Bayesian Adjustments for Nonignorable Nonresponse in an Incomplete Two-way Categorical Table

• Activity Number: 274
• Type: Contributed
• Date/Time: Tuesday, August 9, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Boseung Choi*+ and YouSung Park
• Companies: Korea University and Korea University
• Keywords: Multi-way contingency table ; Non-ignorable ; non-response ; Empirical bayesian methods ; Priors ; EM algorithm

Abstract

We consider a two-way contingency table with nonignorable nonresponse. There are fully classified counts, partially classified counts, and unclassified counts. The partially classified counts often are presented as row and column supplemental margins and called item nonresponse, whereas the unclassified counts are presented in total and called unit nonresponse. When the response mechanism obeys nonignorable nonresponse in a categorical data analysis, the maximum likelihood estimation often yields boundary solutions where the probability of nonresponse is estimated at zero in some cells. To overcome this boundary solution problem, we propose five empirical Bayesian methods by imposing five priors on cell probabilities. The five priors reflect how much the response mechanism between respondents and nonrespondents is different. The five empirical Bayesian methods include the previous studies with a one-way contingency table for special cases. We illustrate the five methods with data on a pre-election survey for the November 1998 Buckeye State Poll. Simulation studies investigate the performance of the five methods and compare them with ML estimation.