Using the Fraction of Missing Information (FMI) to Identify Auxiliary Variables for Imputation Procedures via Proxy Pattern-Mixture Models

• Activity Number: 467
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Jenny Thompson*+ and Rebecca R. Andridge
• Companies: U.S. Census Bureau and Ohio State University
• Keywords: maximum likelihood ; fraction of missing information ; imputation procedures ; proxy pattern-mixture models

Abstract

The survey methodology focus is changing from retrospective analyses towards pre-emptive corrective procedures. Of course, these procedures are not limited to data collection: examples include imputation, adjustment cell weighting, and calibration. In this paper, we consider imputation procedures, which can be used to account for either unit nonresponse or item nonresponse. We propose using a proxy pattern-mixture (PPM) model analysis to evaluate the predictive power of different sets of covariates used in regression imputation models or to determine imputation cells for one or more outcome variables, obtaining the fraction of missing information (FMI) obtained via maximum likelihood estimation from separate PPM models fit to the same data sets. Our variable selection approach is most like that of Särndal and Lundstrom (2010), in that we estimate the FMI and look for the point at which changes in the FMI level off and further auxiliary variables do not improve the imputation model. We illustrate our proposed approach using empirical data from the Service Annual Survey and from the Ohio Medicaid Assessment Survey.