When do pre-election media polls under- or over-estimate the Republican vote? This paper examines the signed forecast errors in pre-election media polls, defined as the signed difference between the poll and the election returns, based on 2,130 state-level polls for U.S. Senate, gubernatorial, and presidential elections from 1990 through 2016. We consider the longstanding Republican perception—and grumble—that the polls tend to under-forecast the Republican vote, then model and estimate the influences on the signed forecast errors made by individual polls. Differential campaign spending turns out to be particularly important—in a theoretically satisfying but unexpected way. Both this and the other estimated effects also look somewhat paler in the most recent elections, reflecting important changes in electoral politics.
