Election Modeling

Election modeling based on historical data

Scope

Model Democratic/Republican vote margins in Massachusetts State Senate and State Representative races.

Resources

• Data sources
  • bwbensonjr/ma-election-db - District-level elections and fine-grained census data
  • MA Commonwealth Election Statistics
  • bwbensonjr/ocpf-cli - OCPF command-line tool
  • Massachusetts Office of Campaign and Political Finance
  • bwbensonjr/mapoli - Scripts for gathering precinct-level election results
• Prior art
  • bwbensonjr/mapoli/model
• Model use cases
  • MA 2022 State Senate General Election Ratings

Current Model

The current mapoli legislative model uses a Bayesian regression model with this structure:

dem_margin ~ PVI_N + incumbent_status + pres_elec

• dem_margin = (dem_percent - gop_percent) * 100
• PVI_N = Partisan Voter Index (PVI) which is Democratic/Republican point margin averaged over the last two presidential elections and normalized by the US-wide point margin.
• incumbent_status
  • 1 if there is a Democratic incumbent in the race
  • 0 if there is no incumbent in the race
  • -1 if there is a GOP incumbent in the race
• pres_elec
  • 1 - The legislative election coincides with a U.S. presidential general election
  • 0 - The legislative election does not coincide with a U.S. presidential general election

Data Availability and Training Cutoff

Every training row needs two things at the same granularity: the Democratic/Republican margin for a legislative race, and the PVI_N built from the two presidential elections preceding that race. State Senate and State Representative districts split municipalities, so PVI has to be aggregated to the district from precinct-level presidential results. Town-level presidential returns cannot be allocated to a district that contains only part of a town. Precinct availability for presidential results is therefore the binding constraint on how far back the data set can reach.

Precinct-level availability at electionstats.state.ma.us

The download/{election_id}/precincts_include:1/ endpoint returns town-level rows with empty Ward/Pct columns for elections before precinct reporting begins. It does not error, so the cutoff has to be detected by inspecting the returned rows.

Office	First year with precinct results
President, Governor, U.S. Senate, U.S. House, State Senate, Governor's Council	2002
State Representative	1980

State Representative is the exception because rep districts are built out of precincts, so results were always broken out that way. Every other office reports one row per municipality (351 rows) through 2000 and roughly 2,170 precinct rows from the 2002 general onward.

Presidential elections with precinct-level results: 2004, 2008, 2012, 2016, 2020, 2024.

Resulting cutoff

The earliest PVI computable from two precinct-level presidential elections is the 2008 PVI (2004 + 2008). A race in year Y uses the PVI from the two most recent presidential elections completed before Y, which puts the floor at the 2010 legislative elections.

Race year	PVI year	Source presidentials	District map
2010	2008	2004 + 2008	2001
2012	2008	2004 + 2008	2011 (remap required)
2014, 2016	2012	2008 + 2012	2011
2018, 2020	2016	2012 + 2016	2011
2022, 2024	2022	2016 + 2020	2021 (remap required)
2026	2024	2020 + 2024	2021

pvi_year in mapoli/pvi/ma_state_leg_pvi_2008_2024.csv names the later of the two presidential elections, except for 2022, which is the 2020 PVI reallocated to the 2021 district map.

Usable range: the 2010 through 2024 legislative elections, eight cycles, plus 2026 as the forward prediction target. 2008 and earlier are out of reach at any useful granularity: they would require 2000 precinct-level presidential results, which do not exist.

Redistricting boundaries

2012 and 2022 are the first elections under new maps, so the prior presidential votes must be reallocated from old precinct geography to new districts. This is what mapoli's ma_precincts_districts_*.csv files do via areal interpolation, and it is why post-2021 vote totals are fractional rather than integers. Rows on either side of a redistricting boundary are not strictly comparable and the remapped PVI values carry interpolation error that the natively-mapped years do not.

