Published on 17 September 2026
ADHD stimulants in bipolar disorder: does the fear of a manic switch survive 939 hospitalisations for mania?
BMJ Mental Health · 2026 · Ermis C, Tanskanen A, Corbeil O, Lieslehto J, Vieta E, Correll CU, Mittendorfer-Rutz E, Tiihonen J, Taipale H
DOI 10.1136/bmjment-2025-302159
PMID 41856665
Scientific 74
Editorial 79
Myth or Reality
The essentials
Among 17,971 Swedish people aged 16 to 65 with bipolar disorder who had received at least one ADHD treatment, followed for a mean of 8.9 years, periods on a stimulant added to an antipsychotic or a mood stabiliser are associated with fewer psychiatric hospitalisations, adjusted hazard ratio 0.89, and markedly fewer substance-related hospitalisations, 0.75. Hospitalisations for mania, which occurred in 939 subjects, are not increased, hazard ratio 0.93, 95% CI 0.72 to 1.20. The design compares each patient with themselves, which neutralises stable differences between individuals. It does not neutralise the bias linked to the fact that treated periods are often more stable periods, and this point limits the reach of the result.
Context
ADHD and bipolar disorder often coexist, about one in six bipolar adults according to the literature cited by the authors, and the situation puts the clinician in a difficult position. The patient is mood-stable, remains inattentive, disorganised, impulsive, and that is enough to block a return to work or study. Treatment for ADHD is indicated. The fear of a manic switch, for its part, is old and widely taught, and it often leads to trying nothing at all.
This fear rests on pharmacological plausibility, the dopaminergic effect of stimulants, and on earlier work reporting an increased risk of psychotic symptoms, hypomanic or manic switch, and mood destabilisation. One point from that earlier work deserves to be kept in mind: the excess risk of manic switch had been observed under methylphenidate monotherapy, not when methylphenidate was added to an antipsychotic or a mood stabiliser. The authors stress that good-quality, long-term data are still lacking in this population, and they explicitly call for randomised trials. In the meantime, registry-based pharmacoepidemiology remains the main available source.
The study at a glance
| Element | Content |
|---|---|
| Population | 17,971 people aged 16 to 65, in Sweden, with bipolar disorder and at least one dispensed ADHD treatment |
| DetailDrawn from a base cohort of 105,495 bipolar patients, 17.0% of that cohort. Swedish national registries, enrolment from 1 January 2006 to 31 December 2021, mean follow-up 8.9 years, standard deviation 4.4. Mean age 32.0 years, 37.6% men, 7.8% aged 16 or 17 at entry. An ADHD diagnosis is recorded for 88.9% of them, inclusion being based on treatment rather than diagnosis | |
| Exposure | Stimulants, methylphenidate, lisdexamfetamine, dexamfetamine, amphetamine, and non-stimulants, atomoxetine and modafinil |
| DetailTreatments added to a mood stabiliser or an antipsychotic. Treatment periods reconstructed from pharmacy dispensations using the validated PRE2DUP algorithm. Only three molecules had sufficient numbers for separate analysis: methylphenidate, lisdexamfetamine and atomoxetine. Clonidine and guanfacine were excluded, for lack of an indication or of sufficient use in adults in Sweden | |
| Comparator | Periods in the same individual on an antipsychotic or a mood stabiliser alone, with no associated ADHD treatment |
| DetailEach patient is their own control. Factors stable over time, genetic, social, personality-related, are neutralised by construction. The reference category is never the absence of any treatment | |
| Endpoints | Primary endpoint, psychiatric hospitalisation. Secondary endpoints, substance-related, somatic, and manic hospitalisations, and a composite endpoint of all-cause hospitalisation or death |
| DetailObjective endpoints drawn from hospital registries and the cause-of-death registry, not self-reported. Mortality is not analysed on its own: it appears only within the composite endpoint | |
| Design and analysis | National registry cohort, within-individual analysis |
| DetailStratified Cox models, each subject forming their own stratum, adjusted for time-varying covariates: temporal order of treatment sequences, time since cohort entry, antidepressant use, use of benzodiazepines and related drugs. Bonferroni correction for 25 comparisons, significance threshold set at p less than 0.0020. Conventional between-individual models served as a sensitivity analysis | |
Quality control
| Point checked | Judgement |
|---|---|
| Within-individual comparison | Major strength |
| Finding With the patient as their own control, fixed confounding factors disappear. This is what sets this work apart from comparisons between treated and untreated patients | |
| Registry completeness | National |
