Propensity Score Analysis with Missing Data: The Comparison of Multiple Imputation Approaches

• Activity Number: 691
• Type: Contributed
• Date/Time: Thursday, August 13, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Social Statistics Section
• Author(s): Eun Sook Kim* and Jeffrey Kromrey and Seang-Hwane Joo and Yan Wang and Jessica Montgomery and Reginald Lee and Patricia Rodriguez de Gil and Shetay Ashford and Rheta Lanehart and Chunhua Cao
• Companies: and University of South Florida and University of South Florida and University of South Florida and University of South Florida and University of South Florida and University of South Florida and University of South Florida and University of South Florida and University of South Florida
• Keywords: missing data ; propensity score analysis ; multiple imputation ; simulation

Abstract

The appropriate treatment of missing data under different missing data mechanisms is essential for unbiased estimates and correct statistical inferences in propensity score analysis (PSA). This simulation study investigates the efficacy of two missing data techniques (multiple imputation and listwise deletion) in PSA. For multiple imputation, four different approaches are considered in combination of two factors: what to impute (covariates only or PS in concert with covariates) and how to combine multiply imputed data (average treatment effects or average PS). Simulation design factors include sample size (500, 1000), treatment effect magnitude (0, .05, .10, .15), correlation between covariates (0, .50), proportion of missing observations (.20, .40, .60), proportion of missing covariates (.20, .40, .60), the number of covariates (15, 30), and missing data mechanisms (MCAR, MAR, MNAR). The missing data treatments serve as a within group factor. Imputing covariates only, combined with averaging treatment effects estimates across imputations, outperforms other methods under MAR, but none of multiple imputation approaches is apt under MNAR.