The Correction You Are About to Apply Is for a Different Poll

"The polls undercount Republicans" is about to be applied to a midterm. That is a reasonable thing to believe. In our poll file the last three presidential cycles all lean the same direction: D+3.68 in 2016, D+4.71 in 2020, D+2.78 in 2024 — one sign, averaging D+3.68 across 7,069 polls. If that is all you have looked at, the inference is not lazy. It is just about a different election.

Take the same file to the seven midterm cycles it covers — 1998 through 2022, 7,582 polls — and the number is D+0.36. Not small in the sense of imprecise. Small in the sense that it is built out of cycles that cancel. The sign flips at five of the six steps; only 2006 to 2010 holds its direction, and even then it decays from R+1.16 to R+0.20.

Seven midterms, five changes of direction in six chances. Whatever produces the recent presidential pattern does not survive here as a correction you can carry into November.

Except in one place, where it does

That would be a tidy piece if we stopped there, and it would be wrong. Split the midterm polls by what they measured and one race type separates from the rest: the national generic ballot. Across its 514 midterm polls the lean is D+2.62, pointing the same way in six of the seven cycles.

So the folk wisdom is right about the generic ballot and wrong about polling. That distinction is the point, because the generic ballot is the only poll number that reaches most people. It leads the segment. It is the push alert. And the correction it teaches — knock two or three points off the Democrat — gets applied to the next poll you read, which is a Senate poll in one state.

Midterm Senate polls lean R+0.20. Apply the generic ballot's D+2.62 to one and you have not removed an error, you have introduced one about three points wide, pointed the wrong way. House-district polls fail on the other axis: 6.66 points of average miss, the largest of any race type here, with a lean of D+0.31 that is indistinguishable from nothing. There is no correction to apply. Only a wide poll.

Sixty-eight days out, our own file is empty

There is a second problem, and it is ours as much as anyone's. This piece publishes 68 days before 3 November. The longest lead time on any poll in the FiveThirtyEight file — all 20,466 of them, every cycle — is 61 days. Not 61 for midterms. Sixty-one, full stop. No poll in the record we rate pollsters against was taken as far from its election as today is from this one.

Across the range it does cover, time helps monotonically. Midterm general-election polls taken more than four weeks out miss by 6.25 points on average across 2,960 polls; two to four weeks, 5.81 across 1,904; one to two weeks, 5.53 across 1,418; final week, 4.70 across 1,169. The last month is worth about a point and a half of accuracy. What the four-plus-weeks bucket cannot tell you is anything about ten weeks, because it does not contain ten weeks.

The instrument that needs no correction

Here is the argument. A poll measures an election that has not happened, and the file says a midterm one misses the margin by 5.76 points on average, in a direction the last cycle cannot tell you. A county result measures an election that did happen; it has no margin of error, it has a canvass. That is not a reason to ignore polls. It is a reason to stop treating a two-point move in a generic ballot as news and a twenty-point move in a county as background.

Florida is the most-polled state in the midterm file at 339 state-level polls, more than any other. Of the twenty states with at least 100, it is the one where the polls have most overstated Democrats: D+3.47, average miss 5.09 points. Hillsborough is Florida's fourth-largest county by 2024 turnout and one of only five in the state that finished inside five points. Its net movement relative to the nation across twenty-four years is 2.25 points. The average miss on a single Florida midterm poll is more than twice that county's entire quarter-century drift.

New York is the opposite case on both counts. Of those same twenty states it is the worst-polled, 8.94 points of average error across 206 polls, and the most Republican-leaning at R+5.29. Those are Senate and governor polls; the chart below is presidential county data. Different elections, different instrument — we are not claiming one explains the other, only that the county series is the more legible of the two. Nassau has moved 22.22 points toward Republicans relative to the nation since 2000, from D+19.54 to R+2.68, and every one of those points is a counted ballot.

None of this says what will happen in November, and we are not going to. It says what to do with the number you will be handed between now and then. If it is a generic ballot, the file gives you a real prior: overstated Democrats in six of seven midterms, average 2.62 points, and 2.33 in 2022 even in the final three weeks. If it is anything else — Senate, governor, a House district — that prior is not yours to use, and the honest reading of our own file is that you do not have one.

Seven midterms, five changes of direction in six chances. The only stable lean in the file belongs to the one poll everybody quotes and nobody is actually voting in.