The Performance of Binary Mixed-effects Models in Settings with Correlation between Random Effects and Model Covariates

• Activity Number: 157
• Type: Contributed
• Date/Time: Monday, August 9, 2004 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): John Neuhaus*+ and Charles E. McCulloch
• Companies: University of California, San Francisco and University of California, San Francisco
• Keywords: clustered data ; conditional likelihood ; misspecification

Abstract

Standard mixed-effects regression models typically assume that the random effects are independent of model covariates but this is frequently not the case in practice. Clustered datasets often exhibit variability in cluster means of the response that cannot be explained by covariate effects and independent random effects. We consider binary models with random intercepts and examine the effect of ignoring such correlations on standard analysis methods. These investigations suggest that while standard mixed-effects models fit by maximum likelihood yield biased estimates of covariate effects, conditional likelihood methods and mixed-effects models that partition covariate effects into between- and within-cluster components often provide nearly unbiased estimates of parameters of interest. We further illustrate the results with model fits to real data.