Dem
47
GOP
53
Yesterday, we had an item about the polls and how well they did. Siena College, the #1 rated pollster on 538, nailed Kamala Harris' vote share very accurately in the swing states, but underestimated Donald Trump's vote share by about 3.5%. How did other pollsters do? Let's look.
At FiveThirtyEight, they looked at two things: statistical error and statistical bias. The error is how far off the polls were and bias is whether the errors were usually in the same direction. They also looked at different ways of getting respondents. It turns out that the most accurate polls ran ads on web pages and cell phone apps, giving people a link to click on. Theoretically, letting people self-select being a respondent is terrible, but imbalances in the sample were corrected mathematically and the results were quite good, with an average error of 1.4 points. Live phone polls and opt-in online panels were quite poor, with errors of 3.3 points and 3.6 points respectively.
Here are the results of polls during the final 3 weeks of the campaigns for elections from 2000 to 2024:
In 2024, the statistical error (how far off the polls were) was better than in previous years, but the poll results were better for the Democrat than the actual results. The same thing was true in 2020 and 2016. This could be due to "shy" Trump voters or undecided voters breaking for Trump very late in the campaign. This is a real phenomenon. (V) once couldn't make up his mind in a primary until he had the ballot in his hand, and even then he couldn't decide for 5 minutes. In any event, 538's result agrees with ours: the polls were fairly accurate but had a noticeable bias for Harris. Ideally, the errors would be random in both directions, but that wasn't the case in either our study or theirs.
Another site,
Split Ticket,
did a similar analysis. Here are the results:
The results here are similar to ours and FiveThirtyEight's except the difference column is done the other way, so Rep +1.0 means Trump did better than the poll by 1 point, which is the same as the poll overestimating Harris. So basically, all three studies are the same: the polls were not far off but they systematically underestimated Trump by a couple of points. (V)