Caffeine helped women's exercise performance by a small amount whatever the cycle phase or contraceptive use: in a three-level meta-analysis of 20 randomised trials (272 women, 144 effect sizes) the pooled effect was g 0.37 (95% CI 0.24 to 0.50; prediction interval −0.13 to 0.87), with g 0.28 to 0.48 in the early follicular, late follicular, luteal and hormonal-contraceptive groups and no difference between them (p = 0.551); certainty low to very low, no trial at low risk of bias. Read in full
Main caveat: No trial at low overall risk of bias; certainty low for the main effect and low to very low for subgroups.
Zhu and colleagues (Huaqiao University, Fujian Fuyao University and University of Malaya) asked whether caffeine improves exercise performance in women and whether cycle phase or hormonal contraception changes the effect, a question most caffeine reviews cannot answer because their samples are mostly men.
They pooled 20 randomised trials in 272 women that let each result be placed in one of four hormonal states: early follicular, late follicular or peri-ovulatory, luteal or mid-luteal, or hormonal contraceptive use. Doses were 3-6 mg/kg, mostly as capsules about 60 min before.
With 144 effect sizes nested within trials, caffeine improved performance overall by g 0.37 (95% CI 0.24 to 0.50), a small effect, with moderate heterogeneity (I² 40.4%) and a prediction interval that crosses zero.
Each hormonal group showed a positive estimate (g 0.28 to 0.48), the contraceptive group only at the margin (95% CI 0.00 to 0.55), and the groups did not differ (p = 0.551).
In exploratory analyses within the early follicular phase, endurance and repeated-output tasks gained more than maximal-output tasks, and trials that did not verify phase hormonally showed larger effects than trials that did; neither pattern appeared in the luteal phase. Dose and timing did not moderate the effect in either phase.
The authors conclude that caffeine is broadly ergogenic in women and that phase- or contraceptive-specific rules are not supported, and call the subgroup patterns hypothesis-generating given the low to very low certainty.
Participants
20 trials, 272 women; 10 to 26 per trial; mean ages 17.3 to 31 years (two taekwondo samples averaged 17.3-17.7 years although the stated inclusion was 18 or older).
Training: 17 trials (k = 136) trained women, 3 (k = 8) untrained or recreationally active; by the review's labels, resistance- or strength-trained women in 7 trials, team-sport players, taekwondo athletes, well-trained triathletes, endurance-trained cyclists and triathletes, and physically active women.
Hormonal state: early follicular 10 trials (k = 44), late follicular or peri-ovulatory 4 (k = 14), luteal or mid-luteal 9 (k = 55), hormonal or oral contraceptive users 5 (k = 31).
Phase verification: none or cycle counting in 15 trials (k = 108), hormonal (LH test, basal temperature or blood) in 5 (k = 36).
Habitual caffeine: low (0-150 mg/day) in 14 trials, moderate (150-300) in 5, high (over 300) in 1.
Our reading of Table 1: three entries (two by Lara and one by Romero-Moraleda) describe 13 trained women of identical age and body mass (the two Lara entries label them well-trained triathletes), probably one cohort; the review does not say how it handled this overlap.
Protocol
Search: started in March 2026 with protocol development, updated 1 May 2026; 1,366 records, 890 screened, 119 full texts, 20 included (99 excluded, among them mixed-sex data without a female split and no phase information).
Inclusion: healthy women, naturally menstruating or using hormonal contraception; one caffeine dose alone (capsule, drink or solution; mouth rinse excluded); placebo or control; objective performance outcomes; randomised trials in English; phase or contraceptive status classifiable.
Caffeine: 3 mg/kg in 12 trials (k = 81), 4 mg/kg in 3 (k = 14), 5 mg/kg or more in 7 (k = 49); capsules in 13, oral supplements in 3, drinks in 3, coffee in 1; 30-90 min before, 60 min in 16 trials.
Outcomes: neuromuscular or maximal output (strength, jump, sprint and agility, Wingate power, bar velocity, maximal contraction; 14 trials, k = 100) and endurance or repeated output (time trials, time to exhaustion, repetitions to failure, total volume, repeated sport tasks; 15 trials, k = 44).
Analysis: Hedges' g, crossover variances with an assumed within-person correlation of 0.5 (0.2 and 0.8 in sensitivity); three-level REML model; prediction intervals; hormonal state as a four-level moderator (omnibus test); exploratory moderators within the early follicular and luteal strata only (dose, task type, verification, timing, habitual intake); leave-one-out and outlier exclusion; Egger's test where 10 or more effect sizes.
Risk of bias and certainty: RoB 2 (crossover version with the period and carryover domain) and a modified PEDro scale (0-12); GRADE by one reviewer, checked by a second.
Funding and interests: no financial support; no conflicts declared; the authors used an image generator for the graphical abstract and icons only.
