The Men's Marathon Record Fell Faster After 2016. A Curve Explains Half.

The men's marathon world record fell faster after 2016 than before it, if you fit a straight line with a bend. The bend is 0.0034 min/km per year (HAC 95% interval -0.0042 to -0.0025). A curved trend with no bend at all fits the same records as well or better. When I add the 2016 bend to that curve, the bend halves and its interval touches zero.

This is a cohort-style time series of 47 year-end record paces, 1980 to 2026, for each sex. It is not a trial. I read the Wikipedia "Marathon world record progression" page as parsed table rows [1]. I did not read any World Athletics or IAAF list, and I did not read a primary source for the 2026 record.

Why I ran it

Shoe claims often rest on one race or one photo. I dislike both. A dated record list is a cleaner test, because it is public and it has a long baseline. The question was simple: after 2016, did the record fall faster than the earlier trend, and how large is the break compared with its interval?

A record list cannot answer "did the shoes do it". I say that now and again at the end. It can only say whether the line bent.

Method

I downloaded the men's and women's tables from Wikipedia on 2026-10-11 [1]. I parsed date and time into a table. Pace is time divided by 42.195 km, in minutes per kilometre.

The parse check had three parts:

• Row count: 53 men's rows and 47 women's rows. They match the page tables (54 and 48 rows, less the header).
• Two paces by hand: Hayes, 2:55:18.4, gives 4.1547 min/km. Da Costa, 2:06:05, gives 2.9881 min/km.
• A threshold set in advance: at least 15 rows per sex. Both passed.

I dropped rows marked "Disputed" (2 men's rows, 8 women's rows, mostly short or unmeasured courses before 1984). I also dropped rows that only cite a newspaper claim. The women's main series is the mixed-race record. The women-only rows are a separate series with 5 records since 2005, so I did not model them.

A running minimum over date gives the progressive records. I then took the record pace at each year end. Each series has 47 points. The list holds 18 record events since 1980 for each sex. The women's count includes early-1980s rows.

The model is a straight line with a level change and a slope change at year T:

$\text{pace} = a + b(t - 2016) + cD + dD(t - T)$

Here $a$ is the pace level at 2016, $b$ is the slope before the break, $D$ is 1 from year T onward, and $d$ is the slope change. A negative $d$ means the record falls faster after T. The unit is min/km per year.

The year-end series is a step function and strongly autocorrelated. So I used two intervals. One is a Newey-West (HAC) interval with 4 lags. The other is a moving-block bootstrap (block of 5 years, 2,000 draws). I also compared models by AIC and leave-one-out error, and I ran a placebo test. The scripts are parse.py, analyse.py and sens_fig.py.

One dead end: my first script was named inspect.py. It shadowed a standard library module under python -I, so I renamed it. It cost nothing but a failed run.

Results: break at 2016

Group	Slope before	Slope after	Slope change d	HAC 95% CI	Bootstrap 95% CI
Men	-0.00366	-0.00701	-0.00335	-0.00419 to -0.00252	-0.00553 to -0.00128
Women (mixed-race)	-0.00838	-0.01512	-0.00674	-0.01030 to -0.00317	-0.01523 to +0.00311

All values are in min/km per year.

For men, $d$ is about 0.2 s/km per year. Over the 10 years to 2026 that is about 0.034 min/km, or 2.0 s/km. Over the full distance, it is about 85 s below the pre-2016 line.

For men, both intervals exclude zero. The break model beats the straight line without a break. AIC is -447.5 against -419.9. Leave-one-out RMSE is 0.0086 against 0.0116 min/km.

For women, the HAC interval excludes zero but the bootstrap interval includes it. The model gain is small: AIC -302.4 against -297.8, and leave-one-out RMSE 0.0400 against 0.0427. I trust the wider interval here, because HAC is a poor fit with 47 correlated points.

A break at 2019, the ratification year of the Kosgei record, gives similar men's results. Men: d = -0.00305 (HAC -0.00435 to -0.00174; bootstrap -0.00591 to -0.00014). Women: d = -0.01041 (bootstrap -0.02368 to +0.00377), and women gain nothing in AIC (-297.2).

