REF-1694 · Hurst et al., 2026 · Journal of the International Society of Sports Nutrition, · Paper · PubMed
A placebo that participants were told was caffeine improved sport performance against their own baseline by d = 0.49 (95% CI 0.28 to 0.69) across 17 studies and 257 healthy adults, 82% men; no study was at low risk of bias (9 high, 8 unclear) and GRADE certainty was low, and because every study compared the placebo with a baseline trial, not with an untreated control, expectancy is not separated from order or learning effects. Read in full
Main caveat: The authors': over half the studies at high risk of bias, with poor allocation concealment and about half not blinding researchers.
Hurst and colleagues pooled studies in which researchers gave healthy adults an inert substance, told them it was caffeine, and compared their sport performance with a baseline or no-treatment condition. From 2,799 records they kept 17 studies published 2006-2026 (one report held two samples), all of which in the end used a baseline condition.
They took one primary outcome per study, such as a time trial, repetitions to failure or mean power, and pooled Cohen's d of placebo against baseline.
Performance improved with the placebo by a small-to-moderate amount (d 0.49, 95% CI 0.28 to 0.69) with low heterogeneity (I² 21%), no sign of publication bias, a prediction interval of 0.01 to 0.96, and leave-one-out estimates of 0.43 to 0.53.
In one-at-a-time meta-regressions the effect was larger when the researchers did not know the substance was a placebo and smaller in samples with more women.
Among secondary outcomes, time to exhaustion (2 studies) and total weight lifted (3 studies) improved most and RPE moderately, while power output, velocity, heart rate, lactate and oxygen uptake did not differ significantly. Risk of bias was poor: nine studies high and eight unclear, with no study at low risk.
The authors rate the certainty low and recommend that practitioners build a positive expectation when they give real caffeine, since deceiving athletes with placebos is unethical.
Participants
257 participants in 17 studies; 210 men (82%); 12 studies men only, 1 women only, 4 mixed (women 32-47% of those samples).
Median age 23.0 years (study means 19.7 to 43.0).
Samples (Table 1): well-trained or competitive cyclists (6 and 14), healthy male volunteers (two samples of 22), resistance-trained men (12), healthy adults (30), physically active people (9, 18, 25), Paralympic weightlifters (4), well-trained middle-distance runners (11), soccer players (8), resistance-trained women (13), healthy runners (22), trained adults (15), recreationally trained long-distance runners (13), trained sprinters (13).
Classified by the review as trained (12 studies) or non-trained (5).
Protocol
Search: Ovid MEDLINE, EMBASE, Web of Science and CINAHL from their earliest records to July 2026, reference lists, and contact with researchers; no grey literature; English, peer-reviewed only.
Inclusion: healthy adults 18 or over; a placebo deceptively described as caffeine; a baseline or no-treatment control (required, since placebo effects cannot be inferred without one); a sport-performance outcome; randomised, quasi-randomised and non-randomised, crossover and parallel designs.
Interventions: capsules (11 studies), decaffeinated coffee (3), saline (2), chewing gum (1); given more than 20 min before testing in 11, 1-20 min in 3, under 1 min in 3; the stated "dose" was low (3.0 mg/kg or less) in 4, moderate (4.0-6.0 mg/kg) in 5, large (over 6.0 mg/kg) in 4, not given in 4.Expectation was built with scientific evidence (10 studies), a verbal promise of better performance (10), anecdotes (3), a presentation or video (2 each) and, in one study, a secretly improved earlier result; median 2 such methods per study (0 to 5).
Designs: 14 crossovers (10 pre-post, 4 balanced-placebo factorial), 3 parallel pre-post; 13 randomised with a familiarisation trial; 5 also gave real caffeine.
Analysis: random-effects meta-analysis (Comprehensive Meta-Analysis v4) of one primary outcome per study, Cohen's d (placebo minus baseline over the pooled SD); for studies with several stated doses or intensities the highest was used; I² and Q; leave-one-out and Egger's test only with 10 or more studies; fail-safe N; random-effects meta-regressions run one moderator at a time.
Risk of bias: Cochrane tool, two reviewers; a study was "low quality" if any domain was high risk, "unclear" if any was unclear.
Certainty: GRADE for the primary analysis (two raters).
Protocol changes: three people instead of two chose the outcomes; effects were computed against baseline instead of against a control because most studies were crossovers.
