Primary Submission Category: Sensitivity Analysis
Reconciling Overt Bias and Hidden Bias in Sensitivity Analysis for Matched Observational Studies
Authors: Siyu Heng, Yanxin Shen, Pengyun Wang,
Presenting Author: Siyu Heng*
Matching is one of the most commonly used causal inference study designs in observational studies. However, post-matching confounding bias typically exists, including overt bias due to inexact matching and hidden bias due to unmeasured confounding. Therefore, in matched observational studies, researchers routinely adopt the Rosenbaum sensitivity analysis framework to assess the impacts of post-matching confounding bias (either overt or hidden) on causal conclusions. In this work, we point out that this routine practice can be overly conservative because solving the Rosenbaum sensitivity model often renders the allocations of hypothetical hidden bias in sensitivity analysis contradict the overt bias observable from the matched dataset. To remove this contradiction, we propose an iterative convex programming approach to conduct a more powerful sensitivity analysis by adapting the solution space of hidden bias in sensitivity analysis to be compatible with the overt bias observed from the matched dataset. Our approach is asymptotically valid and uniformly more powerful than the conventional Rosenbaum sensitivity analysis framework and does not require any modeling assumptions on either the treatment or outcome variable. Our approach has been evaluated through extensive simulation studies and applied to an observational study on educational research.