An Innovative Multiple Imputation Method to Accommodate Complex Sample Design Features

• Activity Number: 363
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Survey Research Methods Section
• Author(s): Hanzhi Zhou*+
• Companies: University of Michigan
• Keywords: complex sample design ; missing data ; multiple imputation ; Pólya's urn scheme ; synthetic data ; Bayesian Bootstrap

Abstract

Despite the fact that incorporating complex sample design features are important to obtain correct Multiple Imputation (MI) inference, this is not typically done in practice. On one hand, existing model-based MI techniques typicaly require strong model assumptions and expensive computation; On the other, there is no sensible way to accommodate survey weight into imputation model. We developed an innovative MI method that is attentive to design features and robust enough to sufficiently capture the correlational structures among survey variables. Under the new method, the complex feature of the survey design (including weights, clustering and stratification) is fully accounted for at the first step through a nonparametric synthetic data generation procedure; conventional parametric MI for missing data is performed at the second step using readily available imputation software designed for an SRS sample. Our simulation study demonstrates that our method has advantages over existing MI methods, particularly in the presence of model misspecification and/or informative sampling. Extensive applications are conducted on survey data from BRFSS and NHIS.