What was never demonstrated over the long term

The largest antidepressant trial ever conducted, re-measured

STAR*D was funded with $35 million from the NIMH — public money, not a company's — to answer what happens to real patients. The 2006 publication reported cumulative remission of 67% after up to four successive treatment steps. It became one of the most widely cited statistics in psychiatry.

In 2023 an independent group took the patient-level data and repeated the analysis faithfully to the study's own original protocol. The result, verbatim from the abstract: “in contrast to the reported 67%, the rate was 35.0% when the HRSD scale specified by the protocol and the inclusion criteria were used”.

The discrepancies the authors document — and which the original investigators dispute: the final publication measured remission with a non-blinded scale, whereas the protocol specified a blinded one and explicitly excluded clinician-administered measures as outcome measures; 931 patients were included without a blind-rated baseline score, among them 99 who already met the remission criterion at entry; 125 were included who were in remission at the start of a subsequent step; and 370 who withdrew after the first visit were excluded.

Two later reanalyses by the same group — preprints, not peer-reviewed — report remission sustained throughout the 12 months of follow-up at 3.1–8.4% (switch arm) and 4.9–12.5% (augmentation arm), the range reflecting stricter or looser criteria for the missing data.

The original investigators published a reply defending the 67%. The controversy is open — but the 2023 reanalysis is peer-reviewed and the data are public, so anyone can check who is right.

[26] Pigott HE, Kim T, Xu C, Kirsch I, Amsterdam J. BMJ Open 2023;13(7):e063095 · doi:10.1136/bmjopen-2022-063095 · [27][28] medRxiv preprints 2025 · [29] reply: Rush AJ et al. Am J Psychiatry 2023

The trials that “prove” long-term benefit are measuring something else as well

Every trial that supports long-term use has the same design: patients stabilised on the drug are randomised into “continue” and “switch abruptly to placebo”. The difference in relapses is recorded as a protective effect.

The problem: abrupt discontinuation causes withdrawal, and no rating scale distinguishes withdrawal from relapse. Part of what is measured as “relapse” may be a withdrawal syndrome.

In an analysis of 14 relapse-prevention trials submitted to the FDA (mean follow-up 38.9 weeks), the drug–placebo difference accumulates very early: 50.3% by week 6, 69.0% by week 12, 101.0% by week 24 — and then it stops. The authors argue that this distribution fits the timing of withdrawal, not protection from relapse, which would be spread evenly.

Hengartner's conclusion, verbatim: “at present there is no reliable evidence that long-term antidepressant treatment is beneficial”.

The other side, fairly stated: the largest withdrawal meta-analysis (79 studies, 21,002 patients) gives a difference of 8 percentage points within randomisation — enough to count, not enough to explain the whole phenomenon on its own. And the Récalt & Cohen reanalysis managed to extract usable data from only 14 of 30 trials, using a threshold they set themselves.

Where both sides now agree: that withdrawal and relapse overlap and are confounded in discontinuation trials. What is contested is how much of the measured difference is explained that way — and that has not been settled.

[30] Hengartner MP. Ther Adv Psychopharmacol 2020;10:2045125320921694 · [31] Hengartner MP, Plöderl M. Ther Adv Psychopharmacol 2021;11:20451253211032051 · [32] Récalt AM, Cohen D. Psychother Psychosom 2019;88(2):105–113 · [13] Henssler et al. Lancet Psychiatry 2024

With antipsychotics the short-term benefit is beyond dispute — the long-term one is not

I have to start from what is strong: a meta-analysis of 65 trials with 6,493 patients showed that antipsychotics reduce one-year relapse from 64% to 27% (RR 0.40; NNT 3). It is among the most solid findings in psychiatry and is not disputed here. The same study also records the cost: weight gain RR 2.07, movement disorders RR 1.55, sedation RR 1.50.

The question is what happens after the first year. In a randomised Dutch trial of first-episode psychosis, 103 of 128 patients were reassessed at seven years: 40.4% (21/52) of the dose-reduction group met the recovery criteria against 17.6% (9/51) of the maintenance group (χ²=8.2, p=0.004; adjusted OR 3.49, p=0.01, with no confidence interval reported). The difference lay entirely in functioning, not in symptoms.

The rebuttal, and it is a serious one: in the first two years that same reduction group had twice as many relapses (43% versus 21%). And a meta-analysis of 11 trials with 2,826 patients showed that patients on placebo deteriorate progressively over time, while those who continue remain stable — evidence in favour of maintenance.

There is also a 20-year observational study in which those not taking antipsychotics had better outcomes. But it is observational, not randomised: continuous prescribing over twenty years is a marker of more severe illness, so the comparison may reflect who became more severely ill rather than what the drug did. I report it with that limitation stated.

What ultimately stands: that dose reduction may bring functional benefit to a selected subgroup of first-episode patients, under close monitoring. Not that antipsychotics should be discontinued generally — that is not supported by the evidence and would be dangerous.

[33] Leucht S et al. Lancet 2012;379(9831):2063–71 · [34] Wunderink L et al. JAMA Psychiatry 2013;70(9):913–20 · [35] Harrow M, Jobe TH, Faull RN. Psychol Med 2014;44(14):3007–16 · [36] Harrow M, Jobe TH, Tong L. Psychol Med 2022;52(13):2681–91 · rebuttals: [37] Leucht S, Davis JM. Br J Psychiatry 2017;211(3):127–9 · [38] Takeuchi H et al. Br J Psychiatry 2017;211:137–43