How does weather change the way a marathon is run?

First calculate each edition’s median outcome among linked runners with a recent benchmark, requiring 100 finishes. Then average edition medians equally within temperature bands, requiring five editions. The profile is normalized by each runner’s own marathon average; performance is relative to the earlier benchmark.

The modeled weather is a start-hour proxy, not each runner’s exposure. Start offsets are absent, temperatures change during the race and humidity, wind, sunshine, terrain and fitness remain potential confounders. Edition counts in the source table are the number of weather exposures; finish counts are not independent weather observations.

For explicitly audited canonical-ID releases, validate unique matching CORE/FULL ID sets and matching recorded edition/name labels, then join by record ID with matching finish and all nine section durations. Legacy exports retain the one-to-one edition, trimmed lowercase name and full-timing join because their IDs are incompatible. Durations are compared to milliseconds. Keep only supplied non-ambiguous runner identities without conflicting recorded gender, inferred birth years spanning more than two years, or duplicate editions. These are candidate cross-race identities, not independently verified people; unlinked runners are absent. The linkage audit records which join was used.

Recompute the benchmark as the fastest eligible finish in the two strictly earlier calendar years. The current race and every other race in its calendar year are excluded. This avoids guessing within-year chronology and prevents current-outcome leakage. It is a recent recorded best, not a fitness measurement, an expected finish, or a lifetime personal best.

Performance change is 100 × (current finish / recent recorded best − 1). Negative is faster. Opening change compares 0–6.21 mi pace with that earlier best’s full-marathon pace. Faster opening: more than 2% faster; similar: within 2%; slower: more than 2% slower. The ±2% and ±5% cutoffs are predefined descriptions, not physiological thresholds.

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.

Full source records are available in public GitHub Releases. These chart tables require at least 100 eligible observations per cell for estimate reliability. Counts refer to finishes, linked pairs or event observations as specified in that answer.

These are observational results. Fitness changes, intentions, training, selection into the dataset and unmeasured conditions can explain differences. Outcome percentiles describe variation among performances, not confidence intervals or advice about the best strategy.

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.

Marathons, years and recorded data. Chart samples may be smaller than the eligible analysis cohort.