Published on 16 September 2026
Resistant schizophrenia: what is a minus 1.53 effect size for duloxetine plus clozapine worth?
In brief
This frequentist network meta-analysis brings together 150 randomised trials, 11,375 patients and 78 drug options or placebo in schizophrenia that has not responded to prior treatment. The authors’ first conclusion is that clozapine remains the reference treatment, superior to haloperidol, chlorpromazine, quetiapine and sulpiride on overall symptoms (SMD 0.35 to 1.00). Nine combinations do better than clozapine alone on the primary outcome, the largest estimate going to intranasal oxytocin added to risperidone, SMD −1.67 (95% CI −2.63 to −0.71). Among the clozapine combinations, duloxetine carries the extreme estimate, SMD −1.53 (95% CI −2.44 to −0.63): it rests on a single double-blind trial of 40 randomised patients. Mirtazapine added to clozapine gives −1.31 (−2.43 to −0.19) on one trial, memantine −0.74 (−1.42 to −0.05) on two trials. Two reservations govern the reading: 104 of the 150 included trials are at moderate or high risk of bias, and Egger’s test shows that small trials overestimate efficacy relative to clozapine (p = 0.0027). The authors themselves write that the sparsity of the network increases the likelihood of chance findings for estimates resting on a single study. At this stage, this is a hypothesis to be confirmed, not a reason to change a prescription.
The context
The question addressed by this synthesis is the one that comes up at a follow-up appointment or in a team meeting: a patient with schizophrenia has already failed one or more treatments, is sometimes already on clozapine, and the improvement remains insufficient. The options then discussed are raising the dose, switching to another drug, moving to clozapine when that has not yet happened, or adding an adjunct to clozapine. These strategies are widely used, but they rest on scattered data: many small trials, heterogeneous definitions of non-response, few head-to-head comparisons between them.
A network meta-analysis is precisely the tool designed for this problem: it links trials that never compared the same treatments with one another and produces a ranking of all the options. It is also the tool whose results are most sensitive to the quality and distribution of the trials that feed it. The two properties go together.
The mechanism
What the augmentation hypothesis assumes, and what the study measured
The rationale for augmenting clozapine rests on the idea that adding a different pharmacological action, noradrenergic and serotonergic for duloxetine, noradrenergic and histaminergic for mirtazapine, glutamatergic for memantine, acts on symptom dimensions that dopamine blockade does not cover. The hypothesis is plausible and long-standing. It is not tested here.
| Item | Content |
|---|---|
| What the study measured | Clinical efficacy and tolerability scores, pooled across trials |
| DetailThe scales vary across the included trials. The published protocol sets a hierarchy: change in PANSS total score, failing that change in BPRS score, failing that endpoint values on either scale. Pooling different scales into a standardised mean difference is legitimate, but it adds a source of variability to the result. | |
| What the study did not measure | No biological or imaging parameter, and no usable plasma level |
| DetailThe authors state that plasma concentrations were not reported in most of the included trials, and that dose variability remains a confounding factor. No marker of receptor occupancy is measured. The result is an effect observed on scores, not a demonstration of mechanism. | |
| A competing hypothesis not ruled out | Part of the effect could be pharmacokinetic rather than pharmacodynamic |
| DetailClozapine is metabolised mainly by CYP1A2, a pathway shared by several drugs that are combined with it: a rise in clozapine plasma levels would also produce an improvement in scores. This is a pharmacological reminder, outside the scope of the study, which does not report plasma levels in most trials and did not attempt to separate the two mechanisms. | |
The study at a glance
| Item | Content |
|---|---|
| Population | Schizophrenia not responding or resistant to prior treatment, mean age 38.0 years, 71% men |
| DetailThe proportion of men is 7,818 out of 11,076 in the 142 trials that report it. Most trials involve working-age adults; three include children or adolescents. Any definition of resistance or non-response was accepted, which the authors acknowledge: the definitions are then sorted into four groups, from the failure of a single antipsychotic to non-response to clozapine. This is clinical heterogeneity, not only statistical heterogeneity. | |
