REF-1245 · Smyth & Muniz-Pumares, 2020 · Medicine & Science in Sports & Exercise · Paper · PubMed · Free full text
How fast can I start a marathon?
No faster than ~94% of critical speed, a ceiling not a target, by association: in ~25,000 recreational marathoners, faster 2-16 km openers more often slowed >25% late (men 26% vs 20.5%). Observational.
Smyth and Muniz-Pumares asked whether critical speed (CS) estimated from raw training data can predict marathon performance and inform pacing. They used Strava logs from runners of the Dublin, London and New York marathons (2014-2017): every logged run in the 16 weeks before the race, sampled every 100 m and grade-adjusted.
For each runner they took the fastest time over 400, 800, 1000, 1500, 3000 and 5000 m and fitted the linear distance-time model to every combination of at least three distances (42 models); the slope is CS and the intercept D′.
All models fitted almost perfectly (R² 0.9999 ± 0.0001) and gave similar CS (3.74 ± 0.08 m/s); the 400-800-5000-m model predicted finish time best (R² = 0.695, leave-one-out error 7.67%; all models 8.01% ± 0.51%).
On average runners covered the marathon at 84.8% of CS; the fraction fell with finish time, from 93.0% for about 150 min to 78.9% beyond 300 min, and only 0.82% averaged faster than their estimated CS.
The speed from 2 to 16 km (base speed) averaged 87.6% of CS; the higher it was relative to CS, the larger the second-half slowdown, and runners above about 94% of CS were more likely to slow by more than 25% over the last 12.2 km (men 26.0% vs 20.5% below it; women 15.6% vs 9.6%). Men slowed this much twice as often as women (20.5% vs 10.7%).
Our reading: a very large, consistent observational picture that gives the base a recreational anchor for "how much of threshold can I hold for a marathon" and a physiological ceiling for the opening pace, close to the elite value of about 96% of CS (REF-809); it does not show that holding the early pace under 94% of CS prevents a collapse, and CS from training logs is only as good as the hard efforts in them.
Protocol
Data: activities logged to Strava by runners of the Dublin (2014-2016), London (2015-2017) and New York (2015-2017) marathons; all logged runs in the 16 weeks before the race; time, distance and elevation every 100 m; grade-adjusted pace computed at 100-m intervals. Data partly provided by Strava under a limited research licence. Anonymised per race, so a runner who ran more than one of these races appears more than once. Only logged running was available.
Who (Table 1): 25,243 runners in the table (our sum), 6,568 women (26.0%, our arithmetic); the Discussion adds that only runners with at least 24 logged activities in the 16 weeks were included; the text's data-set count is 31,190.
Critical speed: for each runner the fastest time over each target distance within the 16 weeks, from a rolling average of grade-adjusted pace (e.g. the fastest 10 consecutive 100-m values for 1000 m), assumed to be time to exhaustion; CS = slope and D′ = intercept of distance against time; 42 combinations of at least three distances. An "adjusted model" that corrected fastest paces judged sub-maximal (supplement) gave similar results and is not reported in the paper.
Prediction: linear regression of finish time on CS and D′, leave-one-out (each runner predicted from all the others); percentage error between actual and predicted time.
Pacing: relative marathon speed = average race speed / CS; base speed = average speed from 2 to 16 km, relative to CS; relative split; a "significant slowdown" = slowing over the last 12.2 km by more than 25% relative to base speed (written as > 1.25).
Subgroups and statistics: men vs women and under 40 vs 40 and over; F-tests for regressions; one-tailed Welch's t-tests between groups; analyses on finish time in 10-min bins and on relative speed in 2% bins; significance at P < 0.01; Python.
Ethics, funding and interests: deemed exempt by the ethics committees of University College Dublin and the University of Hertfordshire (no data collected); no funding stated beyond Strava's partial provision of data; the authors declare no conflict of interest and no financial disclosures.
