data sets and simplifying such data sets for further analyses. They are often heavy on mathematical detail, but they do not assume knowledge of linear algebra. Instead, they demystify eigenvectors, eigenvalues, and similar elements that are somewhat opaque to many ecologists and rarely described fully in the literature. Comparison of similarity and dissimilarity measures is detailed and robust, a welcome contrast to the more cursory treatment these metrics often receive in “biodiversity” texts. Material on symbolic representations of multivariate data for use in exploratory analyses—Chernoff faces and star plots, for example—is novel and compelling. Multivariate methods are often used by biologists who want to explain variation in a large data set but do not have a clear sense of their hypotheses or how they will apply their results. Multivariate approaches are not interchangeable, but there are few resources that equip workers to articulate readily which methods are most appropriate in a given context. Quinn and Keough make substantial progress toward filling that gap.

The book concludes with a pragmatic and entertaining chapter on tabular and graphical presentation of results. For example, the authors cite several of Tufte's neologisms that are eminently suitable in the PowerPoint era, such as “chartjunk … [which is] extraneous ornamentation that puts fancy things all around, but doesn't help explain your results” (p. 499). Sections on displaying summaries of data and error bars are particularly useful.

Experimental Design and Data Analysis for Biologists largely succeeds in its objectives of equipping biologists to design efficient sampling programs, understand the models underlying the most common experimental designs, and appreciate how statistical tools are applied, in theory and practice, to a breadth of pure and applied ecological research. In its straightforward prose and effective examples, it engages readers and goes far toward breaking down barriers between statistics and biology.

Ancillary