Nearly half of the ‘party’ in early-vote trackers is a model

In 21 states, the NBC News early-vote tracker labels its party column “modeled by TargetSmart”. Those states hold 48% of the early votes cast in 2026 as of Oct 23: 10.4 million of 21.7 million ballots. So when the tracker says the national early electorate is 49% Democratic, close to half of the ballots behind that figure carry a modelled party, not one copied from a state record.

The 21 states are Alabama, Arkansas, Georgia, Hawaii, Illinois, Indiana, Michigan, Minnesota, Missouri, Mississippi, Montana, North Dakota, Ohio, South Carolina, Tennessee, Texas, Utah, Virginia, Vermont, Washington and Wisconsin.

That blend sits under one of the most shared lines of the month: this early electorate is far more Democratic than 2024’s, D 49 to R 34 against D 42 to R 40 at the same point. The comparison may well hold. But it compares two mixes of registration and modelled party, and the headline does not say how much of each.

What it means

The national split is a blend of two kinds of number. In some states it comes from registration files. In others it is a model’s best guess. They deserve different levels of trust, and the headline figure hides which is which.

It matters most in the Senate races. Three of the ten battlegrounds we track have modelled party only: Michigan (D+17), Minnesota (D+27) and South Carolina (D+2). Ohio and Texas are modelled in the tracker too, so for those two we use the states’ own primary records instead: D+5.9 and D+3.4.

The obvious alternative is worse. Weighting each county’s early votes by how that county voted in 2024 understated the Democratic margin by 7 to 41 points, with a median of about 16, in the 8 states where we could test it against known party. A vendor model that is close is more useful than a county shortcut that is far off.

Every party figure on this site carries a source tag: reg. for registration files, primary for primary history, model for vendor estimates.

What it doesn’t mean

It does not mean the model is wrong. Where we can check it against registration files, as in Iowa, Maine and North Carolina, the tracker matched the official counts within about one point in both 2022 and 2024.

It does not tell you which way any error runs. In Ohio the model and the state file agree on the margin to a tenth of a point. In Texas the tracker read D+7 on Oct 22 and R+1 on Oct 23, while the state’s records, through Oct 21, give D+3.4: the model was more Democratic one day and more Republican the next.

The larger disagreement is about independents, not the margin. Ohio’s file puts 39.4% of early voters in Other; the tracker puts 8%. Texas records put 34.0% in Other; the tracker puts 9%. A tracker can match the D−R gap and still describe a very different electorate.

And registered party is still not vote choice. Democrats have voted early in larger numbers since 2020, so a Democratic early electorate says more about who votes early than about who wins.

How we checked

Registration states. For Iowa, Maine and North Carolina, which record party, we compared the tracker’s 2022 and 2024 splits with the official files. They matched within about one point.

North Carolina. We tested the usual shortcut, party of the last primary voted, on 1,096,068 early voters. It put the margin at D+21.2 against a true D+20.5, but sorted 24% into Other against a true 34%.

Ohio. The Secretary of State’s absentee file, with party from the last primary, gives D 33.3, R 27.4, Other 39.4: D+5.9. The tracker gives D 49, R 43, Other 8: D+6.

Texas. We joined state early-vote rosters to 2026 primary rosters: D 34.7, R 31.3, Other 34.0, or D+3.4, across 2.16M early voters. The tracker gave D 49, R 42, Other 9 (D+7) on Oct 22 and D 45, R 46, Other 9 (R+1) on Oct 23. Our state-records figure has not been refreshed since Oct 21. Adding 2022 and 2024 primaries in four counties, 25% of votes, moved our margin by only about 0.6 points.

What we publish. Until Election Day we report party from state records wherever they exist, show the tracker figure beside it when the two differ, and tag every number with its source.