The evidence library

Caveat on this rating: With 34 participants the study is underpowered to detect small true effects, and the authors' own Bayesian analysis in the successfully blinded subgroup was inconclusive rather than clearly negative.

Caveat on this rating: The meta-analytic estimates rest on very little randomised data — two parallel RCTs (three comparisons, n=117) for efficacy and two RCTs (n=109) for adverse events — so the confidence intervals are wide and the review is better read as demonstrating absence of evidence than proving absence of effect. Several authors are affiliated with psychedelic research centres and one is chief medical officer of a psychedelics company (disclosed), so the null result is not coming from sceptics.

Caveat on this rating: Participants sourced and weighed their own material, so actual doses were unverified, and the sample was self-selected enthusiasts rather than people with a diagnosed condition. The authors are based at Imperial College's Centre for Psychedelic Research with a co-author from the Beckley Foundation, both proponents of psychedelic research — which cuts against, not toward, a bias explanation for this null result.

Caveat on this rating: Uncontrolled and unblinded: participants knew they were dosing, sourced their own substances, and were self-selected enthusiasts, so nothing here can separate drug effect from expectation. The authors say as much and explicitly call for dose- controlled research.

Caveat on this rating: Open-label by necessity (neither ketamine nor ECT can be masked), so subjective outcomes may favor the treatment patients preferred; population excluded psychotic depression, where ECT performs best.

Caveat on this rating: This umbrella review's STRONG tier reflects its independent, systematic method — its actual conclusion is that the underlying MDMA evidence is weak. Read the tier as confidence in the review, not in MDMA.

Caveat on this rating: The declarations of interest read literally: "KH has previously acted as a temporary consultant for Pfizer (manufacturers of sildenafil). MT has been paid to lecture and received travel expenses from Bristol-Myers Squibb (manufacturers of buspirone) and Otsuka; his spouse is an employee of GlaxoSmithKline (manufacturers of bupropion)." Two of the six authors therefore have financial ties to the makers of the two drugs the review endorses, which is why independence is scored false despite the funding being public. The reviewers also warn that partial reporting of subscale results in the source trials could bias the effect estimates upward.

Caveat on this rating: A regulatory label change is a precautionary act on a safety signal, not proof of incidence, causation, or permanence. The evidence base the PRAC weighed was EudraVigilance reports, literature, social media, and manufacturer reviews. Clomipramine and vortioxetine were part of the same signal assessment but were explicitly not covered by the labelling recommendation. US FDA labelling does not carry equivalent wording.

Caveat on this rating: The authors state that including open-label studies and pooling across different sexual-function scales "could reduce the significance of our findings," so the 25.8-80.3% band is wide partly because the source studies were not uniform. No funding statement or conflict-of-interest disclosure was accessible for this paper.

Caveat on this rating: The high trust score reflects the review's method, funding, and journal, not the strength of the underlying trials: 398 men across 11 studies, several from the 1970s and 1980s, with waitlist controls and confidence intervals touching 1.0. The single trial claiming group therapy outperformed sildenafil alone (WMD -12.40 on the IIEF) had 20 participants and was run by one of the review's own authors, which is a direct conflict on the review's most eye-catching result.

Caveat on this rating: The high trust score reflects the quality of this meta-analysis as a piece of evidence synthesis, not the strength of the effect it found. The pooled benefit rests on a subjective, dichotomised outcome in six trials with demonstrated publication bias, and the authors' own conclusion is the hedged "valerian might improve sleep quality," followed by a call for better studies. Twenty years on, those studies have largely not materialised.

Caveat on this rating: This review reaches the opposite conclusion to Bent 2006 from overlapping trials, and the difference is instructive: Taibi weighted trial quality and recency, and found that the better and newer the study, the smaller the effect. That pattern is the signature of a treatment whose apparent benefit comes from weak methods.

Caveat on this rating: Definitions do the heavy lifting here and are stated rather than assumed. 'Pregnancy-related' means causally related to the pregnancy and within one year of its end - a wider window than the 42-day definition used in international comparisons, which is why UK figures for suicide as a direct cause within 42 days (0.80 per 100,000) look nothing like the 6-week-to-1-year figures quoted. The MMRC category bundles suicide with overdose and poisoning related to substance use disorder, so it is not a suicide count. The 2020 US figure covers 38 states, not the nation, and a 2020 coding change moved substance-use-related overdose deaths into this category, which is part of why 22.5% is not comparable with the 11% from 2008-17. A separate CDC report gives 26% for mental health conditions as a CONTRIBUTING circumstance rather than an underlying cause; that figure answers a different question and is not used. None of these numbers say anything about the risk facing any individual reader.

