Published on 15 September 2026
Clozapine in treatment-resistant schizophrenia: no proven symptomatic superiority at 6-8 weeks
In brief
National and international guidelines place clozapine in treatment-resistant schizophrenia. This individual participant data meta-analysis brought together 19 blinded randomised trials comparing clozapine with other second-generation antipsychotics in treatment-resistant schizophrenia, 1599 participants in all, 1052 of them with data obtained patient by patient. On the primary outcome, change from baseline in the PANSS total score measured after six to eight weeks of treatment, the estimated mean difference is −0.64 points with a 95% credible interval running from −3.97 to 2.63, in favour of the other antipsychotics. Confidence in this evidence is graded very low. The analysis therefore finds no superiority of clozapine on short-term symptoms. Neither does it demonstrate equivalence: the interval remains compatible with an advantage in either direction, and the outcomes that actually drive a clozapine prescription, rehospitalisation, suicidality, mortality, are not reported in the published abstract. The authors conclude in favour of prudent use given the side-effects, not of abandoning the drug.
The context
Clozapine owes its place to a trial from 1988. Kane and colleagues compared clozapine with chlorpromazine in 268 patients selected by a demanding protocol: at least three periods of treatment over the preceding five years, with neuroleptics from at least two different chemical classes, at a dose of at least 1000 mg per day of chlorpromazine equivalent for six weeks, followed by a prospective haloperidol phase without improvement. Thirty per cent of responders on clozapine, four per cent on chlorpromazine. The result was clear-cut, and it grounded forty years of guidelines.
That is not, however, the question that arises in the consulting room. There it runs like this: faced with a patient who is no better after two adequately conducted trials, the threshold set by the TRRIP consensus criteria (Howes et al., 2017), should one move to clozapine with its haematological monitoring, or try a third second-generation molecule? The relevant comparator is no longer a first-generation antipsychotic at high dose, it is olanzapine or risperidone. And in practice clozapine is often started late, well beyond two lines of treatment, which is an observation from the field and not a result of this study.
Three situations are easily run together, and this analysis addresses only one of them. Non-response after a first antipsychotic trial, where the available next steps have been compared inside a single network, is the subject of our analysis of what comes after the first antipsychotic. Established resistance, defined by at least two documented and adequate failures, is the population studied here. Established resistance in a patient who cannot receive clozapine, where the augmentation strategies have been appraised on their own, is covered in our analysis of non-clozapine augmentation. Carrying a result from one of the three over to the other two is the commonest error on this ground.
The study at a glance
| Question (PICO) | |
|---|---|
| Population | |
| Participants with treatment-resistant schizophrenia. The published abstract does not detail the resistance criteria used by each trial. 19 trials, 1599 participants. Individual participant data available for 12 trials and 1052 participants: mean age 37.67 years (SD 11.24; range 10 to 66 years), 348 women (33.08%) and 704 men (66.92%). Data on ethnicity not available | |
| Intervention | |
| Clozapine | |
| Comparator | |
| Other second-generation antipsychotics, “mainly olanzapine and risperidone” in the authors’ own reading | |
| Primary outcome | |
| Change in the total score on the Positive and Negative Syndrome Scale (PANSS) after 6 to 8 weeks of treatment, expressed as a mean difference with a 95% credible interval | |
| Design | |
| Systematic review and individual participant data meta-analysis. Cochrane Schizophrenia Group register searched from inception to 24 January 2024, plus previous reviews. Blinded randomised trials of at least 6 weeks. Bayesian random-effects meta-regression including duration of illness, baseline severity and sex as potential prognostic factors or treatment effect modifiers. Confidence assessed with GRADE. PROSPERO registration CRD42021254986 |
What an individual participant data meta-analysis changes deserves to be said plainly. Instead of adding up published means, it goes back to the patient-by-patient rows and allows adjustment on individual characteristics, so it can look for who responds better to what. It is the most solid device available short of a new trial. It remains dependent on one thing the authors do not control: the willingness of investigators to hand over their data, or not.
