Multilevel Calibration Weighting for Survey Data

Abstract In the November 2016 U.S. presidential election, many state-level public opinion polls, particularly in Upper Midwest, incorrectly predicted winning candidate. One leading explanation for this polling miss is that precipitous decline traditional response rates led to greater reliance on statistical methods adjust corresponding bias—and these failed important interactions between key variables like educational attainment, race, and geographic region. Finding calibration weights account remains challenging with survey methods: raking typically balances margins alone, while post-stratification, which exactly all interactions, only feasible a small number of variables. paper, we propose multilevel weighting, enforces tight balance constraints marginal looser higher-order interactions. This incorporates some benefits post-stratification retaining guarantees raking. We then correct bias due relaxed via flexible outcome model; call approach “double regression post-stratification.” use tools re-assess large-scale voter intention finding meaningful gains from proposed methods. The available multical R package.