Search
Browse By Day
Browse By Time
Browse By Person
Browse By Mini-Conference
Browse By Division
Browse By Session or Event Type
Browse Sessions by Fields of Interest
Browse Papers by Fields of Interest
Search Tips
Conference
Location
About APSA
Personal Schedule
Change Preferences / Time Zone
Sign In
X (Twitter)
Survey researchers working in both academia and industry increasingly rely on multilevel regression and poststratification (MRP) and related extensions to estimate (typically geographic-based) subgroup opinion. Despite a plethora of research using MRP, it is often hard to assess the relative impact of various modelling choices on MRP estimates and the inferences of subsequent analyses using those estimates. Extant evaluations, for example, often rely on held-out survey data — a process that relies on the increasingly problematic assumption that the held-out survey data provides an unbiased estimate of the truth. Using known quantities such as precinct-level, county-level, and state-level election returns from recent elections and several sources of survey data, we investigate how aspects such as overall survey sample size, sample size within each geographic unit, the average sample size within each cell, number of post-stratifying cells, choice of individual covariates, choice of geographic covariates, the choice of geographic level at which to poststratify, and the extent of nonignorable nonresponse bias – affect the error of MRP estimates and how those errors affect subsequent analyses using those estimates. We also consider how the estimation errors we identify in predicting known vote shares may be used to account for measurement errors in other MRP-based regression analyses.