B1192
Title: Studentized sensitivity analysis in paired observational studies
Authors: Colin Fogarty - Massachusetts Institute of Technology (United States) [presenting]
Abstract: A fundamental limitation of causal inference in nonrandomized studies, including observational studies, broken experiments, and synthetic clinical trials, is that perceived evidence for an effect may well be explained away by factors not accounted for in the primary analysis. This deficiency necessitates an additional step, known as a sensitivity analysis, which assesses how strong unmeasured confounding would have to be in order to materially alter the findings of the study. Methods for conducting a sensitivity analysis have been established under the assumption that the treatment effect is constant across all individuals. This assumption is viewed as innocuous in the analysis of randomized experiments, as the variance for the estimated average treatment effect is maximized if the treatment effect is additive; however, in the potential presence of unmeasured confounding, it has been argued that certain patterns of effect heterogeneity may render the performed sensitivity analysis inadequate. We present a new method for conducting a sensitivity analysis in the presence of heterogeneous treatment effects. The method naturally extends conventional tests for the sample average treatment effect in a randomized experiment to the case of unknown, but bounded, probabilities of assignment to treatment. In so doing, we illustrate that concerns about the restrictiveness of the constant treatment effect model are largely unwarranted.
