Why early-vote party splits are not in the forecast

Summary

Before 3 November millions of Americans vote by mail or at early-voting sites, and several states publish how many ballots have come back, split by the party each voter registered with. These figures are widely read as a guide to who is winning. This paper explains why our forecast does not use them, on the evidence of our own data.

The early-vote figures say who has voted, not how. A registered Democrat can vote Republican, unaffiliated voters are a large and growing share, and the mix of early and Election Day voting changes from one election to the next. In the 12 state races of 2022 and 2024 for which we hold early-vote party splits on polling day, reading the split as the result missed the final margin by 14.7 points on average, and by 42.9 in Pennsylvania's 2022 Senate race. Correcting each state's 2024 figure by its own 2022 gap still missed by 7.5 points, against 2.7 for the polls of the final three weeks in the same races. A wider test of 75 races since 2016, published on the site, reached the same conclusion: the one way of using early votes that seemed to help did so because of a coincidence of direction in 2024, not because of anything the early vote said about individual states.

The figures are shown on the early-vote page, compared with the same day in 2022 and 2024, as a measure of turnout. The forecast does not use them, and the live model on election night does not either.

What is held

The database table of early votes holds, for each state, party and day, four counts: mail ballots sent, mail ballots returned, in-person early votes, and the total of the two cast. Party is the voter's registered party: Democratic, Republican, or other (every other party and the unaffiliated). Every figure comes from the state's own election office:

Table: Early-vote figures held, 15 October 2026.

On 15 October, four of the six states with a reader have published 2026 figures, 966,682 ballots in all. Where a state publishes dated records, this year's count is compared with 2022 and 2024 at the same number of days before polling day; Florida and Maryland keep only final counts for past years, so no same-day comparison is possible there.

Why a party split is not a margin

Three mechanisms separate the party split of early ballots from the result.

Registration is not a vote. The split counts ballots by the party the voter registered with, not the party they voted for. In states with many long-registered Democrats who vote Republican, the split points the wrong way.

Method is not intention. A party whose voters move from voting on the day to voting early gains early ballots without gaining a vote; early voting on its own mostly moves votes earlier rather than adding voters (Burden et al., 2014). Pew Research Center found that in 2024, 39% of Republican voters voted in person on Election Day against 28% of Democrats. In North Carolina's 2024 returns, Republicans had all but closed the Democrats' early-vote lead of 2022, and the margin of the result did not move.

Registration moves. In North Carolina, Democrats outnumbered Republicans by 3.7 points of the register in October 2022; on 10 October 2026 the two parties were within 490 voters of each other, and voters with no party or a minor party had risen to 41% of the roll. A smaller Democratic share of early ballots can mean no more than fewer registered Democrats.

The test: early-vote splits against results

For the six states with a full count on polling day in both 2022 and 2024, we set the registered-party split of all early and mail ballots against the margin of the state's top race (in 2022 the Senate race, or in Maine the governor's race; in 2024 the presidential race):

Table: Final early-vote party split against the result. Points; gap = split minus result margin, positive where the early vote was more Democratic than the result.

Read as a forecast, with the gap as its error \(g_s = x_s - v_s\), the split missed by

and the early vote was more Democratic than the result in 10 of the 12 races, by 11.1 points on average. The two exceptions are Colorado's, where almost everyone votes by mail and unaffiliated voters, counted as neither party, cast more of the ballots than either party's voters.

Correcting for each state's habit

A fairer test allows for each state's own habit: if Pennsylvania's mail vote always runs far more Democratic than its result, subtract that. For each state, 2024 is predicted as its 2024 split minus its 2022 gap,

Table: The 2024 split corrected by the state's 2022 gap. Negative errors: the corrected figure was too Republican.

The corrected figure still missed by 7.5 points, and in five of the six states it was too Republican: the early vote became much less Democratic between 2022 and 2024 (by 25.5 points in Pennsylvania, where the number of Republican mail ballots more than doubled), while the results moved far less (6.6 points in Pennsylvania). A correction learned from one election was overtaken by a change in method at the next.

The polls in the same six 2024 races, averaged over their final three weeks, missed by 2.7 points. Even after the most favourable correction, the early vote was nearly three times as far out.

