Comparing the Inferred Preference Model to the Exit Polls

The press has various breakdowns of the 2018 exit polling done by Edison Research. None split things into the same categories we looked at in our post on inferred voter preference in 2018 house elections. But we can merge some of our categories and then compare. For comparison we chose a Fox News post because it had the most detailed demographic splits we found.

We see rough agreement but also some very large discrepancies, particularly around voter preference of men and women that we have yet to explain. NB: In each of the following charts, perfect agreement with the exit polls would leave all the dots in the vertical middle of the chart, at 0%.

After merging men and women, our preference model tracks the exit polls quite well as seen in the chart below.

A fairly large discrepancy appears when we look at sex and race, merging ages. Our model infers similar voting preferences for white men and women and non-white men and women whereas the exit polls show women about 10% more likely to vote for Democrats. This is illustrated in the chart below.

An even larger discrepancy appears when we look at sex and education, merging our age categories. Our inferred result for male college graduates is a full 15% higher than the exit polls, while our result for female non-college graduates is almost 10% below the exit-polls. This is laid out in the chart below.

We’re continuing to investigate these differences and we hope that using other data and methods we can figure out why this model infers higher Democratic voter preference for men and lower for women than exit-poll or other post-election analytics. We welcome your ideas, via email, Twitter or via a github issue.