Did Non-Voters Really Flip Republican? Part III with Updated Data
When we published “Did Non-Voters Really Flip Republican in 2024?” a few months ago, the evidence from the Cooperative Election Study (CES) suggested that non-voters weren’t particularly Republican. We showed that Democrats held a large advantage over the GOP in partisan registration among non-voters in 2024, and that self-identified non-voters supported Donald Trump over Kamala Harris in the public survey data. Specifically, we found that registered non-voters leaned Harris, but unregistered non-voters leaned Trump, with the overall non-voter average being slightly pro-Harris. This Harris support wouldn’t have been enough to swing the election if everyone had voted, but her support over Trump was clear in the data.
Today the CES released its voter validated weights. The updated CES crosschecks voter registration lists to check if survey respondents are actually voters or non-voters. As we said in our earlier piece, we’d update our findings with this new CES update. Specifically, we said we would replicate our analysis using the voter validated weights once they were released.
In the figure below, that is the analysis using voter validated or “VVWeights.” That is, our preregistered replication analysis—using the same code that we wrote back in April—finds that non-voters leaned Harris over Trump, consistent with what we found in our earlier analyses.1 A benefit of these voter validated weights is that we know that these are non-voters, since they are Americans who were successfully matched to government voter registration lists and were confirmed to have not turned out to vote. Furthermore, these survey weights are calculated to create a demographically representative sample of US adults.
However, if we also include all individuals who were not successfully matched to voter registration lists and make the (plausible) assumption that they did not vote, then Trump edges out Harris among non-voters. This is the bar that says “CommonWeight.” This means we might be accidentally assuming some voters are non-voters, since there are other reasons that somebody might not be successfully matched to a voter registration list (such as changing or misspelling a name). On the other hand, it gives us a larger sample size if we include all of these unmatched individuals as non-voters.
There are three takeaways here. First, it’s reasonable to believe that non-voters leaned Trump over Harris in 2024. Second, if we take the CommonWeight estimate as correct, then it is very close to the Pew survey estimate of non-voters, but still very far off from the more aggressive public claims of non-voters being Trump +11 or more.
Third, this is still a highly uncertain question. As we described in earlier pieces, it’s incredibly difficult to survey non-voters; if they were easy to track down and ask about their political attitudes, they’d be voters. This means that survey samples of non-voters are much less certain than survey samples of voters. This sampling uncertainty goes beyond regular statistical uncertainty we see in confidence intervals (for reference, our bootstrapped confidence interval here for the Trump support estimate using CommonWeight is about 2 percentage-points in size).
But beyond that, the fact that different weights—VVWeight vs. CommonWeight—give distinct answers to this question highlight the tough modeling choices that analysts must make, and the way that these choices may impact results.