The wider test

A larger test, set out in a site article of 14 October 2026, paired early-vote party splits with the certified result of the top race in 75 state races from 2016 to 2024, using the US Elections Project's and the University of Florida Election Lab's final pre-election reports for 2020 to 2024 and state election offices for 2016 and 2018. Its findings:

• Read as the margin, the split missed by 11.3 points on average, 7.5 points too Democratic, with a standard deviation of 13.4. In the 58 races with a poll average in our database, the polls missed by 4.1 points and the early vote by 10.8; the early-vote leader won 44 of them, the poll leader 51.
• Corrected by the state's gap at the previous election of the same type, the split missed by 7.1 points (29 races); in the 25 of those with polls, 6.6 against 3.9 for the polls.
• The change in the early-vote split between two elections of the same type barely tracked the change in the result: a correlation of 0.18 across 29 pairs, and 0.03 across 47 consecutive pairs.
• Moving the poll average 30% of the way towards the corrected early-vote figure seemed to help (3.29 points against 3.86), but the gain came from the year-wide pattern of 2024, when the early vote and the polls' miss both moved towards the Republicans in 15 of 17 states. The state-by-state part of the signal added nothing (3.87 against 3.86), shuffling the early-vote changes at random between states within each year did as well in 55% of shuffles, and shifting every poll by the other years' average poll miss, with no early-vote data, did better (2.88).

Figures move within a cycle

The party split also changes through the weeks before polling day, as different kinds of voter return their ballots. At 21 days out and on polling day:

Table: Registered-party split of returned ballots three weeks out and on polling day, points.

In every case the Democratic lead shrank by polling day, by between 7.5 and 25.7 points. A comparison is meaningful only at the same number of days out, which is how the early-vote page shows it.

2026 so far

At 21 days out (13 October 2026), returned ballots stood at 31,581 in North Carolina (D+27.4, against D+32.8 at the same point in 2022 and D+10.1 in 2024), 59,542 in Maine (D+36.3, against D+35.1 and D+23.3) and 347,655 in Pennsylvania (D+48.7, against D+55.3 and D+40.5). Pennsylvania's series has been rebuilt from later files since the site's article on these figures, which gave 277,188 ballots and D+51.6 for the same day.

Against 2024 the Democrats look much stronger this year; against 2022, the right comparison for a midterm, slightly weaker in North Carolina and Pennsylvania and level in Maine. Measured as each party's returned ballots per registered voter, North Carolina's Democrats and Republicans are both at about 70% of their 2022 pace: fewer voters of every party are voting by mail, and the smaller Democratic lead against 2022 is almost wholly the change in registration. None of this says who will win.

What early votes are good for

Two uses survive the evidence. The first is turnout: compared with the same day in the last similar election, on the same mailing calendar and per registered voter, early returns show whether turnout is running ahead or behind. The second is the count: on election night, how many mail ballots remain to be counted, and from which kinds of voters, matters for projecting a partly counted race. Our live model does not use early-vote figures for that; its expected vote in each county comes from 2022 turnout, and it allows for the order in which mail, early and Election Day votes are counted as an uncertainty (From a month out to election night).

The decision

Early-vote figures stay out of the forecast. The rule for every input is that it must improve a forecast that already has polls, on elections it was not fitted on; on our data the early vote does not. Two cycles cannot support a coefficient, and the 75-race test found no state-level signal that survives a shuffle. A signal that survived it in future years would change that, and the tests will be rerun once the 2026 results are certified.

Limits

Few states. Only states that register voters by party and publish dated records can be tested, and they are not a random sample of states.

Registered party. Our figures count registered party, not modelled partisanship. Commercial trackers that model each voter's party might do better; their figures are not used on the site.

Top race only. Each test compares the early vote with one race per state, while voters may split their tickets.

References

Burden, B. C., Canon, D. T., Mayer, K. R. and Moynihan, D. P. (2014). Election laws, mobilization, and turnout: the unanticipated consequences of election reform. American Journal of Political Science, 58(1), 95–109.

Figures as of 15 October 2026. The pages they come from are rebuilt with each new poll.