Published on 14 September 2026
Treatment-resistant schizophrenia without clozapine: which augmentation strategies hold up?
In brief
This pre-registered systematic review retains 78 studies, of which 68 enter the meta-analysis, covering 3,241 patients with treatment-resistant schizophrenia. It reviews nine classes of intervention used alongside ongoing antipsychotic treatment when clozapine is not an option. On the authors’ own reading, none reaches the threshold of evidence that would allow it to be routinely recommended as an alternative to clozapine. The largest effect sizes are also the most fragile: the glycine site agonist trials come almost entirely from a single research group, one of whose authors has held intellectual property rights over D-serine since 2000, and the effect of non-invasive stimulation on positive symptoms loses significance once trim-and-fill correction is applied. The most directly usable result is a negative one: high-dose antipsychotics add nothing in any symptom domain, with the reservation that the authors themselves rate the certainty as low to very low.
The context
The situation is an ordinary one in clinic. A patient has failed two properly conducted antipsychotic trials, clozapine is indicated, and clozapine is impossible. A history of agranulocytosis, refusal by the patient, myocarditis, poorly controlled epilepsy, no practical way of running the haematological monitoring, or simply a care team that will not engage with it. Something else has to be offered. The literature is not short of suggestions: raise the dose, add an antidepressant, add a glycine site agonist, set up repetitive transcranial magnetic stimulation, start cognitive behavioural therapy for psychotic symptoms. Where clozapine remains on the table, the sequence that leads to it is laid out in the NICE schizophrenia decision tree.
These strategies had mostly been assessed separately, each in its own meta-analysis, each with its own inclusion criteria. Carr and colleagues bring them together in a single methodological frame, with the same certainty grid applied to all of them. The authors claim priority nowhere, and they explicitly position their results against the earlier meta-analyses of psychotherapy, ondansetron, electroconvulsive therapy and mood stabilisers. Their question is not the same as the one asked by the network meta-analysis of next-step strategies after a first antipsychotic falls short: here the patient has already reached resistance, and clozapine has been ruled out.
The study at a glance
| Item | Content |
|---|---|
| Population | Adults with schizophrenia or schizoaffective disorder with a documented inadequate response to at least one antipsychotic trial |
| DetailThe inclusion criterion is broader than the TRRIP consensus definition: 32 studies require only one period of non-response. The question the authors ask, that of the patient who cannot receive clozapine, is not itself an inclusion criterion of the trials. | |
| Interventions | Nine classes: high-dose antipsychotics, glycine site agonists, non-invasive brain stimulation, psychotherapy, antidepressants, mood stabilisers, ondansetron, famotidine, sodium nitroprusside and Ginkgo biloba |
| DetailAll are studied as add-ons to ongoing antipsychotic treatment. The authors insist on the point: these are augmentation strategies, not substitutes for antipsychotic treatment. | |
| Comparator | Placebo, sham stimulation or an active comparator depending on the trial. Trials using treatment as usual as the comparator were excluded |
| DetailExcluding treatment as usual is a deliberate methodological choice, meant to avoid inflating effect sizes through a nocebo effect. It partly explains why the effect of psychotherapy on the total score, found in earlier work, does not reappear here. | |
| Outcomes | Positive, negative and total symptoms, expressed as Hedges’ g on change scores |
| DetailThe PANSS is the preferred instrument, except for non-invasive stimulation, where scales specific to auditory verbal hallucinations were extracted instead. The certainty of each estimate is graded with GRADE. | |
| Design | Systematic review and random-effects meta-analysis of randomised trials, protocol pre-registered on PROSPERO, CRD42023418053 |
| Detail78 studies retained in the systematic review, 68 in the meta-analysis for 3,241 patients, 10 reviewed narratively. Searches in PsycINFO, EMBASE, ClinicalTrials.gov and PubMed. The paper carries two different search dates, up to July 2023 in the methods and June 2024 in the results. | |
The quality checks
| Point checked | Judgement |
|---|---|
| Pre-registered protocol | Yes, PROSPERO |
| FindingPre-registration limits the risk of selecting outcomes after the fact, in a field where the temptation is real. | |
| Risk of bias in the included trials | Cochrane RoB 2, trial by trial |
| FindingThree authors rated each trial as low, medium or high risk, with graphical display through robvis. The trial-by-trial detail sits in the supplementary material. | |
| Grading of certainty | GRADE on every result |
| FindingEvery rating reported is low or very low. None reaches moderate. The authors say so rather than dilute it. | |
| Search for publication bias | Possible in two classes only |
| FindingFunnel plots are produced only from eight studies per class upwards, Egger’s test only from ten upwards. Only non-invasive stimulation and psychotherapy meet that condition. For the other seven classes, publication bias remains unassessable. | |
