U.S. Senate

Counties

Every reporting unit in the race. Sort any column. The votes above are this table added up.

Model

1. Raw data

   2024 presidential baseline

   National R−D 1.5

   County presidential file. Not the 2026 result.

2. Raw data

   Generic ballot

   D 50.8 R 42.6 R−D -8.2

   National Wikipedia aggregate, October 5, 2026. State generic ballots, where a public poll exists, replace a uniform national swing: ME, MI, NH, NY, OH. A state with no generic-ballot poll keeps the national number and is not assigned a neighbor's.

3. Raw data

   Party identification

   Pew D+5 and Gallup D+10, blend R−D -7.3

   Pew NPORS 2026, identifiers plus leaners: Republican 43, Democratic 48. Gallup party identification plus leaners, first two quarters of 2026: Democratic 49, Republican 39. Pew is 55% of this blend and Gallup is 45%.

4. Model input

   National block

   R−D -7.8

   60% generic ballot and 40% party identification. The swing versus the 2024 presidential margin is what gets transmitted.

5. Model input

   Swing transmitted

   1.07 × -9.3 = -9.9

   Only the share of the national swing this state took in the last midterm is transmitted. Elasticity is shrunk halfway toward 1.

6. Raw data

   2026 early vote

   None reported

   No verified 2026 returned-mail or in-person ballots for this state in the October 7 file. Nothing was imputed. Day-1 early vote is a signal, not a statewide margin. Applications and ballots that were only sent are not in this count. A day-1 pace gap is not a statewide margin.

7. Model input

   History block

   R−D 15.0

   Weights 2018 5 / 2024 35 / 2020 20 / 2022 40. 2020 and 2022 results for this office are included in every state that had those races, including blowouts. A 2022 election is weighted ahead of 2018. If that office was not on the ballot, the 2020 result for the same office fills the year. Another office is not used. 2018 from 2020 R+41.5 w5, 2024 from 2020 R+41.5 w35, 2020 R+41.5 w20, 2022 R+53.0 w40. This is the last result for the office, not the presidential margin. A 2024 county map is not copied forward unchanged. A 2018 model error is not used.

8. Model input

   Electorate block

   R−D 5.9

   History reweighted by usual midterm turnout, plus half of the transmitted swing. Rural and presidential-only counties still move farther when the result is spread back to counties. The 2026 early-vote count is not inside this margin.

9. Model input

   Turnout block

   R−D 9.1

   Each county is weighted by its 2018 ballots divided by its 2024 presidential ballots. Small counties shrink halfway toward the state median. 2020 mail shares are not used. No verified 2026 returned-mail or in-person ballots for this state in the October 7 file. Nothing was imputed. Day-1 early vote is a signal, not a statewide margin.

10. Model input

    National environment

    R−D 2.9

    History plus the full transmitted swing from the 60/40 national block.

11. Model input

    State environment

    R−D 6.5

    History plus 45% of the transmitted swing.

12. Model input

    State partisan lean

    R−D 16.1

    The 2024 presidential margin of this state, or of this district for a House race. It is heavier than the old state-environment block and is not shrunk toward the polling term.

13. Model input

    Trump popularity

    R−D -4.3

    New York Times approval average, October 7, 2026: 37% approve, 60% disapprove. Net approval -23. That is not a vote share. Four-tenths of the net is a margin, then scaled by this place's partisan lean (factor 0.46), so an unpopular president costs more in a Democratic state.

14. Model input

    Candidate favorability

    R−D -0.4

    Roger Marshall's record minus this place's partisan lean is D+1.1. Thirty-five percent of that gap, capped at 5, is the favorability term, D+0.4. The old margin is not reused.

15. Model input

    Demographics

    R−D 1.0

    County college, rural, and nonwhite shares from the 2018 election-context file change how much of the swing a county takes. No separate microdata file.

16. Model input

    Finance

    R−D 9.4

    No sourced spending edge, so the block starts on history. With a poll, it shrinks toward the polling term with the other non-poll blocks.

17. Model input

    Issues

    R−D 9.4

    No sourced issue score. Same treatment as finance.

18. Model input

    Polling

    R−D -1.2

    Individual polls -1.3 (80%) and the Wikipedia average -0.6 (20%). Named support R−D -1.5. Allocated support R−D -1.2. Named support does not move. Only the undecided line is split, 54.1% Republican and 45.9% Democratic. That is the 2024 two-party Republican share 58.2 moved halfway to 50 and capped at 60/40. A named third-party vote stays third-party. Banned individual rows excluded: 1. State miss shrunk 40% toward zero: 0.00. Public pollsters: 3. Outliers: 0. Recency half-life 16 days. Sample weight n/(n+350). A registered-voter poll is 0.90. An internal poll, or a poll sponsored by one party, is weighted 0.20, and only when a public poll is also in the race. An internal-only file is dropped. A poll that names both parties is not an internal. A poll more than 8 points from the median of the others is cut in half.

19. Model input

    Candidate

    R+1.50 reaches the result

    Roger Marshall keeps the incumbent term of 1.5 after shrinkage. Overperformance is a separate term and is not added again here. Roger Marshall's 2020 race is already in the history block, so it is not applied again.

20. Model input

    Halfway floor

    Not applied

    Not applied. At least three public polls from different pollsters, and none is an outlier.

21. Model input

    Pooling

    35% of the way from each history block toward the polling term

    Otherwise the shrink is 35%.

22. Model input

    Prediction market

    64% win probability

    Roger Marshall 63.8% is a win probability, not a vote share. Republican win probability 63.8 from Kalshi 63.5% Republican; Polymarket 64.0% Republican. A price of 50 is a tie. Each point above 50 is 0.35 margin points, capped at 12, so the raw margin is 4.8. That margin is shrunk 40% toward the pre-market margin 6.2, which leaves 5.4. It enters once, at 2.5%. The polling term is not nudged. The market and the polls differ by more than 15 points of win probability, so the weight is half of 5% and the interval is wider.

23. Model input

    State bend

    None

    Not applied.

24. Model input

    Weights

    polling 21.0% · history 7.6% · candidate 5.8% · electorate 4.8% · national 5.8% · stateEnv 3.8% · turnout 2.8% · demographics 2.8% · finance 2.0% · issues 2.0% · stateLean 17.3% · trump 11.5% · favorability 10.5% · market 2.5%

    History is lighter. State partisan lean, Trump popularity, and candidate favorability are their own terms. A market, when it qualifies, scales every other weight and is not taken out of polling alone.

25. County-sum margin

    6.2

    This is the sum of the reporting units after each unit moved off its own history. It is not the poll average.

26. 80% interval

    -4.5 to 16.9

    10th to 90th percentile of 50,000 draws. Poll, national, state, turnout, and candidate error.

27. 95% interval

    -10.2 to 22.6

    2.5th to 97.5th percentile. Senate and governor state errors in the same state share a correlation of 0.30. The band shown is the marginal interval.