Improving Small-Area Estimates of Public Opinion by Calibrating to Known Population Quantities

Abstract

Multilevel regression and poststratification is widely used to estimate opinion in small geographies and to adjust unrepresentative surveys — and has become a mainstay in the study of dyadic representation. Yet, errors generated by nonignorable non-response and modeling uncertainty make discrepancies between public opinion and policy outcomes difficult to interpret. We propose a principled, data-driven method to leverage auxiliary quantities with known marginal distributions — e.g., election outcomes — to improve estimates of policy attitudes. Our method estimates the geographic correlation between the auxiliary variables and the outcomes of interest, then adjusts the estimates using observed errors in the auxiliary variable estimate. We illustrate our approach using a pre-election poll measuring support for an abortion referendum, finding that the method reduces county-level error by two-thirds. We find similarly dramatic improvements estimating precinct-level opinion. Our method generates new possibilities for accurately estimating policy attitudes at previously unattainable levels of geographic resolution.