Published on 16 September 2026
Antipsychotic polytherapy: does combining two drugs really reduce psychiatric readmission?
In brief
A French nationwide cohort drawn from the national health data system (SNDS). 234,959 adults with a psychotic disorder, identified through at least one psychiatric hospitalisation between 1 January 2014 and 31 December 2020, with a median follow-up of 7 years. The primary outcome is time to psychiatric readmission. The analysis is within-individual: each patient serves as their own control, and different periods of exposure are compared within that patient. What the published abstract reports: being on an antipsychotic goes with a 20 to 50% reduction in the risk of readmission compared with no treatment; 71.6% of patients are readmitted over a median follow-up of seven years; the median cumulative duration of readmission is 67 days; 21.8% have more than four admissions; 71.1% of the cohort were exposed to polytherapy. Among the combinations, those including clozapine (with risperidone or amisulpride) and the combination of loxapine with amisulpride go with a lower risk, with no figure published. Exposure to quetiapine-based combinations is associated with a risk 10 to 20% lower than quetiapine alone. The abstract gives no hazard ratio and no confidence interval. The interpretive sentence of the abstract states that the benefit of polytherapy reduces psychiatric rehospitalisation. PEB does not adopt that verb. The design is observational, and it neutralises only part of the confounding.
The context
Antipsychotic polytherapy is a widespread and unloved practice. Widespread because monotherapy leaves a substantial share of patients symptomatic. Unloved because reference texts discourage it, and because it doubles drug exposure with no guarantee of doubling the effect.
On this point, established practice and the available guidelines point the same way. NICE, in recommendation 1.3.6.10 of its CG178 guideline on psychosis and schizophrenia in adults, advises against initiating regular combined antipsychotic medication other than for short periods, such as when changing medication, a wording dated 2009 and carried over into the March 2014 version. That guideline was published in February 2014, updated in March 2014, and remains in force. In France, the physician guide for long-term condition (ALD) 23 on schizophrenia, from the Haute Autorité de santé (HAS), published in June 2007, states that monotherapy is preferred, orally where possible, and that a combination of antipsychotics may be started after failure of monotherapy, in particular when one antipsychotic is being replaced by another, or in cases of resistance, with its tolerability and relevance to be reassessed very regularly. A point of status is needed: the HAS page devoted to ALD no. 23 currently lists only the document on procedures and services updated in July 2025, which itself states that it is not a good practice guideline and does not constitute a decision aid on diagnostic or therapeutic strategy. The French position on monotherapy therefore rests on a text from 2007 whose currency could not be established on the agency’s website at the date of this analysis.
The challenge to that position came from the Nordic registers. In 2019, Tiihonen and colleagues published in JAMA Psychiatry a Finnish cohort of 62,250 patients, using the same within-individual logic. The combination of clozapine with aripiprazole was linked there to the lowest risk of readmission, better than clozapine alone with a hazard ratio of 0.86 (95% CI 0.79 to 0.94). Across all polytherapy, the gap with monotherapy ranged from 7 to 13%. The authors concluded that current guidelines should revise their categorical stance against maintenance polytherapy. The French work presents itself as a test of the reproducibility and robustness of earlier findings, of which the Finnish cohort is the leading example: the abstract names neither a study nor a country. The French population, for its part, has prescribing habits that differ from those of Finland.
Confounding by indication
Why the causal question cannot be settled here
Nobody randomises polytherapy. The second antipsychotic is added because the first is not enough. Patients who receive one are therefore, by construction, those whose illness is resistant, those who have already relapsed, those whose adherence is a worry. Comparing a period of polytherapy with a period of monotherapy across different people then amounts to comparing two populations that differ in more than treatment. This is what is called confounding by indication, and it is enough to reverse a result.
The within-individual design answers part of the problem, and it answers it well. By comparing a patient’s own periods of exposure, it eliminates by construction every factor that does not vary during follow-up: sex, year of birth, genetic make-up, age at onset, educational level, background severity, a stable social environment. The elegance of the method lies in this: those factors are neutralised without having been measured, or even identified. The advantage is decisive in a health administrative database such as the SNDS. That database does contain clinical information: diagnoses from the hospital activity programme (PMSI), long-term conditions, procedures and dispensings. What it lacks is of another order: a measure of symptom severity, the dose actually taken, and psychosocial data. Conventional adjustment could have captured none of that.