| Finding Prospective collection, with no loss to follow-up through non-response. Hospitalisations are a hard, verifiable endpoint | |
| Correction for multiplicity | Bonferroni applied |
| Finding Five treatment categories crossed with five endpoints, 25 comparisons, and a threshold lowered to 0.0020. The sensitivity analyses, however, are not corrected and should be read as exploratory | |
| Power varies by endpoint | Uneven across endpoints |
| Finding Not all 17,971 subjects contribute to every estimate. In a within-individual model, only subjects who experienced the event and whose exposure varied are counted. Psychiatric hospitalisation involves 7,951 subjects, substance-related hospitalisation 3,860, somatic hospitalisation 4,646, and manic hospitalisation only 939. Hence narrow intervals on the frequent endpoints and wide ones on mania | |
| Healthy-adherer bias | Not eliminated |
| Finding Treatment periods may coincide with periods of better stability, for reasons that precede the prescription. The within-individual design does not protect against factors that vary over time. The authors themselves identify this as the main limitation | |
| Observational nature | No randomisation |
| Finding Possible confounding by indication. The decision to prescribe is made by a clinician who judges the situation to be favourable, which is not neutral | |
| Non-stimulants | Class conclusive, molecules not |
| Finding The non-stimulant class reaches the threshold on the primary endpoint, with a hazard ratio of 0.85. Atomoxetine, modafinil, dexamfetamine and amphetamine salts, however, are not used enough in Sweden to draw molecule-by-molecule conclusions. Absence of a result there does not amount to absence of effect | |
| Scope of the endpoints | Severe events only |
| Finding As the authors note, the endpoints capture only what leads to hospital. Mild somatic adverse effects and hypomanic episodes managed in the community escape this design entirely | |
| Transposition to other jurisdictions | Limited |
| Finding Conditions for prescribing stimulants to adults differ markedly across countries, Sweden and France among them, and should be reverified against the current rules in the reader’s own jurisdiction | |
Results
| Endpoint and exposure | Adjusted hazard ratio |
|---|---|
| Psychiatric hospitalisations, stimulant added | 0.89 [0.85 to 0.93] |
| Reading Primary endpoint. Reduction of about 11%. A modest effect, but estimated precisely on a large sample | |
| Psychiatric hospitalisations, non-stimulant added | 0.85 [0.77 to 0.93] |
| Reading Class-level result, significant after correction, with a p of 0.0008 against a threshold of 0.0020. It does not transfer to the individual molecules: at the molecule level, atomoxetine does not reach the threshold on this endpoint | |
| Lisdexamfetamine, psychiatric hospitalisations | 0.81 [0.75 to 0.87] |
| Reading The most favourable signal among the molecules. A between-molecule comparison in an observational design also reflects prescribing habits | |
| Methylphenidate, psychiatric hospitalisations | 0.92 [0.88 to 0.97] |
| Reading The most prescribed molecule in the cohort. The reduction is real but smaller than under lisdexamfetamine | |
| Substance-related hospitalisations, stimulant added | 0.75 [0.70 to 0.81] |
| Reading A reduction of about 25%, the largest effect in the study, found for lisdexamfetamine at 0.70 and methylphenidate at 0.80, but not for atomoxetine or for the non-stimulant class. Worth noting: this is also the endpoint where healthy-adherer bias weighs most heavily | |
| All-cause hospitalisation or death, stimulant added | 0.90 [0.87 to 0.93] |
| Reading Composite endpoint, occurring in 12,243 subjects, 68.1%. It is very largely dominated by hospitalisations: mortality was not analysed separately, and this figure should not be read as a reduction in mortality | |
| Somatic hospitalisations, stimulant added | 1.00 [0.90 to 1.12] |
| Reading A neutral result. It concerns only somatic events severe enough to lead to hospital, in a population with a median age at entry of 30 years, and says nothing about cumulative long-term effects | |
| Hospitalisations for mania, stimulant added | 0.93 [0.72 to 1.20] |
| Reading This is the result the study was designed to answer. No increase appears, and the upper bound of the interval leaves room for an increase of up to 20% | |
One methodological point deserves to be stated plainly, because it governs the whole reading. The mania endpoint is the one that carries the clinical stakes, and it is also the one with the widest interval. The reason is simple: 939 subjects were hospitalised for mania during follow-up, against 7,951 for any psychiatric reason at all. The correct conclusion is therefore not that stimulants do not trigger a manic switch, but that no increase is detected and that a moderate increase remains compatible with the data.
