Epidemiology: An Introduction

Epidemiology: An Introduction Kenneth J. Rothman New York: Oxford University Press, 2002. ISBN: 0-195-13553-9 The most frequent question I've been asked by colleagues and students about Kenneth Rothman's new book 1 is whether this is “an introduction to Modern Epidemiology”2 (Rothman's now-standard reference, coauthored with Sander Greenland in its second edition). The fact that no one has asked why Rothman would add yet another text to an increasingly crowded field reflects the respect Rothman commands. The question about the book's relation to Modern Epi II reflects an eagerness for a book that would serve as a stepping-stone to that more formidable text. Indeed, in an interview with Roger Bernier in Epidemiology Monitor, 3 Rothman describes his new text as “a ramp” to the earlier, more advanced, book. In that interview he also addresses the issue of the intended audience, namely anyone taking a first course in epidemiology, but especially students (and, one assumes, practitioners) in clinical disciplines. Epidemiology:An Introduction provides the perspective of a thoughtful and influential epidemiologist on the fundamental elements of his discipline. The result is a book that some readers may find too theoretical (Rothman has said some have called it “philosophical”3), and not sufficiently concrete. Indeed, that may be one of the chief problems with using this book as Rothman intends, namely as an introductory text. He acknowledges it probably should not be the sole text for an introductory course and that some elements one usually expects in an introductory text are missing, eg, the history of epidemiology and the role of epidemiology in public health. Rothman's focus is to clarify the “unifying set of ideas” that underlie the epidemiologic enterprise. 3 (That goal perhaps explains why he feels it is important to open the first page of the book with a discussion of confounding.) In the Preface, Rothman proposes that “epidemiology is much more than finely dressed statistics.” He argues that epidemiology has its own foundation, and statistics (though an important tool) is not it. His intent is to present basic “epidemiologic principles and concepts.” The Preface also acknowledges the omission of the history of epidemiology, infectious disease epidemiology, and social determinants of health and disease. Those omissions and the choice to include a chapter on clinical epidemiology (but not public health applications) are indications of the intended audience. This places the text squarely in the tradition of Hennekens and Buring. 4 One implication of this perspective is that the book begins with definitions of epidemiology that focus on the study of disease. Rothman is clearly aware that epidemiologic concepts and methods apply to a broad range of health outcomes. However, the focus on the clinical setting places the emphasis on disease and treatment, rather than on prevention and public health. The organization, content, and pedagogical approach are also more closely akin to the textbook by Hennekens and Buring 4 than to some more recent introductions, such as the text by Leon Gordis. 5 For example, Gordis’ explanation of length-biased sampling (p. 263f) will appeal to the visually oriented, whereas those who prefer more narrative explanations will appreciate Rothman's 1 treatment of the same issue (p. 204). The first chapter introduces epidemiologic thinking, with particular emphasis on confounding. To Rothman's way of thinking, the understanding of why things are not always as they appear is a central feature of epidemiology. Interestingly, he chooses to approach this without a detailed explanation of what is involved, why it occurs, or even of the basic idea of rates. The second chapter introduces Rothman's thinking about causality, which he believes is at the heart of his approach. (This is the one chapter that can be downloaded at the book's Web site: www.oup-usa/epi/rothman.) This is not a chapter that can be dismissed as mere philosophical musings. It lays the foundation for much of what follows. Rothman proposes that an understanding of causality is essential to grasp the meaning of attributable fractions and strength of association, and it provides justification for his focus on additive, rather than multiplicative, assessment of effect measure modification. The chapter includes an important critique of the standard approach to causal criteria. It also challenges the concept of generalizability in terms of statistical representativeness. In the context of that discussion, Rothman makes a telling distinction between applied epidemiology and the science of epidemiology (p. 21). The third chapter provides an explanation of epidemiologic measures, their interpretation and relationships, and a linkage to the measures of effect that build on them. There will be no surprises in this chapter to anyone familiar with Rothman's approach to these issues. Many teachers will welcome the clarifying distinctions among rates, risks, and proportions, and the role of time in understanding all of them. One interesting feature of the discussion of point source and propagated epidemics is the more generalized application of terms often reserved for infectious disease epidemiology. One of the strongest features of the chapter on study design is the unified approach. Positing the cohort study as “the archetype for all epidemiologic studies,” Rothman makes a case that other study designs are actually variations on the cohort design theme. This enhances an understanding of the fundamental goals of epidemiologic studies, and the ways in which the various study designs contribute to achieving those goals. It also justifies the case-control design as a valid scientific approach. Rothman points out that all case-control studies could be viewed as nested within a larger cohort—an important conceptual foundation for understanding the case-control design. In this chapter, he also explains the meaning and impact of latency and induction times, and provides an illuminating discussion of control selection. The discussion of bias in Chapter 5 is clear and to the point. Rather than trying to name and distinguish every conceivable bias, the chapter discusses the familiar broad categories of selection bias, information bias, and confounding, each with clear examples. For those who have read Rothman's previous work, the chapter on “Random Error and the Role of Statistics” will bring a smile of familiarity. Here we find the arguments for estimation rather than statistical hypothesis testing. Confidence intervals (CI) are recommended for describing the precision and location of estimates rather than meager P-values (or, heaven forbid, a mere yes or no judgment of statistical significance). This chapter should be required reading for all epidemiology and biostatistics students, health researchers, authors of journal articles, reviewers, and editors. Just as important, the chapter discusses the limitations of CIs. Lest we forget, CIs (as well as any statistical tests) assume that chance is the only thing causing estimates to deviate from true values. This chapter reminds us that absence of bias is the hidden assumption here—an assumption that is seldom mentioned, perhaps rarely met. The author asserts in the Preface that his emphasis is not on “statistics, formulas, or computation.” This is borne out in the scarcity of algebra. Most of the formula-intensive material is confined to Chapter 7, which includes calculation of confidence intervals, and Chapter 8, where the focus is on calculation of pooled and