Making Inference from Multiply Imputed Data Sets Using Mixture Distributions

• 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): Sana Rashid*+ and Robin Mitra and Russell J. Steele
• Companies: University of Southampton and University of Southampton and McGill University
• Keywords: Missing Data ; Multiple Imputation ; Rubin

Abstract

We focus on strategies of making inferences in the presence of missing data after conducting a Multiple Imputation (MI) procedure. MI's combining rules for inferences (Rubin, 1987) are tested against a new methodology developed using the full posterior density of an estimand of interest. This density is approximated using a particular mixture distribution. Simulated experiments show this approach improves inferences in some aspects, making them more stable over repeated analysis and creating narrower bounds for certain common statistics of interest. Extensions to existing literature have been executed that provide further stability to inferences and also a strong potential to identify ways to make the analysis procedure more flexible. The new approach suggests a variation to the existing mixture approach, utilising Rubin's variance estimates for the statistics of interest to improve its overall performance. The competing methods have been first compared using simulated data sets and then a real data set concerning analysis of the effect of breastfeeding duration on children's cognitive ability.