The paper

Methods

Everything below was fixed in a registered plan before any comparison between arms was computed, except what the last part of this section, "What this version corrects", describes. The plan, its codebooks, both codings and the registered code are under plan/; every addition since is listed in deviations.json.

The question. The FDA pooled 372 placebo-controlled trials of antidepressants in adults. For suicidal behaviour (completed suicide, attempts or preparatory acts) in psychiatric patients it reported odds ratios of 2.30 under age 25, 1.03 at 25 to 64, 0.87 at 25 or over, and 0.06 at 65 or over, declining by 4.6% per year of age (Stone et al. (2009)). In 24 pediatric trials, suicidality had a risk ratio of 1.95 (Hammad et al. (2006)).

Re-analyses of FDA summary data on newer antidepressants disagree about adults:

• Hengartner and Plöderl (2019) found an odds ratio of 2.38 for suicide attempts.
• Kaminski and Bschor (2020), analysing the same data by Mantel–Haenszel, found 1.63.
• Hayes et al. (2019) found no statistical evidence of increased suicide.
• Khan et al. (2018), counting patient-years of exposure, found no difference.

All of these rest on data sponsors submitted to the FDA. ClinicalTrials.gov has since received the results of trials of drugs regulated in the United States. Those results include every serious adverse event (one that results in death, is life-threatening, or requires or prolongs hospitalization), by arm and by MedDRA term, whether or not the trial was published; publications often report these events incompletely (Tang et al. (2015)).

Data. We searched ClinicalTrials.gov's API (version 2.0.5, data as of 2026-10-09) on 2026-10-13. The 39 antidepressant names in data/search.json found 932 interventional studies with posted results, whose records are in data/ctgov/ as returned.

Eligibility. The code kept randomized parallel or factorial trials with at least two adverse-event groups (639 of 932). Two coders then read each trial's titles, conditions, eligible ages, arms and adverse-event groups, without counts. Under a codebook they decided:

• which groups compare an antidepressant with placebo with nothing else different (pooling doses and antidepressants, and marking combination products);
• the design: randomized from the start ("acute"), randomized withdrawal, or other;
• and the indication: depressive, other psychiatric, or non-psychiatric.

The coders agreed on eligibility for 635 of 639 trials (Cohen's kappa 0.987) and on identical comparisons for 395 of the 398 both accepted. A third reader decided the 11 trials on which they differed. The code took each trial's age group from its eligible ages and its mean age from the posted baseline age. It took the completion date from the record. Participants are those at risk in each group's adverse-event tables, its safety population, which need not be every participant randomized.

Outcomes. Two coders assigned each of 4,029 distinct adverse-event terms, under a second codebook, to:

• suicidal behaviour (completed suicide, suicide attempt, intentional self-harm, intentional overdose);
• suicidal ideation;
• completed suicide;
• overdose of unstated intent;
• or nausea.

They gave identical outcome sets to 4,028 terms (kappa 0.995), and a third reader decided 10 terms. All coders and readers were separate instances of Claude Opus 5.5, each working alone.

Suicidal outcomes are counted from the serious-event tables, which list every serious event. Nausea is counted from the non-serious tables, only in comparisons that list it, since those tables omit terms below a frequency threshold. The registered count of an outcome in an arm is the sum over its terms of participants affected, capped at those at risk.

Analyses. The co-primary analyses cover adult trials randomized from the start (minimum eligible age 18) and exclude combination products:

• P1, serious suicidal behaviour;
• P2, any serious suicidality (behaviour or ideation).

The estimate for each is the Mantel–Haenszel risk ratio across comparisons, with the variance of Greenland and Robins (1985), a 95% confidence interval and a two-sided Wald test. The two co-primary p-values are adjusted by Holm's method.