Critical appraisal
| Domain | Judgement |
|---|---|
| Design | Fit for the question |
| Finding The within-individual comparison is the best observational tool for an intermittent exposure in an episodic illness | |
| Endpoints | Objective |
| Finding Hospitalisation is a recorded event, not subject to a rater’s judgement or to the patient’s memory | |
| Time-varying confounding | Main limitation |
| Finding A period on treatment is often a period when the patient is doing better, attends follow-up, adheres to care. The observed association may reflect this state rather than the drug’s effect. No analysis can fully remove this objection. The authors only counter that, over such a long follow-up, clinical monitoring tight enough for every treated period to be a stable phase seems unlikely | |
| Causality | Not demonstrated |
| Finding The direction of the association cannot be established. It is equally consistent to read that stability enables the prescription as the reverse | |
| Power on the manic endpoint | Moderate |
| Finding An interval running up to 1.20 rules out concluding to an absence of effect. The result is reassuring, it is not exonerating | |
| Selection of the population | Relatively stabilised patients |
| Finding The authors say so explicitly: the results hold only for patients already receiving an antipsychotic or a mood stabiliser, hence for a relatively stable subgroup, and do not generalise to unstable or untreated forms | |
| Care setting | Swedish |
| Finding The coverage of psychiatric follow-up and access to outpatient care differ elsewhere, which changes the probability that an episode results in hospitalisation | |
Level of evidence
Confidence is good that no signal of manic worsening appears at this scale. It is weak on the existence of a genuine treatment benefit, because the confounding mechanism at stake, the clinical stability that precedes and enables the prescription, is exactly what would produce these same figures in the complete absence of any pharmacological effect.
What is observed: in these registries, periods on an added stimulant are not accompanied by more hospitalisations for mania or more somatic hospitalisations. What is suggested: a reduction in psychiatric and substance-related hospitalisations, clear for lisdexamfetamine and methylphenidate. What is not established: that this reduction is caused by the treatment.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ In my stabilised bipolar patients who also have ADHD, this gives me something to work with instead of refusing outright. But I’m not going to promise them it protects them: the ones on a stimulant are the ones who were already doing better. ”
What this means in practice: the result shifts the balance of the decision without making it for the clinician. It makes it defensible to try treatment in a stabilised patient with good mood control, under monitoring, after informing them of the actual level of evidence. It does not make it defensible to start a stimulant in a patient in an unstable phase, a situation this study did not examine.
What you can take from this on Monday morning
- Stop opposing a blanket refusal to a request for ADHD treatment in a stabilised bipolar patient already on an antipsychotic or a mood stabiliser. The excess risk of manic switch documented previously had been found under methylphenidate monotherapy, and is not found here when the stimulant is added to existing treatment.
- Set documented mood stability and an established background treatment as a precondition, which matches the population actually studied and echoes the clinical implication the authors themselves draw.
- Agree with the patient and their family on warning signs and on what to do, before starting treatment rather than after.
- Present the level of evidence honestly in consultation: this is registry data, not a randomised trial, and the study establishes the absence of a worsening signal far more solidly than it establishes a genuine benefit.
- Do not transpose the result molecule by molecule. The reduction in psychiatric hospitalisations holds for lisdexamfetamine and methylphenidate, and for the non-stimulant class taken as a whole, but the numbers allow no individual conclusion on atomoxetine, modafinil, dexamfetamine or amphetamine salts.