What the design does not touch remains: everything that changes in the same person at the very moment treatment changes. And that is exactly where confounding by indication reassembles itself. A second antipsychotic is not introduced on just any day, it is introduced when the patient is doing less well.
| Source of confounding | Addressed by the within-individual design? | Consequence |
|---|---|---|
| Factors stable between people (background severity, genetics, social setting) | Yes, by construction | This is the main contribution of the design. No measurement was needed |
| Clinical worsening concurrent with the addition of the second antipsychotic | No | The polytherapy period starts at a higher level of risk. The bias works against polytherapy and underestimates its apparent benefit. This is the argument the Finnish authors were already making in 2019 |
| Regression to the mean after a peak in severity | No | If the combination is started at the worst moment, the spontaneous improvement of the following months will be credited to it. This time the bias works in favour of polytherapy |
| Chronological order of periods, natural course of the illness, change in practice between 2014 and the end of follow-up | Not specified | The abstract does not say whether the analysis accounts for time elapsed since inclusion or for age reached. An absence of information, not an absence of adjustment |
| Life events, comorbid substance use, change in outpatient care | No | The SNDS does not measure them, or only indirectly through hospital diagnoses. They shift within the same person and weigh on readmission |
Two residual biases, two opposite directions. That is why caution is needed as much about the size as about the direction. What is demonstrated here is an association between certain periods of exposure and a lower risk of readmission, in the same patients. What is suggested is that not all combinations are equal. What is expert opinion is the direction and size of the bias that remains, and so the question of whether the observed benefit is inflated or understated. None of these three levels permits the statement that polytherapy reduces readmissions. Although it is the best tool available on claims data, this design remains observational.
The study at a glance
| Question (PICO) | |
|---|---|
| Population | |
| Adults aged 18 and over, with a psychotic disorder, identified in the SNDS through at least one psychiatric hospitalisation between 1 January 2014 and 31 December 2020. 234,959 patients, median follow-up 7 years. The diagnostic codes used and the exclusion criteria do not appear in the published abstract | |
| Intervention | |
| Antipsychotic polytherapy. 71.1% of the cohort were exposed to it during follow-up. The operational definition used (number of drugs, minimum duration of overlap, handling of switches) does not appear in the abstract | |
| Comparator | |
| Antipsychotic monotherapy, and no antipsychotic treatment. The comparator is made up of the same patient’s other periods of exposure, which makes each subject their own control | |
| Primary outcome | |
| Time to psychiatric readmission. No tolerability, mortality or functioning outcome appears in the published abstract, which does not mean they were not measured | |
| Design | |
| Retrospective nationwide cohort on health administrative data, within-individual analysis comparing different periods of exposure in the same patient, in order to limit confounding. The statistical model used and the handling of overlapping periods are not described in the abstract. Published online on 27 April 2026, June 2026 issue · CEBM 2b |
Quality control
| Criterion | Status |
|---|---|
| Completeness of the data source | Sound |
| FindingThe SNDS covers almost the entire population resident in France and enrolled in national health insurance. No sampling, no voluntary recruitment, no loss to follow-up in the usual sense | |
| Size and duration | Sound |
| Finding234,959 patients, median follow-up of 7 years. No French cohort had previously addressed this question on this scale | |
| Confounding between people | Sound |
| FindingThe within-individual design eliminates it by construction, including for factors that were never measured | |
| Confounding within the same person | Unresolved |
| FindingConfounding by indication survives the design. Nothing in the abstract indicates how concurrent changes in treatment and clinical state were handled | |
| Measurement of exposure | Reservation |
| FindingThe SNDS records a dispensing, not an intake. The gap between the two is structural and cannot be measured here. The abstract does not state whether long-acting injectable formulations were distinguished from oral ones | |