| Sample sizes per class | Small, median 29.5 |
| FindingThe median total sample size of the trials is 29.5. Three trials only for antidepressants, seven for high doses, nine for glycine site agonists. | |
| Definition of resistance | Loose in 32 studies |
| Finding32 studies require only one period of non-response. The GRADE summary table quantifies that imprecision class by class: 43% of the high-dose trials meet criteria for treatment-resistant schizophrenia, against 73% for stimulation and 100% for antidepressants. | |
| Patients on clozapine within the samples | Present, proportion poorly documented |
| FindingClozapine augmentation trials are excluded, yet many included trials contain patients taking clozapine. The proportion is reported in only 44 of the 68 meta-analysed studies, and its mean, calculated on the studies where at least one patient was receiving it, is 31%. | |
| Homogeneity of the classes | Broad groupings |
| FindingPlacing drugs with different mechanisms under one label produces a heterogeneity that is not only statistical. The authors acknowledge it for antidepressants, high-dose antipsychotics and mood stabilisers. | |
| Funding of the work | No external funding |
| FindingThe authors state that the work received no external sponsorship. Two of them hold institutional support, from the Wellcome Trust and from the National Institute for Health Research with Maudsley Charity. | |
| Declared interests | Two authors, industry links |
| FindingTwo authors declare consultancy or speaker fees from pharmaceutical companies and the co-direction of a company designing digital resources in mental health. The other eight declare nothing in the published section. | |
The findings
| Strategy | Reported effect, Hedges’ g with confidence interval in brackets |
|---|---|
| High-dose antipsychotics | 7 trials, 467 patients. Positive −0.03 [−0.23, 0.17], I² 3.4%, GRADE low; negative −0.02 [−0.21, 0.17], I² 0%, GRADE low; total 0.00 [−0.18, 0.18], I² 0%, GRADE very low |
| ReadingThe estimate is centred on zero and reasonably precise, and it agrees with the Cochrane data on dose increase after non-response. The authors temper it themselves: high risk of bias in six trials out of seven, different agents from one trial to the next, only 43% of trials meeting criteria for resistance, and above all missing data that bias the result in favour of high doses, since patients who deteriorate drop out of the analyses more often. Their own wording is that high doses may not be effective, and that a benefit in selected patients cannot be excluded. The body of the paper gives g = 0.02 for total symptoms, the GRADE summary table gives 0.00, for the same confidence interval. | |
| Glycine site agonists | 9 trials, 187 patients. Positive −0.56 [−0.81, −0.31], I² 8.5%, GRADE low; negative −1.18 [−1.49, −0.87], I² 26.3%, GRADE low; total −1.17 [−1.75, −0.59], I² 79.9%, GRADE very low |
| ReadingThese are the largest effects in the whole review, and they rest on the narrowest base: 187 patients in total, every trial but two coming from the same research group, one of whose authors acquired intellectual property rights over D-serine in 2000 and signs seven of the nine trials. Excluding those seven trials, the effect becomes non-significant on positive symptoms, −0.41 [−1.3, 0.47], and on total symptoms, −0.80 [−3.37, 1.76], but stays significant on negative symptoms, −1.16 [−1.54, −0.60.] Only two trials remain in that sensitivity analysis, which sharply limits what it can carry. The effect is driven by the full agonists, glycine and D-serine; D-cycloserine, a partial agonist, adds nothing. | |
| Non-invasive brain stimulation | 26 trials, 893 patients. Positive −0.42 [−0.65, −0.18], I² 64.9%, GRADE low; negative −0.13 [−0.35, 0.09], GRADE very low; total −0.13 [−0.38, 0.12], GRADE very low |
| ReadingEgger’s test is significant, z = −2.29, p = 0.022. After trim-and-fill correction, the effect on positive symptoms falls to −0.18 [−0.47, 0.12], and is no longer significant; restricted to the repetitive transcranial magnetic stimulation trials, it falls to −0.14 [−0.46, 0.18.] A second warning signal: the effect is significantly smaller in trials using a strict definition of resistance, z = 2.04, p = 0.04. Positive symptoms here are most often measured with scales specific to auditory verbal hallucinations rather than with the PANSS, so the result reads as an effect on hallucinations. The authors state that no conclusion of ineffectiveness on negative symptoms can be drawn, the stimulation target having been designed for hallucinations. | |
| Psychotherapy | 10 trials, 565 patients. Positive −0.56 [−1.01, −0.10], I² 85.2%, GRADE low; negative −0.12 [−0.40, 0.16], not significant; total −0.56 [−1.17, 0.05], not significant |
| ReadingThe signal sits on positive symptoms and reappears neither on negative symptoms nor on the total score. Heterogeneity is very high and the impossibility of blinding exposes the trials to a substantial risk of bias, hence the low rating. Restricted to cognitive behavioural therapy, the analysis keeps a more modest benefit on positive symptoms, −0.34 [−0.63, −0.06], with clearly reduced heterogeneity, I² 45.9%. Neither the funnel plots nor Egger’s test, z = −0.94, p = 0.35, show asymmetry. | |
| Antidepressants | 3 trials, 187 patients. Negative −0.74 [−1.46, −0.02], I² 76.7%, GRADE very low; total −0.69 [−1.00, −0.38], I² 0%, GRADE low; positive −0.40 [−1.59, 0.79], not significant |