The 18 registered secondary analyses are adjusted by the method of Benjamini and Hochberg (1995). They cover:

• ideation alone, completed suicide, and behaviour counting overdoses of unstated intent;
• indication;
• trials enrolling only people aged 60 or over;
• trials completed after 2017;
• drug class (SSRIs; SNRIs; bupropion; esketamine, zuranolone, brexanolone and dextromethorphan–bupropion; tricyclics, MAOIs and other monoaminergic drugs);
• combination products included;
• randomized-withdrawal trials;
• pediatric trials (behaviour; any suicidality);
• and trials spanning adolescents and adults.

An analysis with fewer than 10 participants affected is reported but not tested.

Every tested analysis is also estimated:

• by Peto's odds ratio (Yusuf et al. (1985));
• by random effects (DerSimonian and Laird (1986), with 0.5 added to every cell of a comparison with a zero cell);
• with a conservative variance, inflated by how far the outcome's count could overstate participants;
• by its most frequent preferred term alone;
• with each trial left out;
• and in two halves of the trials, paired by size and drawn by a seeded generator.

The registered rule calls an estimate robust when p < 0.05 holds in the main analysis, with the conservative variance, under random effects, and with every trial left out.

Two analyses relate effect to age. They regress each comparison's Peto log odds ratio on the trial's mean age, weighted by Peto's variance, across all acute comparisons with events. Each estimate is set against the earlier estimates the plan named for it.

What this version corrects. The first version's methods and adversarial reviewers found that its counts of suicidal outcomes are not counts of distinct participants, that its random-effects analysis changes the estimator as well as the assumption about heterogeneity, and that posting results after 2008 does not make trials independent of the FDA's data. Its domain reviewer found that the FDA's estimate the plan registered for suicidal behaviour at 25 to 64 is the paper's estimate for 25 or over. This version takes each up, after the results were known, with the code and data (data/earlier.json, data/resolved.json) it adds. The registered analysis is unchanged and gives the registered outputs byte for byte; the first version also miscounted the registered secondary analyses as 19.

Counting each participant once. ClinicalTrials.gov gives, for each adverse-event group, the participants affected under each term and the participants affected by one or more serious adverse events (its definitions). A participant with two of an outcome's terms is in both terms' counts, and the registered sum counts them twice. The tables bound the participants with an outcome in each group:

• at least the largest of its terms' counts, and at least the group's participants with any serious event less those listed under its other serious terms;
• at most the sum of its terms' counts, and at most the group's participants with any serious event.

An arm's bounds are the sums of its groups'. The code repeats the registered estimate at every combination of counts these bounds allow, each ambiguous arm independently, and reports the range of the estimate, of its interval's ends and of its p-value. An analysis counts as significant only if it is significant at every combination; for every tested analysis here, the largest p-value comes at the counts least favourable to an increase, the fewest on antidepressants and the most on placebo. The registered Holm and Benjamini–Hochberg corrections are applied at the counts least and most favourable to an increase. Participants with any serious event and deaths are each counted once in the tables and need no bounds. Every group whose count is left open is in results/ambiguous.csv.

What the trials' reports settle. For every group the tables leave open, whichever arm it is in, separate instances of the model searched the trial's public reports (its publications, the FDA's reviews, a sponsor's filing that Japan's regulator posts, a sponsor's dossier for Germany's health technology assessment, and sponsors' briefing documents and study synopses) for how many participants had the outcome's events, and the agent checked each passage against its source. data/resolved.json records what each report says and how it settles the count: by stating which events were one participant's, by counting the distinct participants, or, for pooled trials, by leaving one possibility. Each report's bounds lie within the tables', which the code checks. The analyses are repeated at these narrower bounds, with the same corrections; groups no report settles keep the tables' bounds.

Estimators. The registered random-effects estimate adds 0.5 to every cell of a comparison with a zero cell; with that correction, the DerSimonian–Laird method was among the most biased for odds ratios and risk differences in a comparison of methods for rare events (Bradburn et al. (2007)). To separate that correction from heterogeneity, the correction also fits a random-effects model that needs none (Stijnen et al. (2010)): treating events as Poisson, with participants at risk in place of person-time, the events on antidepressants in a comparison, given its total events, are binomial with log odds log(n₁/n₀) + θ, where θ is normal with mean μ and variance τ². It is fitted by maximum likelihood with 64-point Gauss–Hermite quadrature, with a profile-likelihood interval for exp(μ) and a likelihood-ratio test; with τ fixed at zero it gives a common risk ratio without correction.

