ADHD and Dementia Risk: What the EHR Data Show, and Where the Hazard Likely Comes From

What the All of Us Analysis Measured

The paper is Bauer-Negrini G, Leffa DT, Ferreira PCL, et al. Attention-Deficit/Hyperactivity Disorder and Late-Life Cognitive Impairment. JAMA Psychiatry. Published online October 7, 2026. doi:10.1001/jamapsychiatry.2026.3165. It's open access, and it's worth reading in full before reading anyone's summary of it, including this one.

The design is an electronic health record cohort drawn from the NIH All of Us Research Program, with linked genomic data. Participants were 187,341 adults aged 50 and older at cohort entry, mean age 56.41 years (SD 7.88), 57 percent female, followed for a mean of 10.26 years. ADHD was identified using a published EHR algorithm requiring multiple diagnosis or medication records on separate occasions, and 3,026 participants (1.6 percent) met that definition. The outcome, incident mild cognitive impairment or dementia, was ascertained the same way, from diagnosis codes and dementia medications in the record.

The headline numbers:

Raw incidence was 6.7 percent in the ADHD group (203 of 3,026) and 3.4 percent in the comparison group (6,259 of 184,315), which works out to 8.96 against 3.29 cases per 1,000 person-years. The ratio of those two raw figures is close to 2. The 4.60 figure is what remains after the model accounts for age, and the age difference is substantial: the ADHD group entered at a mean of 52.49 years against 56.47 years for the comparison group. Age is the dominant predictor of cognitive impairment, so the adjusted figure is the more defensible one. Anyone quoting "more than four times the risk" without saying it's an age-adjusted hazard ratio is reporting a number they haven't understood.

This is the second large cohort to report the association. Levine and colleagues reported it in JAMA Network Open in 2023, in a cohort of 109,218 members of an Israeli health maintenance organization born between 1933 and 1952, with an adjusted hazard ratio of 2.77 (95% CI, 2.11-3.63) for dementia. Several health systems on two continents now report the same direction, and the authors note that the published magnitudes are broadly comparable to theirs. The question that matters clinically has moved on from whether the association exists.

The Gap Between the Clinical Hazard and the Genetic Hazard

The most informative comparison in the paper sits in its own results table. A clinical ADHD diagnosis carried a hazard ratio of 4.60. Genetic liability to ADHD, measured across common variants in the same participants, carried 1.35.

A polygenic score is a weak instrument and nobody should expect it to reproduce the magnitude of a clinical diagnosis. The point is the structure of the difference. The polygenic score captures inherited liability and nothing else. The clinical diagnosis captures inherited liability plus everything that happened to a person who had ADHD symptoms for five decades: the schooling that went badly, the injuries, the smoking, the drinking, the untreated hypertension, the years without a primary care doctor. When the version of the exposure that carries lived consequences produces several times the hazard of the version that carries only genetics, the lived consequences are the more likely source of most of the hazard.

The adjustment sequence in the paper points the same way. Each time the authors put measured confounding or mediation into the model, hazard came out.

The headline 4.60 comes from Model 1, which adjusts for sex, race, ethnicity, APOE ε4 status, and cohort entry year, and for nothing else. Somatic comorbidity takes it to 4.04. Psychiatric comorbidity takes it to 3.19. The excess hazard above 1.0 falls from 3.60 to 2.19 across those two steps, roughly two-fifths of the excess removed by comorbidity alone, and the matched analysis removes close to two-thirds of it.

That pattern matters because All of Us measures the relevant exposures badly. Lifetime smoking in this cohort is a yes-or-no item about 100 cigarettes. Body mass index, physical activity, diet quality, alcohol quantity, untreated hearing loss, head injury history, and sleep apnea are measured worse than that or not at all. A covariate set that strips two-fifths of the excess hazard while measured this crudely would be expected to strip considerably more if measured well. Residual confounding and mediation of that shape is the most parsimonious explanation for the hazard that remains.

The baseline table makes the same point from the other direction, and it cuts against the mediation argument in an instructive way. In this cohort the ADHD group was better educated than the comparison group (54.9 percent with a college degree against 46.7 percent) and only slightly more likely to have smoked (48.8 percent against 43.7 percent). All of Us is a volunteer cohort that skews educated and female, so its ADHD participants aren't the population in which the modifiable burden concentrates. The excess hazard nevertheless attenuated by two-fifths in a sample selected against that mechanism, which is what makes the mechanism worth taking seriously in the general population.