Currently materialized vs. reachable

• Materialized. mapoli/pvi holds precinct presidential results for 2012, 2016, 2020, and 2024, normalized into ma-election-db as precinct_presidential_vote. This covers races from 2018 onward.
• Reachable without new sources. The 2004 and 2008 precinct-level presidential results exist at electionstats but have not been fetched. Fetching them extends the training set back to 2010. The 2001-cycle precinct-to-district crosswalk is also absent from the repositories, but it is derivable from the precinct-level State Rep and State Senate result files for 2002 through 2010, since each district's file enumerates its own precincts.
• Out of reach. Races in 2008 and earlier.

The legislative margin is not the constraint

District-level legislative results run back to 1990 in ma-election-db, State Representative precinct results back to 1980, and State Senate precinct results back to 2002. In every case the presidential precinct data runs out first.

Train/test split

Splitting by year rather than randomly matters here because districts repeat across cycles and a random split would leak the same district's behavior between train and test.

A single split on the 2021 redistricting -- train on 2010 through 2020, test on 2022 and 2024 -- was the initial plan. The implemented procedure generalises it to a rolling-origin schedule over every election year from 2014 onward, which scores 424 races instead of 128 and, by including the odd years, puts 24 special elections in the holdout instead of none. The redistricting boundary is still tested: the 2022 fold trains entirely on the old map and predicts on the new one, and redistricting_cycle is a reported segment. See docs/scoring.md.

Published training data

The precinct-level data set described by this section is built and committed by the collection pipeline in docs/pipeline.md, with the column-by-column schema in docs/schema.md.

• data/precinct/ma_precinct_training_set.csv.gz

  • 14,409 rows, one per legislative race per precinct
  • 633 contested State Representative and State Senate races, general and special, spanning 2010 through 2024
  • dem_margin, dem_margin_two_party and PVI_N computed at precinct grain, with incumbent_status, pres_elec, is_special, num_candidates and the eligibility flags carried from the race
  • Against 679 rows in the district-level table the current model uses

  The table is built at the most permissive eligibility rule -- any candidate named in the returns counts, write-in or ballot line -- and carries the flags by which a stricter rule excludes a race, so who counts as a candidate is a downstream selection rather than a property baked in. That is 10 races more than the 623 the ballot-line rule admits. See docs/definitions.md.

Coverage matches the cutoff derived above. Of the 1,651 legislative elections in the window, 1,650 have precinct-level returns; the 2011 3rd Berkshire special reports municipality totals only. 2.4% of rows have no PVI_N, almost all of them precincts created after 2020 that no earlier presidential election covers, and 1.1% sit in precincts split between two districts. Both are flagged per row rather than dropped.

Baseline Model and Scoring

The baseline model is reproduced on the collected data and evaluated by a fixed scoring procedure, so later variables and algorithms are measured against an unmoved yardstick.

• Historical baseline. dem_margin ~ PVI_N + incumbent_status + pres_elec, Gaussian likelihood, fit at race grain with Bambi/PyMC. It remains the stable reproduction and historical benchmark, but it is not the control for a change proposed for the operational forecast. The selected horizon-specific forecast_14d or forecast_60d composite is that experiment's control.
• Scoring. Rolling-origin holdout: each fold trains on every race held strictly before its election date and predicts the races held on that date, over every election date from 2014 onward --- 23 folds, 6 general-election dates and 17 special-election dates. Pooled holdout of 424 races, 24 of them special. RMSE in margin points is the primary score, reported with calibration and win-side metrics and broken out per fold and per segment. See docs/scoring.md.
• Race-grain training table. data/race/ma_race_training_set.csv.gz, 633 races rolled up from the precinct table, described in docs/race_schema.md, with the candidate roster in data/race/ma_race_candidates.csv.gz.
• Campaign finance. data/race/ma_race_finance.csv.gz, one row per candidate per race, carrying receipts and expenditures accumulated to a stated pre-election date. Rolled up into the race table's money columns. Collected by uv run maprecinct finance; see docs/money_results.md.
• Forward forecasts. A response-free 2026 target, exact 60-day and 14-day finance snapshots, immutable prediction directories, and read-only prospective scoring are described in .

