Exercise May Ease Student Anxiety and Depression, but Evidence Stays Weak

A three-level meta-analysis of 43 randomized controlled trials involving 3,199 college students found no statistically significant overall effect of physical exercise on depression, anxiety and stress when all eligible results were included. The pooled estimate favored exercise, but its confidence interval crossed zero and the prediction interval ranged from substantial benefit to substantial harm, indicating that the average result was highly uncertain. The review, published in Frontiers in Public Health, used a model designed to account for multiple dependent outcomes reported within the same study, a technical feature that materially changed the result. The authors concluded that no definitive conclusion regarding efficacy can be drawn.

The primary analysis combined 118 effect sizes in a three-level random-effects model. Hedges’ g was −1.48, with a 95% confidence interval of −3.21 to 0.26 and a p-value of 0.094; the 95% prediction interval extended from −12.95 to 9.99. Because negative values in this analysis represented improvement in negative emotional symptoms, the direction favored exercise, but the evidence did not support a statistically reliable overall effect. By comparison, a conventional two-level model produced a significant estimate of −1.17, with a 95% confidence interval of −1.97 to −0.37. The discrepancy showed that the conclusion depended partly on how statistical dependence among outcomes was handled, since many trials reported several scales or symptom dimensions simultaneously.

Influence diagnostics identified two effect sizes from a single study as both statistically extreme and unusually influential. Their removal changed the three-level estimate to Hedges’ g of −0.56, with a 95% confidence interval of −0.73 to −0.39 and p less than 0.001, suggesting a potential benefit from exercise. However, the corresponding 95% prediction interval still ranged from −1.45 to 0.33, meaning that the estimated effect in a new comparable setting could include no effect. The researchers retained the full-dataset result as the primary overall analysis and treated the adjusted result as exploratory because it depended on removing observations from one study. They did not exclude the values merely because they were large, but because objective indices including Cook’s distance, hat values and leave-one-out diagnostics identified their disproportionate influence on the pooled estimate, confidence interval, prediction interval and heterogeneity structure.

Substantial variation remained among the trials even after the two influential effect sizes were removed. Cochran’s Q was 183.88 with 115 degrees of freedom and p less than 0.001. Variance decomposition attributed 58.21% of total variance to sampling variance, 2.18% to variation among outcomes within studies and 41.79% to variation between studies. A likelihood-ratio test confirmed statistically significant between-study heterogeneity, while within-study variation was not significant. This combination explained why retaining all eligible outcomes preserved more information but also made the pooled average unstable. Robust variance estimation with CR2 adjustment produced almost the same estimate as the primary model, and sensitivity analyses using assumed pre–post correlations of 0.30 and 0.70 also preserved the full dataset’s non-significant conclusion.

The 43 trials were conducted in 16 countries, with 14 studies from China and six from Türkiye. Individual samples ranged from 17 to 206 participants, and reported mean ages fell between 18.41 and 26.06 years where available. Study populations included general students as well as groups with depressive symptoms, anxiety symptoms, psychological disorders, fatigue, mobile phone addiction, or overweight and obesity. Interventions ranged from Pilates, dance training, aerobic exercise, Tai Chi, Baduanjin, yoga, resistance training and high-intensity interval training to ball games and exergaming. Most sessions lasted 30 to 90 minutes, participants commonly exercised two or three times per week, and interventions generally lasted four to 16 weeks. Exercise intensity was inconsistently reported, and control conditions were usually wait-list or usual daily life. Outcome measures included the Beck Depression Inventory, the State–Trait Anxiety Inventory, the Perceived Stress Scale, the DASS-21 and DASS-42, Zung self-rating scales, the HADS, the SCL-90 and the PHQ-9.

Exploratory analyses of the influence-adjusted dataset identified three significant moderators. Baseline population category, geographic region and the family of outcome instrument all appeared related to the estimated effect. Effects were larger among students already experiencing negative emotional symptoms, whose subgroup estimate was −0.96, than among general student populations (−0.47) or other special populations (−0.15). Effects also differed by region, with the estimate significantly smaller in Europe than in Asia, and by instrument, with smaller estimates for the DASS family and other measures than for the Beck Depression Inventory family. These findings were not reproduced in the corresponding full-dataset analyses, however, and they should not be interpreted as causal effects of geography, population status or measurement method. Differences may instead reflect sample composition, intervention implementation, cultural context or psychometric properties.