Key Results
| Measure | Result |
|---|---|
| Fit of the distance-time model (42 combinations) | R² 0.9999 ± 0.0001; CS 3.74 ± 0.08 m/s (SEE CV 1.3%); D′ 149 ± 24 m (SEE CV 19.44%) |
| Best model (400, 800, 5000 m) | CS 3.69 ± 0.57 m/s, D′ 136 ± 39 m; finish-time R² = 0.695, error 7.67% |
| All 42 models / worst (400, 800, 1000 m) | R² 0.67 ± 0.03, error 8.01% ± 0.51% / R² 0.54, error 9.72% |
| CS, men vs women | 3.83 ± 0.54 vs 3.29 ± 0.47 m/s; prediction error 7.67% in both overall |
| Race speed as % of CS | 84.8% ± 13.6% overall; women 85.5% ± 23.2%, men 84.6% ± 7.8% (P = 0.003); under 40 84.8%, 40 and over 84.9% |
| By finish time | about 150 min: 93.0% of CS; beyond 300 min: 78.9% of CS |
| Faster than estimated CS over the race | 0.82% of runners |
| Base speed (2-16 km) | 3.24 ± 0.55 m/s = 87.6% ± 6.9% of CS |
| Slowdown over 25% in the last 12.2 km | men 20.5%, women 10.7% (P < 0.01); under 40 17.9%, 40 and over 18.1% |
| By base speed below / above 0.94 of CS | men 20.5% / 26.0%; women 9.6% / 15.6% (both P < 0.01) |
| By base speed below / above CS | 17% / 23% (P < 0.01) |
Base speed and slowdown: relative base speed correlated with the relative split (R² 0.69-0.77 in men, women and both age groups; P < 0.01 in all but the older runners) and with the share of runners who slowed by over 25% in men (R² = 0.62, P = 0.005) and under-40s (R² = 0.67, P = 0.002), not significantly in women (R² = 0.36, P = 0.13) or over-40s (R² = 0.04, P = 0.7). These R² values come from binned data, not individuals.
Even pacers: runners whose two halves were within 2% of each other ran at about 88% of CS, others at about 84% (Discussion).
Error in minutes: about 8%, or about 18 min for the average finish time of 230 min (authors).
Numbers that do not add up:
Sample: the Methods count 31,190 runners "(Table 1)", Table 1 sums to 25,243 (our arithmetic), and the abstract says ">25,000"; the inclusion rule (24 or more logged activities) appears only in the Discussion, and the analysed n is never stated; the Methods also say every runner in the data set registered a race time, yet Table 1 lists finishers at 87.8-97.6%.
The abstract's "78.9% CS for 360 min marathon times" is ">300 min" in the Results; its "first half of the marathon at >94% of their CS" is the 2-16 km base speed in the Methods and Results, and its slowdown "in the second half" is over the last 12.2 km.
The age difference in CS is given as 8.8% (3.76 vs 3.62 m/s), which is 3.9% (our arithmetic).
Men's late-slowdown rate is 20.5% overall and also 20.5% below 0.94 of CS, with 26.0% above it; that can hold only if almost no men started above 0.94, while a base speed of 87.5 ± 7.0% of CS puts about 18% of men above it and an overall rate near 21.5% (our arithmetic, normal approximation). The women's figures (9.6%, 15.6%, overall 10.7%) are consistent.
"Relative split" is defined twice: second-half time over first-half time (Figure 4) and slowdown over the last 12.2 km against base speed (Methods, Results). The captions of Figures 2-3 invert relative marathon speed ("CS relative to MS").
Mechanisms
Measured: the fraction of CS sustained over the marathon fell as finish time rose; a faster opening relative to CS went with a larger second-half slowdown and more frequent late collapse.
Authors' interpretation: CS itself declines during prolonged exercise (they cite a fall of about 10% in critical power after 120 min below it), so the longer a runner is on the course the lower the fraction of fresh CS they can hold; fitter runners may also resist that decline. Above CS there is no metabolic steady state and D′ is drawn down, which is why an opening near or above CS costs later.
Authors' interpretation: the few runners faster than their estimated CS may be cases where the logs missed a maximal effort, or who were helped by a taper, drafting, a sea-level course or better race-day conditions.
