Published on 16 September 2026
Does methylphenidate raise the risk of psychotic disorder in young people with ADHD?
In brief
In the Finnish birth cohort of 1987 to 1997, 697,289 people in all, 3956 received before the age of 18 a diagnosis of ADHD made in public hospital services from 2003 onwards. Registry follow-up ends on 31 December 2016: participants are then 22.2 years old on average, 29.8 years at most. Sustained methylphenidate treatment is not associated with an increase in the incidence of nonaffective psychosis in the overall sample. The methodological feature that sets this work apart is the use of an instrumental variable, the prescribing propensity of the hospital district of residence, designed to create variation in access to treatment that is independent of individual clinical severity. In a secondary analysis, among children diagnosed before the age of 13, a signal in the opposite direction appears, with a risk difference of −0.24, 95% confidence interval −0.47 to −0.03, p = 0.03 for the three-year window. This signal rests on a subgroup whose size is not given in the article, and it needs replication.
The context
The question comes up at every initiation. A parent has read somewhere that methylphenidate drives children mad, or that it sets off psychosis in adolescence. The fear is not absurd: methylphenidate is a psychostimulant, amphetamines at high doses produce psychotic states, and adolescence is the period when psychotic disorders emerge. The temporal coincidence between treatment and the age of onset is enough to plant a lasting doubt.
The earlier literature is contradictory, and the authors say so. A recent meta-analysis reports an increased risk of psychotic disorder in young people treated for ADHD, as does a Taiwanese registry study with follow-up of up to twelve years, while other work finds nothing. Most of these studies have a follow-up of less than two years, whereas the peak age at onset of psychosis is 20.5 years.
The problem with observational studies on this subject is well known. Children who receive treatment differ from those who do not, in the severity of their ADHD, in their comorbidities, in their family environment, and these differences are themselves associated with psychotic risk. This is confounding by indication. Adjusting for what is measured does not solve the problem of what is not measured. The instrumental variable is one possible answer to this impasse.
What an instrumental variable measures
The idea is to use a factor that influences the probability of receiving treatment without directly influencing the risk of the disease under study. Here, that factor is not a characteristic of the patient but a characteristic of the patient’s area: the prescribing propensity of the hospital district they belong to. Finland is divided into public hospital districts, and each person is assigned to one of them according to their place of residence. These districts prescribe more or less methylphenidate, for reasons of local policy, service habits and resources, unrelated to the severity of a given child’s ADHD.
The instrument is built as the average quantity of defined daily doses filled in the district, for each intervention window, with a leave-self-out procedure so that a person’s own treatment does not contribute to their own estimated exposure. The variability is real: over the first year, propensity ranges from 0.07 to 0.30 across districts.
What this approach measures: the effect of treatment in patients for whom the decision to treat was influenced by local prescribing practice. What it does not measure: the average effect across the whole population, nor the effect in patients who would have been treated whatever their district. The authors say so themselves: their estimates are to be read as a “policy-level local average treatment effect”, which includes the drug’s own effect and the organisational or social spillovers of local practice. What it assumes: that district propensity acts on psychotic risk only through the channel of treatment. This exclusion assumption cannot be proven, but it was subjected to falsification tests, which produced no argument against it.
