A Comparison of Methods for Imputation of Missing Covariate Data Prior to Propensity Score Analysis

• In Event: Bayesian Estimation and Propensity Score Analysis in Educational Statistics

Abstract

Propensity score analysis has become a popular method for quasi-experimental evaluation of educational interventions, because it can remove confounding due to a potentially large number of covariates. However, these covariates may have missing data that have to be dealt with before propensity score analysis. This paper presents a Monte Carlo simulation study comparing imputation strategies for covariates, including single and multiple imputation, imputation with Bayesian linear regression and predictive mean matching, and with propensity score models that include dummy-indicators of missing values. The study focuses on propensity score weighting and evaluates the effects of missing data method choice on occurrence of extreme weights, covariate balance and percent bias reduction.

Authors

• Walter L. Leite, University of Florida

• Burak Aydin, Recep Tayyip Erdoğan University