| ReadingThree trials, three molecules with different mechanisms, mianserin, sertraline and escitalopram. The authors themselves judge that the benefit on the total score probably reflects improvement in mood and anxiety, and they recall that the negative subscale of the PANSS separates primary negative symptoms from depressive symptoms poorly. For total symptoms the GRADE summary table gives −0.55 [−0.89, −0.22], a value different from the one in the body of the paper. | |
| Other classes | Mood stabilisers, 5 trials, 500 patients: no effect. Ondansetron, 2 trials, 159 patients: no effect. Famotidine, 2 trials, 60 patients: total −1.31 [−1.88, −0.74.] Ginkgo biloba, 2 trials, 163 patients: positive −0.48 [−0.79, −0.17.] Sodium nitroprusside, 2 trials, 80 patients: no effect |
| ReadingFamotidine posts the largest effect size in the review on the total score, on two trials and 60 patients. It is the clearest illustration of the general problem: in this literature, effect size is inversely proportional to the size of the base. Ten further studies, not classifiable, are reviewed narratively, among them an oestradiol trial judged large and well conducted that the authors consider worth further work. | |
One reading point carries beyond this paper. The ranking by effect size and the ranking by strength of the evidence base are here almost inverted. The most impressive results, glycine site agonists and famotidine, rest on 187 and 60 patients, and the first does not survive the removal of the trials carrying a conflict of interest. The least spectacular result, the absence of any effect of high doses, is the one with the narrowest interval. Reading a meta-analysis by starting with the g values leads to retaining exactly what should not have been retained.
Critical appraisal
| Domain | Judgement |
|---|---|
| Completeness of the mapping | Strength |
| FindingNine classes of intervention assessed within a single frame, with the same certainty grid applied to all of them, and ten further studies reviewed narratively. The authors claim no priority and compare themselves explicitly with the earlier meta-analyses specific to each class. | |
| Fit between claim and evidence | Caution owned |
| FindingThe abstract states that no intervention reaches the threshold of evidence to be routinely recommended as a viable alternative to clozapine, and the conclusion stays conditional. The authors do, however, name glycine site agonists as the most promising avenue for future trials, precisely the class most exposed to allegiance bias, and they attach the corresponding reservation to it. | |
| Allegiance bias, glycine site agonists | Major |
| FindingOne of the trial authors acquired intellectual property rights over D-serine in 2000 and signs seven of the nine trials in the class. Every trial but two comes from the same research group. Excluding those seven trials removes the effect on positive and total symptoms, but leaves only two trials, which makes the sensitivity analysis itself uninformative. Across the whole review, 12 meta-analysed trials carry a significant conflict of interest. | |
| Publication bias, stimulation | Significant asymmetry |
| FindingEgger’s test significant, z = −2.29, p = 0.022, the effect brought down to −0.18 and no longer significant after trim-and-fill. Funnel asymmetry is compatible with publication bias without formally proving it: it reflects a small-study effect, of which publication bias is only one possible explanation among others. | |
| Power per class | Insufficient |
| FindingWith a handful of trials and a few hundred patients per class, an absence of significance does not demonstrate an absence of effect. That holds for high doses too, where the interval is admittedly narrow around zero, but where GRADE certainty on the total score is very low. | |
| Definition of the population | Imperfect |
| FindingThe inclusion criterion requires only one documented antipsychotic failure, and the gap with the TRRIP criteria is quantified class by class in the GRADE table. Patients on clozapine appear in the samples, with the proportion unknown for 24 of the 68 studies. The available meta-regressions, limited to stimulation and psychotherapy, found no significant effect of that proportion on effect size. | |
| Scope of the results | Augmentation, not substitution |
| FindingEvery included trial adds the intervention to a continuing antipsychotic. The authors explicitly ask that their results be read as augmentation strategies and not as alternatives to antipsychotic treatment. Nor do the results extend to clozapine-resistant schizophrenia. | |
| Measurement of positive symptoms under stimulation | Hallucination-focused scales |
| FindingFor non-invasive stimulation, it is the auditory verbal hallucination scales that were extracted, not the PANSS. The effect on positive symptoms must therefore be read first as an effect on hallucinations, over very short treatment durations. | |
Level of evidence
The design of the work itself is sound: a pre-registered systematic review and meta-analysis of randomised trials, which the Oxford classification places at level 1a. The certainty of the estimates it produces is low or very low for every result reported, which is not a contradiction but a conclusion. A rigorous meta-analysis of a poor literature remains rigorous, and its conclusion is that the literature is poor.