The study at a glance
| Item | Content |
|---|---|
| Population | 3956 individuals diagnosed with ADHD before the age of 18, drawn from the Finnish birth cohort of 1987 to 1997 |
| DetailSource cohort of 697,289 births. ADHD diagnoses coded ICD-10 F90.X made from 1 January 2003, after methylphenidate was brought to market in Finland in December 2002. Median age at diagnosis 14.16 years, interquartile range 11.78 to 15.93 years. Analysis samples of 3888 for nonaffective psychosis and 3953 for schizophrenia, after excluding cases that occurred before the ADHD diagnosis or during the intervention window. | |
| Exposure | Cumulative amount of methylphenidate reimbursed within four intervention windows of 1, 2, 3 and 4 years after the ADHD diagnosis |
| DetailContinuous variable standardised in defined daily doses, ATC code N06BA04. A value of 0 corresponds to no treatment over the window, a value of 1 to 30 mg per day for the whole window. The data come from the reimbursement records of the Social Insurance Institution of Finland (Kela) and from the medical products registry of the Finnish Medicines Agency. Methylphenidate accounts for more than 90% of the stimulant reimbursements recorded in this sample. | |
| Comparator | No separate control group, exposure is continuous and instrumented |
| DetailA descriptive analysis compares treated and untreated individuals, 2728 people, or 69.0%, received at least one reimbursement within four years. But the causal estimate rests on the variation in prescribing propensity between hospital districts, not on a direct comparison of treated and untreated individuals. | |
| Outcome | Diagnosis of nonaffective psychosis, ICD-10 codes F20.x, F22.x, F23.x, F24, F25.x, F28 and F29, by 31 December 2016 |
| DetailA second outcome covers schizophrenia, code F20 alone. Mean length of follow-up 8.47 years, SD 3.00. Mean age at the end of follow-up 22.16 years, SD 2.39, range 19.00 to 29.81 years. | |
| Design | National registry cohort with instrumental variable analysis |
| DetailInstrument: hospital district prescribing propensity, one per intervention window. Estimation by two-stage least squares, ivregress procedure in Stata 18. Significance assessed with the Anderson-Rubin test, robust to weak instruments. Confidence intervals obtained by inverting these tests. Standard errors clustered at hospital district level. First-stage F statistics from 22.68 to 39.01 depending on the window. Analyses conducted from June 2023 to December 2025. | |
Quality control
| Criterion | Judgement |
|---|---|
| Strength of the instrument | F from 22.68 to 39.01 |
| FindingAbove the usual threshold of 10 for the overall sample. The authors nonetheless used Anderson-Rubin tests and a tF ratio adjustment, precisely because these values do not exempt them from inference that is robust to weak instruments. In the stratum diagnosed at age 13 or older, the instrument was too weak to allow the analysis. | |
| Completeness of data capture | Public specialist care only |
| FindingThe Hilmo care register covers only inpatient stays and secondary-level outpatient visits in the public sector. People managed exclusively in primary care or in the private sector are not in the sample. Reimbursements, on the other hand, are recorded regardless of age, wealth or address. | |
| Length of follow-up | 8.5 years on average |
| FindingFollow-up ended on 31 December 2016. Mean age of 22.16 years at the end of follow-up, 29.81 years at most. With the peak incidence of psychosis at around 20.5 years, follow-up extends beyond that peak for only part of the cohort. Diagnoses made after the age of 25 largely escape this design. | |
| Funding and competing interests | Not stated in the source |
| FindingThe full text consulted does not include the funding statement or the declaration of competing interests. These elements therefore cannot be characterised here. Ethics approval was given by the ethics committee of the Finnish Institute for Health and Welfare, and individual consent is not required for register-based studies under Finnish law. | |
| Validity of the instrument | Tested, cannot be proven |
| FindingThe exclusion, independence and monotonicity assumptions were subjected to falsification tests, which the authors report produced no evidence against them. The authors nevertheless acknowledge that they cannot rule out that some unmeasured aspect of a district’s diagnostic culture or care pathway influences both prescribing and psychotic risk. | |
| Size of the positive subgroup | Not given in the article |
| FindingThe number of people diagnosed before the age of 13 and the number of events in this stratum appear in the supplementary material, which could not be consulted. Since the median age at diagnosis is 14.16 years in the whole sample, this stratum is necessarily a minority. | |
| Transfer beyond Finland | To be qualified |
| FindingThe conditions for prescribing methylphenidate, the organisation into hospital districts and diagnostic thresholds differ from one country to another. The Finnish guidelines on the role of primary care in ADHD were, moreover, revised in 2017, after the end of follow-up. | |
The findings
| Outcome | What the data show |
|---|---|
| Excess risk linked to ADHD itself | Odds ratio 3.83 for nonaffective psychosis, 95% CI 3.34 to 4.39, p < 0.001 |
| InterpretationUnadjusted comparison of people with and without ADHD. For schizophrenia, odds ratio 2.37, 95% CI 1.67 to 3.36. This is the background risk that treatment is suspected of worsening. | |
| Overall association, instrumented analysis | Risk difference −0.14 at one year, 95% CI −0.85 to 0.42; −0.15 at four years, 95% CI −0.49 to 0.11 |
| InterpretationThe intermediate windows give −0.09, 95% CI −0.47 to 0.37, at two years, and −0.14, 95% CI −0.49 to 0.27, at three years. All the intervals cross zero. Number of events per window: 193, 164, 136 and 109. Mind the unit: a risk difference of −0.14 is the change in probability, in percentage points expressed as a proportion, associated with one unit of exposure, that is 30 mg per day for the whole window. The very wide intervals of the first window reflect the low precision of the estimate. | |