| Clinical data | Absent |
| FindingNo symptom score, no dose, no data on substance use in the published abstract. It is impossible to check that the periods compared correspond to comparable clinical states, which is precisely the assumption on which interpretation depends | |
| Outcome | Reservation |
| FindingReadmission is an indirect marker of relapse. It also depends on the number of beds, the team’s threshold for admission, the family and social network, and whether an outpatient alternative exists | |
| Figures available | Insufficient |
| FindingThe published abstract reports no hazard ratio and no confidence interval. Reductions are given there as ranges. Any detailed reading of the combinations requires the full text | |
| Independence | Major reservation |
| FindingStudy funded by Eisai. Two authors declare an Eisai affiliation (Eisai France, Eisai EMEA), three work at HEVA, a health data analytics company distinct from the sponsor. The other authors belong to Université Clermont Auvergne, INSERM, Paris-Saclay and Sorbonne Université. The PubMed record carries the authors’ declaration of competing interests: the first author (PML) and another author (PN) declare consultancy, advisory or honoraria links with several companies, including Eisai, Janssen, Lundbeck and Otsuka, and the first author declares expert testimony for Janssen and Otsuka; another author (BF) declares consultancies with many pharmaceutical companies; three authors are employees of HEVA, which received financial compensation from Eisai and from the Association pour la recherche en biologie de physiopathologie des épithéliums humains (ARBPEH) to conduct the study | |
The findings
| Outcome | What the abstract reports |
|---|---|
| Antipsychotic versus no treatment | 20 to 50% reduction in the risk of psychiatric readmission |
| PEB readingThe most robust result and the least discussed. It is a reminder that the question concerns how to treat, not whether treating is worthwhile. A wide range, no confidence interval published | |
| Exposure to polytherapy | 71.1% of the cohort |
| PEB readingA measure of practice and not of effectiveness. Combining was the majority practice in France over this period, which puts real-world use in tension with reference texts | |
| Combinations including clozapine (with risperidone or amisulpride) | Reduction in the risk of readmission |
| PEB readingDirection reported, size unknown the abstract gives no figure. Consistent with the Finnish signal of 2019, where clozapine in combination was also at the top | |
| Loxapine plus amisulpride combination | Reduction in the risk of readmission |
| PEB readingA specifically French signal loxapine is not among the drugs highlighted by the Finnish cohort of 2019. No figure published. This is the result calling for most caution until the full text has been read | |
| Quetiapine-based combinations versus quetiapine alone | 10 to 20% reduction |
| PEB readingThe only quantified comparison between polytherapy and monotherapy in the abstract, and even then as a range and without a confidence interval | |
| Cumulative duration of readmission | Median of 67 days |
| PEB readingDescriptive data giving the order of magnitude of the hospital burden over seven years | |
| Repeated readmissions | 21.8% of patients have more than four admissions |
| PEB readingRoughly one patient in five accounts for a major share of hospital use. That is where the clinical usefulness of the question is decided | |
| Hazard ratios by combination | Not reported in the published abstract |
| PEB readingMissing data the combinations cannot be ranked in any detail from the record alone | |
| Tolerability, mortality, adverse effects | Not reported in the published abstract |
| PEB readingPartial picture an absence from the record does not mean an absence of measurement. But as things stand, we have the benefit side without the risk side | |
A word on the figures in circulation. Precise hazard ratios are sometimes attributed to this work, with no identifiable source. The published abstract reports none. PEB therefore publishes none. Nor does the abstract mention any unfavourable combination. An unverifiable figure is not a cautious figure: it is a figure on hold.
A second reservation, more technical. The inclusion window runs from 1 January 2014 to 31 December 2020, and the stated median follow-up is 7 years. The two pieces of information fit together on one condition only: that data collection continued beyond 31 December 2020. A median covers all the subjects followed, not only the patients included first. The abstract does not say so, and the end date of follow-up does not appear in it.