standardized estimates of effect measures. Chapter 8, on stratified analysis to control for confounding, points out the similarity of goals (control confounding) and method (stratification) of standardized and pooled estimates while describing how they differ (the source of the weights). The argument is made that when the effect varies substantially across strata, standardization can still be used, even though pooling (such as the Mantel-Haenszel approach) may not be appropriate. Because standardization incorporates arbitrary weights that may be external to the data, the standardized estimates are highly influenced by the standard weights used. This variability due to selection of the standard is magnified when there is also variability in the effect across strata, a point not acknowledged in this discussion. Standardized ratios are also artifacts of the standard chosen, but will vary less when the stratum-specific ratios are consistent. An important section of the chapter on stratified analysis is the discussion of “testing” for confounding. The explanation is based on the nature of confounding (that is, what confounding is and why it happens), something that reliance on statistical testing of associations ignores. The chapter concludes with arguments for preferring stratified analysis over multivariable analysis. The chapter on measuring interactions presents another important challenge to the received tradition. Here the author returns to the causality argument to present a case for assessment of additive rather than multiplicative interaction as the analytic approach of choice. The distinction between statistical interaction and biological interaction (and the caution against relying on the implicit multiplicative assumptions inherent in most regression techniques used in epidemiologic research) should stimulate thoughtful discussion even among those who do not entirely agree that, “The reference point for measuring biologic interaction is additivity of risk differences…” (p. 179) Despite his preference for stratified analysis and his caveats concerning regression techniques, Rothman provides a simple, lucid explanation of basic concepts of regression analysis, with an easy transition from linear to logistic regression. He discusses constraints on the outcome variable, meaningful values of “exposure” variables for interpretation, and problems of interpretation of the intercept when it falls out of the plausible range for the independent variables. One basic technique strangely missing is centering of continuous variables, such as age in the example. If the data from Table 10-1 were centered (say on the mean age), the regression line would have the same slope but the intercept would have an interpretable meaning. Centering has other statistical advantages, especially in more complex models. There is a more puzzling issue in the concluding section of this chapter on modeling. The author had previously noted (p. 180) that evaluation of interaction in statistical models using a logarithmic transformation “amounts to an evaluation of departures from a multiplicative model rather than departures from additivity.” In a later discussion of interaction (p. 196), he suggests that a “factored set of terms” created from composite (or cross-classified) exposure variables allows evaluation of departure from additivity “without imposing the multiplicative relation implied by the model.” However, for categorical variables (and his approach involves converting continuous variables to categories or “factoring”), one obtains the same odds ratio estimates from dummy variables for a composite variable and from main effect and product interaction terms for the individual variables (or their dummy variables). This is quite easily seen for the example given of two dichotomous terms. Using the three factored terms for the composite and using the two individual variables and their product will produce exactly the same result in a logistic regression model. The change in coding does not change the multiplicative nature of the logistic transformation and the resulting regression model. Admittedly, the use of the indicator variables from a composite variable will simplify the calculation of the odds ratios and their respective confidence intervals (although many of those calculations can now be obtained from the analysis software). The tradeoff is that one is limited to a single reference category, ie, calculation of the stratum-specific odds ratios is no easier than in the model with the product term. However, evaluating departure from additive interaction using an “observed vs expected” approach may be facilitated by the proposed approach. The final chapter is devoted to epidemiology in clinical settings. This is the only chapter focused on a particular application of epidemiology. There are brief mentions of pharmacoepidemiology and outcomes research, but the bulk of the chapter is devoted to an explanation of screening tests and clinical trials. One of the highlights of the chapter is the box concerning statistical testing for baseline differences after randomization (“An unrejectable null hypothesis,” p. 209). It is Rothman at his best. (Indeed, some of the most interesting discussions in the book are in the boxes, so pay attention to them.) There are a few places where the wording is somewhat loose, which is surprising given the usual concern of this author for precision in discussing epidemiologic concepts. For example, he writes that, “Even unknown risk factors will not confound in a randomized experiment of sufficient size” (p. 109). Elsewhere, he provides examples where randomization does not eliminate confounding. Confounding remains a possibility in real-world studies (although a diminishing one as sample size increases). For any realistically sized study, it would be more appropriate to say that unknown risk factors are unlikely to confound. Even more puzzling is the description of a risk difference as indicating “a benefit of 17% smaller risk” (p. 136). From the earlier discussion of effect measures (eg, p. 48), one would assume the latter phrase referred to a risk ratio of 0.83, not to a risk difference of −0.17. It does not appear on first reading, however, that this book will be plagued by the hundreds of corrections necessary for Modern Epi II. As of this writing, there is only one erratum listed, and I found only one other, although I did not check all the formulas closely. The questions at the end of each chapter are designed to prompt thinking about the concepts rather than repetition of phrases from rote memory. As such, these exercises may prove to be valuable teaching tools, although many instructors may want to supplement them with questions involving application, literature critique, calculation, and interpretation. There is an accompanying website that includes contributed answers to the questions, as well as a discussion forum, corrections, and some useful spreadsheet programs available for download. In sum, this book provides a splendid introduction to our discipline. It is not a public health textbook and many teachers of epidemiology will not find it sufficient in itself—as the author acknowledges. However, it is an important addition to the epidemiology teacher's armamentarium, providing a fuller discussion of fundamental conceptual issues than is typically found at the introductory level. There is ample material for thought here for the experienced researcher as well as the neophyte. The book will most certainly be welcomed by those who would like an entrée to the more challenging world of Modern Epi II.