Abstract:
|
Matching methods such as propensity score matching used to construct artificial treatment and control groups from observational data can be sensitive to model specification, leading to biased estimates of treatment effects. We introduce proximity score matching, a new matching method based on the proximity matrix of a random forest, where treatment and control observations that tend to end up in the same terminal nodes are matched. Experimental results under a variety of treatment selection processes and on the Lalonde dataset demonstrate that our method is able to estimate average treatment effects with less bias than existing methods such as propensity score matching and Mahalanobis distance matching.
|