The earlier analyses' trials. No list of the FDA's adult trials is public; their sponsors submitted them between September 2005 and September 2006, so a comparison of one of their drugs completed by September 2006 is counted as possibly among them. The re-analyses used the counts that Khan et al. (2018) took from the integrated safety summaries of 14 approval programmes. The FDA's medical reviews of the five newest (desvenlafaxine, trazodone extended release, vilazodone, levomilnacipran and vortioxetine) list their studies, which data/earlier.json matches to trials here by the sponsor's protocol number, with each review's address; for vortioxetine, Khan and colleagues used the short-term studies only. No trial here of an older programme's drug was completed by the year of its approval. The FDA's pediatric data were submitted by May 2004.

Earlier estimates. The plan registered 0.87 as the FDA's estimate for suicidal behaviour at 25 to 64, as the abstract of Stone and colleagues gives it; their Table 6 gives 0.87 (0.58 to 1.29) for 25 or over and 1.03 (0.68 to 1.58) for 25 to 64. The registered comparison is reported under its correct label, and the estimate for 25 to 64 is added. Below, "middle-aged adults" are those aged 25 to 64 and "adults past young adulthood" those aged 25 or over. Whether an earlier estimate lies inside our interval is reported at every combination of counts.

Display. Ratios are given to two decimals and p- and q-values to two significant digits, with more where fewer would round a ratio to 1.00 or a p-value to 0.050.

Software. The registered analysis and the correction's run in Python 3.12 with NumPy 2.5.3 and SciPy 1.18.1 in seconds. A third script rounds their outputs. results/analyses.csv lists every registered analysis with its bounds, and results/comparisons.csv every comparison's registered counts.

Results

Trials. Of 932 studies found, 639 passed the structural filter and 393 compared an antidepressant with placebo, in 413 comparisons with 99,730 participants: 56,808 on antidepressants and 42,922 on placebo, since many trials randomized several doses or drugs against one placebo group. The co-primary analyses cover 306 adult trials randomized from the start, with 322 comparisons and 71,643 participants (40,564 on antidepressants, 31,079 on placebo; median trial mean age 44.4 years). Of these, 34,345 were in trials of depressive disorders.

Check of the data. Nausea, listed in 207 adult comparisons with 57,580 participants, affected 6,100 of 33,497 on antidepressants and 2,057 of 24,083 on placebo (risk ratio 2.08, 1.98 to 2.18), as expected. One group lists two nausea terms; counted once, its participants leave the estimate at 2.08. The check says the arms are what they claim to be; it does not test how rare outcomes were coded.

Participants listed twice. In the adult trials, 5 arms list two or more terms of serious suicidal behaviour, and 13 of any serious suicidality, with totals that leave open whether they are one participant or more. In one more group, on antidepressants, the total shows that the two participants its suicidality terms count are one, so the registered count of any serious suicidality on antidepressants, 82, exceeds the most the tables allow, 81.

Serious suicidal behaviour (P1). Serious suicidal behaviour affected 43 to 46 of 40,564 participants on antidepressants and 16 to 19 of 31,079 on placebo. Across the 48 combinations of counts the tables allow, the risk ratio ran from 1.76 (0.9996 to 3.11; p 0.0502), with the fewest on antidepressants and the most on placebo, to 2.20 (1.21 to 3.99; p 0.0097). If every participant a group lists under two terms is one person, it is 2.08 (p 0.017); if none is, 1.86 (p 0.030), the registered estimate. The registered sensitivity analyses, at the registered counts, gave:

• Peto's odds ratio: 1.76 (1.06 to 2.91);
• the conservative variance: p 0.044;
• with each trial left out: estimates from 1.76 to 2.08, the largest p being 0.051;
• random effects: 1.22 (0.78 to 1.90; p 0.38).