Critical appraisal
| Domain | Judgement |
|---|---|
| Level of evidence | Reservation |
| FindingRetrospective cohort on claims data, CEBM 2b. The within-individual design places this work above a conventional cohort for the control of confounding between people. It does not raise it to the level of a randomised trial, and the difference lies precisely in the bias that matters here | |
| Residual confounding by indication | Major limitation |
| FindingThe second antipsychotic is added at a time of worsening. Two residual mechanisms pull in opposite directions, concurrent worsening and regression to the mean. Their net balance cannot be estimated from the available data | |
| Validity of the exposure measure | Reservation |
| FindingA pharmacy dispensing is not an intake. In an illness where partial adherence is the rule, actual exposure is systematically overestimated, and probably unevenly across drugs | |
| Validity of the outcome | Reservation |
| FindingReadmission depends on the illness, but also on the organisation of care. A fall in hospital use may reflect clinical improvement, saturated beds or a shift of the burden towards outpatient care | |
| Statistical precision | Reservation |
| FindingNo confidence interval in the abstract. The size of the cohort suggests narrow intervals, which does not make the estimates more accurate: beyond a certain sample size, systematic error outweighs random error. A tight interval around a biased value is still a biased value | |
| Fit between data and conclusion | Overstatement |
| FindingThe problematic sentence is not the last one in the abstract. The last one reads: “This study confirms the reproducibility and robustness of APP’s effectiveness in real-world settings”, and its verb bears on reproducibility and robustness, which such a design can indeed establish. It is the interpretive sentence that goes beyond what the design allows: “our findings corroborate previous evidence that APP’s benefit reduces psychiatric rehospitalization”. To reduce belongs to the vocabulary of causation. An observational design, whatever its size, establishes an association | |
| Independence | Major reservation |
| FindingIndustry funding, two authors affiliated with the sponsor, three authors employed by the analytics contractor. This disqualifies neither the data nor the design, which are sound. It does call for close attention to the choice of words in the interpretation, and that is exactly where the problem lies | |
| External validity | Sound |
| FindingThe entire French population, French prescribing, drugs in French use. On this count, no foreign cohort can replace this work for a clinician practising in France | |
A remark on how this study connects with the previous one. The Finnish work of 2019 had opened the breach, and its argument was twofold: some combinations do better than monotherapy, and residual bias works against them, so the true effect would be larger than the observed effect. The French work reproduces the direction of the result in a population with different habits, which counts. It brings nothing new to the causal question, because it uses the same design and therefore inherits the same blind spots. Reproducing a result with the same method demonstrates its stability, not its validity.
Two misreadings lie in wait for this dossier. The first would be to take it as a green light to combine, when nothing here compares a combination strategy with a strategy of optimising monotherapy, nor informs the cost in adverse effects. The second would be to dismiss the work on account of its funding, when the data are exhaustive, the design relevant and the result consistent with earlier literature. The right stance fits in one sentence: believe the figures, question the words.
Level of evidence
PEB appraisal: moderate confidence in the existence of an association between certain periods of polytherapy and a lower risk of readmission, low confidence in the size of that association, no confidence in its causal nature. The within-individual design is the right tool for this question on this type of data, and the completeness of the SNDS is an asset few countries possess. What lowers the score comes down to four points: confounding by indication within the same person, which is not neutralised; the measurement of exposure through dispensing; the complete absence of hazard ratios and confidence intervals in the published record; financial dependence on the sponsor, in a work whose interpretation uses the vocabulary of causal demonstration.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ 234,959 French patients, each their own control, and some combinations come out better than the single drug. Fine. Except that the second antipsychotic goes in on the day things go wrong, and no calculation makes up for that. What I take away is that the clozapine combinations do well, and I keep optimising monotherapy first. What I would have wanted is the tolerability side, and it is not in the abstract. ”
What this means in practice: a within-individual design eliminates what does not change in a patient, never what changes at the same time as their prescription. That lens is useful well beyond antipsychotics.
What you can do with this on Monday morning. The course of action does not change. What changes is the quality of what you can say and document.
- Keep the usual order: a well-conducted monotherapy, at an adequate dose and for long enough, before any combination. The French HAS guide of 2007 states that monotherapy is preferred and reserves combination for failure, switching or resistance. NICE, for its part, advises against starting a regular combination outside short periods. This study overturns neither.
- Check that clozapine has been offered before stacking drugs. In this work, among the combinations cited as favourable, two out of four include clozapine, which presupposes that it was introduced first. Adding a third drug in a patient who has never had clozapine is a decision that needs to be justifiable. PEB opinion, not a result of the study.
