Propensity Score Matching for Multiple Treatments using Generalized Boosted Models

Propensity score analysis becomes complicated when there are multiple treatments. Existing studies bypass this complication by focusing on two of the treatments at a time and calculate the pairwise treatment effects, but this pairwise comparison does not allow advanced analysis that investigates how different treatments interact with other variables to influence the dependent variable. In this paper, we show how to obtain a matched sample with balanced pretreatment variables for a dataset with multiple treatments. Using Generalized Boosted Models (GBMs) to calculate the generalized propensity score vector, our proposed procedure match observations on a multi-dimensional space. Researchers to directly conduct complex econometric analysis on the matched sample using standard statistical packages without the necessity of additional programming. Our results also demonstrate the advantage of GBMs over the commonly used multinomial logistic regressions in calculating the generalized propensity score.