Remedying Selection Bias and Omitted Variable Bias With Multilevel Observational Data

• In Event: AERA Poster Session 4
  In Poster Session: Division D Section 2 Poster Session 1

Abstract

Propensity score (PS) analysis has gained popularity for estimating treatment effects with observational data. This analysis requires researchers to measure all confounding covariates that influence treatment selection and outcome processes. When there exist omitted/unmeasured variables, the consistency of parameter estimates is not guaranteed due to violation of the conditional ignorability assumption. To handle selection bias and omitted variable bias, this study assesses the performance of three PS estimators—random-effects, fixed-effects, and hybrid models (centered random-effects models)—under the condition where unmeasured cluster-level covariates exist in multilevel data. A simulation study investigates the finite sample properties of the estimators, and their performance is also evaluated with covariate balance and sensitivity analysis. An empirical example using TIMSS data is provided.

Authors

• Youmi Suk, University of Virginia

• Jee-Seon Kim, University of Wisconsin - Madison

• Hyunseung Kang, University of Wisconsin - Madison