Estimating Health Disparities Using Predicted Race/Ethnicity: Measurement Error in Nonlinear Regression When Predicted Probabilities Are Used as Covariates

• Activity Number: 75
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Rebecca Hubbard*+ and Andy Bogart and Leo Morales
• Companies: Group Health Research Institute and Group Health Research Institute and Group Health Research Institute
• Keywords: measurement error ; non-linear regression ; health disparities ; race

Abstract

In a study of health disparities at Group Health Cooperative, a not-for-profit health system in Washington State, researchers were interested in studying the association between race/ethnicity and use of preventive health services, a binary outcome. Individual level race/ethnicity data were unavailable, but an estimate of the probability of membership in each race/ethnicity group was obtained using data from the U.S. census. Previous research has shown that replacing a categorical covariate with probabilities of class membership produces unbiased, though not fully efficient, linear regression parameter estimates. In this project, we explore properties of estimators arising from using predicted probabilities as covariates in non-linear regression. We propose a method to improve performance and demonstrate the efficiency of this approach relative to the naïve estimator and the full information estimator arising from using the true value of the categorical predictor. Finally, we demonstrate the performance of these approaches in an application to racial and ethnic disparities in use of health services.