Causal Effect Random Forest of Interaction Trees for Observational Data, Applied to Educational Interventions

• Activity Number: 215 - Contributed Poster Presentations: Section on Statistical Learning and Data Science
• Type: Contributed
• Date/Time: Tuesday, August 4, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Juanjuan Fan* and Luo Li and Xiaogang Su and Richard Levine
• Companies: San Diego State University and San Diego State University and University of Texas, El Paso and San Diego State University
• Keywords: propensity score; machine learning; individualized treatment effect; quasi-experimental study design; educational data mining; supplemental instruction

Abstract

Estimating causal treatment effects especially individualized treatment effects (ITEs) using observational data holds great interest in education and other research fields. We propose to incorporate propensity score to extend Random Forest of Interaction Trees to Casual Effect Random Forest of Interaction Trees (CERFIT). As opposed to the counterfactual approach, random forest of interaction trees can estimate treatment effect in one model using all of the data. By integrating propensity score into the tree growing process, subgroups from the proposed CERFIT not only have maximized treatment effect differences, but also similar baseline covariates. Thus it allows for the estimation of the individualized treatment effects using observational data. In addition, CERFIT provides variable importance rankings that may shed light on the variables and subgroups that have the most differential effect from the treatment. Simulation studies for assessing estimation accuracy and variable importance ranking are presented. The method is illustrated in an assessment of a voluntary Statistics supplemental instruction course at a large public university.