Propensity Score Estimation with Boosted Regression

• Activity Number: 212
• Type: Topic Contributed
• Date/Time: Tuesday, August 10, 2004 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Daniel F. McCaffrey*+ and Greg Ridgeway and Andrew Morral
• Companies: RAND Corporation and RAND Corporation and RAND Corporation
• Keywords: causal effect ; bagging ; generalized boosting ; regression trees

Abstract

Causal effect modeling with naturalistic, rather than experimental, data is challenging. Variations in treatment exposure may be confounded with differences in pretreatment characteristics or treatment selection factors. Propensity score methods can theoretically eliminate these confounds for all observed covariates, but accurate estimation of propensity scores is impeded by large numbers of covariates, uncertain functional forms for their associations with treatment selection, and other problems. This paper demonstrates that boosting, a modern statistical technique, can overcome many of these obstacles. We use generalized boosting regression, which maximizes a likelihood function rather than minimizing prediction error. We report the mean-squared error of estimated treatment resulting from alternative stopping rules for the boosting algorithm including methods based on out-of-bag cross-validation and minimization of the imbalance of covariate distributions between groups. We illustrate this approach with a study of adolescent probationers in substance abuse treatment programs. Propensity score weights estimated using boosting eliminated all large pre-treatment group differences.