Published on 16 September 2026
Lithium, valproate, antipsychotics: which is linked to lower suicide risk in bipolar disorder?
In brief
A Korean nationwide cohort covers 44,694 patients with bipolar disorder, followed through healthcare claims data between 2002 and 2020. As monotherapy, the risk of suicide incidents is 39% lower during periods on lithium and 26% lower on valproate, compared with periods without lithium, valproate or an atypical antipsychotic. The main contribution lies elsewhere: the authors also compare each patient with themselves across their treatment periods. In that second analysis, the lowest hazard ratio goes to lithium combined with an antipsychotic, at 0.154, and valproate alone, at 0.251, does better than valproate combined with an antipsychotic, at 0.302. This is an observational study: none of these figures establishes causation.
The context
Suicide shapes the management of bipolar disorder. Whether maintenance treatments genuinely protect against suicidal acts, and which ones, nonetheless remains hard to settle, for a simple reason: the question cannot be put to a randomised trial. Nobody is going to randomise the treatment of patients at high risk of suicide to see what happens.
We are therefore left with observational data, and with their cardinal bias. Confounding by indication: the heaviest treatments preferentially go to the patients perceived as being at highest risk. Comparing the patients who receive them with everyone else then amounts to comparing different populations, and any protective effect is mechanically attenuated by the severity of those who receive it.
The method
Comparing each patient with themselves
The methodological contribution lies in applying two analyses to the same cohort.
The between-individual analysis is the usual Cox model: it compares groups of patients. It remains exposed to confounding by indication, since no adjustment captures real clinical severity.
The within-individual analysis compares each person with themselves, between their treated and untreated periods. Everything that does not vary over time in that person, genetic background, personality, a stable environment, drops out of the equation. What does persist are clinical changes over time: a patient on a combination is often on it because they are doing less well, and the analysis does not erase that.
The study at a glance
| Question (PICO) | |
|---|---|
| Population | |
| 44,694 patients with at least two principal diagnoses of bipolar disorder and at least two prescriptions for lithium, valproate, carbamazepine or an atypical antipsychotic. Mean age 31.1 years, 58.1% women | |
| Exposure | |
| Periods of treatment with lithium, valproate or an atypical antipsychotic (risperidone, quetiapine, olanzapine, aripiprazole, ziprasidone), alone or in combination, compared with periods without any of these treatments | |
| Outcome | |
| Suicide attempts and deaths by suicide | |
| Period | |
| 2002 to 2020, Korean national health insurance data | |
| Design | |
| Nationwide retrospective cohort, between-individual and within-individual Cox models · CEBM 2b |
The quality check
| Criterion | Status |
|---|---|
| Size and duration | Sound |
| Finding44,694 patients, data collected from 2002 to 2020. The record reports neither the mean length of follow-up nor the number of person-years | |
| Dual analytical approach | Sound |
| FindingThe within-individual analysis neutralises constant factors, which no multivariable adjustment does | |
| Nature of the data | Reservation |
| FindingClaims data: neither clinical severity, nor the doses actually taken, nor psychosocial factors are known. A prescription is not a dose taken | |
| Precision of the estimates | Reservation |
| FindingVery wide intervals in the within-individual analysis, up to 0.056 to 0.980 for triple therapy | |
| Pooling of molecules | Reservation |
| FindingResults are reported only for the class of atypical antipsychotics, which groups five molecules with different profiles. Carbamazepine appears in the inclusion criteria but no result is given for it | |
| External validity | Reservation |
| FindingKorean population: prescribing practices and sociocultural context that cannot be assumed to match those of other settings | |
The findings
| Treatment | Between individuals |
|---|---|
| Lithium alone | 0.608 (0.434 to 0.852), i.e. −39.2% |
| Within individualNot reported in the published abstract | |
| Valproate alone | 0.740 (0.577 to 0.949), i.e. −26.0% |
| Within individual0.251 (0.090 to 0.698) | |
| Lithium and antipsychotic | Not reported in the published abstract |
| Within individual0.154 (0.055 to 0.428) | |
| Valproate and antipsychotic | Not reported in the published abstract |
| Within individual0.302 (0.152 to 0.599) | |
| Lithium, valproate and antipsychotic | Not reported in the published abstract |
| Within individual0.235 (0.056 to 0.980), upper bound close to the threshold | |
The published abstract does not report the same treatments in the two analyses, which rules out drawing up a single ranking. It reports the monotherapies in the between-individual comparison, and mainly the combinations in the within-individual comparison. The most parsimonious explanation for this asymmetry is confounding by indication: the heaviest treatments go to the most severely ill patients, which flattens their apparent benefit for as long as different people are being compared. This is an interpretive hypothesis, consistent with the authors’ conclusion, not a result reported as such.