| Intervention | 78 drug options or placebo: antipsychotic monotherapies and combinations |
| Detail69 trials assess an antipsychotic monotherapy, six of them with a placebo arm, and 81 assess a combination. The adjuncts include antipsychotics, antidepressants, antiepileptics, glutamatergic agents and various other products, added to clozapine but also to risperidone or olanzapine. | |
| Comparator | Multiple comparisons within the network, with clozapine as the reference in every plot |
| DetailIn a network of this size, the great majority of treatment pairs have never been tested in a head-to-head trial: the estimate is then derived from the network. See the critical appraisal. | |
| Primary outcome | Change in overall symptom score, expressed as a standardised mean difference |
| Detail116 trials and 8,678 participants provide usable data for this outcome. Secondary outcomes: treatment response, positive, negative and depressive symptoms, discontinuation for any reason, for adverse effects and for inefficacy, overall and specific adverse effects (use of antiparkinsonian medication, sedation, weight gain, prolactin, QTc), quality of life and functioning. | |
| Design | Systematic review and frequentist random-effects network meta-analysis, 150 randomised trials, N = 11,375 |
| DetailFrequentist inverse-variance model, ranking by P-scores, run in R with the netmeta package, not a Bayesian model. Searches of MEDLINE, Cochrane Central, Embase, PsycINFO, ClinicalTrials.gov, the WHO trials registry and the FDA website up to March 2025. The 150 trials correspond to 154 publications, published from 1958 to 2024, 52 of them conducted in the USA; 127 are double-blind. Duration 4 to 52 weeks, median 12 weeks. Protocol preregistered on OSF, conduct declared compliant with PRISMA, network consistency tests, certainty assessed with CINeMA. | |
Quality control
| Criterion | Judgement |
|---|---|
| Preregistered protocol | Yes, OSF |
| FindingOSF registration dated 9 August 2024, reproduced in full in the appendix. Preregistration limits the risk of a retrospective choice of primary outcome. It does not protect against the biases of the included trials. | |
| Conduct declared compliant with PRISMA | Yes |
| FindingPRISMA checklist and flow diagram provided in the appendix: 15,615 records identified, 12,795 screened after removal of duplicates, 150 trials retained. Declared compliance concerns the reporting of the review, not the quality of the data gathered. | |
| Risk of bias of the included trials | 104 trials out of 150 |
| Finding46 trials (30.67%) at low overall risk, 75 (50%) moderate, 29 (19.33%) high, that is 69% at moderate or high risk. This is the heaviest limitation of this synthesis: a network does not correct the biases of the trials that feed it, it propagates them. | |
| CINeMA certainty for the primary outcome | Very low to moderate |
| FindingThe judgements range from very low to moderate, with no comparison graded high; the published abstract states that confidence is low or very low for most estimates. Moderate certainty is not specific to duloxetine: in the assessment table for the primary outcome, more than a hundred comparisons receive this grade, among them memantine plus clozapine versus clozapine alone. | |
| Network consistency | Tested, no global inconsistency |
| FindingFor the primary outcome, global design-by-treatment test Q = 16.594, 19 degrees of freedom, p = 0.617; 15.38% of loops show local inconsistency on separation of direct and indirect evidence. Heterogeneity remains notable: I² = 58.6%, between-study variance 0.0841. These tests say nothing about comparisons for which there is no direct evidence. | |
| Small-study effect | Demonstrated, p = 0.0027 |
| FindingContour-enhanced funnel plot and Egger’s test on the comparison of all treatments with clozapine: t = −3.10, 72 degrees of freedom, p = 0.0027, bias estimate −1.60. The authors conclude that smaller trials tend to exaggerate efficacy relative to clozapine. This result bears directly on the extreme estimates of the network. | |
| Funding | None |
| FindingThe funding section of the published version states that there was no funding source for this work, industrial or otherwise. | |
| Authors’ competing interests | Two authors out of eleven |
