Protective Estimation of Random Intercept Logistic Regression When Data Are Missing at Random

• Activity Number: 635
• Type: Invited
• Date/Time: Thursday, August 2, 2012 : 10:30 AM to 12:20 PM
• Sponsor: WNAR
• Author(s): Sophia Rabe-Hesketh*+ and Anders Skrondal
• Companies: University of California at Berkeley and Norwegian Institute of Public Health
• Keywords: mixed model ; conditional logistic regression ; missing data ; NMAR ; conditional maximum likelihood

Abstract

Protective estimation of random intercept logistic regression when data are missing not at random

We consider the impact of missing data mechanisms on the consistency of estimators that are widely used to analyze binary longitudinal data. Rubin (1976, Biometrika) proposed a classification of mechanisms producing missing data and clarified the conditions for consistent estimation based on maximum likelihood or Bayesian inference. An important extension of Rubin's work was provided by Little (1995, JASA) who explicitly considered the role of covariates and latent variables (random coefficients). The conditions provided by Rubin and Little apply to if a random intercept logistic regression model is estimated by marginal maximum likelihood (MML). Fortunately, conditional maximum likelihood (CML) estimation of the model, treating the random intercept as fixed, turns out to be consistent under considerably more lenient conditions.