uv run legmodel definitions                        # list the data definitions
uv run legmodel score                              # score under the adopted definition
uv run legmodel folds                              # print the fold schedule
uv run legmodel compare baseline baseline_year      # paired comparison of variants
uv run legmodel compare-definitions current two_party_or_strongest
uv run legmodel tenure                            # run the operational tenure replacement
uv run legmodel tenure --legacy                   # reproduce the historical incremental test
uv run legmodel importance                         # variable importance and effect sizes
uv run legmodel target-finance --horizon 60d --dry-run
uv run legmodel forecast --horizon 60d
uv run legmodel forecast-web --horizon 60d       # export the static explorer bundle

The GitHub Pages forecast explorer, its local development commands, and its deployment process are documented in docs/interactive_forecast.md.

Baseline accuracy

Under the pre-change current definition. Superseded by the table in Answered Questions below, which uses the adopted definition; both remain reproducible.

Segment	Races	RMSE	Coverage (90%)	Win accuracy
Pooled	424	15.61	0.889	0.920
General elections	400	14.96	0.900	0.925
Special elections	24	23.98	0.708	0.833
No Democratic candidate	11	28.43	0.364	0.818

First result: does is_special help?

Not overall. Pooled RMSE moves from 15.608 to 15.641, a difference of -0.033 with a 90% interval of [-0.143, +0.078], which contains zero: the data does not separate the two models. The pooled figure hides two real and opposite effects. The term lowers RMSE on the 24 holdout special elections by 1.13 points, and raises it on the other 400 races by 0.14, and because specials are 5.7% of the holdout the second outweighs the first.

On specials it does cut the baseline's large pessimism about Democrats, from -11.41 points of bias to -7.91. The recommendation is to keep the pooled model as the reference and use special_pooled_term --- or equivalently special_split --- when the question is about a special election. That question has since widened into a three-arm comparison; full writeup in docs/special_handling_result.md.

The comparison also surfaced larger failures than the one under test: races with no Democratic candidate (36% interval coverage) and an unabsorbed presidential-year bias swinging from -6.4 to +2.7 points. Both are settled in Answered Questions below.

Model Enhancements

• Incorporate OCPF fundraising data -- docs/money_results.md
• Test a tenure-aware incumbency variable -- docs/tenure_result.md
• Incorporate Census demographic data
• Candidate-committee money is collected; independent expenditures, PAC money and party committee spending are not, and are a separate collection problem

Plan

• Define the data sources, granularity, and variables
  • Data sources - We should look at the data and scripts in these bwbensonjr repositories:
    • bwbensonjr/ma-election-db
      • District-level election results
      • Precinct-level census and demographic data
    • bwbensonjr/mapoli/pvi - Scripts for precinct-level election results
  • Granularity - Our preference is to use precinct-level data, but we may use district-level results for years where we do not have precinct-level data.
• Define model accuracy measurement and scoring - docs/scoring.md
• Gather the data and rebuild the baseline model and evaluate its accuracy - docs/is_special_result.md
• Settle the data definition -- who counts as a candidate and what the margin is measured against - docs/definition_result.md
• Test the deferred variables: presidential-year bias and num_candidates - docs/variant_results.md
• Collect OCPF campaign finance and test it as a predictor - docs/money_results.md
• Derive and test uninterrupted incumbency tenure -

Answered Questions

The five questions deferred from the baseline work are settled, and campaign finance -- the enhancement this README has carried as planned work since the project started -- is answered alongside them. Questions 1 through 3 were decided together, because each rebuilds the training table and invalidates the scorecard; 4, 5 and 6 are variant declarations scored by the existing harness, though 6 also adds a collection stage and new columns.

Full writeups: docs/definition_result.md for the data definitions, docs/variant_results.md for the variants, docs/money_results.md for campaign finance, docs/variable_importance.md for which predictors carry the model and what each is worth, and docs/definitions.md for the mechanism.

