What happens when the race gets warmer?

After checking the raw timings, we excluded 40,901 otherwise plausible finishes from reviewed editions with invalid split grids, incomplete ingestion, unresolved source checks or a selected field. These source exclusions are reported separately from timing errors in the downloadable evidence.

The outcome is each edition's median individual percentage pace change from 0–12.43 mi to 12.43–24.85 mi: 100 × ((time at 24.85 mi − time at 12.43 mi) / time at 12.43 mi − 1). Positive is slowing. These are equal-distance blocks, not measured half-marathon splits; the last 1.36 mi is excluded.

Each city/year must identify exactly one race and one weather row with a valid same-year race date. Keep at least 100 complete eligible finishes per edition and weight each retained edition equally. The number of finishes is not the number of independent weather observations.

Use the producer's observed_at archive hour only when it is within 30 minutes of the supplied scheduled local start; either nearest hour is accepted at an exact half-hour tie. Temperature, dew point, humidity and wind must agree with that hour. Warming is temperature exactly four hourly readings later minus temperature at that start hour; no runner's finish time determines the window. Supplied start times may be approximate or refer to elite rather than mass starts.

Fit one prespecified joint linear model of edition median pace change, with course fixed effects, a centered linear calendar-year term, centered start temperature and its square, start dew point, first-four-hours temperature change and start wind speed. Course denotes the supplied city, not a verified historical route. Course fixed effects are removed by within-course centering.

Report each weather coefficient multiplied by that exposure's observed edition interquartile range (75th minus 25th percentile). This is a conditional regression contrast, not a prediction that sets dew point beyond the physically supported temperature range. The other modeled variables are held constant.

Uncertainty resamples whole courses with replacement, keeping every edition within the drawn course and recomputing the model. The fixed seed is 20260911; 6,000 draws produce percentile intervals at 98.333333% confidence, an approximate Bonferroni adjustment for the three screened exposures. At least 95% of draws must have full rank. Repeated runners across courses are not separately clustered.

Before seeing outcomes, require 30 editions, 10 courses, within-course exposure ranges at least half the overall IQR in at least 10 courses, and variance inflation factor no more than 10. A supported association additionally needs an interval excluding zero, an absolute IQR estimate of at least 1 percentage point and unchanged direction in every leave-one-course-out refit. A precise null requires the entire interval and every leave-one-course-out estimate strictly inside ±1 percentage point. This 1-point threshold is a prespecified practical publication threshold, not a physiological boundary or a claim that smaller changes never matter to runners. All three screen results remain in the audit, including withheld candidates.

This is an exploratory, adjusted edition-level association among eligible finishers. It does not establish a weather penalty, an individual forecast or a causal effect. Field composition, fitness, opening congestion, race organization, route changes and unmeasured conditions remain possible explanations. The opening 0–12.43 mi baseline can itself be affected by weather and congestion. Runners who did not finish or lack valid splits are not represented.

Relative humidity changes with temperature. Dew point is the modeled moisture exposure; the raw humidity percentage is retained for context. Wind speed gives no runner-specific headwind or tailwind without a valid historical route, direction and personal exposure timeline. Scheduled-start weather is not personal wave-start weather.

Use complete, strictly increasing elapsed checkpoints at 3.11, 6.21, 9.32, 12.43, 15.53, 18.64, 21.75, 24.85 and 26.22 mi. Clock strings must parse as H:MM:SS or M:SS. No missing splits are interpolated.

Remove exact duplicate race records, ignoring database IDs, ingestion timestamps and source URLs. Retain finishes from 90 minutes to 12 hours with every section between 3.22 and 32.19 minutes/mi. These quality filters can exclude genuine unusual performances; the analysis describes this eligible cohort, not every entrant.

Apply the reviewed source-quality edition exclusions for this exact export after the timing checks. Known invalid split grids, incomplete ingestion, unreconciled HOLD editions and a selected top-finisher field do not contribute to the analyses or prior benchmarks. Report source exclusions separately; an already invalid timing row is not counted twice. Missing age or recorded gender alone does not exclude an otherwise eligible finish from the overall cohort. Other sparse editions are not declared incomplete merely from their size.

Data: Marathon results, weather and course data. See the marathons, years and recorded conditions. Download all three screening results and edition values. Weather source definitions.

All three weather questions were screened together. The download retains their estimates and uncertainty, including any that were too uncertain for a separate page.