Calibrated Multiple Imputation

• Activity Number: 593 - Multiple Imputation for Complex Health Survey Data
• Type: Invited
• Date/Time: Thursday, August 3, 2017 : 8:30 AM to 10:20 AM
• Sponsor: Government Statistics Section
• Author(s): Joseph Kang* and Yulei He and Precious Esie and Jaeyoung Hong and Kyle Bernstein
• Companies: National Center for HIV/AIDS, Viral Hepatitis, STD, and TB Prevention and National Center for Health Statistics and National Center for HIV/AIDS, Viral Hepatitis, STD, and TB and National Center for HIV/AIDS, Viral Hepatitis, STD, and TB and National Center for HIV/AIDS, Viral Hepatitis, STD, and TB
• Keywords: Multiple imputation ; Calibration ; Complex Survey ; Weights

Abstract

The multiple imputation (MI) method was originally invented under the Bayesian paradigm that ignores survey weights. However, the general practice of using MI for complex survey data clearly involves weights in all analysis. For example, the U.S. Centers for Disease Control and Prevention (CDC) provides five multiple imputed data sets for missing body fat mass data so that analysts may apply survey weights to each imputed data set and combine results using Rubin's rules. In this process, analysts already employ both the Bayesian-driven MI method and the weights that calibrate characteristics of a sample to approximate those of a target population. In this talk, we provide some insights for the MI paradigm for complex surveys and propose a way of calibrating imputation models with survey weights modified by nonresponse mechanism. The purpose of the calibration is to balance distributions of covariates between the missing and the observed data. With the evidence of balanced covariate distributions, calibrated imputed data are generated for complete case analysis. The calibrated MI method is applied for the analysis of NHANES' incomplete sexuality data.