Estimating Subnational Opinion with Cluster-Sampled Polls: Challenges and Suggestions

Multilevel regression with poststratification (MRP) has become widely used in political science to estimate subnational opinion from national polls. This method takes into account both demographic and state-level effects and has greatly expanded knowledge of public opinion at the subnational level. In this paper, I assess the use of MRP on two cluster-sampled polls, the American National Election Studies (ANES) and the General Social Survey (GSS). I discuss concerns that cluster sampling raises and demonstrate that MRP can produce problematic results when used on a typical cluster-sampled poll. I evaluate several ways to improve upon MRP with cluster sampling. I find that adding a state-level predictor to the model, pooling surveys across years and sampling frames, and incorporating additional geographic information can all improve MRP’s performance on cluster-sampled polls, but that caution should still be used when applying MRP to cluster-sampled data. ∗Alissa Stollwerk is a doctoral candidate in the Political Science Department at Columbia University. I would like to thank Jeffrey Lax and Justin Phillips for their insightful comments at all stages of this work, and I thank Andrew Gelman for helpful comments and his assistance with the General Social Survey data as well. I am also grateful to Jim Gibson for his advice on the 1980s joint SRC-NORC sampling frame.