The data definition

Who counts as a candidate, and what the margin is measured against, is now a named definition applied over the published tables rather than a property baked into them. Testing an alternative is a filter and a column choice, not a rebuild.

uv run legmodel definitions
uv run legmodel compare-definitions current two_party_or_strongest

Adopted: two_party_or_strongest — the two-party margin where both major parties stood, the margin against the strongest non-Democrat on that pair's own two-candidate denominator where no Republican ran, races with no Democrat excluded, and the write-in threshold left at the ballot-line rule. 610 races, 413 in the pooled holdout, all 24 holdout specials retained.

Four of the five definition comparisons came back undecided, so the choice rests on a principle fixed before the numbers were seen: the response and the predictor should be measured against the same denominator. PVI_N is a two-party quantity, so the response should be too.

1. A two-party response — undecided, and the pooled figure is a trap. On its own holdout the strict two_party definition posts 12.76 RMSE against current's 15.61, which looks decisive. On the 346 races the two share, the difference is -0.31 [-0.84, +0.38]. The gap is not accuracy: it is the 78 races two_party drops, which current scores at an RMSE of 25.3. Restricting to Democrat-versus-Republican races removes the hard ones, and comparing pooled figures would have credited the model for declining to predict them.

The README's own sub-question answers itself. two_party_or_strongest recovers 93 of the 106 dropped races, keeps 24 holdout specials instead of 22, and costs nothing measurable (+0.111, undecided). There is no case for paying 67 holdout races and two specials for a restriction the data does not reward.

2. The write-in threshold — undecided, left at the ballot-line rule. write_in_5pct against current is -0.020 [-0.050, +0.007]; because both share a response, that comparison isolates the eligibility rule cleanly. What settles it is the other direction: the two races a 5% threshold admits score at an RMSE of 63.2. Admissions also stop changing between 8% and 10%, so any threshold in that range is the same rule under a different name.

The asymmetry the question identified is nonetheless fixed. A definition's threshold now governs both decisions a write-in affects — whether a race is contested and whether the write-in enters the denominator — so a write-in can no longer count for one and not the other.

3. Races with no Democratic candidate — decided: exclude them, from training as well as scoring. This is the only definition comparison the data settles: +0.094 [+0.049, +0.143]. The control makes it informative — keeping them in training while withholding them from scoring is undecided (+0.006), so the gain comes specifically from not training on them. Those 13 races were distorting the fit for the other 610, not merely resisting prediction. Most of the apparent improvement is still removal rather than accuracy, and the writeup says so.

The variants

4. Presidential-date bias — partly fixed. The baseline runs +3.0 points too Democratic on presidential-date general elections and -6.0 too Republican on midterm-date ones, a gap of 9.0 that one binary term cannot correct. (Both figures are over general elections only; every special carries pres_elec = False, so the old blended "non-presidential" figure mixed 237 midterm generals with 24 specials.)

baseline_year, a hierarchical election-date intercept, is the adopted variant: RMSE 14.34, a difference of +0.663 [+0.283, +1.026], coverage 0.935 --- the best of any variant scored --- and a bias gap of -0.56. It is decided under three definitions of four. It cannot shift a forward date's mean, because that date's effect is unobserved and drawn from the hyperprior.

baseline_national_env adds a term signed by the party holding the presidency: -1 off-cycle under a Democratic president, +1 under a Republican one, 0 on a presidential ballot. It still cuts pooled bias from -2.90 to -0.56 and is the only variant that can shift a forward year's mean --- it carries to 2026, where it takes the value +1. But it is undecided under three definitions of four (+0.488 [-0.311, +1.221] under the adopted one), and it is exactly collinear with pres_elec across the first nine folds, because every midterm before 2017 had a Democratic president. It is not adopted on the current evidence; the forward-prediction argument for it is structural rather than measured. See docs/variant_results.md. baseline_pres_incumbent is undecided.

5. num_candidates — undecided; it does not belong in the baseline. A wash under current (+0.026) and under the adopted definition (-0.033), and decidedly worse under the strict two-party definition (-0.064 [-0.107, -0.023]). Once the response is measured on a two-party denominator the term adds variance without adding signal, exactly as anticipated.