| FindingThe first author declares consultancy, advisory or lecture honoraria from Recordati, Lundbeck and Viatris. The senior author declares them from a longer list, including Angelini, Boehringer Ingelheim, Janssen, Johnson & Johnson, Lundbeck, MSD, Otsuka, Recordati, Sanofi Aventis, Sandoz, Sunovion, Teva, Eisai, Rovi and Mitsubishi. The other authors declare no financial relationships with commercial interests. These declarations invalidate nothing in themselves; they justify reading the ranking of options with the same scrutiny as any ranking. | |
The findings
| Outcome | Value and status |
|---|---|
| Clozapine versus several antipsychotics | SMD 0.35 to 1.00 in favour of clozapine |
| FindingSuperiority on overall symptoms over haloperidol, chlorpromazine, quetiapine and sulpiride. On positive symptoms, clozapine slightly outperforms olanzapine, SMD 0.19 (95% CI 0.00 to 0.37), lower bound at zero. On response rates, it outperforms risperidone, OR 0.64 (95% CI 0.41 to 1.01), an interval that crosses one. This is the result the authors put forward: clozapine remains the reference treatment. | |
| Duloxetine plus clozapine versus clozapine alone | SMD −1.53, 95% CI −2.44 to −0.63, a single trial, moderate certainty |
| FindingThe direct evidence comes down to one Italian double-blind trial from 2011, 40 randomised patients, duloxetine 60 mg per day added to clozapine for 16 weeks, in partial responders on clozapine at the maximum tolerated dose for at least one year. CINeMA certainty is graded moderate, downgraded for reporting bias only. | |
| Mirtazapine plus clozapine versus clozapine alone | SMD −1.31, 95% CI −2.43 to −0.19, a single trial, low certainty |
| FindingPoints in the same direction as the previous estimate, with an interval whose upper bound comes close to no effect. Certainty downgraded for within-study risk of bias, reporting bias and indirectness. | |
| Memantine plus clozapine versus clozapine alone | SMD −0.74, 95% CI −1.42 to −0.05, two trials, moderate certainty |
| FindingA more modest estimated effect, carried by two double-blind placebo-controlled trials of 21 and 52 patients, with the same moderate certainty as duloxetine. Amisulpride added to clozapine, supported by four trials, gives the smallest estimate in the series, SMD −0.51 (95% CI −1.00 to −0.02). | |
| The largest estimate for the primary outcome | Intranasal oxytocin plus risperidone, SMD −1.67, 95% CI −2.63 to −0.71 |
| FindingSo it is not a clozapine combination that carries the largest figure in the network, but a hormonal adjunct to risperidone, on a trial of 40 patients. Nine combinations in all do better than clozapine alone on this outcome, with estimates ranging from −1.67 to −0.51. | |
| Positive and negative symptoms | Same drugs, same orders of magnitude |
| FindingAdded to clozapine, amisulpride, lamotrigine and topiramate reduce positive symptoms more than monotherapy (SMD −1.13 to −0.54); duloxetine, memantine and ziprasidone reduce negative symptoms more (SMD −1.98 to −0.99). The same reservations about network sparsity apply. | |
| Tolerability | Analysed, and unfavourable to clozapine on several items |
| FindingClozapine causes more sedation and more weight gain than several comparators, and less use of antiparkinsonian medication. Discontinuations for adverse effects are more frequent on chlorpromazine and on topiramate added to clozapine than on clozapine alone. No difference is shown on quality of life or on functioning. The short duration of the trials makes these outcomes uninformative about long-term safety. | |
| Long-term safety of combinations | Not assessed |
| FindingAn absence of analysis, not a reassuring result. With a median duration of 12 weeks, the included trials do not allow any conclusion about the risks of a combination maintained for months or years. The authors themselves call for studies on long-term outcomes. | |
Critical appraisal
| Domain | Judgement |
|---|---|
| Size of the main effect | A single 40-patient trial |
| FindingA standardised mean difference of −1.53 corresponds to a very large effect, of an order of magnitude well above what the psychopharmacology of schizophrenia usually reports, and this in a population already selected by the failure of previous treatments, that is, where the expected effects are smallest. Verification confirms the feared profile: the direct evidence comes down to a single-centre trial of 40 randomised patients, and the confidence interval, from −2.44 to −0.63, spans almost two standard deviation units. Egger’s test run by the authors also shows that the small trials in this network exaggerate efficacy relative to clozapine. We do not conclude that the result is wrong: we say that a 40-patient trial does not establish a magnitude. | |
