Bayesian Predictive Inference for a Proportion with Nonresponse

• Activity Number: 490
• Type: Invited
• Date/Time: Wednesday, August 3, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Balgobin Nandram*+
• Companies: Worcester Polytechnic Institute
• Keywords: Gibbs sampler ; Log-linear ; Multinomial-Dirichlet ; Patterns of missingness ; Sensitivity analysis

Abstract

In any survey an individual may respond to one or more items, and generally it is not possible to determine whether a nonrespondent differs from a respondent with the same traits. In the Slovenian Public Opinion Survey (SPOS) with the three categorical variables (secession, attendance, independence), there are eight patterns of missingness (eight tables), and the emphasis was in determining the proportion of the population who would attend the plebiscite and vote for independence. One strategy is to center a nonignorable nonresponse model on an ignorable nonresponse model, thereby adding uncertainty to the ignorable nonresponse model. The categorical tables are model simultaneously with different parameters, and these parameters are independent and identically distributed. They differ by one or more sensitivity (or centering) parameters, whose prior specifications give the difference between the ignorable and nonignorable nonresponse models. With these specifications, a Bayesian approach is attractive, and the computation can be done using the Gibbs sampler. Thus, it is possible to study imprecision and ignorance to obtain an uncertainty interval. The SPOS data provide an illus