No significant moderating effect was detected for symptom category, exercise modality, weekly frequency, session duration or intervention duration in the influence-adjusted analyses. That pattern does not establish that these features are interchangeable or unimportant. Some subgroups were small and imprecise: resistance training (g = −0.45) and multicomponent exercise (g = −0.35) did not reach statistical significance, nor did the category of four to five sessions per week (g = −0.31) or sessions of 10 to 30 minutes (g = −0.43). Duration and frequency also captured only part of an exercise prescription, while intensity, progression and adherence were not consistently documented across the primary trials. Overall, the analysis could not determine whether aerobic, resistance, mind–body or multicomponent programs differed, nor whether shorter or longer sessions were superior.

Meta-regression also found no significant linear or nonlinear dose–response relationship for weekly or total exercise duration. For weekly duration, the linear coefficient was 0.0001, with a standard error of 0.0012 and p of 0.956; the nonlinear natural spline model was likewise non-significant, with p of 0.460. Total duration produced similarly null results in linear and nonlinear models, with p values of 0.945 and 0.862. The conclusion is not that dose is irrelevant, but that the available measurements did not reveal a stable pattern. Weekly and total durations were derived from session duration, frequency and intervention length, so incomplete reporting in the original studies may have reduced the precision of these analyses, and accumulated exercise time alone does not capture intensity, progression or adherence.

Methodological weaknesses further reduced confidence in the findings. Using the Risk of Bias 2.0 tool, the reviewers rated 38 of the 43 trials as high risk of bias and the remaining five as having some concerns; none was rated low risk. Concerns centered on deviations from intended interventions, which blinding rarely prevents in exercise research, and on subjective outcome measurement, where awareness of group allocation could influence self-reported responses. Both conventional and multilevel Egger tests suggested possible funnel-plot asymmetry or small-study effects, although trim-and-fill found no missing effect sizes to impute and did not materially change the pooled estimate. GRADE accordingly rated the overall certainty of evidence as very low, reflecting downgrades for risk of bias, inconsistency, imprecision and suspected publication bias. The review was also restricted to English-language publications, potentially excluding relevant studies and introducing language bias, and change-score standard deviations were estimated using an assumed pre–post correlation of 0.50 in the primary analysis.

The search covered Web of Science, PubMed, Embase, the Cochrane Library and Scopus from inception to March 22, 2026, identifying 13,570 records from which the 43 eligible trials were selected, and the protocol was prospectively registered on PROSPERO. The authors note that exercise remains an attractive non-pharmacological option on campuses because it is accessible, low-cost and easy to integrate into daily life, and that plausible mechanisms include modulation of neurotransmitter activity, regulation of the hypothalamic–pituitary–adrenal axis, improved sleep, enhanced self-efficacy and greater positive affect, though these mechanisms were not directly tested in the included trials. Given the very low certainty of evidence, the discrepancy between the full-dataset and influence-adjusted estimates, and a prediction interval crossing the null, the findings should be regarded as suggestive rather than confirmatory. The authors call for rigorously designed randomized trials with complete reporting of randomization, attrition and exercise prescription, more consistent outcome measurement, objective indicators and longer follow-up to clarify heterogeneity, dose–response relationships and underlying mechanisms.

Subject of Research: Effects of physical exercise on depression, anxiety and stress symptoms in college students

Article Title: Intervention effects of physical exercise on negative emotional symptoms in college students: a three-level meta-analysis

Article References: Wang, Z., Dong, L., Rao, Y., Wen, H., Liu, C., & Huang, M. (2026). Intervention effects of physical exercise on negative emotional symptoms in college students: a three-level meta-analysis. Frontiers in Public Health, 14, Article 1911631. https://doi.org/10.3389/fpubh.2026.1911631

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1911631

Keywords: physical exercise, college students, depression, anxiety, stress, meta-analysis, randomized controlled trials, mental health, Hedges' g, dose-response, risk of bias, GRADE