| Direct versus derived comparisons | An essential distinction |
| FindingA network with 78 nodes has more than three thousand possible pairs, and 150 trials can document only a small fraction of them directly: in the certainty table for the primary outcome, the vast majority of comparisons are recorded with zero direct trials. An indirect comparison is not illegitimate, but it inherits the weaknesses of the two branches that produce it and assumes that the trial populations are similar enough to be linked. For duloxetine added to clozapine, a single direct trial feeds the comparison, which is exactly the situation the authors flag as most exposed to chance findings. | |
| Risk of bias of the included trials | High overall |
| Finding104 trials out of 150 are at moderate or high risk of bias. Small trials of lower methodological quality produce larger effects on average, a phenomenon documented here by Egger’s test, and a network is particularly sensitive to it because a thinly supplied node then weighs very heavily on its estimate. | |
| What moderate CINeMA certainty actually means | Not to be overinterpreted |
| FindingModerate CINeMA certainty means that the true effect is probably close to the estimated effect, but that it could differ substantially. It does not guarantee the accuracy of the magnitude. Nor is it a privilege of duloxetine: more than a hundred comparisons for the primary outcome receive this grade, and none receives the high grade. For duloxetine, the one-level downgrade comes from reporting bias alone, which leaves open the question of whether a single trial of 40 patients justified an imprecision judgement with no concerns. | |
| Heterogeneity in the definition of non-response | Unresolved |
| FindingThe trials did not define non-response or resistance in the same way, and the authors accepted all definitions, from the failure of a single antipsychotic to non-response to clozapine. Grouping non-response to a single antipsychotic with documented resistance to two adequately conducted treatments does not describe the same clinical population, and the latter is the one that causes difficulty in practice. The analyses by criteria group do not substantially change the results. | |
| Heterogeneity of scales | A source of variability |
| FindingPANSS and BPRS do not measure the same dimensions with the same sensitivity. Standardisation as a standardised mean difference allows them to be pooled but makes the numerical value dependent on the dispersion of the source populations, which can inflate an effect when samples are small and homogeneous. | |
| Publication bias and small-study effect | Documented by the authors |
| FindingThe contour-enhanced funnel plot and Egger’s test (t = −3.10, df = 72, p = 0.0027) establish that small trials overestimate efficacy relative to clozapine. This is an acknowledged and quantified limitation, which weighs first on estimates resting on one or two small trials. | |
| Robustness to sensitivity analyses | Stable results |
| FindingNine prespecified sensitivity analyses were conducted, including exclusion of the quartile of smallest trials, exclusion of trials at high risk of bias, restriction to double-blind trials and an extreme analysis combining several restrictions. The authors report that the sensitivity, subgroup and meta-regression analyses do not substantially change the results, with the exception of baseline severity: the higher it is, the greater the efficacy of clozapine. One reservation remains: these analyses concern the network as a whole and not the duloxetine comparison alone. | |
| Independence | Competing interests declared |
| FindingNo funding was received for this work, which is favourable. Two of the eleven authors also declare links with several pharmaceutical companies. The two facts coexist and must be reported together. | |
| Claim of priority | Authors’ qualifier, reported verbatim |
| FindingThe authors write, word for word: “To our knowledge, this study is the largest NMA on the pharmacological treatment of TRS.” The claim is therefore bounded by “To our knowledge” and by the scope of treatment-resistant schizophrenia. We report it as it stands and do not adopt it as our own. | |
Level of evidence
The design corresponds to level 1a of the Oxford Centre for Evidence-Based Medicine classification: a systematic review with meta-analysis of randomised trials. That level describes the architecture of the evidence, not its strength. Here, the architecture is first-rate and the materials are uneven.