The halves of the trials gave 1.09 (0.48 to 2.52) and 2.79 (1.26 to 6.19). By the registered rule the estimate is not robust.

Any serious suicidality (P2). Any serious suicidality affected 71 to 81 participants on antidepressants and 36 to 43 on placebo. Across the 36,864 combinations of counts, the risk ratio ran from 1.29 (0.87 to 1.92; p 0.20) to 1.73 (1.16 to 2.60; p 0.0078). With shared participants counted once it is 1.53 (p 0.044), and with none shared 1.47 (p 0.051). The registered estimate, 1.48 (1.01 to 2.18; p 0.045), counts one participant more than the tables allow. At the registered counts, Peto's odds ratio was 1.45, random effects gave 1.10 (0.78 to 1.55), and the conservative variance p 0.075. The halves gave 1.01 and 1.94.

The correction for two endpoints. At the counts least favourable to an increase, the Holm-adjusted p-values are 0.10 for serious suicidal behaviour and 0.20 for any serious suicidality; at the most favourable, 0.016 for both. With every shared participant counted once they are 0.034 and 0.044, and with none shared, 0.060 for both. Whether either outcome passes the registered correction turns on how many participants the tables list twice.

What the trials' reports settle. The trials' public reports settle 11 of the 17 groups the tables leave open, narrow 3 and leave 3 as they were (results/ambiguous.csv). Each of the arms that decided the bounds of serious suicidal behaviour on antidepressants turned out to hold one participant with two of its terms, such as an intentional overdose reported as a suicide attempt. With the reports:

• serious suicidal behaviour affected 43 participants on antidepressants and 17 to 18 on placebo, a risk ratio of 1.86 (1.04 to 3.31; p 0.036) to 1.96 (1.09 to 3.53; p 0.025);
• any serious suicidality affected 72 to 77 against 40 to 41, from 1.38 (0.92 to 2.05; p 0.12) to 1.51 (1.02 to 2.24; p 0.041);
• the Holm-adjusted p-values are 0.071 and 0.12 at the counts least favourable to an increase, and 0.0496 for both at the most favourable.

Serious suicidal behaviour is nominally significant at every count the reports leave. Neither outcome passes the registered correction at the counts least favourable to an increase; both pass, narrowly, at the most favourable.

Estimators. The registered random-effects model estimated the variance between comparisons at 0.00 for serious suicidal behaviour and 0.00 for any serious suicidality, so its estimates are the inverse-variance estimates with the continuity correction: 1.22 and 1.10. That correction reached 37 of the 43 comparisons with serious suicidal behaviour and 46 of the 61 with any serious suicidality, which have no events in one arm. The random-effects model that needs no correction gave, for serious suicidal behaviour:

• at the registered counts, 1.82 (1.08 to 4.41; p 0.025; τ 0.00);
• at the counts least favourable to an increase, 1.73 (1.01 to 3.84; p 0.044);
• at the most favourable, 2.15 (1.23 to 4.53; p 0.0065).

For any serious suicidality it gave 1.50 at the registered counts (1.01 to 2.58; τ 0.26), 1.31 at the least favourable (0.88 to 2.27) and 1.72 at the most favourable (1.16 to 2.62).

Single terms. Suicide attempt, the most frequent term of serious suicidal behaviour, affected 27 participants on antidepressants and 13 on placebo (risk ratio 1.59, 0.79 to 3.20; p 0.19). Suicidal ideation, the most frequent term of any serious suicidality, affected 33 against 20 (1.27, 0.71 to 2.25). No group lists either term twice, so these counts are of distinct participants.