The authors say a version of this themselves, and they say it twice. Their discussion offers two hypotheses for the association. The first is reduced cognitive reserve, in which lower average baseline cognitive performance in ADHD means the clinical threshold for MCI gets crossed sooner as ordinary neuropathology accumulates. The second is that "ADHD increases exposure to cumulative risk factors for cognitive impairment, including chronic stress, social adversity, sleep disturbance, physical inactivity, poor diet quality, cardiometabolic risk, injuries, and substance use." On the genetic result they are equally careful, writing that the polygenic score association "may reflect liability to ADHD and its life-course correlates, shared liability with other psychiatric or neurodevelopmental traits, or broader pleiotropic pathways rather than a direct causal effect on dementia neuropathology."

Those are the two sentences that disappeared from the coverage, and both of them point away from ADHD as a neurodegenerative condition. The reserve hypothesis is worth separating out, because it is neither a neurodegenerative mechanism nor straightforwardly a modifiable one. Cognitive reserve is built largely through education and sustained cognitive engagement, both of which untreated ADHD erodes, so even that pathway runs partly through outcomes that early diagnosis and treatment change.

Mean Age 56 Means Most of These Events Are Mild Cognitive Impairment

The cohort entered at a mean age of 56, the ADHD group at 52.5, and was followed about a decade. Dementia incidence in that age window is low, and the composite outcome reflects it: the MCI-only hazard ratio was 3.83, with 95 percent confidence intervals well away from the null. The dementia-only estimate was 3.07 with much wider intervals (2.12-4.44), which is what a smaller number of events looks like.

Mild cognitive impairment in a 58-year-old is a diagnosis that depends heavily on who is looking and how hard. An adult with ADHD who reports losing his keys, missing appointments, and losing his train of thought in meetings is describing the condition he has had since childhood. In a clinician's note at age 58 the same complaints read as new. ADHD-related executive dysfunction, the cognitive effects of depression and sleep deprivation, and early neurodegeneration are difficult to separate with structured testing and nearly impossible to separate from diagnosis codes.

Detection pressure runs the same direction. People with a documented ADHD diagnosis are, by construction, people in care. They have more clinical encounters, more clinicians asking about cognition, and more chances for a complaint to become a code. The authors are explicit about the hazard this creates. They write that EHR-only ascertainment "is susceptible to biases related to health care access, clinician recognition, and changing diagnostic practices," and that "differential health care contact could also promote greater ascertainment of cognitive impairment among individuals with ADHD." Those two sentences belong in every secondary account of this study.

The lag analyses carry the clearest signal, and they are the numbers most worth knowing. The authors required progressively longer minimum intervals between the ADHD diagnosis and the cognitive impairment diagnosis, from one year out to ten. Notably, the hazard ratio fell from 4.60 to 1.44 (95% CI, 1.12-1.86) once eight years of separation were required, and to a nonsignificant 1.28 (95% CI, 0.98-1.68; P = .07) at nine years. By nine years, 150 of the ADHD cases with cognitive impairment, 74 percent of them, had been reclassified out of the analysis. In other words, most of the fourfold hazard sits in the cognitive impairment diagnoses recorded within eight years of the ADHD diagnosis.

The authors attribute the loss of significance at nine years to reduced power, which is fair. The size of the attenuation is the part to hold onto. A hazard ratio of 1.44 with eight years of separation is a different clinical object from 4.60 with none, and 1.44 is the figure that belongs in any sentence about what an ADHD diagnosis at 55 predicts over the decade that follows. The authors say as much in their limitations: "a contribution from prodromal misattribution cannot be excluded."

This reading has limits. Several analyses in the paper survive it. First, the association persisted in a propensity score-matched cohort balanced on each individual neuropsychiatric condition. Second, it persisted in an analysis restricted to participants with no neuropsychiatric comorbidity at all at cohort entry. Third, it persisted in competing-risk models accounting for death (subdistribution HR, 4.30). The authors' own summary of this is well calibrated: "That a substantial association persisted under this most stringent approach argues against confounding as a full explanation, though residual confounding or mediation cannot be excluded." Nevertheless, no model in the paper moves the hazard as far as the eight-year lag does. The open question is which share of the hazard sits where. Something real is in these data.

