Should We Avoid Random Effects Model When Covariates Correlate with Group Effects?

• Activity Number: 192
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section for Statistical Programmers and Analysts
• Author(s): Ta Liu*+
• Companies: Battelle Memorial Institute
• Keywords: Random Effects ; Hausman Test ; Simulation ; Risk Adjustment ; Multilevel ; Dichotomous

Abstract

For multilevel models where covariates correlate with group effects, random effects model yields biased estimates and inflated standard error. The conventional wisdom is to use fixed effects model when the bias is large enough as measured by Hausman-Wu test. An alternative is to add group averages of correlated covariates in the random effects model. This poster evaluates the effectiveness of this approach for healthcare risk-adjustment models where all variables are binary through simulation. Results show that adjustment is necessary and effective for small data, and when data is large, even the estimate from unadjusted models is close to the true value, which may render the adjustment unnecessary.