Estimation and Evaluation of Propensity Scores with Multinomial Treatment Groups

• Activity Number: 648
• Type: Topic Contributed
• Date/Time: Thursday, August 4, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Mary Slaughter*+ and Beth Ann Griffin and Daniel F. McCaffrey and Cha-Chi Fan and Daniel Almirall and Rajeev Ramchand
• Companies: RAND Corporation and RAND Corporation and RAND Corporation and Department of Energy and University of Michigan and RAND Corporation
• Keywords: propensity score ; balance ; multinomial treatments

Abstract

Use of propensity score weighting when examining more than two treatment conditions has received limited attention in spite of theoretical developments showing that it can be used. Moreover, most applications involving more than two treatment groups depend on parametric estimation of the propensity score model when it has been shown that machine learning methods outperform (in terms of improved balance) the use of simple logistic regression models in the binary treatment case. In this paper, we discuss weighting to compare multiple treatment conditions, describe a method for estimating the multinomial treatment propensity scores that relies on generalized boosted models (GBM), and introduce a useful diagnostic criterion for assessing balance that does not rely on traditional p-values. We illustrate our estimation technique and diagnostic criterion on an example dataset which studies the relative effectiveness of four substance abuse treatment modalities.