Key Results
| Analysis | k (trials) | Hedges' g (95% CI) |
|---|---|---|
| Overall | 144 (20) | 0.37 (0.24 to 0.50), p < 0.001; prediction interval −0.13 to 0.87; I² 40.4%; GRADE low |
| Early follicular | 44 (10) | 0.42 (0.25 to 0.59) |
| Late follicular / peri-ovulatory | 14 (4) | 0.48 (0.26 to 0.71) |
| Luteal / mid-luteal | 55 (9) | 0.35 (0.18 to 0.53) |
| Hormonal or oral contraceptive users | 31 (5) | 0.28 (0.00 to 0.55), p = 0.050 |
| Difference between the four groups | — | F(3,140) = 0.70, p = 0.551 |
Exploratory, within a phase:
| Moderator | Early follicular | Luteal / mid-luteal |
|---|---|---|
| Dose 3 mg/kg | 0.45 (0.04 to 0.85) | 0.21 (0.07 to 0.36) |
| Dose 4 mg/kg | 0.54 (0.02 to 1.07) | — |
| Dose 5 mg/kg or more | 0.65 (0.01 to 1.30) | 0.31 (0.10 to 0.52) |
| Dose as a moderator | p = 0.854 | p = 0.398 |
| Endurance / repeated output vs maximal output | 0.73 (0.45 to 1.01) vs 0.37 (0.12 to 0.62), p = 0.013 | not different (p = 0.602); maximal output 0.26 (0.11 to 0.41) |
| No hormonal verification vs hormonal verification | 0.68 (0.45 to 0.91) vs 0.07 (−0.31 to 0.44), p = 0.008 | not different (p = 0.871) |
| 60 min before | 0.46 (0.19 to 0.72) | 0.28 (0.13 to 0.43) |
Timing did not moderate the effect in either phase; the late-follicular and contraceptive strata were too sparse for these analyses.
Sensitivity: with correlations of 0.2 or 0.8, outlier exclusion and leave-one-out the overall effect stayed positive and significant and the hormonal groups still did not differ; the late-follicular estimate became borderline after removing one trial; in the early follicular phase the task-type contrast held after outlier exclusion and the verification contrast fell to borderline.
Risk of bias: RoB 2, no trial low overall, 1 high (selective reporting), 19 some concerns, concerns mostly in randomisation, period and carryover, deviations and selective reporting; modified PEDro mean 7.9 of 12.
Small-study effects: Egger's test significant in 17 of 27 analyses, the overall model, the early follicular and the luteal strata among them; read by the authors as possible small-study effects.
Individual cycling trials in Table 1: 11 endurance-trained cyclists and triathletes on the pill, 3 mg/kg 90 min before a time trial of about an hour: 3757 ± 312 s vs 3863 ± 419 s on placebo (significant); 19 highly active women on the pill, 3 mg/kg as coffee, 5-km time trial: 516 ± 42 vs 525 ± 47 s (significant); 21 exercising women in a parallel trial, 6 mg/kg, 5-km time trial: not significant.
Practical Application
Caffeine gave women a small gain, and the cycle is not a reason to time or skip it: the pooled gain was small (g 0.37), positive in every phase and in hormonal-contraceptive users, and the phases did not differ. The usual approach applies: 3 mg/kg about 60 min before was the most common protocol here, tried in training first.
No phase-specific dose or timing: higher doses did not add more within the follicular or luteal phase in this pool, and timing did not moderate it (doses were taken 30-90 min before).
This agrees with [SYNTH-37] (the phase is not the lever) and fills a gap SYNTH-67 names (women thin in the supplement evidence). The contraceptive estimate was the smallest (g 0.28) and only just significant (95% CI 0.00 to 0.55), but not different from the others; the review does not split pill types.
Inference, not tested here: the exploratory hint that endurance and repeated efforts gain more than maximal efforts in the early follicular phase comes mostly from trials that did not verify phase hormonally (our reading of Table 1; hormonally verified early-follicular trials pooled at g 0.07, 95% CI −0.31 to 0.44), so it is no reason to change an endurance athlete's plan.
Not settled by this design: individual response (the prediction interval runs from −0.13 to 0.87, so some women may not benefit); endurance-specific effects in women, since most outcomes were strength and power; and side effects or sleep, which were not measured.
Limitations
No trial at low overall risk of bias; certainty low for the main effect and low to very low for subgroups; funnel asymmetry in most tested models, so the pooled size may be inflated by small trials.
Small trials (10 to 26 women); sparse late-follicular (4 trials) and contraceptive (5 trials) strata; contraceptive users pooled across methods and formulations.
Phase verified hormonally in only 5 of 20 trials; the larger effects in unverified early-follicular trials may reflect misclassified phases (the authors' reading).
Outcomes are mostly strength, jump and sprint; few endurance trials, and task types are broad groupings.
No hormones, caffeine levels, genotype, side effects or sleep measured in most trials.
Probable overlap of one triathlete cohort across three included reports (our reading); two included samples were under the stated minimum age of 18.
The search began alongside protocol development in March 2026 and the protocol was registered on 16 March 2026, so registration did not clearly precede the search.
Supplementary appendices not read.
How this entry was checked and changed
Why this level
a meta-analysis of randomised trials whose risk of bias we read in the full text, capped at A- because that assessment is serious: no trial at low overall risk (19 with some concerns, 1 high), GRADE low for the main effect and low to very low elsewhere, and funnel asymmetry in 17 of 27 Egger tests, the overall model included