A county-level model fitting distance alongside a 2024 vote-share covariate
Finding
In an ordinary least squares fit across the 39 county rates, the distance term is associated with 0.361 additional Too Late ballots per 1,000 mail ballots per mile (p=0.0006), with a historical two-ticket vote-share covariate also in the fit but not separable from distance in these 39 rows. A parallel Poisson count model gives a log-rate coefficient for distance of 0.00987 (p<0.001). Both are county-aggregate, ecological specifications that identify no cause and say nothing about any individual ballot or voter.
- Retrospective specification
- A statistical fit over the 39 county rows of one past election's ballot records. Each joint fit carries two terms: distance in miles, and the county's recorded two-ticket share of the November 2024 presidential vote. The rate fits are ordinary least squares on Too Late ballots per 1,000, so a distance estimate is per 1,000 per mile. The count fits use a log link with the returns as an exposure offset and the quasi-Poisson dispersion convention, so a distance estimate there is a log-rate coefficient.
- Recorded two-ticket share, November 2024
- One ticket's recorded county votes divided by the sum of both named tickets' recorded county votes in the November 2024 presidential election, from the state's official county results export. It excludes every other candidate and every write-in, so it is not a share of all presidential votes, of registered voters or of 2026 primary voters. It is a count of votes already cast, used as a covariate in a retrospective fit.
What is the 'historical two-ticket share' used in the county-level model? It is one ticket's recorded county votes divided by the sum of both tickets' recorded votes in the November 2024 presidential election, from the state's official county results, leaving out other candidates and write-ins.
Material qualifications
- The county-level models use 39 county aggregates, not individual voters or ballots, and identify no cause and no individual behavior.
- The historical two-ticket share used in the joint model is one ticket's recorded county votes divided by the sum of both tickets' recorded county votes from the November 2024 presidential results, leaving out every other candidate and write-in and not representing all voters.
- The Spokane facility location used in the distance calculation is approximate: three geocoder candidates span about 4 miles and are road-segment matches, not a confirmed facility address.
- An association between distance and late-ballot rates across county aggregates does not establish that distance, or any other factor, caused any ballot to arrive late.
Exact quantities
Scroll horizontally to see all columns.
| Quantity | Value | Unit | Denominator | Uncertainty |
|---|---|---|---|---|
| Distance term, joint rate fit, mail channel | 0.361 | Too Late per 1,000 mail ballots per mile | One mile of straight-line distance, with the recorded two-ticket share held in the same fit. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
| Conventional p-value for that distance term | 0.00062 | probability | Thirty-six residual degrees of freedom. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
| Two-ticket share term, joint rate fit, mail channel | -31.1 | Too Late per 1,000 mail ballots per unit share | A one-unit change in the recorded share, which is the whole range from none to all, with distance held in the same fit. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
| Conventional p-value for the share term, mail channel | 0.183 | probability | Thirty-six residual degrees of freedom. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
| Distance term, joint rate fit, all channels | 0.095 | Too Late per 1,000 records per mile | The same specification on the all-channel rate. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
| Conventional p-value for the distance term, all channels | 0.043 | probability | Thirty-six residual degrees of freedom. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
| Conventional p-value for the share term, all channels | 0.059 | probability | Thirty-six residual degrees of freedom. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
| Distance term, count fit with a mail-return offset | 0.00987 | log rate per mile | A log-rate coefficient, not a rate difference, under the quasi-Poisson dispersion convention. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
| Pearson dispersion, count fit with a mail-return offset | 65.4 | dispersion | The quasi-Poisson convention scales the standard errors by this factor. | Reproduced from the retained sources by an independent reviewer whose separate implementation agreed with the original run. |
How the evidence supports this finding
Each fit takes the 39 county rows and estimates the distance term with the recorded two-ticket share in the same specification, on the rate scale by least squares and on the count scale with an exposure offset.
Distance and 2024 vote share in the county model
Distance is associated with 0.361 more Too Late ballots per 1,000 mail ballots per mile (p=0.0006) in a county-aggregate fit; this is ecological and identifies no cause.
The county-level models use 39 county aggregates, not individual voters or ballots, and identify no cause and no individual behavior.
The historical two-ticket share used in the joint model is one ticket's recorded county votes divided by the sum of both tickets' recorded county votes from the November 2024 presidential results, leaving out every other candidate and write-in and not representing all voters.
Sources and calculations:
Evidence that could change the finding
A different set of counties, a non-curated facility list, or a different covariate would change these estimates.
- Counterevidence considered
- Reproduced estimates for every published specification
Sources and calculations
Scroll horizontally to see all columns.
| Evidence | Locator | Source | Publisher |
|---|---|---|---|
| The 39 county rows every specification is fitted on | data selection: All 39 rows of model-inputs.csv; no county is excluded, including the county that records zero Too Late ballots. | Penner Strategy derived model input and result tables | Penner Strategy |
| Reproduced estimates for every published specification | data selection: All 8 rows of retrospective-models.csv, plus the ranges recorded in the receipt: no specification that was run is excluded, including the one where the distance term is not conventionally significant. | Penner Strategy derived model input and result tables | Penner Strategy |
| Recorded county votes for the two named 2024 presidential tickets | data selection: 78 rows of the official export, one for each of the two named tickets in each of the 39 counties; every other candidate row and every write-in row is excluded. The resulting historical_two_ticket_share field is one ticket's votes divided by the sum of both. | Official county results export, November 2024 general election | Washington Secretary of State |
| County distance and Too Late rate, all 39 counties | data selection: Columns distance_miles, rate_mail_per1000 and rate_all_per1000, all 39 county rows; no rows excluded. | Penner Strategy derived county distance and Too Late rate table | Penner Strategy |
Every estimate was recomputed from the retained sources in a separate implementation by an independent reviewer, alongside a run of the original project code, and the two agreed; the covariate was recomputed again here from the official county export. This establishes that the arithmetic reproduces. It establishes no cause, nothing about any individual ballot or voter, and no future election.
Data, code and downloads
Every file this finding rests on. Figures carry their own download links beside them.
- Retrospective specifications and their estimates
Every specification this edition publishes, with each term's estimate, standard error, conventional p-value and the note that says what it measures.
- The 39 rows every specification is fitted on
One row per county: distance, both denominators, both Too Late counts, both rates, the recorded two-ticket share of the November 2024 presidential vote and whether the county center is beyond 50 miles.
- Retrospective specification receipt
What was reproduced in each specification, where the covariate comes from, and the list of things these numbers do not establish.
- Downloadable reproduction of the published specifications
Frozen public inputs, pinned dependencies, the calculation, a verifier, a permutation-schedule regeneration script and the expected value for every estimate this edition publishes. Its ballot-count input is coarsened to 480 nonpersonal county groups: no individual or precinct-level record is included.
- The calculation the published estimates come from
The same file the downloadable archive runs: it rebuilds the county rows, fits each specification and prints every estimate.
- The verifier that checks a reproduction against the expected values
Checks each input file's hash, runs the calculation and compares every value with the expected results, refusing with a nonzero exit on any mismatch.
Later changes
Updates to this finding. Corrections and withdrawals are listed with the current answer.
Use this research
Penner Strategy. "A county-level model fitting distance alongside a 2024 vote-share covariate". How often Washington mail ballots are rejected for arriving late. Version 0001. September 12, 2026. https://pennerstrategy.com/research/washington-late-ballots/editions/0001/findings/f-joint-model/. Qualification: The county-level models use 39 county aggregates, not individual voters or ballots, and identify no cause and no individual behavior.
- Print this edition
- Machine-readable release (JSON)
- Supporting files are available from the original source. See the evidence records for each source link.