What the 2024 Meta-Analysis Reported—and What It Did Not
The 2024 systematic review and meta-analysis is the main human evidence source in this series. It asks whether creatine monohydrate changes cognitive performance in adults, but it does not test one single ability or one single kind of user. To read the paper accurately, we need to keep three things together: who was included, how different tests were converted into comparable estimates, and which cognitive domains showed a signal rather than a null result.
The evidence base was narrower than the headline
The review included randomized controlled trials of adults receiving creatine monohydrate as an intervention, either alone or alongside another intervention, compared with placebo or a control condition. The participants were at least 18 years old. Most of the studies involved healthy people; a smaller group involved participants with specific conditions, including fibromyalgia, mild cognitive impairment associated with Parkinson’s disease, and chronic schizophrenia under treatment. That mix makes the review broader than a study of one narrow population, but it also makes “the average adult” a poor description of any individual participant.
The review reported 16 included studies with 492 participants overall. Ages ranged from 20.8 to 76.4 years. Thirteen studies focused on healthy individuals, while three targeted participants with the conditions just described. The intervention periods varied widely. Some were acute or lasted seven days; others lasted four to six weeks, several months, or as long as 18 months. The paper sometimes grouped interventions as shorter than four weeks or at least four weeks, but that grouping is a comparison made for analysis, not a validated rule about how long someone should take creatine.
The intervention itself was more consistent than the population. All included studies used creatine monohydrate. Doses varied across trials, and the review did not establish a single optimal dose for cognitive outcomes. That distinction is important because a pooled result can answer a question about the collection of tested regimens without identifying the regimen that would work best for a particular person. The evidence also does not automatically transfer to another form of creatine or to a commercial combination product.
The review used several conventional tools to combine and evaluate the trials. It expressed continuous outcomes as standardized mean differences, or SMDs, with 95% confidence intervals. It also calculated Hedges’ g, a small-sample correction related to the standardized mean difference. When tests use different scales—one might count recalled items while another uses a reaction-time measure—a standardized effect can put their differences on a common statistical scale. That makes pooling possible, but it does not make the underlying tasks identical.
A standardized mean difference describes separation between the intervention and control groups in units of pooled variability. It is not a percentage improvement and does not directly tell us how many minutes, words, or correct responses changed for a typical participant. A value around 0.3, for example, says that the groups were separated by roughly three-tenths of a standard deviation under the model used. Whether that difference is practically important depends on the task, the person, and the consequence attached to the result. The number becomes especially difficult to interpret when a domain contains several different tests.
The confidence interval adds another essential piece. It represents the uncertainty around the estimated effect under the analysis model. If an interval crosses zero, the data are compatible with no difference as well as with effects in either direction at the chosen confidence level. If it stays on one side of zero, the pooled estimate is conventionally called statistically significant. That label does not prove a mechanism, establish a clinically meaningful threshold, or show that the result will repeat in every population. It tells us how the estimate relates to a particular uncertainty calculation.
The review also assessed risk of bias with the Cochrane RoB 2.0 framework, examined heterogeneity with I², used fixed- or random-effects models according to its stated rules, and conducted subgroup and sensitivity analyses. Where enough studies were available, it used Egger’s test to look for publication bias. It reported low-to-moderate risk of bias, no statistically significant publication bias, and sensitivity analyses that did not materially change the overall conclusions. These are relevant checks, but each answers a different question. None by itself guarantees that all observations entered the model at the right level of independence.
One warning is visible from the paper’s own counts. The review describes 16 studies and 492 participants overall, yet the memory analysis is described as involving 24 studies and 1,000 participants. The intervention table also lists multiple memory subtests within individual trials. A reader should not silently decide that one of these figures is the “real” sample size. The difference may reflect how outcomes were organized, but it means that study count, participant count, outcome count, and independent effect-size count need to be examined separately. This is the point at which a reported result becomes an evidence object that must be audited, not just a number to repeat.
The review’s analysis path can be summarized as follows:
This path shows both the value and the limit of meta-analysis. It can organize a scattered set of trials and make patterns easier to see. It cannot remove differences in participants, tasks, durations, or study quality. Nor can a statistical procedure transform a family of tests into one universal measure of mental ability. The pooled estimate is a summary of the evidence that was selected and combined, not a replacement for the evidence’s structure.
