122 – Topics in Variance Estimation for Complex Surveys
Post-Imputation Calibration Under Rubin's Multiple Imputation Variance Estimator
Michael D. Larsen
The George Washington University
Benjamin M. Reist
U.S. Census Bureau
Multiple imputation has become one of the most popular and successful methods for dealing with missing data in statistical analyses. Multiple imputation allows one to use observed data to model relationships among variables, represent uncertainty in missing values through multiple draws from conditional distributions, and produce both point estimates and variance estimates for parameters. Variance estimates incorporate contributions to variance from both within and between completed data set analyses. Despite the advantages of multiple imputation, it has been noted that multiple imputation variance estimators can be biased. Bias is possible when, in the imputation model, survey weights are not used. Calibration weighting and its familiar forms, including raking and post-stratification, are often used in sample surveys to adjust sample estimates to match control total values and reduce variance. We explore the possibility of using calibration weighting in combination with multiple imputation to remove or reduce bias in multiple imputation variance estimation when survey weights are not used in the imputation model. Methods could apply to both sample survey and more general study design contexts.