A Bayesian approach for handling response bias and incomplete data
Department of Research Methodology, Measurement, and Data Analysis (OMD), Twente University
Project: project financed by NWO
Project running from: 1 July 2008 – 1 July 2012
Supervisors: Prof. dr C.A.W. Glas & dr ir J.-P. Fox (Twente University)
The collection of data through surveys on personal and sensitive issues may lead to answer refusals and false responses, making inferences difficult. Respondents often have a tendency to agree rather than disagree (acquiescence) and a tendency to give socially desirable answers (social desirability). The randomized response (RR) technique has been used to diminish the response bias. Attention will be focused on the usefulness of the randomized response technique. Different settings will be explored, large-scale but also small-scale survey data for binary and polytomous response data. Methodological developments will be made to handle different settings and to test different real-data hypotheses.
Besides the problem of misreporting, respondents may not report an answer to one or more questions. Missing data can also occur due to other causes like, interviewer errors (omitted questions, illegible recording of responses, etc.), and inadmissible multiple responses. In fact, it is not unusual for large data sets to have missing data on a few items. The persons cannot be omitted from the analysis based on the fact that they skipped a few questions since it will result in deletion of a substantial part of the data (these participants provide information on the answered items). In a Bayesian approach, the incomplete data problem can be solved by repeatedly solving the complete data problem. In the setting of large-scale comparative survey data, attention is focused on country-specific imputation methods and/or models for the missing data mechanism.
Date of defence: 6 December 2012
Title of thesis: Bayesian randomized item response modeling for sensitive measurement