Election-Night Models: Calling races from partial returns

In brief

A race can be called before every ballot is counted because the call is an inference about the uncounted vote, not a summary of the votes in hand. The running total is a sample selected by state rules, not by chance, and mail, early in-person and Election Day ballots lean differently and are reported in a set order. An early lead can therefore mislead for hours even when the final call is correct. The wrong call remains possible, but it is rare. The only published live model is for primaries.

How to use this

Before reacting to an early lead, ask which voting modes have reported and which remain. Find the state's published release order and check whether one mode is nearly complete while another has barely started. Compare the current mode's lean with the same mode's lean in a prior election, not with the topline alone. Treat a projected winner as a claim about uncounted ballots, not a verdict from counted ones. If you publish or set counting rules, release the sequence and the mode splits before the count so the drift is legible in advance.

What the story is about

On 5 November 2024, North Carolina reported its presidential votes in the order the state's own rules set. Mail ballots came first, and they showed a Democratic lead of 27.3 points with 5.2% of the count in. Early in-person ballots came next, and once they were added the race was effectively level: a Republican lead of 0.8 points at 79.3% of the count in. Election Day ballots came last and took the Republican margin to 3.3 points at 99.6% in. The canvass that followed left it there. The leader on screen was still wrong for the first few minutes, and the on-screen margin was 2.4 points out at 79% reporting. Both facts were predictable from published rules before anybody voted.

The election-night count is a sample of the electorate, drawn not by chance but by administrative rule, and the rules decide which ballots are in it and when. North Carolina publishes its order in advance: polls close at 7:30pm, absentee-by-mail results are released from 7:30 to 9pm, early voting results from 8 to 9:30pm, and precinct results from 8:30pm to midnight. Military, overseas, provisional and mail ballots that voters were allowed to fix after a problem are added during the canvass afterwards (North Carolina State Board of Elections, no date). Those groups do not vote alike. In the state's official 2024 presidential returns, mail ballots ran 27.30 points Democratic, early in-person ballots ran 2.78 points Republican, and Election Day ballots ran 12.94 points Republican (North Carolina State Board of Elections, 2024). In 2020, when far more people voted by mail, that group ran 43.16 points Democratic and made up 18.0% of the count. Peer-reviewed studies of national returns and of precinct-level returns in California find a related pattern: national returns document the growth and Democratic lean of late-counted ballots and its association with mail and provisional volumes and state partisanship (Foley and Stewart III, 2020), while precinct-level returns show that the shifts are consistent with which precincts and mail ballots are processed first (Li, Hyun and Alvarez, 2022).

The obvious shortcut is to compare tonight's numbers with the same point last time. In North Carolina the gap between Election Day ballots and everything counted before polls closed was 37.85 points Republican in 2020 and 12.12 points in 2024, according to the State Board's precinct-level returns (North Carolina State Board of Elections, 2020, 2024). The benchmark moved 26 points in one cycle. Applied with 79.6% of the Election-Day-inclusive count on screen, the 2020 gap projects a final margin of 8.53 against an actual 3.26, a 5.3-point miss in a state decided by 3.3. County by county it was noisier still: an average gap of 8.4, a range from 18.9 points Democratic to 34.7 points Republican, and 14 of 100 counties running the other way. What survived was the ordering. County gaps in the two cycles showed the same pattern: counties that leaned more Republican in one cycle leaned more Republican in the other, even though the average level shifted by 21.1 points.

Retractions are rare. Decision Desk HQ's own FAQ states that a projection requires greater than 99.9999% confidence, and lists nine publicly documented race-call retractions in national races since 2016 (Decision Desk HQ, 2026). That is the vendor's account of its own record, not an audited tally. The most recent entry is the 2026 Democratic special primary in New Jersey's 11th district, where the desk projected Malinowski and Mejia overtook him on late returns from Morris County. That was a special primary rather than a general election. So a benchmark error usually costs a desk hours of numbers that look wrong, rather than a winner announced wrongly. The wrong call remains possible; it is just the rare outcome.

