Multiple Imputation for Missing Covariates in Regression Models with Interactions

• Activity Number: 80
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Soeun Kim*+ and Catherine Ann Sugar and Thomas R. Belin
• Companies: University of California at Los Angeles and University of California at Los Angeles and University of California at Los Angeles
• Keywords: missing data ; multiple imputation ; regression ; interaction

Abstract

Inference based on data from surveys is often complicated due to incomplete data which can lead to bias in the results. Standard missing data procedures for regression models do not reflect possible interactions in regression relationships among variables. We consider a regression model with one fully observed covariate (either dichotomous or continuous) and one continuous variable whose values may be missing. We derive the conditional distribution of the missing covariate given the observed covariate and the outcome variable, and conduct a simulation study comparing the performance of imputation under this correct conditional distribution to that obtained using standard methods which assume multivariate normality of the covariates. Several imputation methods are compared in terms of confidence interval coverage, widths and bias, and the results suggest that imputing the interaction term without constraining it to be a product of covariates performs better than methods that preserve the relationship of interaction term as a product.