The authors’ conclusion is cautious and deserves to be quoted exactly: “Suicide risk decreased during treatment with lithium or valproate alone. In patients with bipolar disorder who are at a high risk of suicide, atypical antipsychotics were associated with a reduced risk of suicide when combined with lithium or valproate.” Associated with, not responsible for.
Critical appraisal
Three limitations weigh on the interpretation, in this order of importance.
The design remains observational. The within-individual analysis neutralises what is stable in a person, not what changes with them. An episode of decompensation leads both to treatment being stepped up and to a rise in suicide risk: confounding persists, in a direction the study does not allow us to quantify.
The data are administrative. They show what was claimed, not what was taken, nor at what dose, nor with what adherence. The method used to identify suicide attempts and deaths is not detailed in the published abstract.
Two major treatments are missing: lamotrigine, a commonly discussed option in bipolar depression, and antidepressants. Their absence from the model limits direct transfer of the results to any given clinical caseload.
Level of evidence
PEB appraisal: moderate confidence in the direction of the effects, low confidence in their magnitude. The dual analysis strengthens the credibility of the observed direction without allowing a hazard ratio to be treated as the result of a randomised trial. On a question where a randomised trial is impossible, this is probably the best level of evidence attainable, and that does not make it causal.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ Lithium is still lithium, that does not change. What interests me is understanding why combinations looked ineffective in all the earlier studies: we were comparing severely ill patients with less severely ill ones. That said, a Korean claims database with no lamotrigine is not going to make me rewrite my prescriptions. ”
What this means in practice: the study corrects an optical illusion more than it establishes a hierarchy. It supports lithium and warns against concluding too quickly that a combination is useless.
What you can do with this on Monday morning.
- Reaffirm the place of lithium in suicide prevention: it is the monotherapy with the lowest relative risk in the between-patient comparison, nearly 40% lower, with valproate alone also significant at 26% lower.
- In a high-risk patient already on a mood stabiliser, consider adding an atypical antipsychotic as a supported option, without turning it into a rule on the strength of this study alone.
- Stay wary, for good, of the reasoning that concludes a treatment is ineffective because the people who receive it are doing worse. That is the most transferable lesson of this work, well beyond bipolar disorder.
- Bear in mind that the reference category for all these comparisons is the period without lithium, valproate or an atypical antipsychotic. The observed reductions in risk relate to those periods, which argues for particular vigilance during windows without maintenance treatment.
A reservation that overrides everything else. Prescribing valproate and its derivatives in bipolar disorder, during pregnancy and to girls and women of childbearing potential, is governed by the regulatory text in force where you practise, and that text is what to check before any prescription. The cohort analysed here is 58% women and does not take this constraint into account. No antisuicidal signal justifies setting it aside.
Frequently asked questions
Does lithium plus an antipsychotic reduce suicide risk by 85%?
No. That figure comes from an observational within-individual analysis, with a confidence interval running from 0.055 to 0.428. It indicates a direction, not a magnitude that can be carried over to an individual patient.
Why do between-individual and within-individual analyses give opposite results?
Because they do not answer the same question. The first compares different people, the second compares periods within the same person. When the heaviest treatments go to the most severely ill patients, only the second shows the effect of the treatment rather than the effect of severity.
Do these Korean cohort results apply to routine practice elsewhere?
Partly. The methodological reasoning is universal. The ranking of the drugs, on the other hand, comes from a Korean population, with neither lamotrigine nor antidepressants in the model, which limits direct transfer.
Is an atypical antipsychotic alone enough to lower suicide risk in bipolar disorder?
The published abstract reports no result for atypical antipsychotics alone. What the authors state is that the reduction in risk is associated with them when they are combined with lithium or valproate.
Annotated bibliography
Park SA, Son S, Tae BS, Choi H, Jeong JH, Yoon HK, Shin C, Kwon DY, Ko YH (2026). Long-term treatment with lithium, valproate, and atypical antipsychotics on suicide risk in patients with bipolar disorder: a nationwide retrospective cohort study. Journal of Affective Disorders, 392, 120264. DOI 10.1016/j.jad.2025.120264 · PMID 40930182. The source study analysed here.