Inference: CS from training logs is a floor where the logs hold no hard 400-m, 800-m or 5-km effort, which makes a runner's relative pace look higher than it is; the 94% figure is tied to this estimation method and should not be moved onto a laboratory or test-based CS without that caveat.
Practical Application
How fast can I start a marathon? No faster than ~94% of critical speed, a ceiling not a target, by association: in ~25,000 recreational marathoners, faster 2-16 km openers more often slowed >25% late (men 26% vs 20.5%). Observational.
Goal check against critical speed: recreational runners averaged 84.8% of their training-derived CS over the marathon, about 93% for a 2.5-h finish and about 79% for 5 h or more, and more than 99% ran below CS (B+, observational). A goal pace at or above CS is unrealistic; the elite ceiling is about 96% (REF-809).
Use 94% as a ceiling for the opening, not a target: for a runner with a CS of 4:30/km (3.70 m/s), 94% is about 4:47-4:48/km (our arithmetic). For anyone outside about 2.5-3 h, the race average will be well below this, so the opening should sit nearer the planned average (the base's practice in SYNTH-64: first half at goal pace, never faster).
Estimate CS from the logs, then sanity-check it: the fastest 400 m, 800 m and 5000 m from the last 16 weeks predicted finish time with about 8% error (about 18 min for a 230-min marathon). Inference: without a recent hard 5 km and some fast reps the estimate is low; a 5-km race in the build makes it usable.
Limitations
Observational. Faster openers may differ in fitness, experience, heat, crowding or targets; the study cannot show that a lower opening prevents collapse. The 94% figure is an observed point on a continuous relationship, not a pre-specified or validated cut-off.
CS from training logs. Fastest segments were assumed to be maximal; no heart rate or effort data; only logged running; GPS sampled at 100 m; no comparison with a test-based CS in these runners. D′ was poorly estimated (SEE CV 14.56-19.44%) and came out higher in women than men, which the authors suggest may reflect estimation error.
Analysis on bins. The near-perfect R² values for pacing describe 10-min and 2% bin means; individual scatter is wide (relative race speed SD 13.6 points).
Sample. The analysed n is not stated; repeat runners may appear more than once; three big-city sea-level marathons, about three quarters men; no weather, drafting or fuelling data.
Reporting. The sample counts, the 360-min and "first half" wording in the abstract, the 8.8% age difference and the men's slowdown percentages do not reconcile (see Key Results).
Interests. No conflicts declared; data partly from Strava under a research licence.
How this entry was checked and changed
Why this level
large observational cohort, n = 25243 by the sum of Table 1 (our arithmetic; the text gives 31,190 runners in the data set and ">25,000" analysed), read from the full text (open access, CC BY-NC-ND 4.0). Associations only; critical speed was estimated from efforts not known to be maximal and was not checked against a laboratory or test-based value in these runners; the pacing curves are fitted to binned group means
Full text (checked 2026-10-27)
Activity logged to Strava by 31,190 runners of the Dublin, London, and New York marathons from mid-2014 to late 2017; only runners with at least 24 activities logged in the 16 weeks were included (full text)
The best performing model was correlated with MT (R2 = 0.695) and was able to predict MT with a 7.67% error; the 42 models gave an error of 8.01% ± 0.51% (full text)
Runners completed their marathon at an average speed that corresponded to 84.8% ± 13.6% of their CS (full text)
Runners with finishing times of ~150 min completed the marathon at 93.0% of their CS, whereas runners with finishing times >300 min completed the marathon at 78.9% of their CS (full text)
A small number of runners (0.82%) completed the marathon at speeds faster than their estimated CS (full text)
Slowdown greater than 25% in the last 12.2 km: males 20.5% vs females 10.7% (full text)
Relative BS < 0.94: 20.5% (men) and 9.6% (women) slowed significantly, compared with 26.0% and 15.6% for relative BS > 0.94 (full text)
Estimates of CS for younger athletes were 8.8% higher than for older athletes (3.76 ± 0.59 vs 3.62 ± 0.54 m/s) (full text)
The authors declare no conflict of interest; data were, in part, provided by Strava under limited research license (full text)