The Impact of Imbalance in Two-level Logistic Models

• Activity Number: 374
• Type: Topic Contributed
• Date/Time: Wednesday, August 1, 2007 : 8:30 AM to 10:20 AM
• Sponsor: Section on Health Policy Statistics
• Author(s): Huanyu Chen*+ and Roslyn A. Stone
• Companies: University of Pittsburgh and Veterans Affairs Pittsburgh Healthcare System
• Keywords: random intercept (RI) model ; random coefficient (RC) model ; racial disparities research

Abstract

We extended explicit variance formulas for fixed effects in two-level balanced binary data (Moerbeek et al. 2002, 2005) to account for imbalance both between and within clusters. Derivations were based on a linearization of 2-level RI and RC logistic models using first order marginal quasi-likelihood (MQL1). In a study of racial disparities in 30-day mortality among black and white veterans under age 65 following hospitalization for pneumonia at 149 VA hospitals (n=37,111), the site-specific proportions black ranged from 0% to 89.9% (29.1% overall). The observed between- and within-site imbalances in these VA data increased the variances of the race coefficients in the RI and RC models by factors of 1.23 and 1.16, respectively, relative to the balanced case. Compared to simulation results using second order penalized quasi-likelihood, analytic variances based on MQL1 are biased downward.