Randomization Inference and Sensitivity Analysis with Binary Outcomes in Matched Observational Studies Through Integer Programming

• Activity Number: 246
• Type: Contributed
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 5:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Colin Fogarty* and Dylan Small and Mark E. Mikkelsen and Pixu Shi
• Companies: The Wharton School and University of Pennsylvania and University of Pennsylvania and University of Pennsylvania
• Keywords: Randomization Inference ; Observational Studies ; Integer Programming ; Sensitivity Analysis ; Matching ; Binary Outcomes

Abstract

We develop methods to address a common problem in the analysis of matched observational studies: how can one use randomization inference to perform hypothesis testing and to form confidence intervals for meaningful causal estimands when outcomes are binary? We show that for many causal parameters of interest with binary outcomes, inference under no unmeasured confounding can be conducted through the solution of an integer linear program. We further show that through solving an integer quadratic program, one can perform a sensitivity analysis assessing the robustness of a study's findings to unmeasured confounding. We highlight the strength of our integer programming formulation, which allows for an expedient solution of the required optimization problem even with large data sets and large matched set sizes. We present an observational study comparing the impact of two post-hospitalization protocols on 60-day hospital readmission rates to illustrate our methods.