Using multilevel regression with poststratification to obtain regional health estimates from a Facebook-recruited sample

Abstract

PurposeWe assess the effectiveness of multilevel regression with poststratification (MRP) as a tool to mitigate selection bias from online surveys of small geographical regions. MethodsWe collected self-reported health information from an Internet-based sample of adults residing within the St. Louis, MO, metropolitan area in 2017. We created Bayesian hierarchical models with three sets of predictor variables for each of six common health behaviors and outcomes, with results poststratified using the American Community Survey to estimate region and ZIP Code Tabulation Area–level prevalence. ResultsWhen comparing MRP estimates with a population-based sample as a reference, we found that adjustment using MRP can reduce bias in prevalence estimates and provide estimates for local area prevalence. 14 of 18 adjusted estimates were closer to the benchmark than the unadjusted estimates and MRP using all three covariate sets resulted in better overall agreement with the benchmark compared with the unadjusted estimates. ConclusionsMRP can improve prevalence estimates from self-selected Internet-based samples, although a nonnegligible amount of bias may remain. Illustrating the utility and limitations of this method will help researchers develop relevant estimates of the local public health burden, helping local health officials better understand and reduce poor health outcomes.