A Comparison of Covariate Selections for Causal Forests Based on the Multilevel Propensity Score (Poster 17)

• In Event: AERA Poster Session Eleven
  In Poster Session: Propensity Score, Bayesian Techniques, and Applications in Multilevel Modeling

Abstract

Causal Forests model based on the multilevel propensity score has been introduced to estimate the average treatment effect (ATE). This simulation study was designed to assess the accuracy for ATE in multilevel data under a variety of conditions including: ratios of treatment to control group members (1:1, 1:3), the number of groups (50, 100, 200) and the average group sizes (20, 40, 50). The accuracy of ATE estimates was compared among the seven types of covariate selection: T (covariates related to treatment), O (covariates related to outcome), C (covariates related to both treatment and with the outcome), T+C, O+C, T+O, and T+O+C. The relative biases were acceptable including covariates with T+C or T+O+C were included.

Authors

• Hyewon Chung, Chungnam National University
  Presenting Author

• Hyeonjeong Jeon, National Youth Policy Institute
  Non-Presenting Author

• Yeheun Baek, Chungnam National University
  Presenting Author

• Eunah Jang, Chungnam National University
  Presenting Author