Caveat on this rating: Every study in this row is observational and self-selected, and that cuts against the row's own direction as much as for it. Cohen 2006 recruited at three specialist perinatal psychiatry centres, so its population is women with severe recurrent illness rather than a general one, and its hazard ratio of 5.0 cannot be read as the causal effect of stopping. Grote 2010 measured antenatal depression rather than treatment status, and its effect sizes swing with the measurement instrument and the country. Jarde 2016's own conflict-of-interest subgroup nearly doubles the preterm-birth odds ratio, which is a warning about the literature and is printed here for that reason. No randomised trial of continuation versus discontinuation exists, so 'stopping is riskier' and 'continuing is riskier' rest on the same grade of evidence - which is the point of the stop, and why it closes on a conversation rather than a conclusion.

Caveat on this rating: What this row does not claim: that SSRIs in pregnancy are risk-free. Adjustment is not randomisation, and the propensity-score models can only correct for what was measured in claims data - smoking, alcohol, illness severity and body-mass index are poorly captured in Medicaid records, and residual confounding runs in both directions. The primary-PPHN result still clears one after adjustment. Brown 2017's sibling analysis has wide intervals rather than a null, and its authors wrote that a causal relationship cannot be ruled out. Sujan 2017's surviving preterm-birth association may itself reflect illness severity within a family rather than the drug. Levinson-Castiel is a single centre with 60 exposed infants. The defensible reading is that the large early relative risks shrank substantially under better designs, that the absolute excess risks are small where they persist, and that neither 'proven harmful' nor 'proven safe' is what the evidence supports.

Caveat on this rating: The uncertainty here is real and is printed rather than resolved. Patorno's adjusted risk ratio for cardiac malformations is 1.65 with a lower bound of 1.02, and the right ventricular outflow tract estimate is 2.66 with a lower bound of exactly 1.00 - both only just clear chance, and the lithium numerator for that outcome is suppressed under the Medicaid data-use agreement, so this row makes no claim about how many Ebstein cases the study did or did not observe. Patorno excluded Ebstein anomaly as a named outcome on purpose, because clinicians may be likelier to code a defect as Ebstein in an infant known to have been lithium-exposed. Munk-Olsen 2018's first-trimester malformation result is significant while its any-time and cardiac-specific results are not. Fornaro 2020's meta-analysis and Diav-Citrin 2014 point the same direction with wide intervals. The honest summary is a small absolute increase, dose-related, on a very small base - not the register's 400-fold, and not nothing.

Caveat on this rating: Named limits. EURAP is a prospective registry of women in epilepsy care, not a randomised comparison, and women prescribed valproate differ from women prescribed lamotrigine in ways registries cannot fully adjust away - the strongest single caution on every figure here. NEAD is observational and its age-6 arm retained 224 of 311 children. Christensen's whole-cohort autism estimate (4.42% vs 1.53%) is confounded by indication and the epilepsy-restricted hazard ratio of 1.7 (0.9-3.2) is NOT statistically significant; both are printed so the reader can see the difference the comparison group makes. The North American registry's 9.3% valproate figure is consistent with EURAP but its published abstract carries no confidence interval, so it is not quoted here. What is not in dispute across registries, cohorts and three regulators is the direction, the dose-dependence and the rough size.

Caveat on this rating: What is NOT established: that the narrative PLLR labels are better understood by clinicians in practice. No study shows that, and the Namazy survey points the other way - 68% of respondents did not find the narrative concise, and 95% still used the letters. This row asserts the documented change and the documented reasons the agency gave for it, not that the replacement works. The Namazy survey is also a 12% response rate (184/1500) among allergy/immunology clinicians, not a general prescriber sample. The Addis comparison is from 2000 and concerns three systems, two of which still exist.

