Design of observational studies (Q5918888)

Observational studies are a fundamental part of research, e.g. in epidemiology, social sciences, psychology. They are called observational studies because the investigator observes individuals without manipulation or intervention. This is in contrast with experiments, such as randomized controlled trials, where each subject is randomly assigned to a treated group or a control group. From the preface to the first edition [Zbl 1308.62005]: ``An observational study is an empirical investigation of effects caused by treatments when randomized experimentation is unethical or infeasible. The quality and strength of evidence provided by an observational study is determined largely by its design. Excellent methods of analysis will not salvage a poorly designed study. The line between design and analysis is easier to draw in practice than it is in theory.'' The author on page 4: ``At the most elementary level, a well designed observational study resembles, as close as possible, a simple randomized experiment.'' The present monograph is the second edition of Rosenbaum's book from 2010. Already the first edition was reviewed very friendly by statistical journals. For example, \textit{L. W. Miratrix} from Havard University writes in Psychometrika 2014 [Psychometrika 79, No. 3, 540--542 (2014; Zbl 1308.00023)]: ``Design of Observational Studies (DOS) is an introductory textbook on analyzing observational data built on permutation-based inference, matching, and sensitivity analysis. The book is both a pragmatic how-to manual and a delightfully written manifesto. It is quite refreshing to read a textbook that has any personality whatsoever, not to mention one with a philosophy quite different from what one typically encounters.'' Small critizism can be found also, e.g. P. G. Gilbert, University of Washington writes in Biometrics 2010: ``A limitation -- or simply a fact -- of the book is that it almost exclusively considers the technique of matching for minimizing observable biases, so much so that the book may have been appropriately titled, `Design of matched observational studies'. While the principles and concepts translate well to unmatched techniques such as regression, the absence of discussion about how the translation is made may make it hard for some readers to apply the book's ideas to the design of unmatched studies.'' The second edition differs from the first in two ways: There is additional material (four new chapters, 7, 19, 20, 21, and numerous additions and revisions to existing chapters, including several new sections) and there is a new relationship with the R statistical language, including a companion R package DOS2. The book is structured in five parts: `Beginnings', `Matching', `Design sensitivity', `Enhanced design' (new) and `Planning analysis'. The first part outlines an overall approach for tackling observational studies and gives background material on analyzing randomized trials. Part II goes into matching in more detail. Following a detailed overview of matching, including propensity score matching and exact matching, there are several short chapters that provide a wide array of matching approaches with discussion on how these various techniques might be more or less appropriate in different contexts. Part III discusses sensitivity analysis and its central role in observational studies. This part teaches how to use sensitivity analysis to compare alternative designs for an observational study, discusses the role of heterogeneity in observational studies, suggests how to evaluate uncommon but dramatic responses to treatment (by contrasting different approaches) and shows how expected patterns of responses may affect sensitivity to unmeasured biases. Part IV discusses evidence factors (a study has two evidence factors if it permits two statistically independent tests of hypothesis of no treatment effects, where biases that would invalidate the first test have no effects on the second) and the construction of groups (comparison among several groups provide strengthened evidence). Part V is the shortest part of the book but conveys a very important message by showing how to improve design sensitivity and how to apply Ronald A. Fisher's suggestion `to make theories elaborate' to disambiguate the association between treatment and outcome. The preface to the first edition ends with: ``To end well, how should we begin?'' A possible answer could be: By reading this book.