Start · the problem
Everything changed at the same time
A plain before-and-after comparison credits the shoe with everything that moved over those fifteen years: pacing, course choice, prize money, training and the depth of East African fields. The study compares the pre-Vaporfly era (2010–2016) with the post-Vaporfly era (2018–2024) and leaves out 2017 as a transition year. Each framework holds a different thing constant:
- Difference-in-differences. Road marathon improvement minus track 10,000 m improvement, with the track change rescaled to marathon time. Track racing took up carbon plates later and with smaller claimed gains, so anything that sped up both events cancels.
- Within-athlete paired. Sixteen runners with at least two elite marathons in each era. Comparing a runner with themselves holds genetics and training history fixed.
- Cohort survival. The shift in the top-30 median between eras, with 55% of it attributed to shoes, plus a changepoint search on yearly sub-2:10 and sub-2:25 counts.
Km 10 · the answer
Three methods, a wide spread, one pooled answer
Pooled across the three frameworks. 95% CI −4 to 150 seconds.
At 2:05 pace (5.63 m/s), 73 seconds is about 409 meters of road: a runner in 2018 shoes would finish that far ahead of the same runner in 2016 shoes. The shaded band is the 95% interval, which runs from 23 meters behind (drawn from the start line) to 842 meters ahead.
The frameworks disagree by a lot. Within-athlete pairing is the lowest at 44 seconds, with an interval that crosses zero. Difference-in-differences is the highest at 199 seconds, built on one year of women’s track results as the pre-era control. Cohort survival sits between them at 71 seconds.
The pooled value weights each estimate by the inverse square of its interval width, so cohort survival, with the tightest interval, carries about 93% of the weight. The pooled interval comes from the spread of the three estimates (the estimate ± 1.96 times their standard deviation over √3), with no bootstrap at this step. The cohort figure depends on an assumption the analysis makes, that 55% of the cohort shift is the shoe. The slider changes it.
At a 55% shoe share, the pooled estimate is 73 seconds
Each framework's estimate and 95% interval in seconds of shoe-attributable improvement,
re-pooled with the formula in synthesize() as the cohort share moves.
Estimates and intervals from outputs/analysis_summary.json. The cohort interval
scales with the share, as in the script (raw_ci_low and raw_ci_high times the share).
28 October 2026: the bootstrap now resamples the same statistic as the estimate.
Km 21 · the control
Track improved too, eight times less
Women’s road marathon top-50 median time fell 228 seconds (2.57%) between the eras. Women’s track 10,000 m fell 0.32%. Rescaled to marathon time, the track change accounts for 29 of the 228 seconds, which leaves 199 for road-specific causes. A placebo split inside the pre-era (2010–2013 against 2014–2016) gives 40 seconds.
10 of 16 runners got faster after 2017, by a median of 44 seconds
Change in each runner's mean elite marathon time, post-era minus pre-era. Left of the line is faster. Hover, tap, or focus the chart and use the arrow keys.
From data/athlete_career_arcs.csv, rebuilt without wheelchair athletes. Raw times, no age adjustment.
The paired t-test gives p = 0.39, so the direction is suggestive and the size is not settled. Until 28 October 2026 this chart had 17 runners. One was Shelly Woods, a wheelchair racer whose pre-era mean of 2:11:04 was faster than the women’s running world record of the time. Checking the source pages showed the scraper had read every “Wheelchair women” and “Handcycle women” table as women’s results: 106 rows, 54 athletes. All are now excluded.
Km 30 · the timeline
Elites first, the field three years later
The changepoint search puts the break in the sub-2:10 and sub-2:25 counts at 2020, three years after the Vaporfly went on sale. The repo’s hand-compiled shoe timeline shows why that lag is plausible: Nike had the field to itself until 2020, when three other brands released carbon-plated racers within four months.
- Nike Vaporfly 4% prototype; Kipchoge wins Olympic gold in Rio
- Breaking2 at Monza, 2:00:25 (exhibition, not ratified)
- Vaporfly 4% released
- Kipchoge 2:01:39 world record, Berlin
- Kosgei 2:14:04 world record, Chicago
- World Athletics caps stack height at 40 mm with one plate
- Saucony Endorphin Pro; New Balance FuelCell RC Elite (06); Adidas Adizero Adios Pro (08)
- Changepoint in sub-threshold counts (this analysis)
- Asics Metaspeed Sky; Puma Deviate Nitro Elite (09)
- Hoka Rocket X 2
- Assefa 2:11:53 world record, Berlin
- Kiptum 2:00:35 world record, Chicago
Release and record dates from data/shoe_timeline.csv. Bold rows are world records.
Men under 2:10: 23 a year before the Vaporfly, 29 after
Performances under the threshold per year in the dataset, with each era's mean as a line. 2020 is the pandemic year. Hover, tap, or focus the chart and use the arrow keys.
Counted from data/elite_marathon_times.csv with the thresholds in framework3_cohort_survival()
(under 7,800 s for men, 8,700 s for women); the era means match analysis_summary.json.
More fast times followed. Sub-2:10 men went from 23 a year to 29 (1.25×) and women under 2:25 from 13 to 25 (2.0×). At the majors the count is capped by the size of the lead packs. Most of the post-era rise comes late: 2024 alone has 53 men under 2:10 and 43 women under 2:25.
Km 35 · limitations
What this doesn’t show
The track control has 27 rows, and its pre-era side is one year (2016) of women’s results. Men have no pre-era track data at all. That is the input to the highest estimate, so the 199 seconds is best read as an upper bracket.
The race tables have no ages, so the within-athlete change is not age-adjusted. The paired runners were typically four to six years older after 2017, which pushes that estimate toward zero. No race records which shoe anyone wore; era stands in for shoe use.
The 55% shoe share is an assumption. At 40% the cohort estimate is 52 s and the pooled value 53 s; at 70%, 90 s and 91 s. The script has no switch for this, so those values come from re-pooling the summary file with the script’s formula, which is what the slider above does.
The sensitivity runs cover only the difference-in-differences: 183 s with the top 25 per year instead of the top 50, and 199 s with Kipchoge or time-trial events removed (the point estimate is women-only, so dropping Kipchoge cannot move it). The interval is 74 to 327 s for the top 25 and the baseline 83 to 338 s for the two exclusions, which remove no women’s results.
Finish · how it was built
Where the numbers come from
- Elite times
- 1,908 rows from major-marathon Wikipedia result tables (Berlin, London, Chicago, Boston, Tokyo, New York, a few Dubai entries), 2010–2024; 106 wheelchair and handcycle rows are excluded, leaving 1,802
- Control
- 27 track 10,000 m performances, 2016–2024
- Shoe timeline
- 19 hand-compiled rows: releases, first major wins and the 2020 World Athletics rule
- Methods
- Difference-in-differences with a placebo window, paired t-test and 1,000-draw bootstrap, slope-change changepoint search, inverse-CI-width pooling
- Stack
- Python, pandas, NumPy, SciPy, matplotlib, seaborn
- Biomechanics
- Hoogkamer et al. 2018, Kipp et al. 2019, Senefeld et al. 2023
- Reproduce
python src/analysis.pyregenerates the 8 figures and the JSON summary in under a minute; the current outputs are from the 28 October 2026 rerun- Related
- World-Record Half-Lives models how quickly records fall, including the marathon in this era