- Have a useful figure to hand for a family: in this French cohort, 71.6% of patients were readmitted over a median follow-up of seven years, and 21.8% more than four times. Saying that relapse is common and expected relieves guilt, and it opens the conversation on continuity of care rather than on the number of tablets alone.
- Document, for every combination in progress: date the second antipsychotic was introduced, reason, expected clinical goal, planned reassessment date. The French HAS guide asks for tolerability and relevance to be reassessed very regularly. It is also, very concretely, what these cohorts cannot see in the SNDS, and what protects the prescriber.
- Monitor the side this work does not inform: weight, waist circumference, blood glucose, lipid profile, corrected QT, cumulative anticholinergic burden, sedation. A combination doubles exposure, not necessarily benefit.
- Take the reading lens with you. Faced with any real-world study comparing two strategies, ask three questions: who decides on the exposure, what changes at the same time as it, and does the outcome measured depend on anything other than the patient. Here, all three answers govern the interpretation.
- The course of action is set out in the NICE decision tree for schizophrenia in adults.
Frequently asked questions
Does this study justify combining two antipsychotics?
No. It describes a statistical association between certain periods of polytherapy and a lower risk of readmission in the same patients. It does not compare a combination strategy with a strategy of optimising monotherapy, and it does not inform the cost in adverse effects. Reference texts continue to place monotherapy first.
Does a within-individual design remove confounding by indication?
It does not remove the part that matters. Everything stable in a person disappears by construction, including factors nobody has measured. Nothing that varies at the same time as the prescription disappears. Yet the second antipsychotic is introduced precisely when the clinical state is deteriorating, which is the very definition of confounding by indication, transposed inside an individual.
Why are there no hazard ratios for antipsychotic combinations in this analysis?
Because the published abstract contains none. Precise values circulate with no identifiable source, they do not appear in the publication. A figure that cannot be shown is not cited. The full text probably contains them, and this analysis will be updated if reading it warrants.
Is psychiatric readmission a good outcome measure?
It is the best available in a claims database, and it has real clinical value: a patient who does not go back to hospital lives better. But it measures the organisation of care as much as the illness. The number of beds, a team’s threshold for admission, the presence of family support or of a mobile team weigh on this outcome without saying anything about the effectiveness of a drug.
Does industry funding invalidate this study?
No, and that shortcut would be lazy. The data come from the SNDS, the design is relevant, the result goes in the same direction as earlier literature. The funding calls for scrutiny of one precise point: the way the conclusion is worded. Here, the abstract states that the benefit of polytherapy reduces psychiatric rehospitalisation, whereas an observational design establishes an association. The gap is small on paper, and considerable in what it licenses clinicians to do.
What do guidelines say about antipsychotic polypharmacy?
NICE, in its CG178 guideline published in February 2014, updated in March 2014 and still in force, advises against starting regular combined antipsychotic medication, other than for short periods such as a switch. The French HAS physician guide for ALD 23, in its June 2007 version, prefers monotherapy and reserves combination for failure or resistance, with regular reassessment. A useful clarification: the current HAS page on ALD no. 23 lists only the July 2025 document on procedures and services, which is not a good practice guideline. The French preference for monotherapy is therefore established practice, resting on an old text whose current status could not be confirmed.
Annotated bibliography
Llorca PM, Falissard B, Baloche E, Bournane R, Schmidt A, Cals-Maurette M, Panes A, Nuss P (2026). Real-world effectiveness of antipsychotic polytherapy on rehospitalization in psychotic disorders: A French nationwide cohort analysis. Comprehensive Psychiatry, 148, 152704. DOI 10.1016/j.comppsych.2026.152704 · PMID 42096968. The source study analysed here. Published online on 27 April 2026, June 2026 issue. Funded by Eisai. Affiliations: Université Clermont Auvergne and CNRS, INSERM and Université Paris-Saclay, Eisai France, Eisai EMEA, HEVA, Sorbonne Université and AP-HP. The PubMed record carries the authors’ declaration of competing interests: the first author (PML) and another author (PN) declare consultancy, advisory or honoraria links with several companies, including Eisai, Janssen, Lundbeck and Otsuka, and the first author declares expert testimony for Janssen and Otsuka; another author (BF) declares consultancies with many pharmaceutical companies; three authors are employees of HEVA, which received financial compensation from Eisai and from the Association pour la recherche en biologie de physiopathologie des épithéliums humains (ARBPEH) to conduct the study. The published abstract reports no hazard ratio, no confidence interval, no end date of follow-up and no operational definition of polytherapy.