Caveat on this rating: The declarations of interest read literally: "KH has previously acted as a temporary consultant for Pfizer (manufacturers of sildenafil). MT has been paid to lecture and received travel expenses from Bristol-Myers Squibb (manufacturers of buspirone) and Otsuka; his spouse is an employee of GlaxoSmithKline (manufacturers of bupropion)." Two of the six authors therefore have financial ties to the makers of the two drugs the review endorses, which is why independence is scored false despite the funding being public. The reviewers also warn that partial reporting of subscale results in the source trials could bias the effect estimates upward.

Caveat on this rating: A regulatory label change is a precautionary act on a safety signal, not proof of incidence, causation, or permanence. The evidence base the PRAC weighed was EudraVigilance reports, literature, social media, and manufacturer reviews. Clomipramine and vortioxetine were part of the same signal assessment but were explicitly not covered by the labelling recommendation. US FDA labelling does not carry equivalent wording.

Caveat on this rating: The authors state that including open-label studies and pooling across different sexual-function scales "could reduce the significance of our findings," so the 25.8-80.3% band is wide partly because the source studies were not uniform. No funding statement or conflict-of-interest disclosure was accessible for this paper.

Caveat on this rating: The high trust score reflects the review's method, funding, and journal, not the strength of the underlying trials: 398 men across 11 studies, several from the 1970s and 1980s, with waitlist controls and confidence intervals touching 1.0. The single trial claiming group therapy outperformed sildenafil alone (WMD -12.40 on the IIEF) had 20 participants and was run by one of the review's own authors, which is a direct conflict on the review's most eye-catching result.

Caveat on this rating: Both arms improved similarly on the primary desire and arousal outcomes, so this trial does not show that mindfulness specifically works — it shows that eight weeks of structured group attention works. There was no no-treatment or waitlist arm, so regression to the mean and expectancy effects cannot be separated out. All 148 participants were cisgender women, and the authors note the need to diversify samples.

Caveat on this rating: Every pooled effect is against a waitlist, not an active or placebo control, so attention and expectancy are baked into the d = 0.58 figure. The search ends at 2009, predating the internet-delivered treatments that now dominate access. No total participant count is reported in the abstract, and the MEDLINE record lists the publication type "Research Support, Non-U.S. Gov't", indicating unnamed external support.

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: This row asserts an absence, which is the claim most easily overstated. What is established is that the delayed onset is not explained by the immediate neurochemical change, and that the field's own reviews present competing accounts rather than a settled one. What is NOT established is that either account here is correct: the emotional-processing model and the neuroplasticity model are both active research programmes with supporting and non-supporting findings, and this row deliberately reports that they are unresolved instead of ranking them. The claim that no clinical test measures a chemical imbalance is a statement about available assays, not a claim that no biological contribution exists. A narrative review is also a lower grade of evidence than a systematic one, and it is cited here for what the field says about its own uncertainty rather than as a pooled result.

Caveat on this rating: This literature does not point one way, and the row would be dishonest if it implied otherwise. A randomised study from the same research group as the bogus-test experiment found self-blame and perceived helplessness GREATER after a cognitive-behavioural explanation than a biological one (Lee et al. 2016, J Soc Clin Psychol 35(7):571-588), and a pre-registered sham-genetic-test experiment found no effect on beliefs about depression at all (Lamontagne et al. 2023, PMID 36869258). Both are named here and neither is quoted, because their full texts were not obtained. The Kvaale effect sizes are small to moderate, the dangerousness result may reflect publication bias by the authors' own account, and the samples across this field are dominated by undergraduates and online panels rather than people in treatment. The absence of any study of debunking, or of any behavioural outcome, is stated as an absence found by search - it is not proof that no such work exists anywhere.

Caveat on this rating: Every efficacy number here is contested at the edges, and none of the disputes touches the conclusion this stop rests on. Cipriani 2018's own authors graded certainty moderate to very low and 9% of included trials at high risk of bias; its between-drug ranking is not established and is not used here. Stone 2022's three-class result comes from a statistical model, not an observed grouping, and its reading has been argued over since publication - though its co-author list includes the field's most prominent placebo-effect sceptic, which cuts against overstating benefit rather than for it. Both meta-analyses are dominated by short, largely industry-sponsored acute-phase trials in adults meeting formal criteria for major depression, and neither measures long-term outcomes. Cowen and Browning's emotional-processing account is a hypothesis with supporting experiments, not a settled mechanism. The ANTLER figures describe trial conditions, not any individual reader's risk.

