Creatine and jump performance: 119 trials, fragile effects

A network meta-analysis is the most ambitious thing evidence synthesis attempts. It puts every supplement that has been tested against a placebo on one scale, then estimates the comparisons nobody has run. One published in October 2026 in the Journal of the International Society of Sports Nutrition did this for lower-limb power and jump height across 119 randomised trials and roughly sixty supplement nodes. Creatine is in it. It beats placebo in two of the four outcome networks.

That is the headline, and it is the least interesting part. The useful material is in the sensitivity analyses the authors ran afterwards, and in one line of the dose section: when the model was restricted to trials whose supplement dose could be verified from the published report, creatine’s advantage in countermovement jump fell from 2.26 cm to 0.30 cm and the interval crossed zero. The ingredient did not change. The documentation did. That is a sourcing problem before it is a science problem.

What the review actually did

Four databases were searched to 8 September 2026, returning 36,338 records in the update alone. After screening, 119 randomised trials were included: 106 parallel-group and 13 crossover, published between 1992 and 2026, with intervention durations from 1 to 210 days. Participants were healthy and physically active. Seventy-five trials enrolled only men and 17 only women.

Rather than pooling every jump into one outcome, the authors built four separate Bayesian random-effects networks: peak power (47 trials, 1,464 participants), relative peak power (33 trials, 939), vertical jump (31 trials, 1,002) and countermovement jump (44 trials, 1,226). Risk of bias was assessed with RoB 2 across 156 result-level records — 70 low, 79 some concerns, seven high — and certainty with CINeMA.

One structural feature frames everything that follows. Direct evidence is thin. Of the direct comparisons in the four networks, 49 of 61, 22 of 28, 44 of 48 and 35 of 44 rested on a single trial. Most of the 1,636 pairwise estimates the review produces are therefore indirect, inferred through the placebo node rather than observed.

Where creatine appeared, and by how much

Against placebo, creatine improved relative peak power by 0.36 W·kg⁻¹ (95% credible interval 0.03 to 0.66), supported by ten direct trials, and countermovement jump by 2.26 cm (0.46 to 3.94), supported by seven. In the peak power and vertical jump networks its interval included zero. The largest estimates in the review belong to other compounds: HMB at 64.70 W (19.15 to 110.50) for peak power and 0.84 W·kg⁻¹ (0.42 to 1.28) for relative peak power, and betaine at 3.72 cm (1.24 to 6.18) for countermovement jump.

Credible intervals describe the average effect. Prediction intervals describe what a new trial under similar conditions might find, and they are wider. For creatine in relative peak power the prediction interval ran from −0.29 to 0.97 W·kg⁻¹, and in countermovement jump it also included zero. Only HMB in relative peak power, betaine and palmitoylethanolamide kept prediction intervals clear of zero.

The rankings are worse than the estimates. The three highest SUCRA values for peak power went to L-carnitine (0.888), a ketoanalogue node (0.871) and arginine alpha-ketoglutarate (0.862). Each rested on a single trial, and all three had credible intervals that included zero. A high ranking in a sparse network is a statement about how little is known, not about how well something works.

How much of it survives

The authors re-ran each network under a long list of restrictions. The creatine result in countermovement jump did not survive most of them.

All five estimates come from the same review and the same 44-trial network; they differ only in which trials and which extracted values the model was allowed to use. 95% credible intervals throughout.

The pattern repeats in the power network. Restricted to parallel-group trials, creatine in relative peak power was 0.35 W·kg⁻¹ (−0.01 to 0.68). Restricted to trials at or above the median sample size, 0.17 (−0.39 to 0.59). Restricted to trials with a verifiable dose, −0.04 (−0.53 to 0.44). Dropping any one of four individual trials also pushed the interval across zero.

Two further details are worth stating plainly. In countermovement jump the test for small-study effects was statistically significant (coefficient −0.72, 95% CI −1.21 to −0.23; adjusted p = 0.023), meaning less precise trials reported smaller effects — and omitting one trial, a 2026 study of creatine gummies in female beach volleyball players, removed that signal. Separately, a second 2026 review in the same journal assembled 41 soccer-specific creatine studies, declined to pool them, and reported mixed vertical jump findings at very low certainty. Two independent syntheses, the same verdict: suggestive, not established.

What it means for a formulator

1. A dose that cannot be verified is not a dose. The review built 101 comparisons between dose nodes of the same supplement. Twelve had a direct trial behind them, and exactly one excluded zero. Restricting any network to trials with verifiable doses collapsed creatine’s advantage in both outcomes. If the published literature cannot say what was given, a purchase specification is the only place where that question gets answered. That means an assay method that separates creatine from creatinine, a numerical acceptance limit per batch rather than a bare “99.9% purity”, and a certificate of analysis issued for the batch being shipped — which is how CREAPRIME™ is specified.

2. The format is part of the dose. The single trial carrying most of the countermovement jump result tested creatine gummies; removing it drops the estimate to 0.61 cm. Gummy and ready-to-drink matrices combine low pH, high water activity and process heat, and creatine converts to creatinine under those conditions. A label dose is therefore a statement about the end of shelf life, not about the moment of packing, and the specification has to say which of the two the assay refers to. Overage is calculated against the matrix, not against a generic table.

3. A multi-ingredient estimate describes a product, not an ingredient. The multi-ingredient node in the jump network returned 1.79 cm (0.33 to 3.24) from three direct trials, one of which tested a post-workout product containing carbohydrate, leucine-enriched whey protein, creatine, HMB and vitamin D₃. Nothing in that estimate can be assigned to any single component. For a brand designing a blend — a LACTOPRIME™ protein system with a creatine fraction, for instance — the consequence is that the evidence travels with the finished formula, while the auditable part travels with each ingredient’s own specification. Those are two different documents and they are not interchangeable. If you want the specification and a batch certificate before you decide anything, ask us for them.

Limits of this evidence

The review is retrospectively registered (PROSPERO CRD420261283835), which weakens the protection against selective outcome reporting. It included only peer-reviewed publications in English, and the authors note that trials whose dispersion data were missing could not enter the synthesis — a data-availability bias they flag themselves. Several nodes rest on one trial, including the largest estimate in the whole review, iron at 6.81 cm for vertical jump, which came from eleven participants.

Between-trial heterogeneity looks small at the centre and large at the edges: the descriptive medians were 3.6%, 14.6%, 7.2% and 11.1% across the four networks, with upper credible bounds reaching 34.1%, 62.1%, 45.0% and 46.7%. The reporting bias domain was rated “some concerns” for every comparison in the review. And the sex imbalance is substantial: 75 male-only trials against 17 female-only, with no detectable sex interaction in any network — which is not the same as evidence of none. The full text obtained from PMC carries no funding or conflict-of-interest statement, so none is reported here.

The honest summary is the one the authors give. Creatine’s direction of effect is consistent with its mechanism and with the wider literature; its magnitude in these networks is small, its certainty is low or very low, and it depends on which trials the model is allowed to see. That is a reasonable place to be with an ingredient. It is not a place from which to write a claim.