Desks start reasoning from the count while it is still coming in. Decision Desk HQ's own description is that its live product "adjusts continuously based on reported vote patterns and the characteristics of the remaining uncounted ballots", starting from the pre-election forecast (Decision Desk HQ, 2026). That is the inference in one sentence: the number on screen is an input, and the output is a claim about ballots nobody has counted yet. The one published model comes from a peer-reviewed paper and belongs to Decision Desk HQ: its Live Primary Model. It works from voting collection units, meaning batches of votes from one place, that are complete or nearly complete, and uses regression, which estimates how strongly one measured quantity moves with another, and copula methods, copulas being a way of joining separate statistical estimates, to produce win probabilities and expected shares (Donnini et al., 2024). Two limits matter. Donnini et al. (2024) describe the only published live model, and it is for primaries.

So a call is an inference, not a summary of the votes in hand. It is a claim about the votes still out: who cast them, in which mode, and where they live. The partial count that feeds it is a sample, and not a random one, because state rules decide what has been counted by any given hour. Those rules also decide the order, and the modes lean differently, so the running total drifts even when nothing is wrong. A lead can therefore mislead for hours while the call itself turns out correct. A desk that reads the uncounted vote correctly can call a race while ballots remain uncounted.

So what

The composition of the count is what matters on the night. An early lead tells you which ballots have been counted, and on its own it says little about where the count is going. What you can use is the mode. Mail, early in-person and Election Day ballots lean differently, and the order is set by state rules. If a lead comes from a group that has nearly finished reporting while another group that leans the other way has barely started, expect the lead to move. The number that decides the race on the night is the one nobody has counted yet, so that is the one worth asking about.

For political parties

For a party, the trap is treating a mode-driven lead as a verdict. The first numbers are a sample chosen by counting rules, so a strong hour for you may be a mirror, and building a message on it means telling your supporters something the count will contradict. Ask instead which ballots are still out, in which mode, and how those groups leaned last time. A party that reads only the topline will find out at the end of the count.

For government

Government sets the rules that produce the count the public sees, so it is the one actor that can make the sequence legible before it happens. Publishing what will be released and when, and splitting the totals by how people voted, lets anyone watching see that the drift is built into the order rather than evidence of a problem. That is what the North Carolina timeline does. Better to say that before the count, when it is information, than after it, when it sounds like an excuse. A government that publishes the rules gives its citizens the means to read the numbers correctly.

References

Decision Desk HQ (2026) Frequently asked questions. Available at: https://www.decisiondeskhq.com/faq (Accessed: 10 September 2026).

Donnini, Z., Louit, S., Wilcox, S., Ram, M., McCaul, P., Frank, A., Rigby, M., Gowins, M. and Tranter, S. (2024) 'Election night forecasting with DDHQ: a real-time predictive framework', Harvard Data Science Review, 6(4). Available at: https://doi.org/10.1162/99608f92.ccb395f0 (Accessed: 10 September 2026).

Foley, E.B. and Stewart III, C. (2020) 'Explaining the blue shift in election canvassing', Journal of Political Institutions and Political Economy, 1(2), pp. 239–265. Available at: https://doi.org/10.1561/113.00000010 (Accessed: 10 September 2026).

Li, Y., Hyun, M. and Alvarez, R.M. (2022) 'Why do election results change after election day? The "blue shift" in California elections', Political Research Quarterly, 75(3), pp. 860–874. Available at: https://doi.org/10.1177/10659129211033340 (Accessed: 10 September 2026).

North Carolina State Board of Elections (2020) Official precinct-level results, 3 November 2020 general election [data file]. Available at: https://s3.amazonaws.com/dl.ncsbe.gov/ENRS/2020_11_03/results_pct_20201103.zip (Accessed: 10 September 2026). Calculations from this data were made for this article and are available upon request.

North Carolina State Board of Elections (2024) Official precinct-level results, 5 November 2024 general election [data file]. Available at: https://s3.amazonaws.com/dl.ncsbe.gov/ENRS/2024_11_05/results_pct_20241105.zip (Accessed: 10 September 2026). Calculations from this data were made for this article and are available upon request.

North Carolina State Board of Elections (no date) Election night reporting timeline. Available at: https://www.ncsbe.gov/results-data/election-results/election-night-reporting-timeline (Accessed: 10 September 2026).