Difference-in-differences(DID) is a popular method to evaluate the effect of a treatment with multiple subpopulations (treatment and control groups) and outcomes that are measured in each group before and after the treatment.
To obtain Abadie(2005)'s DID estimator, two stage estimation are performed:(stage-1) estimating propensity score, (stage-2) computing DID estimator using estimated propensity score. Same as many causal estimators, performance of Abadie's DID estimator depends on estimated propensity score, so the model assumption for propensity score is substantially important. However, it is difficult to specify propensity score model properly in practice.
In this presentation, we propose a method to estimate DID by random forest, extending the result of Wager and Athey(2017), without estimating propensity score. We show random forest based DID estimator is consistent and asymptotically normal, and illustrate random forest based DID estimator works well under moderately small observations through simulation experiments. At last, we illustrate the result of comparing random forest based DID and other causal effect estimators using Lalonde data.
|