Abstract:
|
When choosing a design for an observational study, one hopes not only to minimize bias associated with observed covariates but to ensure that resulting inferences are as robust as possible to unmeasured confounders. For matched studies a measure called design sensitivity, which describes the asymptotic power of a sensitivity analysis for unobserved confounding, allows researchers to compare possible designs on these terms. However, similar tools are not available outside the matching context. We fill this gap by constructing design sensitivity for weighting estimators, using a pilot sample to estimate nuisance parameters not present in the matching case. By comparing design sensitivities, we interrogate how key features of weighted designs, including estimands and model augmentation, impact robustness to unmeasured confounding. We support and illustrate our results through extensive simulations and re-analysis of canonical data examples from the literature on causal inference.
|