Leveraging Time-Series Information to Improve Small-Area Estimation

• In Event: Measuring Public Opinion

Abstract

Multilevel regression with poststratification (MRP) has quickly become the gold-standard method to handle small-area estimation in public opinion surveys, and many scholars have proposed further modifications to MRP to improve its performance. At the same time, the proliferation of survey programs—both in the United States and cross-nationally—has endowed social scientists with an extensive library of public opinion data. These programs frequently ask the same or similar survey items in multiple years, but scholars rarely use this temporal structure to its fullest potential. I propose leveraging this time-series information to improve subnational opinion estimation in years or units where data is scarce.

I test a variety of approaches, ranging from expanded multilevel models with random intercepts by year to more complex dynamic models that explicitly impose a temporal structure on the parameters of interest. I test these approaches on population-level data from the Cooperative Election Study, which repeatedly asked the same policy preference items from 2006-2021. Overall, results suggest two takeaways: First, scholars can increase the accuracy of MRP models simply by allowing the model to borrow information from other time periods. Second, the partial-pooling advantages of MRP models also enable scholars to construct time-series estimates of public opinion at the subnational level, with a degree of accuracy above and beyond that which can be achieved with a more common no-pooling approach.

Author

• Isaac Mehlhaff, Texas A&M University