Design of Observational Studies

Introduction.- Matching to control bias from measured covariates.- Addressing bias from covariates that were not measured.

From the reviews:

"I should begin by noting that the first time I read the book's title I thought that it would address the topic of classical epidemiological designs like case-control studies, cohort studies, etc. However, I was wrong. Rosenbaum's book addresses the crucial topic of designing and analyzing empiric non-randomized investigations to prove causal relationships between treatments and outcomes. These types of studies, the observational ones, are prone to present two well-known selection biases: overt biases, i.e., differences in outcomes between treatments may reflect measured pre-treatment differences between groups rather than effects of treatments, and hidden biases, i.e., the same situation but with pre-treatment differences that were not recorded in the study.

To overcome overt biases, applied statisticians usually advocate the use of model-based adjustments and to overcome hidden biases they usually recommend designing a randomized experiment. However, Rosenbaum's book addresses these two biases with an alternative approach: propensity score matching for overt biases and sensitivity analysis for hidden biases. Practitioners may think that sensitivity analysis means performing several analyses of the same data set; however, this is not the Rosenbaum's approach.

Overall, the book is written in a clear and concise way, merging theoretical and practical aspects. Small examples are provided to develop the understanding of key issues in parallel with real examples of practical size from both the economics and the biomedicine areas. Moreover, although the book is not intended as a statistical software oriented book, the text includes some code in R and SAS. For example, Chapter 13 is devoted to matching in R.

Finally, the book covers all the relevant issues in designing and analyzing treatment effects in observational studies, with the exception of observer bias, i.e., the bias present when the assessment of the