ADDRESSING DESIGN and ESTIMATION CHALLENGES WHEN USING MRP in PUBLIC HEALTH and BEHAVIORAL SCIENCE APPLICATIONS

• Activity Number: 295 - SPEED: Big Data, Small Area Estimation, and Methodological Innovations Under Development, Part 1
• Type: Contributed
• Date/Time: Tuesday, July 30, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Robert Petrin* and Alexa DiBenedetto and Luke Vaicunas
• Companies: Ipsos Public Affairs and Ipsos and Ipsos Public Affairs
• Keywords: Survey; Nonprobability Sampling; MRP; Bayes; Trend Analysis

Abstract

Multilevel regression with poststratification (MRP; e.g., Gelman, 2007; Ghitza and Gelman, 2013) provides a flexible method for improving inferences from probability and nonprobability sample surveys. MRP has been applied in a range of public health and behavioral science applications (e.g., Barbour et al., 2016; Eke et al., 2016; Lei, et al., 2018; Warshaw and Rodden, 2012; Zhang et al., 2015). In spite of its potential, researchers have limited formal guidance regarding its performance, particularly when designing studies focusing on inferences for small subgroups and / or nonstandard estimands. This paper assesses MRP’s performance under scenarios common in public health and social research, as well as the potential of specialized designs and weakly-informative priors to overcome challenges in MRP estimation and inference. The paper presents results from comprehensive empirical simulations based on applying MRP to estimate variation in trends in economic outlook among key subgroups of small business owners in the US. We conclude with observations on the potential and limitations of MRP, while offering strategies for expanding its use in other areas, such as data integration.