Of the roughly three thousand one hundred county-level election offices in the United States, I have so far found six hundred and twelve that publish precinct results in a format a computer can read without a human retyping them. The rest publish PDFs. A non-trivial number of those PDFs are photographs of paper.
I bring this up because of a finding you have read a hundred times: split-ticket voting has collapsed. The share of voters picking a president from one party and a senator from another has gone from something like a quarter of the electorate in the 1970s to a rounding error. It is the most repeated quantitative claim in American politics. It is in every strategy memo I have been shown. It is the load-bearing premise under the argument that persuasion is finished and only turnout is left.
I believe it. I want that on the record before I complain, because the complaint is not that the finding is false.
The complaint is that I cannot reproduce it below the state level across most of this country, and neither, in the strict sense, can the people who assert it with the most confidence.
The mechanics are dull and they matter. At state level the calculation is easy: two statewide totals, subtract, done. That is where the famous series comes from and it is solid. But the interesting form of the question is not how many states split their tickets. It is what sort of person splits, and where, and under what local conditions — and that question lives at the precinct, because a precinct is small enough that its answer is not the average of two opposing things.
To get to the precinct you need the canvass. The canvass exists; it is a public record in every state. In something like eight of ten counties it exists as an image. So the national precinct datasets that researchers actually use are assembled by a small number of underfunded projects that transcribe, normalise and reconcile those images by hand, county by county, cycle by cycle. They are careful people doing unglamorous work and I have relied on it for years. The coverage is still uneven. Boundaries move between cycles. The reconciliation involves judgment calls that are documented honestly and, in practice, not read by anyone quoting the output.
None of that makes the headline finding wrong. It makes the precision spurious. When somebody tells me split-ticket voting fell by a specific number of points among a specific group in a specific kind of place, I have learned to ask which counties are in the sample, and the honest answer is usually: the ones that were easy.
The easy ones are not a random draw. A county that publishes machine-readable results is a county with a funded elections office, which tracks population, which tracks nearly everything else. The measurement is at its best exactly where the phenomenon is least surprising, and at its worst in the small rural counties where a genuine holdout tradition of ticket-splitting would have survived longest if it survived anywhere.
I have run my own version and got the expected answer, which is precisely why I trust it less than I would like to. My sample was assembled by the same convenience I am describing. I know what I would have done if the number had come out the other way: I would have gone looking for the error. That is not a procedure. It is a preference with a spreadsheet attached.
So the conclusion is small and unsatisfying, which is usually a sign it is the real one. The finding is probably true. The confidence attached to it is manufactured downstream of a filing-format decision made by a few thousand county clerks who were not thinking about political science when they made it. And a claim that is used to justify abandoning persuasion entirely — to stop talking to anyone who might be moved — ought to rest on something firmer than the counties that could afford a spreadsheet.
Six hundred and twelve. The number goes up every cycle, slowly, mostly because somebody calls a clerk and asks nicely. It is the only part of this I can personally do anything about and I spend more evenings on it than I am willing to defend.
