Multilevel Regression and Post-Stratification for Small-Area Estimation of Population Health Outcomes Using BRFSS: An Evaluation of Cross-Level Inference

• Activity Number: 690
• Type: Contributed
• Date/Time: Thursday, August 13, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Survey Research Methods Section
• Author(s): Xingyou Zhang* and James B. Holt and Hua Lu and Paul I. Eke and Kurt J. Greenlund and Janet B. Croft
• Companies: CDC and CDC and CDC and CDC and CDC and CDC
• Keywords: multilevel regression and poststratification ; small area estimation ; cross-level inference ; Behavioral Risk Factor Surveillance System ; American Community Survey ; Small Area Health Insurance Estimates

Abstract

Multilevel Regression and Poststratification (MRP) has been shown to be a flexible and valid approach to generate small area estimates (SAEs) of health outcomes for local levels in the United States such as census tracts and counties. We applied tract-level poverty to county-level regression coefficients in model prediction. We examined whether this cross-level inference produces any substantial bias in SAEs. In this study, we fitted a multilevel prevalence model using 2012 Behavioral Risk Factor Surveillance System data and generated county-level SAEs of uninsured adults aged 18-64 years via two multilevel prediction models with county-level versus census tract-level poverty and compared them with estimates from the 2012 American Community Survey (ACS) for 814 counties and Small Area Health Insurance Estimates (SAHIE) for 3,142 counties. SAEs of uninsured adults from the prediction model using census tract poverty were almost the same as SAEs using county poverty (correlation coefficients >0.99). Both were strongly correlated with ACS and SAHIE SAEs (correlation coefficients >0.86 and > 0.80 respectively). Thus the cross-level inference in MRP results in very minor bias in SAEs.