Caveat on this rating: This stop is the dispute, so the caveat is about what it does NOT establish. It does not establish that the serotonin hypothesis is disproved: the review's own framing is absence of support, and Jauhar and colleagues argue that its method manufactures that absence. It does not establish that the rebuttal is correct either. The two sides also disagree about what tryptophan depletion is for - whether the test is inducing depression in healthy people, or worsening it in people with a history - and that disagreement is unresolved rather than adjudicated here. Four further published critiques and a second author reply appeared in 2024 and are named in the body only as "a second exchange", because their full texts are paywalled with no PMC deposit and this library does not quote a paper it has only read about. Declared interests are printed for both sides precisely because neither set of interests decides the science.

Caveat on this rating: The central dispute in this row is unresolved and is left that way on purpose. Pies argues that no professional body, textbook or peer-reviewed consensus ever advanced a chemical-imbalance theory and that critics cherry-pick quotations; Ang and colleagues report that highly cited reviews and every textbook they sampled did support the serotonin theory between 1990 and 2010. Two of the three authors on that side are also authors of the 2022 umbrella review, which is a stake in the outcome and is stated rather than hidden. Their paper was a sample of highly cited work, not a census, and its full text could not be obtained, so nothing beyond its published abstract is quoted here. The France 2007 survey is 251 US undergraduates in one place at one time and does not describe the general public. Nothing in this row establishes that anyone intended to mislead; intent cannot be documented from outside, and the essay itself does not claim it.

Caveat on this rating: Definitions do the heavy lifting here and are stated rather than assumed. 'Pregnancy-related' means causally related to the pregnancy and within one year of its end - a wider window than the 42-day definition used in international comparisons, which is why UK figures for suicide as a direct cause within 42 days (0.80 per 100,000) look nothing like the 6-week-to-1-year figures quoted. The MMRC category bundles suicide with overdose and poisoning related to substance use disorder, so it is not a suicide count. The 2020 US figure covers 38 states, not the nation, and a 2020 coding change moved substance-use-related overdose deaths into this category, which is part of why 22.5% is not comparable with the 11% from 2008-17. A separate CDC report gives 26% for mental health conditions as a CONTRIBUTING circumstance rather than an underlying cause; that figure answers a different question and is not used. None of these numbers say anything about the risk facing any individual reader.

Caveat on this rating: The screen's accuracy depends on the reference standard: the 0.88/0.85 figures are against semistructured clinician interviews; against fully structured or MINI interviews the authors report different values. The prevalence gap (31% vs 17%) is across heterogeneous populations and does not by itself show that any one clinic over-diagnoses; it shows what happens when a screen is read as a verdict.

Caveat on this rating: The trials pooled here predate most collaborative-care programmes; the USPSTF's B grade rests on evidence that screening programmes with support improve outcomes, which is not in conflict with Gilbody's null for screening alone. Both are stated. Neither shows that screening harms.

Caveat on this rating: Measure 134 is a process measure and its follow-up plan definition is broad by design; a documented referral counts, and the specification itself warns against reflexive prescribing. Measure 370 is an outcome measure and remission is a legitimate goal. The critique in this trail is about incentives around a screening score, not about the measures' stated intent, which is quoted.

Caveat on this rating: The 0.388 factor comes from a CMS table whose model version and date were not printed on the pages read; it is a pre-2024 numbering and the 2024 model reclassified the category, so treat it as the order of magnitude a coded diagnosis added, not a current price. MIPS adjustments are budget-neutral and most clinicians land near zero; the plus-or-minus 9 percent is the statutory range, not a typical outcome. The OIG figure covers diagnoses of all kinds recorded on HRAs, not depression alone.

Caveat on this rating: One trial of 120 outpatients in a single setting, with pharmacotherapy restricted to two drugs; the review's 'virtually all' spans heterogeneous designs. The effect size is large and consistent with the wider literature, but this is not the same evidence tier as the meta-analyses in stops 2 and 3. The claim that the measures do not pay for the conversation is an observation about the measure definitions quoted in stop 4, not a finding of any study.

Caveat on this rating: Funding a tool is not the same as biasing it: the PHQ-9's accuracy has since been tested in an independent individual-participant meta-analysis of 58 studies (stop 2), and the tool is free precisely because of the grant. What the grant bought is reach, not a rigged score. The GAD-7 funding statement was read from the instrument's own attribution, not from the 2006 paper's text.