What actually works

Caveat on this rating: Small to moderate effects over mostly short trials in clinical populations, with publication bias untestable; the review found no advantage over other active treatments. The published abstract and the authors' manuscript disagree on the participant and citation counts.

Caveat on this rating: A small effect (0.09 SD) in self-selected users paid to deactivate during an election; mood measured by daily texts did not move significantly. The companion Hunt trial is 143 students over three weeks.

Caveat on this rating: Observational cohort with self-reported symptoms dichotomised at a cutoff and imputed missing data; the authors' own hypothesis was that externality would predict psychotic experiences, not depression. The 1988 meta-analysis it sits beside could not be read and is cited without figures.

Caveat on this rating: Short laboratory experiments on momentary feeling by the originators of the theory being tested; sample sizes and funding not read. The effect includes lowered positive excitement, and only chosen solitude produced relaxation.

Caveat on this rating: Observational experience sampling in a self-selected 2010 iPhone sample; time-lag analyses suggest direction but cannot establish cause. The headline percentages are from Harvard's press release, not from paper text we could read.

Caveat on this rating: Scale-development work in mostly student samples; it distinguishes two traits and says nothing about whether reflection can be taught or whether it reduces distress. Scale details are relayed through a 2011 secondary source because the original full text was not readable.

Caveat on this rating: The psychotherapy effect was measured mostly in people who were also taking antidepressants; the authors call that a serious limitation and lean toward combination in severe OCD. The behavioural-therapy arm is 287 people. Treatment-resistant patients were excluded. The abstract does not separate OCD subtypes.

Caveat on this rating: Fifteen of sixteen trials were in nonclinical samples, so the pooled effect is mostly about mood in the general population. g 0.16 is small. Two errata were published and were not read for this row.

Caveat on this rating: A prevention trial in overweight adults without current major depression; the food arm was a behavioural therapy, not a whole-diet change; so it does not test what the small treatment trials tested. Twenty four percent did not complete. Several authors disclose pharmaceutical fees.

Caveat on this rating: Primary result not significant (HR 0.78, 0.55 to 1.10); the quoted 0.59 is a diabetes subgroup. Older adults at high cardiovascular risk; depression was a secondary outcome. The olive oil arm is not reported in the abstract.

Caveat on this rating: The very large effects are against waiting lists and placebo, the weakest comparators; the authors themselves list methodological problems in the trials. This analysis found no significant gain from adding medication to CBT, which is the opposite lean from the 54-trial network meta-analysis; the disagreement is unresolved.

Caveat on this rating: n=67, 56 with complete data; single blind; 55 of 67 already on therapy or medication; control was social support, not treatment; NNT interval 2.3 to 27.8; remission counts of 10 versus 2. A 2018 correction disclosed an unreported recruitment channel.

Caveat on this rating: Open label with a self-rated primary outcome. The abstract gives completers (72), not the number randomised. Control was befriending, not treatment. Men aged 18 to 25 only.

Caveat on this rating: Observational associations, confounded by everything that goes with eating well. The Mediterranean estimate rests on four longitudinal studies. One author discloses industry support from food companies. Two corrections were published and were not read for this row.

Caveat on this rating: Self-reported depression. 152 recruited, 95 with three-month data, 85 with six-month data. Diet, cooking classes and fish oil in one arm against social groups in the other, for different durations, so the diet cannot be isolated. No effect size in the abstract.

Caveat on this rating: One trial, 122 people, one drug. Thirty-eight percent of people who started the therapy did not respond, and the treated-versus-completer gap shows dropout. The combination had the highest treated response even though the difference from therapy alone was not significant. The therapy was intensive, which most clinics do not deliver.

Caveat on this rating: A survey threshold, not a clinical assessment: the 2.3% is DSM-IV criteria applied by interview. The authors argue the subthreshold quarter carries real burden, so the gap between 25% and 2.3% is not a reassurance and should not be read as one.

Caveat on this rating: Usual care rather than an attention-matched control, single-blind, and a student sample with insomnia rather than a clinical psychosis population; the authors flag the generalisation question themselves. The mental-health effects are small in size even though they are precise. Retention at follow-up is not stated in the abstract.

Caveat on this rating: Forty-four percent follow-up at six months is a serious limitation and the abstract states it. The prevention question the trial was named for returned a null that the trial was too small to answer (22 cases). The symptom result stands against an attention-matched placebo, which is a fair comparison, but it is a symptom score, not a diagnosis.

Caveat on this rating: A non-inferiority design proves "not worse by more than 1.9 points", not equivalence or superiority. Open-label treatment with masked assessors. Roughly a fifth of participants were missing at 12 months in the mITT analysis. The result is about trained, supervised junior workers in a trial, not about self-help.