The earlier analyses' trials. The FDA's reviews list 16 of the adult trials, with 8,487 participants, among the studies of the programmes the re-analyses used: 7 to 9 of their participants on antidepressants and 1 on placebo had serious suicidal behaviour. Two more trials, academic studies of bupropion completed before 2006, could be among the FDA's adult trials by drug and date; they had no serious suicidal events. In the remaining 304 comparisons (288 trials, 62,652 participants):

• serious suicidal behaviour affected 36 to 37 on antidepressants and 15 to 18 on placebo, a risk ratio of 1.64 to 1.98 (lower end of the interval 0.91 to 1.07; p 0.030 to 0.097);
• with the trials' reports, serious suicidal behaviour affected 36 against 16 to 17 (1.73 to 1.83; p 0.0497 to 0.070);
• any serious suicidality affected 60 to 67 against 35 to 42 (1.17 to 1.53; p 0.044 to 0.44).

Trials completed after 2017, a registered secondary analysis that none of the earlier data could hold, gave 9 against 4 (2.16, 0.66 to 7.05).

Against earlier estimates. At every count, the interval for serious suicidal behaviour contains both re-analyses' estimates for suicide attempts (2.38 and 1.63) and lies above the FDA's estimate for adults past young adulthood (0.87), the value the plan registered as its estimate for middle-aged adults. Whether it contains the FDA's estimate for middle-aged adults, 1.03, depends on the counts: the interval's lower end runs from 0.9996 to 1.21. The same holds in the trials neither data could hold. At every count the interval for any serious suicidality lies above the FDA's suicidality estimate for middle-aged adults (0.79). Computed before registration, blind to arm, the co-primary behaviour analysis had power 0.93 to detect the larger re-analysis estimate and 0.50 to detect the smaller.

Secondary analyses. Table 1 shows the tested secondary analyses at every count the tables allow. At the counts least favourable to an increase none reached the registered false discovery threshold; at the most favourable, four did: behaviour counting overdoses of unstated intent, behaviour in trials of depressive disorders, behaviour with combination products included, and pediatric suicidality. With the trials' reports (Table 2), none did at any count. Too few events remained to test:

• bupropion (2 against 2 to 4);
• tricyclic, MAOI and other monoaminergic drugs (4 against 3 to 4);
• other psychiatric indications (4 against 3 to 5);
• non-psychiatric indications (5 against 2);
• completed suicide;
• randomized-withdrawal trials;
• trials of older adults;
• and trials spanning adolescents and adults.

Table 1. Tested secondary analyses at every count the posted tables allow: participants affected on antidepressants and placebo, risk ratio, p, and Benjamini–Hochberg q, each from its value at the counts least favourable to an increase to its value at the most favourable where they differ.

Table 2. The secondary analyses the trials' reports change, at every count they leave, as in Table 1.

Pediatric trials. The 30 pediatric trials had 7,094 participants, with a median trial mean age of 13.5 years, and every one was completed in 2006 or later, after the FDA's pediatric data were submitted. As in the adult trials, one group's total shows that two participants its suicidality terms count are one, so the registered count of any serious suicidality on antidepressants, 38, exceeds the most the tables allow, 37. At every count, their interval for any serious suicidality contains the FDA's pediatric estimate (1.95). The trials' reports settle every pediatric group the tables leave open: any serious suicidality affected 35 children and adolescents on antidepressants and 8 on placebo (2.20, 1.02 to 4.75; p 0.044).

Age. At every count the tables allow, the odds ratio changed by 0.8% to 2.3% per year of mean age across the 51 comparisons with serious suicidal behaviour and a known mean age (p 0.21 to 0.65), and by −1.8% to −0.6% across the 73 with any serious suicidality (p 0.12 to 0.61). At the registered counts the slopes were 1.5% (−1.9% to 5.1%) and −1.3% (−3.5% to 1.1%). Every comparison with events took its age from a posted mean, not a median.

All serious events and death. Any serious adverse event, which the tables count once per participant, affected 913 participants on antidepressants and 743 on placebo (1.09, 0.997 to 1.19; p 0.057). Death from any cause, in 121 comparisons that reported it for every group (11,053 participants on antidepressants and 9,690 on placebo), affected 159 against 136 (1.17, 0.98 to 1.39; p 0.076).