6. Campaign finance — the largest gain measured so far. baseline_money_logratio adds the log ratio of Democratic to opponent receipts, measured 14 days before each race's own election, and lowers pooled RMSE by +1.700 [+1.128, +2.255] — roughly twice what the year intercept was worth. The effect is concentrated in open seats (+3.686 [+2.481, +4.872]) and undecided against a Democratic incumbent.

Money is never read from OCPF's published cumulative figure, which is a full calendar-year total fetched after the fact and therefore includes money raised after the polls closed — 28% and 70% of two cycles' totals for one Boston filer. Every figure is reconstructed from report line items over a window ending at a stated pre-election date. The result is not evidence that spending changes outcomes; see docs/money_results.md.

7. Incumbency tenure — operational replacement undecided; retain the selected forecasts. Replacing incumbent_status with signed cap-four tenure in the 14-day operational composite changes general-election RMSE from 12.039 to 12.122. The control-minus-replacement difference is -0.083 with an election-date-clustered interval of [-0.345, +0.181]. The 60-day sensitivity is also undecided. The earlier baseline-plus-tenure test remains a historical incremental experiment, not the operational selection result. See docs/tenure_result.md.

Accuracy under the adopted definition

Variant	Segment	Races	RMSE	Coverage (90%)	Win accuracy
baseline	Pooled	413	15.01	0.896	0.915
baseline	General elections	389	14.19	0.910	0.915
baseline	Special elections	24	24.76	0.708	0.833
baseline_national_env	Pooled	413	14.52	0.915	0.915
baseline_national_env	General elections	389	13.55	0.931	0.915
baseline_national_env	Special elections	24	22.68	0.792	0.917
baseline_year	Pooled	413	14.34	0.935	0.927
baseline_year	General elections	389	13.45	0.938	0.933
baseline_year	Special elections	24	23.25	0.750	0.833
baseline_money_logratio	Pooled	398	13.34	0.894	0.920
baseline_money_logratio	General elections	376	12.81	0.904	0.928
baseline_money_logratio	Special elections	22	20.73	0.727	0.773
baseline_spend_logratio	Pooled	398	13.05	0.895	0.920
baseline_spend_logratio	General elections	376

These supersede the baseline table above but are not the same measurement: different races and a different response. current stays registered and scorable, so the earlier figures remain reproducible.

The money row is scored on 398 races, not 413. It excludes the 15 holdout races where a candidate could not be resolved to an OCPF filer, because a missing filer and a candidate who raised nothing are different facts and the variant may not impute one into the other. The paired comparison above scores both models on those same 398 races, so it is a like-for-like difference.

The same four contrasts over expenditures rather than receipts are registered and scored too. baseline_spend_logratio is the most accurate variant on this holdout, but it beats the receipts term by only +0.322 [+0.056, +0.599] and that margin is undecided under two of the four definitions, so baseline_money_logratio stays the adopted money variant — spending is also the more endogenous measure, being the quantity a campaign adjusts in response to how close the race looks. See docs/money_results.md.

Of the money and pre-money variants, baseline_money_logratio is much the most accurate, including on special elections, which are the model's worst segment. It loses a little interval calibration relative to baseline_year and gives up the ability to score the 15 unmatched races at all. baseline_year is best calibrated; baseline_national_env is the only one of the older pair usable for a forward prediction, and the money term is too — a figure measured 14 days out is knowable before the election it predicts.

Open Issues

Known defects and limitations carried forward, as distinct from the planned work in Model Enhancements and Plan.

pres_elec costs accuracy in the money model

Dropping pres_elec from baseline_money_logratio improves pooled holdout RMSE by 0.705 [−1.100, −0.299] — decided — even though its coefficient is +7.09 [+5.24, +8.97] and nowhere near zero. The finding survived the refold essentially unchanged: it was 0.722 [−1.116, −0.320] on year folds.

pres_elec is a property of a race's election date, and a fold is one election date, so its value is constant across every race in a holdout by construction — a wrong shift lands on all of them in the same direction rather than averaging out. Across the folds its coefficient ranges from 7.09 to 13.54 while PVI_N holds between 1.40 and 1.65, and the presidential years themselves disagree (2012 +27.5 mean margin, 2020 +13.9).