Caveat on this rating: Enhanced usual care as the comparator, in a previously untreated population in one Indian state, so the gain partly reflects how little the control group otherwise received; an attention-matched control would be a harder test. Three-month primary outcome. The suicidality result is a secondary outcome. Generalisation to well-resourced settings is an open question the trial does not answer.

Caveat on this rating: A superiority trial that found no difference is not proof of equivalence; the interval allows a modest advantage either way. Participants knew their allocation. The population was people with three or more episodes on a stable maintenance dose, with their own doctor supporting the taper. This row is a question to bring to a prescriber, not a plan to make from a page.

Caveat on this rating: Non-active controls only, so the comparison is against nothing rather than against another treatment; high heterogeneity; and the abstract itself reports that blinded trials found smaller effects, which means the pooled 0.66 is likely inflated by expectation. Mixed samples of healthy and ill participants, with symptom scores rather than diagnosis as the outcome.

Caveat on this rating: Small effect by the authors' own numbers; self-selected online sample; self-report outcomes; adolescents only; three-month horizon. Included as the lower bound of what a free intervention can do, and it should be read as that, not as an alternative to care for a young person with significant depression.

Caveat on this rating: Metaregression compares across trials, so the frequency effect is observational even though the trials were randomised; confounding by trial design, population or therapy type is possible and the authors flag the finding as needing replication. The g = 0.45 figure is a model estimate, not the result of a trial that assigned frequency.

Caveat on this rating: Small trial, about thirty per arm, wide confidence intervals, eight weeks only. The fluoxetine-versus-placebo null is underpowered and should not be read as the drug failing in general; the combination that included fluoxetine did best. Modified intention-to-treat analysis. Light boxes cost money, and outdoor light was not what was tested.

Caveat on this rating: A meta-analysis of observational routine-care studies, so the null association is descriptive, not causal. The confidence interval on the main estimate is wide (-0.70 to 0.36), which is consistent with a moderate association in either direction as well as with none. The authors themselves say results depended on the methodological criteria applied.

Caveat on this rating: This row is the contest. The 2008 effect size was disputed in print in 2010 for a calculation error, heterogeneity and weak comparators; any reply by the original authors was not read for this row. Both abstracts concern complex, chronic disorders, so neither supports nor refutes long therapy for common problems. Scored below a meta-analysis of randomised trials on purpose.

Caveat on this rating: One trial, 326 people, Finland; outcomes are self-report questionnaires on functioning and quality of life rather than diagnosis or remission, and the long-term arm received many more sessions, so the comparison is not cost-matched. The result genuinely points both ways depending on the horizon and the measure, and this row keeps both.

Caveat on this rating: Mendelian randomisation depends on the genetic instruments affecting depression only through activity; the authors ran sensitivity analyses for that, but it cannot be proven. European ancestry only. The protective estimate is per standard deviation of measured acceleration and does not translate into a dose. The self-report null means most survey-based evidence in this area is on weaker ground than it looks.

Caveat on this rating: A routine-care data set, not a trial: patients were not assigned a number of sessions, and people who dropped out early may have been counted differently from those with planned endings. The 14-session figure is the point by which most responders had responded, not a ceiling on who can respond. Two symptom questionnaires define response; functioning and quality of life were not measured here.

Caveat on this rating: A single university counselling centre with a young, mostly White sample; frequency was chosen, not assigned. The effect on speed is between small and medium by the authors' own description, and the paper reports no difference in total change, so it argues for efficiency rather than for a better end result.

Caveat on this rating: Observational, single country, primary care, twelve sessions at most. Clients chose when to end, so session count is an outcome of improvement as much as a cause of it; that is the paper's own point, and it means the falling percentages cannot be read as later sessions doing harm. The sample was 92.4% White and 73.1% female, as the abstract reports.

Caveat on this rating: Dose-response studies in routine care are observational: nobody was randomised to a number of sessions, so the curve describes what happened, not what a fixed dose would do. The review is explicit that chronic and severe disorders are under-studied. The 4 to 26 range is a spread across settings, not a prescription for any one person.

Caveat on this rating: Small trials pooled across six protocols that differ in target and dose; the pooled safety comparisons (k = 7 for mania, k = 10 for headache) are thin. Favourable does not mean settled.

Caveat on this rating: The strongest thing this paper says is about its own evidence base: 138 of 141 trials at high risk of bias or with concerns. Any ranking built on that is provisional, which the authors state. The 2019 BMJ network in this wave, with a different trial set, likewise found overlapping intervals among the TMS forms.

Caveat on this rating: Four randomised trials is a thin base for a head-to-head conclusion, and the null difference has a wide interval (-1.62 to 0.27) that does not rule out either direction. The maintenance signal rests on short follow-up, as the authors say.

