Published on 16 September 2026
Lithium and suicidality: what can fifteen randomised trials show when events are this rare?
In brief
Fifteen randomised trials: 1698 patients in the eight placebo-controlled trials and 1338 in the seven open-label trials, 3036 participants in all. For suicide attempts, the odds ratio on lithium is 0.73 with a 95% confidence interval running from 0.41 to 1.31, not significant, with 25 events on lithium against 63 on placebo. For completed suicide, it is 0.61 with an interval from 0.25 to 1.48, not significant, with 4 deaths against 13. Both estimates point towards protection, and neither reaches statistical significance. The authors attribute this lack of significance to a type II error, in other words to insufficient power, and not to an absence of effect. No quantitative analysis of suicidal ideation could be carried out, because methods and measurement instruments varied too much from one trial to the next. The authors list five limitations that weigh on the sensitivity of the whole: small sample sizes, diagnostic heterogeneity, inadequate treatment fidelity, subtherapeutic lithium levels and differences between participants in prior suicidality.
The context
Few questions come up as consistently in the consulting room as this one: does lithium protect against suicide? Patients ask it when they hesitate over the burden of monitoring, families ask it, and so do colleagues who have to choose between mood stabilisers. The usual answer rests on consistent observational data, with a counter-argument the demanding reader knows well: patients who stay on lithium are not the same as those who stop it.
A meta-analysis of randomised trials is supposed to settle precisely this point. This one does not, and that is what makes it instructive. It shows the limit of a methodological device faced with a rare event, and it forces a distinction between two statements that a quick reading runs together: the trials do not demonstrate protection, and the trials demonstrate an absence of protection.
The study at a glance
| Question, population, outcomes | |
|---|---|
| Population | |
| Patients enrolled in randomised trials assessing lithium on suicidality, 3036 participants in all, the sum of the two reported subsets. The version consulted does not detail the diagnostic breakdown; it lists diagnostic heterogeneity among the limitations. | |
| Intervention | |
| Lithium. The version consulted specifies neither the formulations nor the doses used in the included trials. | |
| Comparator | |
| Placebo in the eight controlled trials. The nature of the comparators in the seven open-label trials is not specified in the version consulted. | |
| Outcomes | |
| Suicidal ideation, suicide attempts and suicide mortality. | |
| Design | |
| Systematic review and meta-analysis of randomised trials, conducted according to Cochrane methods and PRISMA guidelines. OVID databases searched, Embase, MedLine and PsychINFO, from January 2013 to July 2024, supplemented by manual reference searching for earlier studies. Of 1793 articles identified, fifteen studies met the eligibility criteria: eight placebo-controlled, 1698 patients, and seven open-label, 1338 patients. Data extraction and quality assessment by two independent reviewers. |
Quality control
| Criterion | Judgement |
|---|---|
| Quality assessment | Conducted |
| FindingThe supplementary material presents the risk of bias and quality assessment of the fifteen included studies using the National Institutes of Health Quality Assessment Tool, in fourteen items. Ten studies are rated Good, five are rated Fair, none is rated Poor. | |
| Publication bias | Not documented |
| FindingThe version consulted reports no formal test of asymmetry and no funnel plot. This is an absence of reported analysis, not a demonstrated absence of bias. The distinction matters all the more because such tests have low sensitivity anyway when the number of trials is small. | |
| Raw event counts available | Yes |
| FindingEvent counts are given arm by arm, which lets readers judge the power for themselves. That is a mark of transparency. | |
| Statistical power | Insufficient |
| FindingSeventeen completed suicides in all, 4 on lithium and 13 on placebo. The authors themselves conclude that the lack of significance most likely results from a type II error. | |
| Lithium levels achieved | Subtherapeutic |
| FindingThe authors cite subtherapeutic lithium levels among the limitations affecting the sensitivity of the analysis. The version consulted gives neither a numerical threshold nor trial-by-trial values. | |
| Homogeneity of designs | Partial |
| FindingSeven trials out of fifteen are open-label. The risk of detection bias on outcomes that are partly self-reported is not negligible, and the version consulted does not state the respective contribution of the two sets to each estimate. | |
| Length of observation | Not detailed |
| FindingThe version consulted does not report follow-up durations trial by trial. The authors recommend that future research focus on longer trials, which suggests that they regard the available durations as insufficient. | |
The findings
| Outcome | Result |
|---|---|
| Suicide attempts | Odds ratio 0.73 (95% CI 0.41 to 1.31), not significant |
| Reading25 events on lithium against 63 on placebo. The version consulted does not state the number of trials contributing to this estimate. The interval covers both a large reduction and a moderate increase. | |
| Completed suicide | Odds ratio 0.61 (95% CI 0.25 to 1.48), not significant |
| Reading4 deaths on lithium against 13 on placebo. Here again, the number of contributing trials is not stated in the version consulted. The point estimate is favourable, the precision is very low. | |
| Suicidal ideation | Pooling not possible |
| ReadingHeterogeneity in the methods and instruments used to quantify suicidal ideation prevented quantitative analysis. This outcome was therefore not analysed, which is not the same thing as a negative result. | |
| Quality of included studies | Ten rated Good, five rated Fair |
| ReadingA result read from the table in the supplementary material. The included studies range from 1973 to 2023, which spans very different standards for recording suicidal events. | |