What this does not show: it does not show that the bend is caused by anything. It shows only that a line with a bend fits better than a line without one.

The placebo test

I placed the same slope break at every year from 1990 to 2010 (21 years) on the same window. Then I asked where 2016 ranks.

For both sexes the 2016 break is more negative than all 21 placebos. The rank is 1 of 22. With 22 candidate years, that is a permutation p of about 0.045 at best. The tests are not independent, so I read that p as loose.

The placebos differ by sex. For men, the placebo median is -0.00234, so most placebo breaks also point downward. That means the record was already falling faster in most later windows. A bend is the normal shape of this series. For women, the placebo median is +0.0039, so 2016 stands out more, but see the next section.

What this does not show: a rank of 1 does not mean 2016 is special for a reason. I chose 2016 before I ran the test, which helps. But the placebo years cover only 1990 to 2010, and the series is short.

Checks that weaken the result

This is the part I would not skip.

• Curvature. A quadratic trend with no break has AIC -453.1 and leave-one-out RMSE 0.0083 for men. That beats the linear-break model (AIC -447.5). If I add a 2016 break to the quadratic, d falls to -0.00172 (HAC -0.00345 to +0.0000036). That is about half the linear-model value, and the interval touches zero. So the men's break depends on the model.
• End date. Without the 2026 record (data to 2025), d = -0.00276 (HAC -0.00376 to -0.00176). With data to 2023, d = -0.00355 (HAC -0.00484 to -0.00226). The result does not rest on the last record. I did not check the 2026 record against a primary source.
• Women's start year. If the window starts in 1970, the women's sign flips: d = +0.0020 (HAC -0.0048 to +0.0088). The women's series is flat from 2003 to 2019 and then falls in steps.
• Women's quadratic. The quadratic with a break gives d = -0.0149 (HAC -0.0245 to -0.0054) and AIC -308.6. It is the best women's model by AIC. It rests on 14 to 18 events, so I treat it as a hint only.
• Level change. The level-change term is not stable. Its bootstrap interval includes zero for men. At 2016 it is positive for women.

My reading: the data support a faster men's fall after 2016 under a straight-line frame. A curved frame keeps about half of it. For women, the data support no firm break estimate.

What a record list cannot separate

• Shoes from pacing teams. Pacing teams and rabbits changed over the same years.
• Shoes from courses. Flat, fast courses and the Berlin and Chicago circuits produce many of the records.
• Shoes from more athletes. More and better-paid athletes would bend the line without any shoe effect.
• Shoes from list rules. The source column mixes IAAF, ARRS and press reports, and list rules changed.
• The best from the rest. A record is the single best result among many. The list says nothing about the top 10 or the median elite runner.

This is my old complaint in a new form. A lab result on running economy and a record-list result are both far from "this shoe makes this runner faster in a race". Neither one closes the gap.

My present view

A faster fall in the men's record after 2016 is likely under a linear-trend model, and it stands out against placebo years. A curved trend explains about half of it. A record list cannot name a cause. I put about 0.6 on "some real bend in the men's line after 2016" and much less on any share being due to shoes, because the list does not test shoes. I hold no firm view on the women's break.

What would change my view:

1. Top-10 or median-elite times by year from a public results database. If the whole field bent at 2016, that points away from a few pacing-team records.
2. A primary-source record list from World Athletics, to replace the Wikipedia secondary list.
3. A larger trial of shoes in the same runners, which would test shoes directly. The shoe-injury trials I have covered before do not answer this, because they measure injuries, not race pace.

What I would do next

The next study I would trust uses race results, not records. It would take every finisher in a few large marathons from 2010 to 2026, with year and course fixed, and ask whether the whole distribution moved at 2016. A shoe effect should show up in the middle of the field. A pacing-team or course effect should not. I will try that with public results next.

Files

The outputs are parsed_records.csv, events files, results.json and break_table.csv, plus the three scripts. No app was published.

Break estimates (min per km per year) with HAC and block-bootstrap 95% intervals, AIC and leave-one-out error.