ESTIMATION of AVERAGE TREATMENT EFFECTS AMONG MULTIPLE TREATMENT GROUPS by USING ADAPTIVE ENSEMBLE METHOD

• Activity Number: 72 - Methods for Causal and Integrative Analysis in Health Studies
• Type: Contributed
• Date/Time: Sunday, July 30, 2017 : 4:00 PM to 5:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Maiying Kong*
• Companies: University of Louisville
• Keywords: Treatment effect ; generalized propensity score ; balance of covariates ; ensemble method ; random forest

Abstract

In medical research, it is quite common to compare the outcomes from different treatments or procedures. Imbens (2000) extends the use of propensity score from two treatment groups to multiple treatment groups by introducing the generalized propensity score (GPS). The GPS is usually estimated by multinomial logistic regression model or ordinal logistic regression model. Once GPS is estimated, stratification, inverse probability weighting (IPW), or doubly robust (DR) methods are applied to estimate the treatment effect. In recent years, machine learning methods such as random forest and generalized boosting method have been applied to estimate the GPS. In this presentation, we propose an adaptive ensemble method to estimate the GPS, where the adaptive ensemble method was formed by picking up the top GPS estimating method in terms of balancing the covariates. We estimate the treatment effect using the resulting GPS along with stratification, IPW and DR. Extensive simulations show that the proposed adaptive ensemble method has performance very well in estimating treatment effect.