Discussion

In these trials, serious suicidal behaviour was more frequent on antidepressants at every count the posted tables allow, by a ratio between 1.76 and 2.20. The tables alone leave its significance open, because in 5 arms they do not say whether a participant listed under two terms of suicidal behaviour, such as an intentional overdose and a suicide attempt, is one person or two. The trials' own reports answer that for most of them: each arm on antidepressants held one participant with both terms. With the reports the excess is 1.86 to 1.96, nominally significant at every count they leave, but it does not pass the registered correction for two endpoints at the counts least favourable to an increase, and passes only narrowly at the most favourable. That is close to what the first version reported from its double counts, now on counts that hold.

The first version called these trials independent of the data behind the earlier analyses because they posted results after 2008. That does not follow, and for the re-analyses it is not so: Hengartner and Plöderl (2019) and Kaminski and Bschor (2020) used the approval data of the newest programmes, whose short-term trials are among those here. Those trials, with 8,487 of the 71,643 participants, hold 7 to 9 of the participants with serious suicidal behaviour on antidepressants and 1 on placebo, so part of the agreement with the re-analyses comes from the same trials. In the trials neither set of earlier data holds, the estimate is close to the smaller re-analysis estimate and its interval still contains both, but whether it is significant depends on the counts. The trials completed after 2017, which no earlier analysis could hold, point the same way and are few.

The earlier estimates differ from these in what they count, and so cannot be set against them as tests:

• The FDA's adult analysis (Stone et al. (2009)) counted events its sponsors adjudicated blind to treatment, serious or not, in psychiatric patients, by each participant's age. For suicidal behaviour at 25 to 64 it found 1.03 (0.68 to 1.58), in Table 6, not the 0.87 its abstract gives for that range, which is its estimate for 25 or over. Our adult trials include participants aged 18 to 24, for whom the FDA found 2.30, and non-psychiatric indications.
• The re-analyses counted suicide attempts in the approval programmes' summaries, which also hold open-label and active-controlled studies.
• An analysis of clinical study reports for five antidepressants found an odds ratio of 0.81 (0.51 to 1.28) for suicidality in adults and 2.39 (1.31 to 4.33) in children and adolescents (Sharma et al. (2016)).

What these data add is an estimate of serious suicidal behaviour from posted results, which leans the same way as the re-analyses at every count, and an account of how much of that agreement rests on shared trials.

The registered random-effects analysis seemed to show the estimate shrinking toward one. It did not show heterogeneity: the variance between comparisons was estimated at zero, and most comparisons with events have none in one arm, so the 0.5 the method adds to each of their cells, and the weights that come with it, pulled the estimates toward one. A random-effects model that needs no correction gives estimates close to the Mantel–Haenszel ones. Suicide attempt alone, the outcome's most frequent term and one the tables count once per participant, does not differ significantly between the arms; the composite's excess also rests on intentional overdoses, self-harm and completed suicides.

Three readings of the excess remain open:

1. A real but modest increase. The crude rates were 10.6 to 11.3 serious suicidal events per 10,000 adults on antidepressants and 5.1 to 6.1 on placebo, over the weeks these trials last. That would be consistent with the activation or akathisia some patients experience early in treatment.
2. Unequal observation. Placebo groups in antidepressant trials often lose more participants to lack of efficacy. If a trial records adverse events only while participants take study medication, the antidepressant groups are watched longer and accumulate more events of every kind. The small excess of all serious adverse events (1.09) is consistent with some such imbalance. It is smaller than the excess for suicidal behaviour, but posted results give no person-time to correct it, unlike the patient-year analysis of FDA data by Khan et al. (2018).
3. Chance. The estimate rests on few events, its significance on how events are counted, and the halves differ.

Posted person-time, or individual participant data from the larger programmes, would tell the first two apart.

In pediatric trials, any serious suicidality was higher on antidepressants at every count, close to the FDA's estimate (Hammad et al. (2006)) and to the estimate for children and adolescents from clinical study reports, in trials all completed after the FDA's analysis. The warning for young people rested on that estimate, and these data agree with it, though they pass the correction across the secondary analyses only at the counts most favourable to an increase. Serious suicidal behaviour alone did not differ in these trials, so the pediatric difference lies mostly in serious ideation.

