Multiple Imputation for Incomplete Background Variables in Large-Scale Assessment Surveys

• Activity Number: 658
• Type: Contributed
• Date/Time: Thursday, August 4, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Yajuan Si*+ and Jerome P. Reiter
• Companies: Duke University and Duke University
• Keywords: Baysian ; missing ; imputation ; categorical ; education

Abstract

Large-scale surveys of educational progress are group-level assessments targeting policy relevant subgroups. In addition to cognitive data based on tasks that are designed to elicit what students know and can do, a large number of background variables are collected. However, background variables are often collected through self-reports, where missing values are common and non-negligible. Imputation methods for continuous data cannot be applied directly when all the background variables are categorical. Modeling multivariate categorical data of high dimensions is even more challenging. We would like to propose an efficient Bayesian imputation approach and evaluate different imputation methods. We implement a nonparametric approach using Dirichlet process mixture of product multinomial distributions for imputation. Measuring the influence of missing values of background variables on subgroup reporting will meaningfully contribute to the initial survey design and the latter operation analysis. Synthetic and real data analysis will be examined.