DIAGNOSTICS FOR MULTIPLE IMPUTATION BASED ON THE PROPENSITY SCORE

Abstract

Multiple imputation (MI) is a popular approach to handling missing data, however, there has been limited work on diagnostics of imputation results. We propose two diagnostic techniques for imputations based on the propensity score (1) compare the conditional distributions of observed and imputed values given the propensity score; (2) fit regression models of the imputed data as a function of the propensity score and the missing indicator. Simulation results show these diagnostic methods can identify the problems relating to the imputations given the missing at random assumption. We use 2002 US Natality public-use data to illustrate our method, where missing values in gestational age and in covariates are imputed using Sequential Regression Multiple Imputation method.