The quality check
| Criterion | Status |
|---|---|
| Registered protocol | Sound |
| FindingPROSPERO CRD42021254986, a number that appears in the published abstract | |
| Identification of trials | Sound |
| FindingThe Cochrane Schizophrenia Group specialised register searched from inception, 13,876 references screened | |
| Access to individual participant data | Reservation |
| Finding12 trials out of 19, 1052 participants out of 1599. The abstract does not state whether the primary estimate combines individual and aggregate data | |
| Primary outcome | Sound |
| FindingTotal PANSS, a 6 to 8 week window, the standard measure in the field | |
| Length of observation | Reservation |
| FindingTrials of at least 6 weeks, outcome read between 6 and 8 weeks, while resistance is a chronic situation | |
| Study withdrawals | Reservation |
| FindingThe authors explicitly list premature study discontinuation among their sources of uncertainty | |
| GRADE confidence | Very low |
| FindingThe lowest level of the scale. An estimate at this level can be overturned by future data | |
| Representativeness | Reservation |
| Finding33.08% women, no data on ethnicity, an age range of 10 to 66 years as it stands in the abstract. The lower bound would imply a participant aged 10 in a trial of treatment-resistant schizophrenia, which is unlikely: barring a data entry artefact, to be confirmed against the full text | |
| Funding | Reservation |
| FindingThe published abstract declares the German federal ministry of education and research. The funder index of the record carries six entries: Bundesministerium für Bildung und Forschung (01KG2015), Federal Ministry of Education and Research Berlin Office, China Scholarship Council (202006240091), Eli Lilly and Company, Bundesministerium für Bildung und Forschung (01EE2303B), Novartis. The most telling point: Eli Lilly markets olanzapine and Novartis markets clozapine, so the marketing authorisation holders of the main comparator and of the product under assessment both appear. Neither industrial entry carries a grant number, which may reflect the supply of trial data rather than funding. The full declaration must be read before any conclusion is drawn | |
| Competing interests | Reservation |
| FindingThe record carries no declaration of competing interests: it remains to be consulted in the full text. The author list includes investigators from the original trials, which partly explains the availability of individual participant data. That is both the strength of the design and a point of independence to document | |
| Involvement of people concerned | Sound |
| FindingPeople with lived experience of mental illness took part in the work, as the abstract states | |
The findings
| Outcome | Result |
|---|---|
| Overall symptoms, total PANSS | Mean difference −0.64 (95% CrI −3.97 to 2.63) |
| PEB readingNo superiority demonstrated on a scale running from 30 to 210 points, both bounds of the interval stay below what a clinician perceives in consultation | |
| Heterogeneity between trials | τ = 2.68 |
| PEB readingSubstantial residual variability not explained by the factors taken into account | |
| Prognostic factors | Shorter duration of illness and higher baseline severity associated with a larger reduction in symptoms |
| PEB readingConsistent with clinical experience but these factors say who will improve, not which molecule to choose | |
| Treatment effect modifiers | Neither those factors nor sex modify the relative effect between clozapine and the comparators |
| PEB readingNo favourable subgroup identified a negative search for modifiers is not the same as their absence: this kind of analysis is almost always underpowered | |
| Confidence in the evidence | GRADE very low |
| PEB readingMajor uncertainty this is the most important item in the abstract, ahead of the point estimate itself | |
| Outcomes not reported | Rehospitalisation, suicidality, mortality, treatment continuation: not reported in the published abstract |
| PEB readingInformation unavailable the abstract describes only the primary outcome and says nothing about the presence or absence of secondary outcomes. Yet those are the commonest reasons for starting clozapine | |
Two things overlap here and are worth separating. Precision first. The interval runs from four points in favour of the comparators to two and a half points in favour of clozapine, on a scale with an amplitude of 180 points: it rules out a moderate effect in either direction. The estimate is therefore precise, only its sign remains undetermined. Confidence next. GRADE rates it very low, because of risk of bias, missing data and heterogeneity between the original trials. An estimate that is precise but unreliable can be shifted by future data, which is not the same thing as a question left open for want of precision.
What is demonstrated. On symptoms measured after six to eight weeks of treatment, the whole set of available blinded randomised trials shows no superiority of clozapine over the other second-generation antipsychotics, and the point estimate leans very slightly towards the comparators.
What is suggested. Part of this null result probably comes from the definition of the population and from the length of follow-up. The authors themselves list the unavailability of some data, uncaptured sources of variance and premature study discontinuation among their limitations.
What is opinion. The idea that clozapine retains an advantage on outcomes not reported in the published abstract, rehospitalisation, suicidality, mortality, remains a hypothesis. It is supported by observational cohorts, it is not demonstrated by a randomised trial.