One finding cuts against the coding story and should be reported alongside it. The polygenic score was more strongly associated with dementia alone (1.62) than with MCI alone (1.21). Miscoding of executive dysfunction as MCI would not produce that pattern, and it leaves room for a genuine shared-biology contribution to late-life neurodegeneration. Both pathways are present in these data, with the attenuation pattern and the diagnosis-versus-score gap indicating the modifiable pathway carries the larger share.

Ten of the Fourteen Lancet Commission Risk Factors Are Elevated in ADHD

The 2024 Lancet Commission on dementia prevention identified fourteen potentially modifiable risk factors and estimated that eliminating all of them could prevent up to 45 percent of dementia worldwide. Set that list beside what is known about the course of untreated ADHD and the overlap is close to complete.

Ten of the fourteen are elevated in ADHD on evidence ranging from strong to indirect. For three of the remaining four, midlife LDL cholesterol and hearing loss and vision loss, the Commission's risk factor is the untreated or unmanaged version, which is an adherence and access problem of exactly the kind ADHD produces. Air pollution is the one factor studied mainly as a cause of ADHD rather than a consequence of it. This is the same mediator structure that drives the life expectancy gap in ADHD, where the mortality excess runs through accidents, suicide, substance use, and cardiovascular risk, each of which is treatable.

What the Polygenic Score Can and Cannot Rule Out

The genetic result is the strongest card in the paper, and it is why the coverage reached for neurobiology. A polygenic score is fixed at conception, so it can't be caused by prodromal dementia, which removes reverse causation as an explanation. That's a real methodological advantage.

Although that advantage is genuine, it leaves the behavioral pathway untouched. Genetic liability to ADHD is genetically correlated with educational attainment, smoking initiation, body mass index, and substance use disorder. A polygenic score for ADHD is therefore also, in part, a score for the exposures on the Lancet list. A hazard ratio of 1.35 across quintiles is well within the range that correlated lifetime exposure would produce, with no direct neurodegenerative mechanism required. The absolute figures make the scale concrete: 3.67 cases per 1,000 person-years in the high genetic risk category against 3.04 in the low, a difference of roughly 0.6 cases per 1,000 person-years.

The causal-inference literature has looked directly at this. Pagoni and colleagues ran two-sample Mendelian randomization using the largest available European-ancestry GWAS for both conditions, 20,183 ADHD cases and 71,880 Alzheimer's disease cases. They found an odds ratio of 1.00 (95% CI, 0.98-1.02; P = .39) for ADHD liability on Alzheimer's disease, published in Translational Psychiatry in 2022. Multivariable models adjusting for educational attainment and IQ gave the same null. The reverse direction was also null.

That estimate rested on ten genetic instruments and on clinically diagnosed Alzheimer's disease rather than the broader MCI-or-dementia composite, so it isn't a definitive negative, and Alzheimer's disease is only one dementia subtype. It is, nevertheless, the most direct causal test available, and it doesn't support a direct genetic route from ADHD to neurodegeneration. Readers interested in what the ADHD genetics literature establishes may find the heritability overview useful.

What the Israeli Cohort Added About Treatment

Levine and colleagues stratified their 2023 analysis by psychostimulant exposure and reported no clear association between adult ADHD and dementia risk among participants who had received psychostimulant medication. They framed it carefully, and so should anyone citing it. The comparison is observational, the treated subgroup was likely small, and confounding by indication runs in several directions at once, since the adults who get diagnosed and treated for ADHD in their sixties differ systematically from those who don't.

The All of Us authors couldn't test this at all. Their limitation is explicit: "We could not address treatment effects here due to limited data on lifetime treatment for ADHD." They call whether treatment modifies risk "a key unanswered question," which is the correct status for it.

What treatment does to the mediators is far better established, and that evidence is where the clinical confidence comes from. Medication periods are associated with substantially fewer motor vehicle crashes and criminal convictions in the self-controlled designs we reviewed in JAMA Psychiatry in 2025. Swedish registry data link pharmacological treatment to roughly 19 percent lower all-cause mortality, and treatment doesn't increase substance use risk in most populations. Each of those outcomes sits on the Lancet list. Accordingly, reducing head injury, smoking, and substance use in a 40-year-old with ADHD is worth doing on its own evidence, and if the modifiable pathway carries most of the dementia hazard, it is also the most plausible available lever on that hazard.

Where the Coverage Went Past the Data

The study's senior author, Tharick Pascoal, said directly that the work doesn't prove ADHD causes dementia, and the paper's limitations section is appropriately blunt. The secondary coverage was looser. One widely shared summary described genetic risk as "independently accelerating" cognitive decline, which converts an association with a polygenic score into a mechanism the paper never tested, and dropped both the prodromal-misattribution and detection caveats the authors raised.

