Last week, I took issue with a few of the claims embedded in the notion of “wave” elections. (I’m not the only political scientist pondering the substance behind all the talk of “waves,” “tsunamis,” and the occasional “earthquake”–check out political scientists Daniel DiSalvo and James Ceasar writing here). In this post, I want to consider another possibility that commonly comes up in discussions of wave elections. Do polls in the days before the election systematically understate support for the out-party?
I start with a random subset of 90 Senate or Gubernatorial elections from 1990 to 2006. In these elections, the major-party nominees are both white males, to sidestep the issue of race– or gender-related biases. From these elections, we can track down 218 polls that occurred within a month of the election, and compare the actual share of support for the GOP candidate to the support projected by the polls. On average over this period, the GOP’s share of the two-party vote is 0.7 percentage points lower than its share of the polling. Here’s the graph:
The black diamonds indicate the median gap between how the GOP candidate performed and how he polled. Positive numbers indicate better performance than expected. The smaller gray circles represent 1,000 bootstraps, to provide a sense of the (significant) uncertainty inherent in the small samples. In 1994, the archetypal “wave” election, the GOP candidates do outperform their polls by 1.3 percentage points. But there is essentially no evidence of systematic out-performance by the Democrats in 2006, when the polls were dead-on. Also, the year that truly stands out is 1998, which was an unusual midterm but not a “wave” by anyone’s definition. (I suspect what is going on here is simply that the 1998 sample includes several lopsided races, where polls typically overstate support for the front-runner.) From what we’ve seen in the past few decades, there isn’t much reason to think that polling accuracy gets unusually battered by the waves.