This extends the presidential-date bias already recorded under question 4 rather than contradicting it. Not yet acted on: a variant dropping the term is a registry entry and a rescore, and no alternative is adopted until that is compared under every scored definition.

What the special-election work added to this. special_split's general component excludes special elections from training, which changes what pres_elec is estimated from, since every special carries pres_elec = False. It beats both pooled arms on presidential-date general elections under all four definitions, and loses to them on midterm-date ones. So part of what looks like a special-election question is really a question about the timing term. See docs/special_handling_result.md and docs/variable_importance.md.

national_env is collinear with pres_elec for nine folds --- resolved

Resolved. The confounding refusal now runs predictor-versus-predictor, not only predictor-versus-grouping-factor, and baseline_national_env is refused on exactly the nine folds 2014-01-07 through 2017-07-25.

The defect was real and is now measured rather than argued. national_env is pres_elec - 1 identically on every training window whose midterms all fell under a Democratic president, so the two terms were one column carrying two parameters. Nothing refused it, because the harness only ever tested a predictor against a grouping factor --- which is why baseline_year_pres was refused for this exact shape while baseline_national_env was not. The check is now pairwise over the expanded design columns, the separating count is published per fold, and the refusal reason travels on each row of fit_diagnostics.csv.

What the measurement showed: the term becomes identified on 2017-10-17, and for the five folds from there to 2018-11-06 its entire basis for distinguishing a Republican from a Democratic presidency is between one and six special elections --- races coded as a midterm-backlash electorate they are not. Only from 2020-03-03, once the 71 races of the 2018 general enter the training window, does the separating count reach 77. The variant's holdout falls from 413 races to 253 under the adopted definition and to 240 under generals_only; its previously published figures over 413 races are superseded by the refusal, not by a rescore.

Its replacement, baseline_timing, keeps all 413. See docs/timing_result.md.

Special elections are entangled with the timing terms --- resolved

Resolved. Ballot timing is now one declared four-level categorical, ballot_timing, with presidential as the reference; special elections take their own level rather than being coded by the president's party.

Every special election in the record carries pres_elec = False, because none has ever fallen on a presidential general date, and that produced three consequences, none deliberate: any "non-presidential" figure blended midterm generals with specials; specials were the only races separating pres_elec from a per-date effect; and specials received a national_env value describing an electorate they are not. Restricted to general elections there are exactly three timing groups --- presidential ballot (231 races), midterm under a Democratic president (271), midterm under a Republican one (71) --- and pres_elec plus national_env saturated them, with the third group resting on the single 2018 election.

What the measurement showed: baseline_timing beats baseline by +0.771 [+0.394, +1.163] under the adopted definition and +0.754 [+0.343, +1.169] under generals_only --- decided under both --- and the gain is almost entirely the 71 races of 2018, where it is +4.265. It narrows the presidential-date bias gap from 8.99 to 6.60 and halves pooled bias, without the collinearity baseline_national_env bought its bias reduction with. Against baseline_year it is undecided.

The scorecard's ballot_timing segment now takes its levels from the predictor's declaration, so a segment row and a coefficient describing the same population cannot disagree. midterm_general has split into midterm_dem_pres and midterm_gop_pres; any figure quoted against the old three-level names describes a different population.

baseline_year fits do not converge --- resolved

Resolved. Every fold under every definition now samples cleanly: zero divergent transitions across all 40 fits, worst R-hat 1.0027, lowest bulk ESS 2436. The variant is adopted alongside baseline_national_env.

The fix is not the one this section originally proposed, and the way it was wrong is worth keeping. It suggested a non-centred parameterisation first --- but bambi builds group effects non-centred by default and the project never overrode it, so the sampler had always been running the recommended fix. There was no funnel to straighten.

