Marginal Mean Weighting Through Stratification for Identifying Subpopulation-Specific Optimal Treatment Sequences

• Activity Number: 10
• Type: Invited
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Committee on Women in Statistics
• Author(s): Guanglei Hong*+
• Companies: The University of Chicago
• Keywords: Causal inference ; non-parametric ; Peters-Belson method ; propensity score ; time-varying treatments

Abstract

There have been intense debates about how many years of English language learning (ELL) services would be optimal for language minority students in the US schools. Evaluations are challenging because ELL services are allocated on a yearly basis often according to a student's current English proficiency. Moreover, a treatment sequence optimal for "fast" English language learners might not be optimal for "slow" learners. Past research has provided strategies for evaluating dynamic treatment regimes in experimental settings and for evaluating time-varying treatments in non-experimental settings. We extend marginal mean weighting through stratification (MMW-S), a propensity score-based non-parametric weighting method, to evaluations of alternative treatment sequences. The MMW-S results are generally robust when the functional form of the propensity score model is misspecified. Adapting the Peters-Belson method, we identify optimal treatment sequences for subpopulations of individuals who would display distinct response patterns when assigned to the same treatment sequence. We use simulations to assess the performance of our analytic procedure in comparison with the existing approaches.