Caffeine and sprint times: what a 21-sprinter crossover trial can detect
A crossover trial finds 3 and 6 mg/kg caffeine cut 60-m sprint times. We check its power sensitivity analysis and estimate how many sprinters smaller effects would need, with fibub's sample size calculator.
Caffeine is one of the few supplements with good evidence for helping athletes. A new trial tested it on 21 male 100-m sprinters, each of whom ran after a placebo, a moderate dose and a higher dose, on different days and without knowing which. Both doses cut sprint times, and the higher dose cut them more. A study this size raises a question every researcher should ask: how big an effect could 21 people reliably detect, and how many would a follow-up need? This article answers it with the sample size calculator, the same maths the authors’ check relies on. The uncertainty propagation calculator helps when you combine timing errors, and the publication plot maker draws your own power curves.
This is general education about a study’s statistics, not advice on taking caffeine. High doses can cause side effects; check with a doctor or sports dietitian.
The paper in brief
The trial was randomised, double-blind, placebo-controlled and crossover. Crossover means every sprinter tried every condition, so each one is compared with himself. That removes the large differences between athletes and makes small changes easier to see. The doses were 3 and 6 mg of caffeine per kg of body weight, about 210 and 420 mg for a 70 kg sprinter.
The main results, as reductions in time against placebo:
- 10 m: 0.056 s with 3 mg/kg, 0.075 s with 6 mg/kg.
- 60 m: 0.123 s and 0.175 s. All were clear (p < 0.001 after correcting for several comparisons).
- 6 against 3 mg/kg: clearly better at 30 and 60 m, but at 10 m the 0.019 s gap was not significant (p = 0.085).
Reaction time at the start and jump height also improved. The authors note that they did not calculate the sample size in advance. Instead they report a sensitivity analysis: the smallest effect 21 sprinters could detect with 80% power. That is the number we check.
Checked with the tools
We used the paper’s sample size, effect sizes and confidence intervals with the Sample size calculator’s maths: exact power for a paired t test from the noncentral t distribution, as in R’s power.t.test.
Step 1: the sensitivity analysis. The effect size for a paired design is dz: the mean difference divided by the standard deviation (SD) of the differences. With 21 sprinters and 80% power, the tool gives the smallest detectable dz as 0.643 at the usual 5% level, and 0.761 at 0.05 ÷ 3 = 0.0167, the stricter level for three comparisons. The paper says 0.64 and 0.76: an exact match (Figure 1).
Step 2: how big were the real effects? A paired confidence interval is the mean ± t × SD ÷ √n. Working backwards from the 60-m interval for 3 mg/kg (0.086 to 0.160 s), the SD of the differences is about 0.081 s, so dz ≈ 1.5. For 6 mg/kg it is about 2.2. These are far above 0.64, so the study had near 100% power for its main findings. This is our estimate: the paper’s intervals come from its own models, not a plain paired t test.
Step 3: the small effect it could not confirm. At 10 m, 6 mg/kg beat 3 mg/kg by only 0.019 s. If the SD of that difference is like the 10-m SD for 3 mg/kg against placebo (about 0.049 s, our assumption), dz ≈ 0.39. With 21 sprinters the power to detect that is only about 39%. To reach 80% a follow-up would need about 56 sprinters (Figure 2). So “not significant” here means “too small for this study to see”, not “no difference”.
| Check | Paper | Tool result |
|---|---|---|
| Detectable dz, 80% power, α = 0.05 | 0.64 | 0.643 |
| Same at α = 0.0167 | 0.76 | 0.761 |
| Effect size, 60 m, 3 mg/kg | (as g, range 0.90 to 1.66 for all timed outcomes) | dz ≈ 1.5 from the CI |
| Power for 6 vs 3 mg/kg at 10 m | p = 0.085 | about 39% (our SD estimate) |
| Sprinters for 80% power at that effect | (not reported) | about 55 to 56 |
How we made the graphs
The charts and every number above come from scripts/blog-figures/caffeine-100m-sprinters-crossover-trial-power-sensitivity.mjs in fibub’s public code. It calls the functions behind the sample size calculator, from src/lib/sci/power.ts:
powerTgives the exact power of a paired t test for a sample size, effect size and significance level;solveNfinds the smallest sample size that reaches a target power;tInvfromsrc/lib/sci/stats.tsgives the t value used to turn the paper’s intervals back into SDs.