What is best supported, without reaching high certainty: raising antipsychotic doses beyond standard doses improves no symptom domain, the authors concluding that high doses may not be effective in treatment-resistant schizophrenia, without excluding a benefit in selected patients. What is suggested without being demonstrated: an effect of antidepressants on negative and total symptoms, an effect of psychotherapy and of non-invasive stimulation on positive symptoms, an effect of glycine site agonists across all three domains. What weakens those last three signals: allegiance bias for the glycine site agonists, publication asymmetry for stimulation, very high heterogeneity for psychotherapy.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ So when I cannot use clozapine, I still have nothing solid to offer. What this study really teaches me is that there is nothing to expect from pushing the dose. And I will be wary of the numbers that shine: they are the ones resting on the fewest patients. ”
What this means in practice: the message is not discouraging, it is protective. It names a widespread practice that is costly in adverse effects, dose escalation, and finds no symptomatic benefit in it. It also invites a fresh look at the obstacles to clozapine rather than a search for a substitute whose efficacy would be established.
What you can do with this
- Expect no symptomatic gain from exceeding standard antipsychotic doses. The estimate is centred on zero across all three domains and consistent with the earlier Cochrane data, even if certainty remains rated low to very low.
- Work methodically back through the obstacles to clozapine before looking for an alternative. The authors point out that none of the strategies assessed reaches the threshold of evidence to be recommended routinely as an alternative.
- If an augmentation is tried, treat it as an individual trial with a defined outcome measure and a stopping date, not as an established treatment.
- Be ready to answer a patient or a family who has read something about D-serine or about magnetic stimulation: positive results do exist, they rest on very small samples, and they weaken or vanish as soon as the conflict of interest or the publication asymmetry is taken into account.
- Remember that all these strategies were tested as add-ons to the ongoing antipsychotic, never in its place.
- Keep the reading lesson. Faced with a meta-analysis, read the certainty column and the number of patients before the effect size column.
- The course of action is set out in the NICE decision tree for schizophrenia in adults.
Frequently asked questions
Does this study say that no alternative works?
No. It says that none reaches the threshold of evidence to be routinely recommended as a viable alternative to clozapine, which is a different statement. For most classes, the samples are too small to rule out a real effect. The clearest negative result concerns high doses, and it is itself rated low to very low in certainty.
Why does the effect of glycine site agonists weaken under scrutiny?
Because the base is very narrow, 9 trials and 187 patients, because all but two come from the same research group, and because one of the authors of seven of those trials acquired intellectual property rights over D-serine in 2000. The sensitivity analysis that excludes those seven trials pushes the effect on positive and total symptoms below the significance threshold, while the effect on negative symptoms remains significant. It then rests on two trials only, which is a reason not to over-read it in either direction.
Is repetitive transcranial magnetic stimulation useless in schizophrenia?
These data do not allow that conclusion. They show that the available literature on positive symptoms carries significant publication asymmetry and that the corrected mean effect is no longer significant. A real effect of small magnitude remains possible. The authors also state that no conclusion of ineffectiveness on negative symptoms can be drawn, since the included protocols targeted hallucinations and the treatment durations were very short.
What should be done in practice when the patient refuses clozapine?
That question falls outside the scope of the meta-analysis, which does not assess strategies for revisiting the decision. On the substance of its result, it argues for working on what drives the refusal rather than substituting a strategy whose efficacy is not established, and for checking the monitoring requirements in force at the time and place of prescribing.
Annotated bibliography
Source study. Carr R, Cannon A, Finelli V, Bukala B, Cimen Y, Filipova P, Cummings C, Pillinger T, Howes OD, McCutcheon RA. Non-Clozapine interventions in treatment-resistant schizophrenia: a systematic review and meta-analysis. Molecular Psychiatry. 2025;31(1):526-544. DOI 10.1038/s41380-025-03255-y. PMID 41044402. PMCID PMC12700803. PROSPERO protocol CRD42023418053. Work declared free of external funding; two authors declare links with the pharmaceutical industry.
What was consulted. References verified against the full text of the version published, on 12 August 2026. The supplementary material of the study could not be consulted. This analysis underwent an independent double reading.