Key Results
Primary outcome (17 studies, 257 participants):
| Measure | Result |
|---|---|
| Pooled effect, placebo described as caffeine vs baseline | d = 0.49 (95% CI 0.28 to 0.69), p < 0.001 |
| Prediction interval | 0.01 to 0.96 |
| Heterogeneity | I² = 21.03%, Q = 19.30, p = 0.21 |
| Leave-one-out | d 0.43 to 0.53 |
| Publication bias | Egger 0.84 ± 1.31, p = 0.53; fail-safe N 107 |
| Moderator: researchers blinded | b = 0.52 ± 0.24 (95% CI 0.01 to 0.99), p = 0.03 — larger effect when researchers did not know |
| Moderator: proportion of women | b = −0.93 ± 0.38 (95% CI −1.68 to −0.19), p = 0.01 — smaller effect with more women |
| Other moderators (exercise type and duration, training status, age, design, dose, timing, placebo type, real caffeine in the study, number of expectation methods, risk-of-bias score) | not significant (details in the supplement, not read) |
Secondary outcomes (positive d = better with placebo):
| Outcome | Studies (n) | d (95% CI) | p |
|---|---|---|---|
| Time to exhaustion | 2 (17) | 1.07 (0.38 to 1.75) | < 0.01 |
| Total weight lifted | 3 (56) | 0.73 (0.21 to 1.25) | 0.01 |
| RPE | 11 (167) | 0.52 (0.26 to 0.78) | < 0.01 |
| Oxygen uptake | 3 (30) | 0.51 (−0.14 to 1.16) | 0.13 |
| Number of repetitions | 5 (84) | 0.48 (0.06 to 0.90) | 0.03 |
| Time to complete | 4 (54) | 0.38 (0.00 to 0.75) | 0.05 |
| Lactate | 3 (29) | 0.23 (−0.30 to 0.75) | 0.40 |
| Heart rate | 6 (64) | 0.21 (−0.13 to 0.55) | 0.23 |
| Power output | 5 (60) | 0.05 (−0.30 to 0.41) | 0.77 |
| Velocity | 4 (60) | −0.15 (−0.51 to 0.20) | 0.51 |
RPE: more expectation methods went with a larger effect (b = 0.21 ± 0.10, 95% CI 0.01 to 0.41, p = 0.04), and the effect was larger in running than weightlifting (b = −0.90, p = 0.02), not significantly different from cycling (p = 0.09).
Risk of bias: 9 of 17 studies high (53%), 8 unclear (47%), none low; allocation concealment high risk in 9 (53%); participants or researchers not blinded in 8 (47%); manipulation check unclear in 13 (76%); randomisation reported in 16 (94%); all outcomes reported in 17.
GRADE: low for the primary outcome; low to very low across all analyses (supplement).
Practical Application
Belief that one has taken caffeine improved performance on its own, by about half a standard deviation (d 0.49) in laboratory tests, so part of what athletes feel from caffeine may be expectation. The evidence is low certainty: no study was at low risk of bias, and each compared the placebo with a baseline trial rather than with an untreated control, so practice and order effects are mixed into the number.
This adds to, and does not change, [SYNTH-67]'s training-trial rule. Inference, not tested here: a with-and-without caffeine trial on matched sessions that the athlete knows about measures caffeine plus expectation together, and an athlete who wants to know whether caffeine itself works has to compare it with a look-alike capsule they cannot tell apart, which most athletes cannot arrange alone.
The authors' practical suggestion (not tested in this review): when real caffeine is taken, a positive expectation of benefit is worth keeping, because deceptive placebos are unethical in practice. Inference, not tested here: telling an athlete that caffeine "will not do much" may cost some of its effect.
Do not read the moderators as rules: the smaller effect with more women rests on five studies with women and 82% men overall, and the larger effect when researchers were blinded comes from one-at-a-time meta-regressions in a set of studies with serious risk of bias.
No outcome here is a race: the pool is mostly strength tasks (9 of 17 studies), with three cycling and five running tests; power output and velocity did not improve with the placebo, while time to exhaustion (2 studies) and lifting volume did.
Limitations
The authors': over half the studies at high risk of bias, with poor allocation concealment and about half not blinding researchers; few women (5 studies, 18% of participants); English, peer-reviewed studies only; six studies gave the placebo less than 20 min before the task, before caffeine would normally peak, which may have made the deception less believable.
Baseline, not control: all studies used a baseline condition and the review computed the placebo-minus-baseline difference; 10 of the 14 crossovers were plain pre-post tests, so learning the test, order and familiarity can all raise the placebo-trial result. Thirteen studies used a familiarisation trial and randomised order, which reduces but does not remove this (our reading).
Mixed outcomes: one primary outcome per study pooled across tasks as different as a 40-km time trial, finger flexion, bench-press velocity and RPE (one study's primary outcome), and the highest stated dose or intensity was taken where there were several.
Small studies: 4 to 30 participants each; secondary pools have 2 to 11 studies, and meta-regressions were run one variable at a time.
Reporting: the results give 14 crossovers, the protocol-change note "13 of the 17".
Funding and interests: none received; no conflicts declared. One included study (Hurst 2019, 1000-m time trial in well-trained runners) carries the first author's surname; whether it is the same author is not stated.
Not read: the supplementary material (search strings, study summaries, risk-of-bias table, moderator results and GRADE tables for the secondary outcomes).
How this entry was checked and changed
Why this level
risk of bias read in full and serious (no study at low risk, 9 of 17 high; GRADE low for the main outcome, low to very low for the rest), which caps a review at A-; held one step lower because every pooled contrast is the placebo trial against a baseline trial in the same people (10 of the 14 crossovers were plain pre-post tests), not against a no-treatment control, so it cannot separate expectancy from learning, order or familiarity with the test