| Diagnosed before age 13, three-year window | Risk difference −0.24, 95% CI −0.47 to −0.03, p = 0.03 |
| InterpretationA signal in favour of reduced risk, from stratified secondary analyses. The article’s abstract gives −0.45 for this lower bound and the main text −0.47, an internal discrepancy in the publication that does not change the conclusion. It must not be presented as an established benefit. | |
| Diagnosed before age 13, four-year window | Risk difference −0.21, 95% CI −0.48 to −0.07, p = 0.02 |
| InterpretationAt the one-year and two-year windows, the trend is there but does not reach the threshold: −0.27, 95% CI −0.62 to 0.04, p = 0.07, then −0.28, 95% CI −0.55 to 0.01, p = 0.06. The consistent direction across the four windows argues against an isolated artefact, without settling the question of power. | |
| Diagnosed at age 13 or older | Analysis not possible, instrument not strong enough |
| InterpretationThis point is essential and often misreported. It is not an absence of signal in adolescents, it is an absence of result. The authors state explicitly that they cannot rule out an increased risk in this stratum, all the more so because the overall result includes the childhood stratum, where risk appears reduced. | |
| Schizophrenia | No relationship found, only 32 cases |
| InterpretationRisk differences of −0.06 to −0.07 depending on the window, all intervals crossing zero. Cases were too few to stratify by age at diagnosis. | |
| Short term | 6 nonaffective psychoses at 3 months, 11 at 6 months, none on methylphenidate |
| InterpretationNone of these people had been dispensed methylphenidate within that window. A descriptive finding, on numbers too small to draw conclusions, but useful in the face of the fear of an immediate trigger. | |
Two reading precautions. First, the absence of association observed in the overall sample is not a demonstration of absolute safety. It means that, in this cohort, with this method and this power, no excess risk appears, and that the overall estimate mixes a childhood stratum where risk seems reduced with an adolescent stratum that could not be analysed. Second, the protective signal is the most appealing and the most fragile result of the study. A subgroup defined by age at diagnosis, two p values just below 0.05, a size not given in the article: these are the usual conditions of a result that fails to replicate. It deserves to be reported, not to be put forward.
Critical appraisal
| Domain | Judgement |
|---|---|
| Control of confounding | The best available |
| FindingThe instrumental variable is the observational design closest to randomisation when a trial is impossible. It does not equal it. The models are also adjusted for individual variables and for district variables: population size, use of child and adolescent psychiatry services, parental education level. | |
| Data quality | Dispensing registers |
| FindingNeither self-reported nor retrospective. The measure concerns medication actually dispensed and reimbursed, which remains an approximation of actual intake, as the authors acknowledge. The median time between diagnosis and first treatment is 7 days, which limits bias from delayed treatment. | |
| Robustness of the main result | Consistent across four windows |
| FindingThe absence of association is stable across the four windows and the two outcomes. It does not, for all that, converge with the whole literature: several earlier studies, including a 2025 meta-analysis and a Taiwanese registry study, reported an increased risk, without controlling for confounding by indication. This is the point of disagreement to which this study contributes an argument, by controlling for confounding by indication, without settling it on its own. | |
| Nature of the estimated effect | Local policy-level effect |
| FindingThe authors specify that the instrument captures a composite of district practices: local policies, organisation and resources. The estimate therefore includes the drug’s own effect and any organisational spillovers. It answers the question of the effect of shifting prescribing practices, not exactly that of the isolated pharmacological effect. | |
| Status of the subgroup analyses | Secondary, no correction |
| FindingThe strata are listed in the methods section (sex, education level, age at diagnosis, birth year, family history of inpatient admission), which argues for pre-planning. The authors describe them as secondary analyses. No protocol registration is cited in the text, and no correction for multiple comparisons is mentioned. Two p values of 0.02 and 0.03 across a series of strata call for caution. | |
| Exclusion assumption of the instrument | Tested, cannot be proven |
| FindingThe falsification tests and robustness checks, an instrumental variable probit model and the tF ratio adjustment, are in the supplementary material, which could not be consulted. Their figures therefore cannot be verified here; only the authors’ account of them in the main text is available. | |
| Applicability to your practice | Indirect |
| FindingThe arrangements for initiating and renewing methylphenidate prescriptions vary between health systems and may differ from the Finnish context; they should be checked where you practise, at the date of reading. The results moreover concern neither diagnoses made in adulthood, nor amphetamines, which were prescribed too rarely to children in Finland to be studied. | |
Level of evidence
Confidence is good in the main result: in a national birth cohort followed to 22 years of age on average, with a design built to limit confounding by indication, methylphenidate is not associated with an excess of nonaffective psychosis. Confidence is low in the protective signal, which rests on a stratified secondary analysis, without correction for multiple comparisons and without replication. It is nil in adolescents, where the instrument was too weak to produce a result, and where the authors write that an increased risk cannot be ruled out.
