Superficially an interesting paper, but chatGPT seems to have found some flaws.
chatGPT(5.6max paid):
The paper finds an interesting association between nitrogen dioxide and slower marathon times, but its novelty is incremental, and several methodological and reporting problems weaken its stronger conclusions. I reviewed the uploaded paper, its supplement and the publicly released analysis code.
The study analysed 2,756,553 marathon performances across 75 race events in Berlin, Boston, Chicago, London, New York and Tokyo during 2010–2024. These are performances, with some runners potentially appearing more than once.
Each participant received their race’s estimated exposure to nitrogen dioxide, NO₂, and fine particulate matter, PM₂.₅. Mixed-effects models related pollution to finishing time, adjusting for weather and calendar year.
The principal results were:
| Pollutant | Men: change in finishing time per additional 1 µg/m³ | Women: change per additional 1 µg/m³ |
|---|---|---|
| NO₂ | +1.43 minutes; 95% CI 0.66–2.21 | +1.56 minutes; 95% CI 0.63–2.48 |
| PM₂.₅ | −0.13 minutes; CI −0.67–0.40 | −0.14 minutes; CI −0.76–0.49 |
Positive values indicate slower finishing. Moving from the lower to upper quartile of modelled NO₂ exposure—a difference of approximately 5.1 µg/m³—corresponded to about 7–8 minutes slower finishing, roughly 3% of average finishing time.
The authors also report:
- Larger NO₂ associations among slower runners: approximately 2.12 minutes per µg/m³ in their “recreational” category versus 0.34 minutes in their “elite” category.
- Larger associations in younger runners: among men, 2.17 minutes at ages 18–29 versus 1.32 minutes at ≥60.
- Almost identical proportional associations in men and women: approximately 0.56% and 0.55% of mean finishing time per µg/m³.
They propose inflammation, oxidative stress and impaired respiratory or vascular function as explanations. These mechanisms were not measured in this study.
The main novelty is the updated multinational dataset and detailed NO₂ subgroup analysis. Much of the underlying finding already existed:
| Earlier research | What was already established or investigated |
|---|---|
| El Helou et al., 2012 | Analysed approximately 1.79 million marathon performances and found associations involving NO₂, alongside stronger weather effects. |
| Hodgson et al., 2022 | Compared elite and amateur half-marathon runners’ responses to environmental conditions. Significant pollution associations chiefly involved ozone. |
| Fleury et al., 2025 | Analysed approximately 2.56 million performances across 140 race events, finding PM₂.₅ associations that varied across finishing-time percentiles. |
Thus, this is principally an extension and attempted replication, with interesting subgroup hypotheses. It provides no new experimentally demonstrated biological mechanism.
The study has useful strengths: objective finishing times, multiple cities, explicit modelling of race-level clustering, ground-monitor comparisons and several sensitivity analyses. However, my main concerns are substantial.
-
The pollution evidence comes from 75 race events.
Millions of finishers precisely describe each race’s performance distribution, but pollution varies between only 75 events. The mixed-effects model appropriately recognises clustering; nevertheless, its random race intercepts do not automatically eliminate confounding by differences correlated with pollution.
Course difficulty, qualification requirements and participant composition differ between cities. The reported primary model does not include city fixed effects to isolate changes within the same marathon across years. Those comparisons would provide a more convincing basis for attributing differences to pollution.
-
The claimed weather adjustment for the distributional analysis is problematic.
Page 5 describes adjusting finishing times for weather within each race event. Yet everyone in an event has the same assigned heat index and wind speed. Their effects therefore cannot be estimated from differences between runners within that event.
The released R code performs precisely this regression, then adds the race mean back to the residuals. Given the stated data structure, that returns the original finishing times: it does not remove weather effects.
Unless different final code was used, this undermines the supposedly weather-adjusted distributional results. The ordinary Kolmogorov–Smirnov tests also lack an adjustment for clustering by race. This criticism concerns the distributional analysis; the main mixed-effects regression uses a separate adjustment procedure.
-
Figure 4 does not support the text’s description.
The paper says higher NO₂ consistently shifts finishing-time distributions towards slower performance. However, Figure 4 reports median differences of:
- −7.8 minutes in the fastest 10%.
- −3.4 minutes in the next performance category.
- +2.1 minutes in the “recreational” category.
- +0.3 minutes in the slowest 10%.
These mixed directions conflict with the claimed uniform shift. A significant Kolmogorov–Smirnov test establishes a difference between distributions; it does not establish that every part of the distribution shifts towards slower times.
-
Exposure measurement and reporting need clarification.
Several descriptions disagree:
Issue Discrepancy or concern Exposure window The main paper describes 06:00–18:00, while supplementary analysis tables repeatedly specify a three-hour window centred on race start. Spatial resolution Methods specify 0.75° × 0.75°, consistent with approximately 80 km resolution; the discussion instead says approximately 11 km. Wind speed Table 1 gives Tokyo’s mean wind speed as 43.1 m/s, equivalent to approximately 155 km/h. This is implausible as an average across marathon days and suggests a units or data problem. Calendar-year adjustment The manuscript specifies a linear trend, while the released primary code uses categorical year effects. The exposure-window discrepancy appears throughout the supplement. Copernicus confirms that EAC4’s resolution is approximately 80 km. A wind-unit labelling error alone would not necessarily change the pollution coefficient, but it needs resolving.
Moreover, the supplement reports substantial differences between modelled NO₂ and selected ground monitors, including a mean bias of −40 µg/m³ in London. Correlation with monitors can support relative comparisons while leaving the precise minutes-per-µg/m³ calibration uncertain.
-
The subgroup findings do not establish biological susceptibility.
“Elite” means the fastest 10% of each race, a broad category whose reported mean finishing time is about 3 hours 7 minutes. Conclusions about professional elite athletes therefore overreach this definition.
Performance categories are also defined using the outcome being studied: the finishing time achieved during the pollution exposure. These comparisons describe different portions of race results, but cannot cleanly distinguish pre-existing ability from pollution response.
Likewise, the age estimates have overlapping confidence intervals, and formal tests establishing differences between age-group coefficients are not reported. Missing age data for about one-third of performances and the selection of unusually fit older marathon finishers further complicate interpretation.
-
The PM₂.₅ findings are unstable, and causality remains unresolved.
PM₂.₅ is null in the main models, apparently beneficial after adjustment for NO₂, and positively associated with slower times when Boston is excluded. The authors reasonably reject the apparent benefit as causal, but this instability also warrants caution about confidently ranking pollutants.
Analysing finishers alone can miss pollution-related withdrawals. NO₂ may also mark a broader traffic-pollution mixture. Without physiological measurements or a stronger causal design, the study cannot establish that NO₂ itself produced the reported delay.
I would treat the primary NO₂ association as suggestive evidence requiring further analysis. Before relying on its numerical effect sizes or “sixfold susceptibility” claim, I would want consistent final code and exposure definitions, corrected distributional analyses, and estimates that isolate variation within each city. My assessment is based on document and code inspection; I have not reproduced the results from the underlying runner data.