What the domains actually showed
The clearest positive result in the 2024 paper was memory. The review reported a standardized mean difference of 0.31, with a 95% confidence interval from 0.18 to 0.44, and a Hedges’ g of 0.3003, with a 95% confidence interval from 0.1778 to 0.4228. The authors interpreted this as a statistically significant improvement in memory and rated the certainty of evidence as moderate. In plain language, the creatine group’s pooled performance was higher than the control group’s on the collection of memory outcomes included in that analysis.
That sentence needs two qualifications immediately. First, the effect is an average standardized difference, not a promise that every participant will remember more. Some participants may show little change, and the estimate does not identify a single memory task that should be expected to improve. Second, the number of reported memory observations exceeds the number of unique randomized participants in the paper’s overall description. The positive estimate is therefore a result to inspect carefully, especially because the later commentary challenges the independence assumption behind the pooling.
Attention produced a split pattern. The review did not find a statistically significant effect on attention scores. It did report a significant effect on attention time: the standardized mean difference was −0.31, with a 95% confidence interval from −0.58 to −0.03, and Hedges’ g was −0.3004, with a 95% confidence interval from −0.5719 to −0.0289. The negative direction is meaningful because lower time represents faster completion in these outcomes. The result therefore concerns how quickly participants completed certain attention tasks, not a general increase in attentional capacity.
A faster time can be useful, but speed is not the only property of attention. A participant who responds more quickly but makes more errors has not necessarily improved in the way a real-world task requires. The review’s separate treatment of scores and times is therefore valuable: it prevents a time-based result from being restated as if accuracy, sustained focus, and speed all changed together. It also shows why a domain label should never replace the outcome definition.
Processing speed showed the same contrast. For processing-speed scores, the pooled result was not statistically significant. For processing-speed time, the review reported a significant standardized mean difference of −0.51, with a 95% confidence interval from −1.01 to −0.01, and a Hedges’ g of −0.4916, with a 95% confidence interval from −0.7852 to −0.1980. Here again, the negative sign corresponds to less time to complete the task. The estimate is larger in standardized terms than the reported memory and attention-time effects, but its interval is wide and the review rated the certainty of processing-speed evidence as low.
The distinction between speed and score changes the possible interpretation. Processing-speed time may capture rapid information handling under the conditions of a test. It does not establish that participants understood more difficult material, made better choices, or sustained performance over a working day. A pooled result can be statistically real within a task family while still being too narrow to support a broad claim about cognition.
The review did not find a statistically significant improvement in overall cognitive function. It also found no significant effect for executive-function scores or executive-function time. Those null results matter because they resist the temptation to treat the positive domains as evidence of a uniform cognitive gain. If creatine’s effects were broad and consistent across the measured functions, we would expect a clearer pattern across overall cognition and executive outcomes. The absence of significance does not prove that no effect exists, particularly with small and heterogeneous studies, but it limits what the review can claim.
The paper similarly reported no significant effect on attention scores or processing-speed scores, despite the time-based signals. This is not a contradiction that needs to be smoothed away. It is a description of a mixed evidence pattern. The most accurate summary is that the review found positive pooled results for memory and selected completion-time outcomes, while several broader or score-based domains remained statistically uncertain or null.
Subgroup analyses made the pattern more conditional. For attention time, the review reported a significant effect among participants with illnesses and among adults aged 18–60, but not among healthy participants or adults over 60. For processing-speed time, the reported effect was significant among female participants but not male participants. The study also found no significant difference between shorter and longer intervention groups in the attention-time analysis. These patterns may suggest that health status, age, or sex modifies response, but they were based on limited study numbers and should not be turned into personalized rules.
The review’s GRADE judgments reinforce that caution. Memory received moderate certainty, while processing speed, overall cognitive function, executive function, and attention ranged from low to very low certainty. The paper’s sensitivity analyses and publication-bias tests support the internal consistency of the reported analysis under its chosen methods. They do not answer the later question of whether multiple correlated outcomes were treated as independent. Methodological reassurance is not all-or-nothing; one set of checks can be reassuring while a different assumption remains unresolved.
Taken together, the 2024 review reported a specific pattern rather than a simple verdict. It provides a plausible and statistically positive memory estimate, along with positive results for the time needed to complete some attention and processing-speed tasks. It also provides nonsignificant results for overall cognition, executive function, attention scores, and processing-speed scores, plus lower certainty for most domains. Before deciding how much weight to give the positive memory result, the next question is not merely whether the p-value crossed a threshold. It is whether the observations being pooled were independent enough for the confidence interval and significance label to mean what readers assume they mean.