To understand when polls are accurate and when they fail, we adopt a Bayesian hierarchical modeling approach that separates poll bias and variance at the election level, and links error components to a broad range of election features including mobi- lization, candidacies, polarization, and electoral conduct. An empirical study of 9,298 pre-election polls across the 367 U.S. Senate elections, 1990-2022, reveals an over- all trend toward smaller but more uniform errors over time, a negative association between poll variance and mobilization and polarization, and a tendency to underes- timate more ideologically extreme Republican candidates. Moreover, Republican poll bias has a modestly positive link with the level of state democracy. Contrary to the- oretical expectations, we find little evidence that female or minority candidates are overestimated in polls. While large parts of the variance remain unaccounted for, the empirical approach is promising and easily extended to include other potential error sources.
A) Number of polls over the legislative period , B) Prior specificatioon, C) Minority status based on facial characteristics and name, D) Results by election cycle