MULTIPLE IMPUTATION FOR MISSING ITEMS IN MULTI-THEMED QUESTIONNAIRES

Open Access
Author:
Liu, Rong
Graduate Program:
Statistics
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
February 26, 2010
Committee Members:
  • Joseph Francis Schafer, Dissertation Advisor
  • Joseph Francis Schafer, Committee Chair
  • Murali Haran, Committee Member
  • Runze Li, Committee Member
  • D Wayne Osgood, Committee Member
Keywords:
  • factor model
  • EM algorithm
  • Markov chain Monte Carlo
  • incomplete data
Abstract:
Questionnaires used in survey-based research are often arranged in multiple sections. Each section contains items that are closely interrelated, serving one or more themes. Even with a modest number of sections, the resulting dataset may have a large number of variables, which poses special analytic challenges for dealing with missing values. Current procedures for multiple imputation may fail because the underlying models do not take into account the thematic nature of the questionnaire and are over-parameterized. Attempts to simplify the model--for example, by assuming that the items within a theme are conditionally independent given a small number of latent factors--may fail to capture special features of the data if the specified model does not fit. In this dissertation, I develop new multiple-imputation procedures for multi-themed questionnaire data based on a fexible class of confirmatory factor models. I present PX-EM algorithms for maximum-likelihood estimation in exploratory and cofirmatory factor analysis with incomplete data. The factor model is then relaxed by adding an additional random component which allows the covariance structure to deviate from the assumed model. I present an MCMC algorithm for generating Bayesian multiple imputations under this extended model. These techniques are illustrated using data on emotional distress from a large adolescent health survey.