Critical appraisal
| Domain | Judgement |
|---|---|
| Question and protocol | Sound |
| FindingPre-specified question, registered protocol, a single primary outcome | |
| Literature search | Sound |
| FindingSpecialised register, no lower date limit, search closed on 24 January 2024, 13,876 references screened | |
| Missing individual participant data | Major limitation |
| FindingSeven trials out of nineteen did not supply their data. There is no reason for that non-supply to be random: it depends on how old the trial is, on the sponsor, on whether the team still exists. Nothing guarantees that the trials that were handed over resemble the others | |
| Definition of the exposure | Reservation |
| FindingThe published abstract does not detail the resistance criteria used by each trial. What follows is a hypothesis of critical reading and not a result of the study: the broader the resistance criterion, the more the population includes patients who are not truly resistant, in whom no specific advantage of clozapine is expected, and the more the estimate is pulled towards zero. Neither the variability of the definitions nor the dilution it would produce can be quantified from the abstract | |
| Nature of the outcome | Reservation |
| FindingA change in PANSS score is a surrogate for the clinical outcome. It says nothing about staying on treatment, about admission to hospital, or about survival | |
| Power and precision | Reservation |
| Finding1052 participants with individual data. The credible interval, from −3.97 to 2.63 points on a scale whose amplitude is 180 points, is narrow relative to that scale: the estimate is precise and a moderate effect is excluded. What is very low is the GRADE confidence, because of risk of bias, missing data and heterogeneity between the original trials. The uncertainty therefore bears on the reliability of the measurement, not on its precision | |
| Publication bias and competing interests | Reservation |
| FindingSome of the trials comparing clozapine with a newer antipsychotic were run at a time when the comparator was still in development. This point is not addressed in the published abstract: we flag it as context to be checked trial by trial, not as a result of this meta-analysis. The record moreover carries no declaration of competing interests, which remains to be consulted in the full text. The author list includes investigators from the original trials, which partly explains the access to individual participant data: that is the strength of the design as much as a point of independence to document | |
| Fit between claim and evidence | Sound |
| FindingThe authors write that they did not provide evidence for superior efficacy, not that the treatments are equivalent. The wording is exact, and it is more cautious than the reading that will be made of it | |
Level of evidence
PEB appraisal: high confidence in the observation, very low confidence in the conclusion one would be tempted to draw from it. What is firmly established is that there is today no randomised demonstration of a short-term symptomatic superiority of clozapine over the other second-generation antipsychotics. The observation is an old one, and this analysis confirms it with the best tool available. What remains highly uncertain is the direction of the true difference and the reliability of the estimate itself, graded very low, and above all what happens beyond six to eight weeks and on the outcomes that matter. The high scientific score rewards the method and the honesty of the wording, not the certainty of the result.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ This does not make me stop a clozapine, it makes me check what I am calling resistance. Two adequately conducted trials, at a proper dose, for long enough, with adherence verified: if I do not have that in the notes, I am not looking at resistance, I am looking at an incomplete file. And six weeks of PANSS tells me nothing about what I prescribe clozapine for, which is to keep the patient out of hospital and alive. ”
What this means in practice: the point of vigilance moves from the molecule to the indication. This analysis does not call into question the use of clozapine in a genuinely resistant patient, it is a reminder that the quality of the evidence depends on the quality with which resistance is defined, in trials as in clinical notes.
What you can do with this on Monday morning. The course of action does not change, the way it is documented does.
- Write down, before any initiation, the two trials that establish resistance: molecule, dose, a duration of at least six weeks at an effective dose, adherence verified where possible by a plasma level or by a long-acting injectable formulation. That is exactly what was missing from some of the trials analysed here.
- Check that olanzapine and risperidone have been tried at a sufficient dose before concluding that a patient is resistant, since those are the comparators that carry most of the signal in this analysis.
- Do not delay clozapine in a patient whose resistance is established. The absence of evidence of superiority on a symptom score at six to eight weeks is not evidence of equivalence, and the published abstract does not report the outcomes that motivate initiation. The converse holds too, and it is the authors’ own position: they write that, given the side-effects of clozapine, the observed uncertainty regarding its superiority “warrants prudent use and further research”. Caution therefore applies on both sides, to the indication as much as to the monitoring.
- Set up haematological monitoring before the first dose, following the scheme in force where you practise. Neutrophil thresholds, the frequency of counts and the rules for restarting after an interruption are laid down by the applicable national text, and that text is what to check rather than a remembered schedule.
- Keep the rest of the framework in place: the monitoring record, slow titration, vigilance for myocarditis over the first two months, active screening for constipation. Bowel obstruction is a serious and underdiagnosed gastrointestinal complication whose lethality exceeds that of agranulocytosis in several published series.
- Prepare a short answer for the patient or family who will have read a headline announcing that clozapine does no better than the others. The accurate sentence fits on one line: it has not been shown to do better on symptoms at two months, and it has not been shown to do the same on everything else.
- The course of action is set out in the NICE decision tree for schizophrenia in adults.
Frequently asked questions
Does this study mean that clozapine is not superior to the other antipsychotics?
No. It shows that there is no evidence of superiority on a short-term symptom outcome, with confidence graded very low. The estimate is nonetheless precise: the credible interval rules out a moderate effect on the PANSS. What is fragile is the reliability of the data that produce it, between risk of bias, missing data and heterogeneity between trials. An absence of a significant difference is not a demonstration of an absence of effect.
Should clozapine be offered sooner, or later?
Nothing here justifies delaying an initiation. The operational message concerns the rigour of the preceding step: two documented trials, at an effective dose, for long enough, with adherence verified. Poorly established resistance exposes the patient to two symmetrical errors, prescribing clozapine to someone who does not need it, or giving it too late to someone who does.