Caveat on this rating: A 48-person trial with a one-month primary endpoint, from the developers, following a 29-person trial from the same group. The FDA clearance record in this wave notes that effectiveness of the device "has not been established beyond the timepoints evaluated" and that the trials behind it were single-site. Early, promising, and not yet independently replicated.

Caveat on this rating: The pooled effect and the null RCT point in different directions only if uncontrolled studies are read as evidence of efficacy, which the authors do not do. The evidence base in adolescents is small, and the one adequately powered sham-controlled trial was negative.

Caveat on this rating: Seven small RCTs with widely different doses cannot be added up, and the authors do not try. The absence of a pooled number is the finding.

Caveat on this rating: A 510(k) is a substantial-equivalence determination, and the summary says so. The efficacy figures come from a 29-person trial stopped early at an interim analysis and three open-label studies, all at one site, assessed at four weeks. The record itself flags generalisability and blinding as unevaluated or uncertain.

Caveat on this rating: The rankings are network estimates with wide intervals (bitemporal ECT 2.57 to 30.91), and 83% of trials were at low or unclear risk of bias by the authors' own count, so the ORDER of treatments is far less certain than the finding that most beat sham. The 2026 network meta-analysis of rTMS modalities in this wave reaches the same conclusion about ordering: overlapping intervals and mostly low-confidence rankings.

Caveat on this rating: Confidence in these numbers is low by the authors' own grading: 1 of 218 studies met the Cochrane criteria for low risk of bias, and CINeMA confidence was low for walking or jogging and very low for every other modality. The credible intervals for the five modalities overlap heavily, so the order above is not an established ranking and must not be read as one. The authors name the blinding problem themselves and call for future trials to blind participants and staff. Results did appear robust to publication bias. A correction was published (BMJ 2024;385:q1024) and is logged as unverified - the BMJ site refuses automated fetch - but nothing quoted here depends on a precise ranking. Effect sizes drawn from different literatures are not directly comparable, so this row deliberately does not set these numbers against antidepressant trial results.

Caveat on this rating: Observational. The correlations are small - 0.05 within persons, 0.08 between - and the authors state plainly that the practical effect is comparable to other everyday activities. People who move more may differ from people who move less in many ways this design cannot adjust away. The association runs in both directions, so "exercise lifts mood" is only half of what was measured: feeling better also predicts moving next. Considerable heterogeneity across individuals, only partly explained by sociodemographic moderators, means the average says little about any one person. The negative association with calmness is a real finding rather than a nuance to skip - activity tracked energetic arousal most strongly, so someone seeking calm may not find it here.

Caveat on this rating: Read this as absence of evidence, not evidence of absence - the body says so and the authors insist on it: "Contrary to expectations, nature-based interventions did not significantly outperform controls. However, this finding should be interpreted cautiously... studies in this group were small, at moderate-to-high risk of bias and conceptually diverse... This conceptual heterogeneity probably diluted the pooled effect estimate and limits interpretability." The abstract adds that the evidence "was limited by conceptual and methodological heterogeneity". Across the 183 trials risk of bias was "frequently moderate to high" and funnel asymmetry suggested publication bias, though sensitivity analyses held. Separately: the field's most-cited experiment, Bratman et al. 2015 PNAS (38 people), did not hold exercise constant - its nature route climbed 155m and its urban route 4m over the same 5.3km - and its brain finding has never been replicated.

Caveat on this rating: Strong design, still not causal, and the authors say so: "Despite the strong longitudinal design of our study, our risk estimates fundamentally only show correlations." Three limits they name: residential choice may be genetically confounded; unmeasured socioeconomic factors such as green-space quality and crime rates may remain; and NDVI "provides no information about the use of green space", so a child who only sees a park scores the same as one who plays in it daily. Our own added critique: there is no sibling or within-family analysis, so families who move somewhere greener are never compared with their own relatives. The 55% figure is the paper's "up to" across various disorders, not a single headline estimate.

Caveat on this rating: THE WIDELY-QUOTED VERSION OF THIS FINDING IS RETRACTED. Gu et al. 2019 (Br J Psychiatry 215(2):456-467, doi 10.1192/bjp.2018.295), the source of the much-repeated "19% higher depression risk per 10 ug/m3 PM2.5", was retracted - confirmed three ways on 2026-08-28: the PubMed record for PMID 30719959 is flagged RETRACTED ARTICLE; a separate retraction notice exists at PMID 32349797 / doi 10.1192/bjp.2020.87; and Crossref carries a type-retraction update-to relation from both the publisher and Retraction Watch. The published notice gives no reason we could read, so none is asserted here. Do not use the 19% figure; 1.102 [1.023-1.189] is the surviving estimate. It rests on five observational studies with, in the authors' words, "heterogeneous outcome definitions, exposure assessment, and residual confounding" - and poverty tracks bad air, so the inversion is about consistency, not about proof.

