Primary Submission Category: Difference in Differences
Survival after hospitalization: Constructing counterfactual time-to-event outcomes for difference-in-differences studies
Authors: Laura Hatfield, Bret Zeldow,
Presenting Author: Laura Hatfield*
Improving the health care and outcomes of hospitalized patients is the goal of many health policy interventions. To estimate the effects of these interventions in non-randomized settings, many researchers turn to difference-in-differences designs, which contrast observed outcomes with counterfactual outcomes imputed under a “parallel trends” assumption. Inspired by difference-in-differences designs, we propose a novel strategy to impute a counterfactual survival curve by leveraging the evolution of a comparison group’s survival curve. Our strategy can identify survival time estimands like the difference in median survival. We contrast our novel approach with estimands based on event risks/rates, which can be identified with simple additive and multiplicative versions of parallel trends. In an analysis of survival data for hospitalized patients, we compare feasibility, plausibility, and usefulness of the methods. We show that although estimands based on event risks/rates are simple to identify and estimate, they can obscure important patterns in the presence of censoring. By contrast, our proposed strategy identifies the whole survival curve and estimands based on any summary of it and therefore can correctly account for censoring using simple Kaplan-Meier-based estimation methods.