Confidence is high on three points: the question asked is clear, the methodological conduct is complete and verifiable (preregistered protocol, PRISMA, consistency tests, nine sensitivity analyses, CINeMA assessment comparison by comparison, investigation of the small-study effect), and the central result, the superiority of clozapine over several antipsychotics, rests on the best-supplied nodes of the network, with up to eleven trials for the comparison with olanzapine.
Confidence is low on the magnitude of the combination estimates. An SMD of −1.53 in schizophrenia not responding to treatment would, if confirmed, exceed what is usually observed with an antipsychotic versus placebo in an unselected population. We now know what carries it: one trial, 40 patients, one centre. The same remark applies to mirtazapine, one trial, and to a lesser extent to memantine, two trials. These values should be treated as hypotheses and not as measurements, and the Egger’s test published by the authors indicates the likely direction of their correction.
Finally, an isolated estimate does not make a proof. The authors themselves conclude that the sparsity of the network increases the likelihood of chance findings for estimates based on a single study, and call for research comparing combinations without clozapine with high-dose clozapine monotherapy. The present work should be read as one piece of evidence to add to a contested file, not as its conclusion.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ Minus 1.53 in patients who have already failed everything, frankly, that makes me raise an eyebrow. So you look at what is behind it: one trial, forty patients, sixteen weeks. There you are, that settles it. It is not nothing, it is double-blind, but it is not enough to change the way I prescribe either. What speaks to me more is that mirtazapine and memantine point the same way, so there may be something on the augmentation side. For my part, with a patient on clozapine who has plateaued, before adding an antidepressant I first check adherence, the clozapine level and smoking. And if I had to combine, I would not do it without an ECG and close monitoring. ”
What this means in practice: this work justifies reopening the question of clozapine augmentation and raising it in team meetings, not putting it into practice on the strength of a 40-patient trial. Before considering an adjunct, the ordinary causes of an insufficient response (adherence, dose, plasma level, comorbid substance use, smoking and its effect on clozapine metabolism) remain the first things to explore.
What you can do with this
- What to take away first. The main message of this synthesis is not duloxetine, it is that clozapine remains the reference treatment for treatment-resistant schizophrenia, superior to haloperidol, chlorpromazine, quetiapine and sulpiride, and that its advantage over olanzapine and risperidone is real but narrow.
- How to understand the augmentation signal. Several clozapine combinations do better than monotherapy on overall symptoms, which makes the hypothesis interesting. The magnitudes reported, however, often rest on one or two small trials, and the published Egger’s test shows that these trials pull the effects upwards.
- What to say to a patient or family who has read something. That recent research is exploring the addition of a second medicine to clozapine when it is not enough, that the first results are encouraging but sometimes rest on a single study of a few dozen patients, and that there is currently no reason to change a treatment that is partially working on the basis of these results.
- What to watch if the question of a combination arises nonetheless. In France, for example, adding an antidepressant to clozapine is a use outside the marketing authorisation, and the Public Health Code makes such prescribing conditional on the absence of an appropriate authorised alternative. Wherever you practise, the rules governing this kind of prescribing should be checked in their current wording before any decision.
- A pharmacological reminder, outside the scope of this study. Clozapine is metabolised mainly by CYP1A2. Fluvoxamine is a potent and well-characterised inhibitor of it; duloxetine is itself a CYP1A2 substrate, which makes a pharmacokinetic interaction plausible. A combination of this kind therefore requires anticipating a possible rise in concentrations, with the attendant risk of sedation and lowering of the seizure threshold. These points are a pharmacology reminder and not a result of the study, which does not report plasma concentrations in most trials.
- The two non-negotiable forms of monitoring. Haematological monitoring of clozapine, because of the risk of agranulocytosis, applies regardless of any adjunct, and its terms are set by the product information that applies where you practise, to be consulted in its current version. The question of QTc prolongation arises for clozapine as for the drugs combined with it, which justifies a baseline electrocardiogram and a check of electrolytes.