What is demonstrated, within the limits of the design: the absence of association between sustained treatment and the incidence of nonaffective psychosis in the overall sample. What is suggested: a possible favourable effect of sustained treatment in children diagnosed before the age of 13. What remains hypothesis: the idea that early treatment would protect through a developmental mechanism. The authors put it forward on the basis of imaging studies and animal models that they cite, but this study contributes no mechanistic data of its own.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ At last, something to tell parents. A national cohort, follow-up into the early twenties, and no excess of psychosis. The protective signal in the younger ones, I keep to myself until someone finds it elsewhere. And in teenagers, they could not conclude, so I say nothing. ”
What this means in practice: the fear of a psychotic turn no longer needs to weigh in the foreground on the decision to initiate treatment in a child, nor to justify postponing it. It remains something to listen to and discuss, not a reason not to treat. The possible benefit of early treatment is not presented as an argument for treating earlier. In adolescents, these data allow neither reassurance nor alarm.
What you can do with this
- Answer the parents’ question with a figure and a source, rather than with general reassurance. A national birth cohort of 697,289 people, 3956 of them with ADHD followed through registries to 22 years of age on average, is an argument that can be heard.
- Do not delay a justified initiation in a child on the grounds of psychotic risk. Postponement has a cost, at school, in relationships and in accidents, that is itself documented.
- Do not extend the conclusion to diagnoses made in adolescence. The study could not test this stratum, and an absence of result is not a negative result.
- In the discussion, draw a clear line between the risk of a lasting psychotic disorder and transient psychotic symptoms linked to overdose, which are a different question calling for a different course of action.
- Keep up the usual monitoring: blood pressure, heart rate, growth curve, sleep, appetite. This work addresses only one of the risks to be followed.
- Do not turn the protective signal into an argument for prescribing. It is presented here as a hypothesis awaiting replication, not as an expected benefit.
Frequently asked questions
Can an observational study tell whether methylphenidate causes psychosis?
The authors explain that a randomised trial would be both unethical, since one arm would have to be denied an effective treatment, and impractical, since it would need to run for more than a decade to assess an effect on the developing brain. The instrumental variable is the best approximation available. It rests on an assumption that cannot be proven, only subjected to falsification tests, and that limit must be stated alongside the result.
Can methylphenidate still trigger psychotic symptoms?
Transient psychotic symptoms on psychostimulants are described, particularly with overdose or misuse. They are a reason to stop and reassess. The question addressed here is a different one: the incidence of an established nonaffective psychotic disorder, recorded in the national registries during follow-up.
Should children with ADHD be treated earlier and for longer to prevent psychosis?
No, not on the basis of these data. The signal comes from a secondary analysis in a stratum whose size is not given in the article, without correction for multiple comparisons and without replication. The authors themselves conclude that the mechanisms underlying a potentially reduced risk “will require replication and further study”. The decision to treat remains based on the functional impact of the ADHD.
Does this result apply to lisdexamfetamine and other stimulants?
No. The exposure studied is methylphenidate, which accounts for more than 90% of the stimulant reimbursements recorded in this sample. The authors state that amphetamines could not be studied, because of too few prescriptions for children in Finland, and point out that earlier research suggests a higher rate of psychosis on amphetamines than on methylphenidate.
Why do the risk differences in this study look so large?
Because they are not to be read as a gap between two groups. The coefficient is the change in probability associated with one unit of exposure, defined as 30 mg of methylphenidate per day for the whole intervention window, and only in patients whose treatment depends on the practices of their district. It is a local quantity, estimated with a linear probability model, with wide intervals. The authors also repeated the estimation with an instrumental variable probit model to account for out-of-range predictions.
Annotated bibliography
Source study. Healy C, O’Hare K, Lång U, Metsälä J, Pulakka A, McGrath J, Migone M, Keating D, Romaniuk L, Gyllenberg D, Kajantie E, Perrett G, Hill J, Elwert F, Kelleher I. Methylphenidate Treatment and Risk of Psychotic Disorder. JAMA Psychiatry. 2026;83(6):611-619. doi:10.1001/jamapsychiatry.2026.0152. PMID 41879751. PMCID PMC13019342.
Supplementary material. Supplements 1 and 2, which contain the tests of the instrumental variable assumptions, the detailed stratified analyses, the robustness checks and the modified STROBE-MR checklist, could not be consulted. Elements that appear only there are flagged as unverifiable in the body of the analysis.