How much do exercise-cognition effect sizes shrink after adjusting for publication bias and small-study effects?

How much do the numbers actually drop?

The clearest answer comes from a meta-meta-analysis — a study that pools the results of 215 separate meta-analyses, covering 2,239 effect-size estimates from randomized controlled trials of exercise and cognition [1]. Before bias adjustment, such syntheses tend to report encouraging average benefits. After adjusting for selective reporting, the average effect for general cognition was a standardized mean difference (SMD) of 0.227, with a 95% credible interval from 0.116 to 0.330 [1]. SMD is a unit-free measure: 0.2 is conventionally a small effect, so this is a real but modest average benefit.

Memory and executive function tell a starker story. Their adjusted averages were SMD = 0.027 and SMD = 0.012, respectively, with credible intervals that include zero [1]. In plain terms, once you correct for the tendency of small positive studies to get published and small null studies to disappear, the average effect on these two domains is barely distinguishable from nothing. The same analysis found wide prediction intervals spanning both negative and positive effects, meaning individual studies or settings can plausibly show harm as well as benefit [1].

Why correcting for bias pulls the average down so far

Publication bias is the tendency for studies with exciting, statistically significant results to be published while null or negative ones languish in file drawers. Small-study effects are a related pattern: small trials often report larger effects than large ones, which can happen because small studies are more likely to be published only when their results look impressive. When a field is riddled with both, the naive average is inflated. The meta-meta-analysis found strong evidence of selective reporting and large between-study heterogeneity — meaning the studies genuinely disagree with each other, not just randomly [1].

The exercise-snack review offers a concrete illustration of the same correction in a narrower setting. It applied trim-and-fill, a standard method that estimates how many missing studies would be needed to make a funnel plot symmetric and adds them back in. The authors concluded that the apparent effect was driven by substantial small-study effects or publication bias [2]. That is a qualitative warning rather than a precise shrinkage figure, but it points the same direction: the headline effect was partly an artifact of which studies got published.

Where do real benefits survive the correction?

Shrinkage is not uniform. The meta-meta-analysis found that specific population-intervention combinations showed more consistent benefits, even though broad claims of generalized cognitive enhancement looked premature [1]. The practical implication is that exercise probably helps cognition in some groups and formats more than others, and the average across everything masks that variation.

A meta-analysis of adults with serious mental illness hints at what those pockets can look like. In 15 chronic-exercise trials with 936 participants, exercise showed large effects on reasoning and problem solving and small effects on executive functioning [3]. Broken down by diagnosis, people with depression showed moderate effects on executive functioning and large effects on processing speed, while people with schizophrenia showed large effects on reasoning and problem solving [3]. These are unadjusted headline effects from a small evidence base, so they should be read as promising signals in specific populations rather than proof that survives every bias correction. Still, they fit the broader pattern: targeted recommendations hold up better than sweeping ones.

What should a reader do with these numbers?

The honest summary is that exercise is not a reliable, general-purpose cognitive enhancer of the size often advertised. The best-corrected estimates put general cognition at a small average benefit and memory and executive function near zero [1]. That doesn't mean exercise is worthless for the brain — it means the average hides real variation, and the strongest claims rest on evidence that shrinks under scrutiny.

Two caveats belong right next to those numbers. First, the adjusted estimates come with wide credible and prediction intervals, so the true average could be somewhat larger or smaller, and individual contexts could differ in direction [1]. Second, several of the narrower reviews assessed publication bias but did not report a corrected effect size, so we can't say precisely how much their estimates would move [4][5]. What we can say is that when a correction is applied, the effect tends to get smaller — sometimes much smaller — and that the most defensible recommendations are the ones tied to a specific population and a specific kind of training.

Sources used in this answer

Related research questions