Handling Missing Outcomes in Propensity Score: Comparing Multiple Imputation and Multiple Imputation-then-Deletion

• In Event: AERA Roundtable Session Saturday 1:45 pm - Gold 4
  In Roundtable Session: RT3: Imputation Strategies and Synthesis/Meta-Analytic Designs with Mediation Effects or Complex Structures

Abstract

In propensity score analysis (PSA), much research has focused on missing covariates, but outcome missingness remains understudied. This study uses a Monte Carlo simulation to evaluate four strategies for handling missing outcomes: MI-within, MI-across, MID-within, and MID-across. These strategies combine two methodological choices: (1) using standard Multiple Imputation (MI) versus MI-then-Deletion (MID), an alternative that discards imputed outcomes before analysis; and (2) using within versus across pooling methods.
Results indicated that within strategies consistently outperformed across approaches. While MID-within sometimes offered lower bias, the traditional MI-within strategy produced the lowest variance and MSE, making it the most accurate method overall. These findings provide practical guidance for researchers facing missing outcome data in PSA.

Authors

• Xinchang Zhou, University of Illinois at Urbana-Champaign

• Yan Xia, University of Illinois at Urbana-Champaign