Creatine for female athletes: 34 trials, small effects

Creatine is the most studied ergogenic ingredient on the market, and most of that literature was generated in mixed-sex or male-dominant trials. Brands launching female-focused products have had to borrow those numbers. A multilevel meta-analysis published in Frontiers in Nutrition in September 2026 asked what the female-specific evidence, on its own, supports.

The answer is narrower than the category's marketing suggests. Thirty years of trials, pooled with a model that keeps every eligible outcome instead of one endpoint per study, give a small average effect and low certainty. For anyone writing a product specification or a claim, that gap between a statistically positive estimate and a usable recommendation is the part worth reading.

What the review actually pooled

Lin, Zhu and Hong searched PubMed, Scopus, Web of Science, SPORTDiscus and CINAHL for studies published between 1 January 1996 and 5 March 2026. The search returned 13,244 records. After removing 4,481 duplicates and screening 8,763 titles and abstracts, 160 reports were assessed in full text and 34 studies entered the systematic review, covering 692 female or female-eligible participants in swimming, soccer, volleyball, futsal, dance, wrestling, resistance training and laboratory exercise.

Only 21 of those 34 studies provided female-specific, model-compatible data. They contributed 170 effect sizes from 424 analysable participants to the primary multilevel model: 68 performance effects, 62 body-composition effects, 33 physiological effects and 7 cognitive or reaction-time effects. The remaining 13 studies reported mixed-sex results without female-specific estimates, or crossover data without the within-participant correlations needed to pool them. That is the first finding, and it is a reporting problem rather than a biological one.

The numbers

The overall model returned Hedges g = 0.23, with a CR2 robust 95% confidence interval of 0.09 to 0.37 (p = 0.003). Domain estimates were more informative. Exercise and sport performance gave g = 0.30 (95% CI 0.06 to 0.54) across 18 studies and 68 effects. Body composition gave g = 0.24 (95% CI 0.06 to 0.43) across 18 studies and 62 effects. Physiological outcomes — biomarkers, cardiorespiratory and neuromuscular responses — gave g = 0.10 with an interval spanning zero (−0.17 to 0.38, p = 0.361).

The estimate is stable but not strong. Leave-one-study-out analyses ranged from 0.20 to 0.25 with every lower confidence limit above zero, and imposing within-study sampling correlations of 0.30, 0.50 and 0.80 moved the estimate to 0.23, 0.23 and 0.22. Heterogeneity was substantial at study level (I² = 70.95%), and 72% of the estimated random-effect variance sat between studies rather than between outcomes within a study — protocols and populations differ more than endpoints do.

Duration was analysed descriptively, without a between-stratum test. Protocols of seven days or less gave g = 0.20 (95% CI −0.13 to 0.53); longer protocols gave g = 0.25 (95% CI 0.08 to 0.43). The intervals overlap heavily. The data do not isolate duration as the explanation.

How it compares with the wider creatine literature

Placed next to syntheses that pool mixed-sex evidence, the female-specific estimate is roughly half the size — with far lower certainty attached.

The pattern is consistent rather than contradictory. Creatine's effect concentrates where phosphocreatine availability limits the task — short, repeated, high-intensity work — and thins out elsewhere. What the female-specific literature adds is uncertainty, not a different mechanism. The one female population that already has its own pooled dataset behaves the same way: a 2026 meta-analysis of seven trials in postmenopausal women found a lean-mass gain of +0.37 kg, and only at 5 g/day alongside resistance training. Effect sizes from Wang and Bai are quoted as published in their indexed abstracts, which report point estimates and p-values without confidence intervals.

What it means for a formulator

1. There is no female-specific dose to engineer around. The pooled trials used everything from 2 g/day for six weeks to 5 g four or five times daily for less than a week, plus body-weight protocols of 0.075 to 0.3 g/kg. Those regimens produced overlapping estimates. A product built on a proprietary ramp or a "women's dosing protocol" is claiming a resolution the evidence does not have. The defensible design target is the intake pattern the EU already recognises — 3 g of creatine per day — and then the engineering problem becomes delivering that amount at the end of shelf life, not at the time of blending. How much of the declared creatine is still creatine by that date is a purity question, and the distinction it turns on is set out in creatine and serum creatinine: what the 2026 evidence shows — the degradation product on a certificate of analysis is not the blood analyte that shares its name.

2. Form and purity belong in the specification, not the copy. Five of the 34 reviewed studies never state which creatine they used; two used creatine citrate, one di-creatine citrate, one monohydrate gummies. If a literature base this size cannot say what material it tested, a purchase order that says "creatine, 99%" says even less. What closes the gap is a specification that fixes the form, names the assay method — HPLC that separates creatine from creatinine, rather than a total-nitrogen proxy — and sets per-lot acceptance limits for creatinine and process-related impurities. That is what a batch certificate of analysis for CREAPRIME™ is for: the specification travels with the material instead of being re-typed by an intermediary.

3. Body-composition endpoints are measurement-sensitive. The authors are explicit that short-term fat-free-mass increases may partly reflect total-body-water shifts rather than contractile tissue, and that measurement method should be reported per endpoint. A body-composition promise resting on a five-day DXA result is, in part, measuring water. For a brand designing a female-focused SKU — creatine alone or creatine alongside a protein system, where a 2026 network meta-analysis of protein blends versus whey puts whey first for lean body mass and the blend node fifth of seven — the position that survives scrutiny is the authorised performance claim plus a verifiable specification. If you want the specification and analytical methods reviewed against your own formulation before committing to a palletised order, our technical team can work through the documentation with you.

Where the evidence stops

The review is candid about its own limits, and they are real. No included study met low-risk criteria across all risk-of-bias domains. Of the 33 randomised or allocated studies, the randomisation process was rated "some concerns" in 32, missing outcome data in 27, and selection of the reported result in all 33 — protocols and trial registrations were usually unavailable. One study was explicitly non-randomised.

Certainty was rated very low for the overall and physiological estimates and low for performance and body composition. Residual heterogeneity was concentrated in the performance domain (Q = 129.78, df = 67, p < 0.001). Publication-bias assessment was exploratory and underpowered at 21 study-level aggregates (regression z = 0.67, p = 0.505). The review was registered on OSF after the primary synthesis was completed, which the authors note does not provide the safeguards of prospective registration; the extraction files, decision rules and analysis code were archived publicly. Trials rarely recorded baseline muscle creatine, diet, menstrual phase or contraceptive use, so none of those factors can be identified here as modifiers of response. The full text was retrieved from PubMed Central and carries no funding or conflict-of-interest declaration in the article body.

Cognitive and reaction-time outcomes came from two studies and seven effects. They were retained in the overall model but not fitted separately, and the authors state plainly that they do not support cognitive claims. That is the correct reading of thin data, and it is the one we follow.