Critical appraisal
| Domain | Judgement |
|---|---|
| Type II error | Very likely |
| FindingThis is the cardinal point of the appraisal, and it is also the authors’ explicit position. With so few events, a real protective effect would remain undetectable. The lack of significance here carries no information about an absence of effect. | |
| Plasma concentration | Bias towards the null |
| FindingThe authors retain subtherapeutic lithium levels among the explanations for the low sensitivity. If the effect depends on concentration, including such trials pulls the estimate towards no effect. This mechanism is an interpretive hypothesis; it is not tested in the version consulted. | |
| Mixing of designs | To watch |
| FindingCombining placebo-controlled trials with open-label trials complicates the interpretation of a pooled result. The exact distribution of outcomes between the two sets is not given in the version consulted. | |
| Consistency with observational data | Concordant |
| FindingThe authors note that the direction of the effect matches that of observational studies. This concordance is not proof; it makes the protective hypothesis more plausible than it would be with divergent results. It is an argument from consistency, not a causal argument. | |
| What not to write | Caution |
| FindingThat the trials show lithium to be ineffective against suicide risk. They show that they are not sized to answer the question. | |
Level of evidence
A meta-analysis of randomised trials, which corresponds to level 1a on the Oxford scale. This classification is an editorial reading of the design; it does not appear in the publication. The level of evidence of a design is not the same as the solidity of the answer: here the design ranks high and the question remains open. What is demonstrated is that the available randomised trials do not allow a conclusion on this outcome. What is suggested is the existence of a protective effect, supported by the convergent direction of the two estimates and by consistency with observational data. What is expert opinion is the weight given to these long-term follow-up data in clinical decision-making.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ Four deaths against thirteen, and I am told it is not significant. Of course it is not significant, there are not enough events for it to become so. It changes nothing about what I do for a bipolar patient at risk of suicide. ”
What this means in practice: this meta-analysis brings nothing that would lead to reconsidering the place of lithium. It changes the way we talk about it, by accepting that on this precise point the experimental evidence is lacking and the available data are mainly observational.
What you can do with this
- Do not change a lithium prescription on the basis of this study alone. It provides no argument for a lack of efficacy on suicide risk.
- Be ready to answer a patient who has read that lithium does not reduce suicide risk: the comparison rests on too few events to settle the question, the two estimates point towards protection, and the experimental evidence remains to be established.
- Check the lithium level before interpreting an unfavourable course. A concentration that has stayed below the therapeutic range describes insufficient exposure, a situation from which no lack of efficacy of the treatment can be inferred.
- In any woman of childbearing potential, apply the rules on information, contraception and monitoring set out in the lithium product information in force where you practise, independently of this analysis.
- In teaching, this article is a textbook case for showing that a wide confidence interval is information, not a missing result.
Frequently asked questions
Does lithium reduce the risk of suicide?
The randomised trials gathered here cannot demonstrate it, for want of a sufficient number of events. The two estimates point towards a reduction, and the authors note their consistency with observational studies, which suggests an effect without establishing it.
Why not simply run a larger lithium trial on suicide?
It would take very large sample sizes and prolonged follow-up to observe enough events. The authors indeed recommend longer trials using standardised suicide risk assessment tools.
What does a non-significant odds ratio of 0.61 mean?
That the point estimate corresponds to a risk reduction, but that the data remain compatible with a range of values running from a large reduction to a moderate increase.
Why was suicidal ideation not meta-analysed?
Because the methods and measurement instruments differ too much from one trial to another to be pooled. No conclusion, in either direction, can be drawn on this outcome.
Should another mood stabiliser be preferred for suicide risk?
This study provides no comparison between drugs on this outcome. The question is handled with current guidelines and the patient’s profile.
Annotated bibliography
Source study. Wang JX, Le GH, Wong S, Teopiz KM, Kwan ATH, Rosenblat JD, Rhee TG, Ho R, Lo HKY, Goldberg JF, Vinberg M, Grande I, Mansur R, Meyer JM, McIntyre RS. The efficacy of lithium in the treatment of suicidal ideation, behavior and suicide: An updated systematic review and meta-analysis of randomized controlled trials. Journal of Affective Disorders, 2025, volume 387, article 119487. Volume 387, dated 15 October 2025, carries no issue number in the version consulted. DOI 10.1016/j.jad.2025.119487, PMID 40441661. Open access publication under a Creative Commons licence, copyright notice 2025 The Authors, published by Elsevier B.V.
Protocol registration. No protocol registration number appears in the version consulted. The authors state that they followed Cochrane methods and PRISMA guidelines.
Funding. The funding statement does not appear in the version consulted. Its absence does not allow the conclusion that the work was unfunded.
Competing interests. The declaration of competing interests does not appear in the version consulted. Its absence does not allow the conclusion that there are no competing interests. The make-up of the author list, which brings together several clinicians who are very active in the psychopharmacology of bipolar disorder, is reason enough in itself to consult the declaration in the full version before any reuse.