- The course of action is set out in the NICE decision tree for schizophrenia in adults.
Frequently asked questions
Should duloxetine be added for patients with an insufficient response to clozapine?
No, not on the basis of this work. The estimate, SMD −1.53 with a confidence interval of −2.44 to −0.63, comes from a single trial of 40 patients randomised over 16 weeks. It is a hypothesis to be confirmed by larger trials, and in France, for example, the combination remains outside the marketing authorisation.
Why be wary of such a favourable result in resistant schizophrenia?
Because the magnitude is far greater than what is usually observed in the psychopharmacology of schizophrenia, and it appears in a population where the expected effects are smallest. In this synthesis, Egger’s test shows that small trials overestimate efficacy relative to clozapine, and the authors themselves point out that the sparsity of the network favours chance findings when a comparison rests on only one study.
Is an indirect comparison in a network meta-analysis as reliable as a direct one?
It carries less weight. An indirect comparison links two treatments never set against each other in the same trial by going through a common comparator. It assumes that the populations of the linked trials are similar enough for that route to be valid. In a network of 78 options, most comparisons are of this kind: the published certainty table shows that most of them rest on no direct trial.
Is the tolerability of clozapine combinations reassuring?
No. The tolerability outcomes were analysed and are not neutral: clozapine causes more sedation and more weight gain than several comparators, and discontinuations for adverse effects are more frequent on topiramate added to clozapine. Above all, with a median duration of 12 weeks, these trials do not allow the safety of a combination maintained over the long term to be assessed. This is an absence of data, not to be read as an absence of risk.
What should I conclude if another meta-analysis reaches the opposite conclusion?
That the question remains open. This work does not close the debate on the strategies to adopt in treatment-resistant schizophrenia; it confirms the place of clozapine and adds combination estimates whose robustness will have to be judged against the number of trials behind them and against converging or diverging syntheses.
Annotated bibliography
Source study. Samara M, Lappas AS, Pinioti E, Glarou E, Fober I, Christogiannis C, Siafis S, Christodoulou N, Helfer B, Mavridis D, Leucht S. Efficacy and tolerability of pharmacological interventions for schizophrenia non-responsive to prior treatment: a systematic review and network meta-analysis. eClinicalMedicine, 2025; 84: 103291. DOI 10.1016/j.eclinm.2025.103291 PMID 40535000 Contribution: bringing 150 randomised trials and 78 drug options or placebo into one network on a question where direct comparisons are rare, with a 279-page appendix giving certainty comparison by comparison, nine sensitivity analyses and Egger’s test. Limitations: 104 of the 150 trials are at moderate or high risk of bias, no comparison reaches high certainty, the definitions of non-response are heterogeneous, several combination estimates rest on a single small trial, and the median duration of 12 weeks rules out any conclusion on long-term safety.
Samara MT, Dold M, Gianatsi M et al. Efficacy, acceptability, and tolerability of antipsychotics in treatment-resistant schizophrenia: a network meta-analysis. JAMA Psychiatry, 2016; 73(3): 199 to 210. DOI 10.1001/jamapsychiatry.2015.2955 PMID 26842482 Contribution: the reference network meta-analysis on treatment-resistant schizophrenia, cited by the authors as the work that cast doubt on the superiority of clozapine over the other second-generation antipsychotics, which sets out what is at stake in the present synthesis. Limitations: it covers only antipsychotic monotherapies and does not address combinations.
Huhn M, Nikolakopoulou A, Schneider-Thoma J et al. Comparative efficacy and tolerability of 32 oral antipsychotics for the acute treatment of adults with multi-episode schizophrenia: a systematic review and network meta-analysis. Lancet, 2019; 394(10202): 939 to 951. DOI 10.1016/S0140-6736(19)31135-3 PMID 31303314 Contribution: the reference that gives the usual order of magnitude of antipsychotic effects in multi-episode schizophrenia, essential for judging how unusual an SMD of −1.53 is. Limitations: it concerns a population not selected for resistance, so the comparison is not like for like.