Case studies of polling failures focus on within-election differences in poll accuracy. The crucial question of why polls fail in one election but not in others often remains a matter of speculation. To develop a contextual understanding, we review and unify theories of election features suspected of encouraging polling errors, including mobilization, candidacies, polarization, and electoral conduct. We extend a Bayesian hierarchical modeling approach that separates poll bias and variance at the election level and links error components to electoral predictors. Investigating 9,298 pre-election polls across 367 U.S. Senate elections, 1990-2022, we find an overall trend toward smaller but more uniform errors. Poll variance is negatively associated with mobilization and polarization. Until 2004, frontrunners and incumbents were overestimated, but there is little evidence that polls are biased for female or minority candidates. Finally, Republicans in states with lower levels of state democracy have been underestimated in recent years.
A) Number of polls over the legislative period , B) Prior specification, C) Empty model results with reduced time window D) Minority status based on facial characteristics and name, E) Detailed results