The two real causes were:

• An auto-scaled group-SD prior. Bambi derived 1|election_year ~ Normal(0, HalfNormal(135)) from the intercept's scale, against a response whose own standard deviation is 25.2 points, and asked it to inform a between-year SD estimated from as few as four year groups. The variant now declares HalfNormal(5).
• pres_elec collinear with the year grouping. pres_elec is a property of a race's election date, not of its calendar year: it is computed as election_date in PRESIDENTIAL_ELECTION_DATES, and a special election held in a presidential year but on its own date carries False. What made it collinear with a per-year intercept is narrower --- every special election in the record falls off the presidential general date, so specials are the only races that separate the two. Across 610 races pres_elec varies within a year only in 2016 and 2020, on 8 races, all of them specials; in the 2010-2013 training window, not at all. The variant now drops it. The refusal rests on that count measured per fold, not on the calendar.

The improvement grew rather than shrank once the fits were clean: RMSE 14.209 against the superseded 14.503, a paired difference of +0.804 [+0.410, +1.179] against the superseded +0.510, and coverage holding at 0.927.

Those figures are themselves now superseded. Under the date schedule the variant groups on election_date rather than election_year --- a calendar year spans several fold dates, so a year-level effect would leave the holdout's own level partially observed --- and posts 14.34, +0.663 [+0.283, +1.026], coverage 0.935, still the best-calibrated variant scored. The bias gap closes from 8.99 to -0.56, driven by the term removed rather than the one added, since the baseline's fitted pres_elec coefficient was applying a shift that miscalibrated out of sample.

baseline_national_env remains the only variant that can move a forward prediction: a date intercept cannot shift a future date's mean, because that date's effect is unobserved and drawn from the hyperprior, whereas the national environment is known before the votes are cast. That argument is structural and stands. Its measured advantage does not --- it is undecided under three definitions of four, for reasons recorded in Open Issues above. The two answer different questions and are reported together rather than ranked.

Three things carried forward from the fix:

• A variant now declares its own priors and sampler settings, and fit_diagnostics.csv records what each fit actually ran under, so a result obtained at a raised target_accept is distinguishable from one obtained at the default.
• A predictor constant within every level of a variant's own grouping factor is refused per fold, with the refusal published. This catches a predictor against a grouping factor only — it does not catch two collinear fixed effects, which is how baseline_national_env carries national_env alongside pres_elec for nine folds without complaint (see Open Issues).
• baseline_year is no longer nested in baseline, and its comparison is labelled non-nested, naming the term added and the term removed.

Full writeup: docs/variant_results.md.

Special elections remain the worst segment

24 holdout races at 24.77 RMSE against a pooled 15.01, with 0.708 interval coverage against a nominal 0.90. The model is overconfident about them, and no arm tested fixes it.

Three arms were scored --- specials pooled with an is_special term, pooled without one, and split into a general model that never sees a special plus a special model fit on everything. Pooled, the three are indistinguishable under three definitions of four. The segments disagree consistently: both distinguishing arms beat the plain one on specials and lose to it on midterm generals, and the split arm beats both on presidential-date generals under every definition.

Under the date schedule baseline_year is the only variant here that improves the segment (23.68 RMSE, 0.792 coverage). baseline_national_env reverses --- best of the four on the year schedule, worst now at 26.40 --- which is consistent with the entanglement recorded above: every special gets a nonzero national_env, describing an electorate it is not.

Whatever is left is a thin-data problem: 24 holdout races spread over 17 fold dates is not much to diagnose from, and the arms differ by less than the Monte Carlo noise between two fits of the same model. Full writeup: docs/special_handling_result.md.

The adopted definition rests on a principle, not on a measurement

Four of the five definition comparisons came back undecided, so two_party_or_strongest was adopted on the pre-registered tie-break rather than because the data preferred it. That is the honest outcome at this holdout size, but it means the decision is provisional: more cycles, or a variable that interacts with the response definition, could settle it empirically. Every superseded definition stays registered and scorable so the question can be reopened without re-running collection.

The write-in threshold is provisional for the same reason. It sits at the ballot-line rule because the two races a 5% threshold admits score at 63.18 RMSE, not because a threshold is wrong in principle. If genuine write-in campaigns become more common, this is worth revisiting.