Improving Subnational Opinion Estimation from Cluster-Sampled Polls

Abstract The development of multilevel regression and poststratification (MRP) has allowed scholars to more accurately estimate subnational public opinion using national polls. However, MRP generally recovers less accurate estimates from polls whose respondents are selected using cluster sampling – also called area-probability sampling. This is in part because cluster-sampled polls rely on a complex form of random sampling focused on national representativeness that may result in small or unrepresentative subsamples in subnational geographies. This has limited MRP’s usefulness in subnational opinion estimation in several contexts, including historical polls in the US, where cluster-sampling was common into the 1980s, and large academic studies in many countries today. In this paper, I propose two approaches to improve estimation from MRP with cluster-sampled polls. The first is pooling data from multiple surveys to produce a larger sample of clusters. The second is clustered MRP (CMRP), which extends MRP by modeling opinion using the geographic information included in a survey’s cluster-sampling procedure. Using simulations, I show that both methods improve upon traditional MRP, and I validate them using historical polls in the US