From the paper: 21 sprinters, the reported sensitivity values, and the reductions with their 95% intervals. Our assumptions, labelled in the script: the intervals are treated as paired-t intervals, and the SD of the 6 against 3 mg/kg difference at 10 m is taken to equal the SD of 3 mg/kg against placebo at 10 m.
How we test the maths
On every change to fibub, tests/sci.mjs checks the power maths against R and SciPy:
- A two-group t test with effect size 0.5 needs 63.77 per group for 80% power (R’s
power.t.testgives 63.77), and 64 per group gives power 0.80146. - A paired t test with 10 pairs and dz = 1 has power 0.80310, as SciPy’s noncentral t gives.
- Two proportions, 50% against 75%, need 76.7 per group for 90% power.
Check it yourself
On the sample size calculator:
- Set Comparing to Paired or before-after means.
- Type 0.643 in Difference to detect and 1 in Standard deviation (so the difference is in SD units). Leave Significance level α on 0.05 and Power on 0.8.
- The answer reads 21 pairs, with Power with that n about 0.800: the paper’s sensitivity analysis.
- Now type 0.019 in Difference to detect and 0.049 in Standard deviation. The answer reads 55 pairs. (Our script uses 0.0494 s and gets 56.)
- The calculator’s α menu offers 0.01, 0.05 and 0.1, so it cannot do the 0.0167 check directly. Choose 0.01 for a stricter bound: the same 0.643 then needs 32 pairs.
What researchers can use it for
- Plan a crossover or before-after study: type the smallest change that matters and the SD of the changes into the sample size calculator.
- Write a sensitivity analysis when the sample size was fixed by availability, as here: find the smallest effect the study could detect.
- Read a “not significant” result fairly: work out the power for the effect that was seen before calling it “no effect”.
- Recover an SD from a confidence interval in a paper, then use it to plan the next study.
- Combine timing errors (gate timing, reaction) with the uncertainty propagation calculator before deciding what difference is worth detecting.
The fine print
- The sample size, sensitivity values and confidence intervals come from the paper. The SDs and dz values in Steps 2 and 3 are our estimates from those intervals, assuming paired-t intervals.
- The SD of the 6 against 3 mg/kg difference is an assumption; the true one could be larger or smaller, and the 56 would change with it.
- Power with several comparisons depends on the correction used; the paper uses Holm’s method, which is less strict than dividing α by three.
- General education only, not advice on caffeine use.
Sources
- Mei, T., Li, J. and Yang, X. (2026), Acute doses of 3 and 6 mg/kg caffeine produce outcome- and phase-specific improvements in male 100-m sprinters: A randomized, double-blind, placebo-controlled crossover trial, PLOS ONE 21(11): e0361864, doi:10.1371/journal.pone.0361864. Licence CC BY 4.0, checked on the journal page, 2026-11-16. Values used here, from the Methods and Results: 21 male sprinters, no dropouts; no a priori power calculation; sensitivity: 80% power for dz = 0.64 at two-sided α = 0.05 and dz = 0.76 at α = 0.0167; reductions in time vs placebo with 95% CI: 10 m, 0.056 s (0.033 to 0.078) and 0.075 s (0.053 to 0.097); 60 m, 0.123 s (0.086 to 0.160) and 0.175 s (0.138 to 0.212) for 3 and 6 mg/kg; 6 vs 3 mg/kg at 10 m 0.019 s (Holm p = 0.085).
- Paired t-test power with the noncentral t distribution: Cohen, J. (1988), Statistical Power Analysis for the Behavioral Sciences, 2nd ed.; R's power.t.test.
If you are an author or publisher and want a credit or a passage changed or removed, tell us and we will act quickly.