The Placebo Arm Is Not Neutral: Who Leaves, and Why, Across Nine Psychiatric Diagnoses
- Systematic review and random-effects meta-analysis of placebo arms only, with no active-arm comparison anywhere in the paper. For each of nine psychiatric diagnoses the authors took the most recent high-quality systematic review, then drew the ten highest-quality placebo-controlled RCTs from it, giving 90 trials; 86 reported total dropouts, covering 10,056 participants who received placebo. The protocol was preregistered (OSF u469a).
- Total dropout under placebo differed significantly across diagnoses (Q = 82.2, df = 8, p < 0.001). Schizophrenia was highest at 0.41 (95% CI 0.35–0.48). Major depressive disorder was among the lowest at 0.17 (95% CI 0.15–0.19), as was ADHD at 0.17 (95% CI 0.11–0.25).
- The dominant reason for leaving was lack of effect, not adverse events. Dropout for lack of effect ran from 0.19 in schizophrenia (95% CI 0.15–0.23) down to 0.04 in major depression (95% CI 0.03–0.06), Q = 71.3, df = 8, p < 0.001. Dropout for adverse events was lower in every diagnosis but one: it peaked at 0.07 in OCD (95% CI 0.05–0.09) and bottomed out at 0.02 in ADHD (95% CI 0.01–0.03), Q = 32.1, df = 8, p < 0.001. OCD is the exception the authors single out, the one diagnosis in which adverse events were the commonest stated reason for leaving.
- The honest limit of the design: these are aggregate published data from placebo arms, not individual-participant data, and no active treatment was analysed, so nothing here can speak to drug efficacy. Heterogeneity for total dropout was high, with I² above 75% in five of the nine diagnoses, and the Egger test for total dropout gave p = 0.004.
Every trial you read has a control arm, and most of us never ask what happened to the people in it. This meta-analysis asks nothing else. It takes the placebo groups of 90 high-quality psychiatric trials and measures what an inert treatment costs the person assigned to receive it.
What the placebo arms actually showed
Bschor and colleagues ran no new trials. Working from a preregistered protocol and the Cochrane Handbook, they identified the most recent high-quality systematic review for each of nine psychiatric diagnoses, pulled the ten highest-quality placebo-controlled RCTs from each review, and analysed the placebo arms alone. Ninety trials entered; 86 reported total dropouts across 10,056 placebo recipients, 82 reported dropouts attributed to adverse events, and 71 reported dropouts attributed to lack of effect. Rates were pooled with a random-effects model (DerSimonian and Laird) on logit-transformed proportions. There is no active comparator in this analysis. That absence is the single most important thing to carry into the numbers.
Total dropout differed sharply by diagnosis (Q = 82.2, df = 8, p < 0.001). Schizophrenia trials lost 0.41 of the placebo group (95% CI 0.35–0.48), with panic disorder and mania also near the top. At the other end sat PTSD, major depressive disorder at 0.17 (95% CI 0.15–0.19) and ADHD at 0.17 (95% CI 0.11–0.25). Splitting by stated reason sharpens the picture considerably. Dropout for lack of effect ranged from 0.19 in schizophrenia (95% CI 0.15–0.23) to 0.04, reported as 3.8%, in major depression (95% CI 0.03–0.06), Q = 71.3, df = 8, p < 0.001. Dropout for adverse events peaked at 0.07 in OCD (95% CI 0.05–0.09) and bottomed out at 0.02 in ADHD (95% CI 0.01–0.03), Q = 32.1, df = 8, p < 0.001. Across the set the authors report that more than twice as many placebo recipients left for lack of efficacy as for adverse events, but they single out OCD as the exception, the one diagnosis in which adverse events were the commonest stated reason for leaving. In multivariable meta-regression, no factor other than diagnosis remained significantly associated with any dropout category.
The pooled estimates are not tidy, and the paper says so. Heterogeneity for total dropout was substantial, with I² above 75% in five of nine diagnoses; for lack-of-effect dropout it exceeded 75% in four. Only the adverse-event analysis was comparatively stable, with I² below 75% across all diagnoses. The Egger test for total dropout was p = 0.004, which is consistent with a small-study effect, and a trim-and-fill analysis imputing 18 studies shifted the overall rate from 0.25 to 0.30, which the authors describe as small although it is a fifth of the original estimate in relative terms. One measurement caveat deserves naming, because the authors themselves rest their design on it: dropout is a proxy for nocebo effects, chosen because it is reported consistently enough to compare across diagnoses. A low adverse-event dropout rate does not mean placebo recipients reported few unpleasant symptoms. It means those symptoms rarely reached the threshold of leaving the study.
For your practice
Read the direction of the finding carefully. The paper does not show that placebo harmed people. It shows that the commonest stated reason for leaving an inert treatment was that it was not working, which in a psychiatric trial means the illness went untreated, and that this is priced very differently by diagnosis. Schizophrenia, mania and panic disorder sat at the top of the table; in schizophrenia specifically, roughly four in ten placebo participants were gone before the trial ended. The authors suggest that symptom severity may make waiting under placebo hard to sustain in schizophrenia and mania, and that the distress of an attack may drive it in panic disorder, but this design cannot test either explanation. In major depression, where the episode is more often self-limiting, fewer than one in five left.
It matters just as much not to run the argument backwards. Nothing in this analysis licenses the claim that psychiatric drugs are merely placebo, because no active arm was examined at any point. Nor does it show that placebo is inert in the consulting room. The same group's earlier analysis in JAMA Psychiatry in 2024 found the largest symptom improvement under placebo in major depression and generalised anxiety disorder, and the smallest in mania, OCD and schizophrenia. The dropout pattern here lines up with that: where the placebo response was large, people stayed, with OCD again the exception the authors flag. The two analyses draw on the same 90 trials from the same project, so their agreement is partly structural rather than independent replication, but the direction is at least internally consistent.
The practical residue is about time, and it is an inference rather than something the paper tested. Dropout for lack of effect is a crude index of how long a patient will tolerate a treatment that is doing nothing, and across these diagnoses it varies by roughly a factor of five. That is a reasonable prior for setting your own review intervals. In acute psychosis and mania, treat a quiet non-response as an urgent finding and put an explicit early checkpoint in the plan rather than letting the first review drift to week six. In an uncomplicated depressive episode, a longer initial window is more defensible, and the corollary is that early improvement in that setting carries less information about your specific intervention than it feels like it does.
What goes wrong in a placebo arm is usually not a side effect but time spent not getting better, and how expensive that time is depends on the diagnosis.
Only placebo arms were analysed, with no active comparator, so the paper says nothing about how any drug performed. Dropout is a coarse proxy for nocebo effects rather than a direct measure of them, and only eight to ten trials per diagnosis were included. Heterogeneity was high for total dropout, with I² above 75% in five of the nine diagnoses, and the Egger test for total dropout gave p = 0.004, consistent with small-study effects. The authors note that differences between diagnoses may partly reflect variation in trial methodology and in how hard sites worked to retain participants, rather than the conditions themselves.