Estimating Difference-in-Differences in Time-To-Event Models: Some Simulation Results

Abstract

This paper uses a simulation experiment to compare the distributional properties of three estimators for the difference-indifferences in time-to-event models. When events are observed in continuous time and subject heterogeneity are observed, the Cox (1972) estimator is the appropriate estimator. When events are observed in fixed time intervals, and subject heterogeneity is not observed, the same estimator is biased. The non-parametric maximum likelihood estimator (which corrects for fixed interval sampling and unobserved heterogeneity) is nearly unbiased. The linear probability estimator is severely biased and can have the wrong sign in some cases.

Keywords: time-to-event models, survival models, difference-in-differences estimation, competing risks, maximum likelihood estimation, proportional hazard models