Tiihonen J, Taipale H, Mehtälä J, Vattulainen P, Correll CU, Tanskanen A (2019). Association of antipsychotic polypharmacy vs monotherapy with psychiatric rehospitalization among adults with schizophrenia. JAMA Psychiatry, 76(5), 499-507. DOI 10.1001/jamapsychiatry.2018.4320 · PMID 30785608. Finnish cohort of 62,250 patients, 1996 to 2015, same within-individual logic. The combination of clozapine with aripiprazole has the lowest risk of readmission, with a hazard ratio of 0.86 (95% CI 0.79 to 0.94) against clozapine alone, and 0.82 (95% CI 0.75 to 0.89) under the conservative definition excluding periods shorter than 90 days. Across all polytherapy, the gap ranges from 7 to 13%. The authors explicitly argue that residual bias underestimates the benefit, and call for guidelines to be revised. The closest earlier reference to the French work, which announces that it tests the reproducibility and robustness of findings of this kind without naming any study or country.
Højlund M, Köhler-Forsberg O, Gregersen AT, et al. (2024). Prevalence, correlates, tolerability-related outcomes, and efficacy-related outcomes of antipsychotic polypharmacy: a systematic review and meta-analysis. The Lancet Psychiatry, 11(12), 975-989. DOI 10.1016/S2215-0366(24)00314-6 · PMID 39547246. A recent synthesis covering both the tolerability side and the efficacy side of polytherapy, that is, the ground the French study documents only by half. Flagged here as further reading: its abstract could not be consulted at the date of this analysis, and none of its results is used.
National Institute for Health and Care Excellence (2014). Psychosis and schizophrenia in adults: prevention and management. Clinical guideline CG178, published in February 2014, updated in March 2014 and still in force at the date of this analysis, recommendation 1.3.6.10 dated 2009, which advises against initiating regular combined antipsychotic medication except for short periods, for example when changing medication. nice.org.uk. A normative source, produced by expert consensus on the basis of a literature review. It predates the register cohorts discussed here, which is in itself an argument of the Finnish authors.
Haute Autorité de santé (2007). Guide médecin, affection de longue durée n° 23, schizophrénies (physician guide, long-term condition, schizophrenias). June 2007. In France, the guide states that monotherapy is preferred, orally where possible, and that a combination of antipsychotics may be started after failure of monotherapy, in particular when one antipsychotic is being replaced by another, or in cases of resistance, with its tolerability and relevance to be reassessed very regularly. guide_ald23_schizophr_juin_07.pdf. Reservation on status: the HAS page devoted to ALD no. 23 no longer lists this guide, only the document on procedures and services updated in July 2025, which states that it is not a good practice guideline and does not constitute a decision aid on diagnostic or therapeutic strategy. Direct access to the 2007 guide on the agency’s website now redirects to a login page, and whether it is still in force could not be established.
What was consulted. References checked on 10 August 2026 on Crossref (metadata, volume, article number, funder) and on OpenAlex (PMID, affiliations, text of the published abstract, reproduced word for word); the Europe PMC query failed that day, the service returning an HTTP 429 error on every attempt, and so could not be carried out. The article is open access under a CC BY licence, but its full text could not be consulted on the publisher’s website at the date of this analysis: the findings therefore rest on the published abstract. Guidelines checked on 10 August 2026 against the CG178 guideline of the National Institute for Health and Care Excellence, the physician guide for ALD no. 23 of the Haute Autorité de santé of June 2007 and the ALD no. 23 document on procedures and services updated in July 2025. This analysis underwent an independent double reading.