How does this result sit alongside the cohorts reporting lower mortality on clozapine?
By holding both, without merging them. The Finnish FIN20 cohort, which follows 62,250 patients over twenty years, associates clozapine with the best mortality figures of all the molecules studied, with an adjusted hazard ratio of 0.39 for all-cause mortality and 0.21 for suicide. These are associations, not demonstrated effects: being on clozapine depends on who was selected, followed and monitored, and haematological monitoring imposes regular medical contact by construction, itself a factor in better health. The methodological detail matters: in that work, the morbidity analyses compared each patient with himself, which neutralises much of the selection bias, whereas the mortality analyses rested on between-individual comparisons, where confounding by indication is hardest to rule out. The abstract of the randomised trials analysed here reports no mortality data, and their six to eight week duration would not lend itself to it. The two bodies of evidence do not answer the same question, and neither refutes the other. This is a real tension in the literature, not a debate already settled.
A difference of 0.64 points on the PANSS, what does that represent?
Nothing perceptible. The scale runs from 30 to 210 points. Even taking the bound least favourable to clozapine, around four points, one stays well below what is visible in consultation. An effect was indeed estimated, and estimated precisely, since the interval excludes a moderate effect. The debate is therefore not about the size of the observed effect, it is about the confidence that can be placed in the data producing it, graded very low.
Should this be raised with a patient already stable on clozapine?
Not spontaneously, and certainly not as a challenge to the treatment. In a stable patient, the question of comparative efficacy at the start of treatment no longer applies. Stopping an effective clozapine exposes the patient to a relapse risk that clinical experience judges to be high: that point is expert opinion here, it does not come from the analysis under discussion. If the patient raises the subject, the distinction between absence of evidence and evidence of absence is exactly what needs to be conveyed.
Annotated bibliography
Schneider-Thoma J, Hamza T, Chalkou K, et al. (2025). Efficacy of clozapine versus second-generation antipsychotics in people with treatment-resistant schizophrenia: a systematic review and individual patient data meta-analysis. The Lancet Psychiatry, 12(4), 254-265. DOI 10.1016/S2215-0366(25)00001-X · PMID 40023172. The source study analysed here. Declared funding: German Ministry of Education and Research. PROSPERO registration CRD42021254986.
Kane J, Honigfeld G, Singer J, Meltzer H (1988). Clozapine for the treatment-resistant schizophrenic. A double-blind comparison with chlorpromazine. Archives of General Psychiatry, 45(9), 789-796. DOI 10.1001/archpsyc.1988.01800330013001 · PMID 3046553. The founding trial: 268 patients randomised after at least three periods of treatment over the preceding five years, with neuroleptics from at least two different chemical classes, at a dose of at least 1000 mg per day of chlorpromazine equivalent for six weeks, then a prospective haloperidol phase. 30% of responders on clozapine against 4% on chlorpromazine at six weeks. Two major differences from the meta-analysis analysed here: the comparator is a first-generation antipsychotic, and resistance is defined far more strictly, with a prospective verification phase. This study therefore does not answer the same question, which explains part of the gap between the two results.
Howes OD, McCutcheon R, Agid O, et al. (2017). Treatment-Resistant Schizophrenia: Treatment Response and Resistance in Psychosis (TRRIP) Working Group Consensus Guidelines on Diagnosis and Terminology. American Journal of Psychiatry, 174(3), 216-229. DOI 10.1176/appi.ajp.2016.16050503 · PMID 27919182. Consensus criteria for defining and documenting treatment resistance. This is an agreement between experts and not an experimental result, but it is the tool that most of the trials gathered in the meta-analysis lacked. It is also the text that sets the threshold of two adequately conducted antipsychotic trials before resistance can be spoken of.
Taipale H, Tanskanen A, Mehtälä J, Vattulainen P, Correll CU, Tiihonen J (2020). 20-year follow-up study of physical morbidity and mortality in relationship to antipsychotic treatment in a nationwide cohort of 62,250 patients with schizophrenia (FIN20). World Psychiatry, 19(1), 61-68. DOI 10.1002/wps.20699 · PMID 31922669. The observational counterpoint. Over a median follow-up of 14.1 years, clozapine is associated with the lowest mortality hazard ratios: 0.39 (95% CI 0.36 to 0.43) for all causes and 0.21 (0.15 to 0.29) for suicide, with a cumulative mortality at twenty years of 15.6% on clozapine against 25.7% on any antipsychotic and 46.2% with no treatment. Limits that are decisive for interpretation: a register study without randomisation and, above all, morbidity analysed by comparing each patient with himself while the mortality results rest on between-individual models, therefore exposed to confounding by indication and to the selection bias of the adherent, monitored patient. These figures do not demonstrate a causal effect of clozapine on survival, they justify not concluding from six to eight week trials alone.