The reported functional findings were also flattened. Among participants who went on to develop cognitive impairment, comorbid ADHD was associated with worse self-reported mental health (OR, 1.69), lower quality of life (OR, 1.66), lower social satisfaction (OR, 2.39), and about 20 percent more total visits. Self-reported general health (OR, 1.05) and physical health (OR, 1.31) were not significantly different, and emergency department and inpatient visits were not either. A summary saying that people with ADHD and dementia "fare worse" across the board is describing a broader result than the one reported.

The pattern is familiar and it carries a cost in the exam room. Adults with ADHD already arrive worried that their memory problems mean something degenerative. "I've always lost my keys. Is this how it starts?" A headline promising a fourfold risk of dementia, stripped of the attenuation data and the ascertainment caveats, produces a conversation that starts from fear rather than from blood pressure, cigarettes, alcohol, sleep, and whether anyone has treated the ADHD.

What I Tell a 55-Year-Old With ADHD Who Read the Headline

The clinical response to this literature is unglamorous and almost entirely actionable:

• Treat the ADHD. Whether treatment lowers dementia risk is unknown. That it lowers crash risk, injury risk, and all-cause mortality, and that it makes every item below easier to sustain, is well supported.
• Get blood pressure, LDL cholesterol, and hemoglobin A1c measured and treated. Midlife hypertension and midlife LDL are on the Commission's list, and both are routinely under-detected in people who miss appointments.
• Stop smoking, with pharmacological help. Cessation in ADHD is harder and is more likely to succeed when the ADHD is treated concurrently.
• Keep alcohol well under guideline limits and address any substance use disorder directly. See the evidence on ADHD treatment and substance use.
• Have hearing and vision tested, and correct what is correctable. Untreated hearing loss carries one of the larger population attributable fractions on the Commission's list, untreated vision loss was added in 2024 on meta-analytic evidence, and both are adherence problems more than diagnostic ones.
• Take head injury seriously. Driving, cycling helmets, contact sport decisions, and fall prevention later on. This is the one factor where the ADHD link is strong and the intervention is immediate.
• Get screened for sleep apnea if there is snoring, daytime sleepiness, or treatment-resistant inattention. Sleep apnea and ADHD overlap and are frequently confused, and untreated apnea has independent cognitive and cardiovascular consequences.
• Treat depression. It is a Commission risk factor, a common ADHD comorbidity, and a cause of cognitive complaints that get misread as early decline.
• Maintain aerobic activity and social connection. Both appear on the list, and both tend to erode in untreated adult ADHD.
• If cognitive symptoms are new, get them evaluated properly. Lifelong executive dysfunction isn't mild cognitive impairment, and a real change from a personal baseline deserves formal neuropsychological testing, not reassurance and not a diagnosis code.

On the Call for Cognitive Monitoring

The paper closes by calling for clinical guidelines recommending cognitive monitoring in adults with ADHD. The case for monitoring is reasonable on its own terms, and the implementation problem is specific enough to state now. Screening instruments for early cognitive change were built for populations whose baseline attention and executive function are normal. Run a brief cognitive screen on a 60-year-old with lifelong ADHD and the instrument will register inattention, slow processing on timed tasks, and weak delayed recall, all of which may have been present at 25.

Given that the instrument registers traits that may have been present at 25, monitoring that works in this population has to be longitudinal against a personal baseline. The useful version starts in midlife, before anyone is worried, and repeats. A one-time screen at 65 will generate false positives at a rate that does real harm, because an MCI label changes insurance, employment, driving, and how a family treats someone for the rest of their life. If guideline bodies take this recommendation up, the baseline requirement needs to travel with it.

What Would Actually Settle This

Four studies would move this question from plausible to established, and all four are feasible with existing infrastructure:

1. Treatment-stratified analyses in registry data with dispensing records adequate to characterize lifetime ADHD treatment exposure, in a within-individual design that handles confounding by indication.
2. Formal mediation analysis quantifying how much of the ADHD-dementia hazard runs through each Lancet factor in a cohort where those factors are measured rather than inferred from codes.
3. Biomarker-adjudicated outcomes. This group studies amyloid and tau imaging, so the natural next step is whether adults with ADHD and incident cognitive impairment carry the neuropathology, or whether the phenotype is executive and vascular. That single result would separate the two explanations more decisively than any amount of EHR adjustment.
4. Prospectively measured cognition in aging ADHD cohorts, with a personal baseline, so that lifelong inattention isn't read as new decline.