- Keep up routine cardiovascular and metabolic monitoring. The study shows no excess of somatic hospitalisation, but it captures neither mild adverse effects nor effects that accumulate beyond its follow-up horizon.
Frequently asked questions
Does this study prove that stimulants do not trigger mania?
No. It shows that no increase is detected among the 939 subjects hospitalised for mania during follow-up. The confidence interval leaves room for an increase of up to 20%, and the design does not allow causal attribution of what is observed.
What is healthy-adherer bias, in concrete terms?
The periods during which a patient takes a treatment are often the periods when they attend follow-up, are doing better, and take care of their health. These periods would be associated with fewer hospitalisations even if the drug had no effect at all. The within-individual design does not correct for this.
Is lisdexamfetamine preferable?
The signal is more favourable for it in this study, 0.81 on psychiatric hospitalisations against 0.92 for methylphenidate. But the molecules were not compared with each other: each was compared with periods without ADHD treatment in the same patients. The comparison is not randomised and also reflects the profiles of patients to whom each molecule is prescribed. This data is not sufficient to establish a hierarchy.
What about atomoxetine?
It was analysed separately, in 2,470 patients, and does not reach the significance threshold on either psychiatric or substance-related hospitalisations. The authors attribute this result to a lack of power after correction for multiple comparisons, and call for larger studies. This does not mean the molecule has no effect.
Are these results applicable outside Sweden?
Conditions for prescribing stimulants to adults differ between countries, and in France, for instance, the context is more restrictive. The result informs the discussion of benefit against risk, it does not establish a new practice on its own, and readers should check the specifics of their own jurisdiction.
Annotated bibliography
Source study. Ermis C, Tanskanen A, Corbeil O, Lieslehto J, Vieta E, Correll CU, Mittendorfer-Rutz E, Tiihonen J, Taipale H. Reduced risk of cause-specific hospitalisations and all-cause hospitalisation/mortality during treatment with attention-deficit/hyperactivity disorder medications in the course of bipolar disorder: a Swedish registry-based within-subject cohort study. BMJ Ment Health. 2026 Mar 19;29(1):e302159. doi: 10.1136/bmjment-2025-302159. PMID 41856665. PMCID PMC13034203.
Declared funding. H. Taipale was funded by the Sigrid Juselius Foundation, J. Lieslehto by the Finnish Medical Association, grant 7709. Data came from the REWHARD consortium, supported by the Swedish Research Council, grant 2021-00154. Study approved by the regional ethics committee of the Karolinska Institutet, Dnr 2007/762-31 and Dnr 2021-06441-02. Article not commissioned, externally peer reviewed. Conflict-of-interest statements verified against the freely available full text. Four of the nine authors declare extensive and varied links with the pharmaceutical industry, outside the submitted work: E. Vieta for thirty-seven companies, as grants, consulting or continuing medical education; C. U. Correll for at least sixty-seven companies in consulting or honoraria, in addition to expert testimony for four laboratories, participation on data monitoring committees for five trials, grants from three companies, UpToDate royalties, and holdings in six biotechnology companies. J. Tiihonen declares honoraria from Janssen, Lundbeck and Otsuka, and several consulting assignments. H. Taipale declares honoraria from Gedeon Richter, Janssen, Lundbeck and Otsuka. Four authors, Tiihonen, Taipale, Tanskanen and Mittendorfer-Rutz, also took part in research projects funded by grants from Janssen-Cilag paid to their institution, outside the submitted work. The remaining three authors, Ermis, Lieslehto and Corbeil, declare no conflicts of interest.
Base cohort. Ermis C, Taipale H, Tanskanen A, et al. Real-world effectiveness of pharmacological maintenance treatment of bipolar depression: a within-subject analysis in a Swedish nationwide cohort. Lancet Psychiatry. 2025;12:198-207. doi: 10.1016/S2215-0366(24)00411-5. Cohort of 105,495 bipolar patients from which the 17,971 subjects analysed here were drawn.
Supplementary material. An online supplementary file, bmjment-29-1-s001.docx, contains tables S1 to S5 and figures S1 to S13, including the sensitivity and subgroup analyses. It was not included in the version consulted and could therefore not be reviewed.