Caveat on this rating: The famous number is far weaker than its fame. The threshold was not pre-specified - it emerged from cutting the data into 60-minute blocks - and the curve is not a staircase: adjusted odds for good health run 1.59 [1.31-1.92] at 120-179 minutes, then FALL to 1.44 at 180-239, rise to 1.55, and fall again to 1.33 at 300+, with overlapping confidence intervals throughout. The data cannot distinguish two hours from five. The authors call the threshold and peak "more as a starting points for discussion and further investigation, than clearly established findings", report that nature time explained "approx. 1%" of variance, and state they "are unable to rule out the possibility that the association is, at least in part, due to healthier, happier people spending more time in nature."

Caveat on this rating: One city, 112 people, one week of measurement, analysed cross-sectionally - this is a signal to take seriously, not a settled result, and the badge reflects that. The direction is what matters: the authors write that these findings "contrast many studies showing positive relationships between time outdoors and mental health and highlight the need for research with diverse populations and better measures of outdoor conditions and intentionality of greenspace contact." Reverse causation is unexcluded here too: people already anxious may move differently. The reading this row supports is that "outdoors" is not a single exposure, not that being outdoors is harmful.

Caveat on this rating: These dropout figures come from trial conditions - supervised, scheduled, usually free - which is not the same as starting alone, and the review's own finding that professional supervision lowered dropout says so. The 18.1% is trim-and-fill adjusted for publication bias; the unadjusted rate differs. Trial participants had volunteered for an exercise study, so they were more willing than depressed people in general by selection. The finding that higher baseline depression predicted greater dropout is an association from meta-regression across studies, not a prediction about any individual. Behavioural activation is adjacent to exercise but not the same thing: it schedules valued and rewarding activity of any kind, delivered by a trained worker, and no trial found here tests exercise as behavioural activation against behavioural activation as usual. COBRA compared it with CBT, not with exercise.

Caveat on this rating: The 2011 trial carries a published rebuttal in the same journal: Coen, Lawlor and Kenny, "Failure to demonstrate that memory improvement is due either to aerobic exercise or increased hippocampal volume", PNAS 2011;108(18):E89, author reply E90, doi 10.1073/pnas.1102593108. A 2% change in a subregion sits near the limit of what structural MRI resolves reliably in 120 people, and the anterior hippocampus is a smaller compartment measured more noisily than total volume. The meta-analysis found the left-hippocampus effect only, and attributed it post-hoc to prevented decline - a post-hoc subgroup result, weaker than the headline sounds. Neither study enrolled depressed people or measured mood, so neither supports any claim about exercise and emotional health. Animal models show hippocampal growth consistently; the human evidence is called equivocal by these authors.

Caveat on this rating: Observational, not a trial: 15 prospective cohorts, so causation is not established. The standing objection is reverse causation - early, undiagnosed depression reduces activity, so low activity may be an early symptom rather than a cause. The 3-year minimum follow-up mitigates this but does not remove it. Heterogeneity was large and significant (I2 = 74%), and physical activity was self-reported in the contributing cohorts, which over-estimates it. The paper's 11.5% population attributable fraction assumes a causal relationship the design cannot establish. The 45-minute, three-to-five-times figure comes from Chekroud 2018 - a cross-sectional survey of self-reported poor mental health days in 1,237,194 US adults, not a trial, so it cannot establish that any type or dose is best; its senior author is affiliated with Spring Health, a commercial mental-health company, and Microsoft provided cloud computing.

Caveat on this rating: The "as effective as antidepressants" claim cannot be made from this review and is refused here. Its head-to-head comparisons are underpowered: exercise vs medication is 5 trials and 330 people (SMD -0.11, 95% CI -0.33 to 0.10, low certainty); exercise vs psychotherapy is 10 trials and 414 people (SMD 0.03, -0.16 to 0.23, moderate certainty). Absence of a detected difference in small trials is not evidence of equivalence - those intervals are wide enough to hold a meaningful advantage either way. A real non-inferiority trial: COBRA (Richards, Lancet 2016) randomised 440 people against a pre-specified 1.9 PHQ-9 margin. Nobody has done that for exercise versus antidepressants. Also recorded here: adverse events included musculoskeletal injuries and depression among those exercising. The 2013 version put the robust-trial subset at -0.18 and non-significant; the 2026 update puts it at -0.46. The effect attenuates under good methods; it no longer disappears.