Several things are not supported by these data:

• An age gradient. The FDA found the risk falling with each participant's age. These trials show no such trend in their mean ages at any count. A trial's mean age poorly stands in for its participants' ages, and adult trials cluster in middle age, so this is not evidence against the FDA's participant-level gradient.
• Differences between drug classes. The class estimates rest on few events each, so none can be ranked.
• A conclusion about completed suicide, of which there were 7.
• Advice to individual patients. Effects on suicidal behaviour in short trials are not a measure of antidepressants' net effect on suicide, which also depends on their benefit for depression. Nothing here shows that people taking an antidepressant should stop it; stopping abruptly carries its own risks.

A larger registry analysis, with person-time and with each participant counted once, would narrow these bounds and this interval. It should combine ClinicalTrials.gov with the European and other registries that post results, and add trials completed after 2026.

Limitations

• Counts are bounds. Posted tables count participants by term. Where a group lists two or more of an outcome's terms, the number of distinct participants is known only within bounds, and the significance of the main estimates depends on where in those bounds it lies.
• The trials' reports. The reports that settle shared participants were searched for after the results were known. They differ in kind (publications, regulators' reviews, sponsors' documents), some count a period or a category slightly differently from the registry, and one levomilnacipran trial's reports give one participant with serious suicidal ideation where its registry record gives two. Three groups found no report.
• Serious events only. Only serious adverse events are counted. Suicidal ideation that never needed hospital care is not captured, so the outcome is serious suicidality, not all suicidality.
• Few events. Serious suicidal events are rare in trials lasting weeks: tens of events in each analysis, so the intervals are wide.
• No person-time. Posted results give participants at risk but not the time each was observed. Differential dropout and on-treatment reporting could bias the comparison, as the Discussion says.
• Event classification. Adverse-event terms are recorded by investigators and not adjudicated, and trials differ in what they call serious and in how they ask about suicidality.
• Trials included. Trials excluded people at high risk of suicide, except the esketamine trials that enrolled them. Only trials with posted results on ClinicalTrials.gov are included, and many older and non-US trials are missing.
• Overlap with earlier data. Trials in the re-analyses' data are identified only for the programmes whose FDA reviews list their studies; the FDA's adult trials are not listed anywhere, and are matched by drug and date.
• Age. Age enters only at the level of the trial (its eligible ages and its mean age), never the participant's.
• Pooling. Comparisons pool doses and, in some trials, more than one antidepressant. A combination product enters only one sensitivity analysis.
• Coders. The coders, the third readers and the author are models of one family. They worked blind to the counts, but share whatever priors that family holds.
• Earlier estimates. Risk ratios here are compared with odds ratios from earlier analyses, which is reasonable for events this rare, but those analyses differ in population and in how they counted events.

Provenance

A person asked for a new study outside mathematics that could matter. The agent, a Claude-family model (claude-opus-5-5), proposed this one with two alternatives; the person chose it and approved registering the plan. The agent designed the study, searched ClinicalTrials.gov, wrote both codebooks, the plan and the code, and registered the plan. Separate instances of the same model screened the trials and coded the terms, two working independently and a third deciding where they differed. The agent then ran the analysis, wrote the claims and the first version's paper, and applied the hazard screen.

After the first version's reviews, the person asked whether its reviewers' findings could be addressed; the agent assessed them and proposed this correction, which the person asked for. The agent wrote the correction's code and paper. Separate instances of the same model searched the earlier papers and the FDA's reviews for the facts data/earlier.json records, and the trials' reports for the facts data/resolved.json records, quoting each source; the agent checked the estimates it uses against Stone and colleagues' Table 6 and Hammad and colleagues' abstract, checked each report's passage against the source, and matched trials to the reviews' study lists by protocol number. The data are the ClinicalTrials.gov records retrieved on 2026-10-13, carried unchanged in data/ctgov/. No person wrote, coded, analysed or reviewed any part of the work.