Until then, the defensible statement is the one the association supports and no more. Adults with ADHD reach late life carrying a heavier load of established dementia risk factors, and they're diagnosed with cognitive impairment at two to four times the rate of their peers. The burden they carry is mostly made of things that medicine already knows how to treat. Clinicians seeing midlife ADHD patients should be vigilant about blood pressure, smoking, alcohol, sleep, head injury, and mood at every visit, and careful about the inference a fourfold hazard ratio invites.

Frequently Asked Questions

Does ADHD cause dementia?

No study has shown that. The October 2026 JAMA Psychiatry analysis found that adults with an ADHD diagnosis had a 4.60 hazard ratio for mild cognitive impairment or dementia, and the authors state that the design can't establish causation. A Mendelian randomization study of genetic liability to ADHD and Alzheimer's disease found no causal effect (OR 1.00; 95% CI, 0.98-1.02). The most likely explanation for the association is the accumulation of established dementia risk factors across an untreated ADHD course.

How much higher is dementia risk with ADHD?

Raw incidence of mild cognitive impairment or dementia over a mean 10.3 years was 6.7 percent in adults with ADHD and 3.4 percent in adults without. The age-adjusted hazard ratio was 4.60, falling to 3.08 after adding smoking, education, and health care use to the model, to 2.26 in a matched cohort, and to 1.44 (95% CI, 1.12-1.86) once eight years were required between the ADHD diagnosis and the cognitive impairment diagnosis. The earlier Israeli cohort reported an adjusted hazard ratio of 2.77 for dementia.

Does ADHD medication lower dementia risk?

Unknown. The All of Us study could not test it, and its authors call treatment modification "a key unanswered question." The Israeli cohort found no clear excess dementia risk among adults with ADHD who had received psychostimulants, which is suggestive and not conclusive given the small treated subgroup. Treatment does reduce several of the mediators, including motor vehicle crashes and all-cause mortality.

Is my ADHD forgetfulness early dementia?

Lifelong inattention, losing objects, and losing a train of thought are ADHD symptoms and they aren't mild cognitive impairment. What matters is change from a personal baseline, and a genuine decline deserves formal neuropsychological testing, not a diagnosis code. Depression, sleep apnea, alcohol, and medication effects all produce cognitive complaints that mimic early decline and that respond to treatment.

Why are researchers studying ADHD polygenic scores and dementia?

A polygenic score is fixed at conception, so it cannot be caused by prodromal dementia, which makes it useful for ruling out reverse causation. Genetic liability to ADHD is also correlated with educational attainment, smoking, and body mass index, so an association between the score and dementia can still run through modifiable exposures instead of shared neurobiology.

What should I do if I have ADHD and I'm worried about dementia?

Work the modifiable list. Get ADHD treatment if you don't have it, get blood pressure and LDL and A1c measured and treated, stop smoking, keep alcohol low, get hearing and vision corrected, protect your head, get screened for sleep apnea, treat depression, stay physically and socially active. Ten of the fourteen factors the Lancet Commission associates with up to 45 percent of dementia worldwide are elevated in ADHD, and nearly all of them respond to ordinary medical care.

Further Reading

• ADHD and Life Expectancy — The same mediator structure in the mortality data: the UK matched cohort, the Barkley estimates, and the Swedish registry evidence on medication and mortality
• Adverse Outcomes of Untreated ADHD — Pillar resource: suicide, substance use, accidents, and academic failure, with the NCS-A adjusted odds ratios
• ADHD Pharmacology and Natural Course — Pillar resource: medication evidence and functional outcomes, including the Li self-controlled case series
• Stimulant Medications and Protection from Adverse Outcomes — Which functional outcomes change with treatment, and by how much
• ADHD Medications and Cardiovascular Risk — The large-scale evidence on stimulants, blood pressure, and cardiac events
• Is ADHD Genetic? — Twin studies, GWAS, and how to read a polygenic risk score
• ADHD or Sleep Apnea? — Overlapping presentations and why the distinction matters for cognition
• ADHD and Sleep — Circadian delay, sleep debt, and cognitive consequences
• ADHD Psychiatrist NYC — Evaluation and medication management for adults and adolescents