Regency B
Multilevel Regression with Poststratification for Local Estimation: An Example and Lessons Learned (303940)
Lauren Arnold, Saint Louis University
*Travis Loux, Saint Louis University
Erik Nelson, Indiana University - Bloomington
Mario Schootman, SSM Health
Enbal Shacham, Saint Louis University
Keywords: selection bias, model-based adjustments, secondary data, social research, Bayesian methods
Multilevel regression with poststratification (MRP) is becoming a popular way to adjust for selection bias in social science research. To perform MRP, the researcher models an outcome in a sample data set, then applies that model to population data. The population predicted outcomes are aggregated to the level of interest, weighting by population stratum size. In this presentation, we describe an example of MRP to estimate metro area and ZIP code level prevalence of various health behaviors and outcomes. Participants were recruited through Facebook to take a general health survey. Raw and MRP-adjusted estimated were computed and compared to results from a population-based study. The MRP estimates reduce bias by 20 to 50% depending on modeling choices. We will focus on tools and application of MRP as well as lessons learned in this project.