Eleven ways a dementia statistic can mislead you — and how to check

Short answer: Almost every misleading dementia statistic goes wrong in one of eleven ways — it ignores how rare the condition is, it ignores that people die of other things, it compares two different kinds of life, it quotes a population number as if it were your number, it counts who is alive rather than who newly gets ill, it mistakes an early symptom for a cause, it uses the wrong comparison group, it gives a relative figure without the absolute one, it ignores how wide the uncertainty is, it rounds a broader outcome up to “dementia”, or it rests on a paper that has since been retracted. None of these are lies. They are mostly accurate numbers being asked to mean something they don't. Below is each one, with a real example from real research, and the question that catches it.

This page exists because we kept needing it. Every post we write runs into one of these, and re-explaining them each time is worse than putting them in one place.

1. The base rate: a test's accuracy is not your answer

The trap. A screening test advertises "94% accurate." You score badly. It feels like a 94% chance something is wrong. It isn't — and it usually isn't close.

The real example. A 2024 Cochrane review of self-administered cognitive tests examined eleven studies and 2,303 participants. One widely used test showed pooled sensitivity of 94% — it catches most people who do have dementia — alongside specificity of 66%, meaning it wrongly flags about a third of people who don't.

Now put that into a population where about 5 in 100 people have dementia. Out of 1,000 people: roughly 52 have it, and the test catches about 49 of them. Of the 948 who don't, about a third — some 322 people — get flagged anyway. So around 371 people "fail," and fewer than 50 of them actually have the condition.

Roughly seven out of every eight people who fail that test do not have dementia.

The check to run: How common is this condition in people like me? Sensitivity is what tests advertise. Specificity multiplied by rarity is what decides what your bad result means. The rarer the condition, the more of the alarms are false.

More on this in are online dementia tests accurate?

2. Competing mortality: you can lower your "lifetime risk" by dying sooner

The trap. A study reports that a high-risk group has a lower lifetime chance of dementia than a low-risk group, and someone concludes the risk factors are protective.

The real example. The ARIC study followed 12,409 adults for around 26 years and produced a genuinely useful headline: people with none of three midlife vascular risk factors stayed dementia-free for about 30.1 years from age 55, against 17.5 years for those with all three — a gap of roughly 12.6 years.

The part that almost no coverage carried: the people with all three risk factors also had about 5.6 times the hazard of dying without dementia. Because dementia is largely a disease of late life, dying earlier removes the years in which you would have been most likely to develop it. Their cumulative lifetime incidence by age 95 came out lower than that of the healthiest group.

They did not avoid dementia. They developed it earlier when they developed it, and often did not live long enough to.

The check to run: What else was happening to this group over the same period? Any "lifetime risk of a late-life disease" number is silently competing with death from everything else. This is the single strongest reason we prefer dementia-free years to lifetime risk percentages — see what midlife blood pressure, diabetes and smoking cost you in dementia-free years.

3. Healthy-user bias: comparing two different kinds of life

The trap. People who take a supplement, a medicine or a habit are compared with people who don't — as if the only difference between the groups were the thing being studied.

The real example. Hormone therapy is the cleanest case on record. Pooled observational studies suggested women who took it had around 22% lower Alzheimer's risk (RR 0.78). Pooled randomised trials of the same therapy pointed the other way (RR 1.38). Same treatment, opposite directions.

The explanation is not that one set of researchers was incompetent. Women prescribed hormone therapy were historically more affluent, better educated, more engaged with healthcare, less likely to smoke and more physically active — every one of which independently lowers dementia risk. Comparing users with non-users in the real world partly compares two different kinds of life. Randomised trials exist precisely to break that link, and when the link is broken, the benefit goes with it.

The check to run: Was this randomised — and if not, who chooses to do this thing? If the people doing it are systematically healthier, wealthier or more health-engaged, expect the observational estimate to flatter it. More in HRT and dementia risk.

4. A population attributable fraction is not your personal risk

The trap. You read that a risk factor accounts for "2% of dementia cases" and conclude it barely matters. Or that fourteen factors account for "45% of cases" and conclude you personally can cut your risk by 45%.

The real example. Smoking carries a population attributable fraction of roughly 2% in the 2024 Lancet Commission's accounting. That number is small largely because smoking has become much less common — not because smoking does little to an individual smoker's brain. A PAF multiplies how much harm a factor does by how many people are exposed to it. Change the prevalence and the PAF changes, while the personal risk stays exactly where it was.

It runs the other way too. "Up to 45% of cases could be prevented or delayed" is a statement about what would happen to a whole population if fourteen exposures were eliminated everywhere, over a lifetime. It is not a personal discount code, and nobody can promise you a percentage.

The check to run: Is this number about a population or about a person? If it is a share of cases, it is telling you where public-health effort pays off — not what your own odds are. See the 14 modifiable risk factors, in plain English.

5. Prevalence is not incidence: who is alive with it vs who newly gets it

The trap. A statistic counts the people currently living with a condition and gets read as a statement about who develops it.

The real example. Almost everyone has heard that two-thirds of people living with Alzheimer's are women. That figure is real, and it is a prevalence figure — it counts who is alive with the diagnosis at a point in time. Women live longer on average, and age is by far the strongest risk factor for dementia, so a longer-lived group accumulates more cases.

When researchers look at incidence instead — who newly develops dementia at a given age — the Alzheimer's Association's own 2026 report states that most US studies have found no meaningful difference between men and women at any given age.

The two-thirds number is, to a large degree, a statement about longevity. Which does not make women's dementia risk uninteresting: what genuinely differs is which modifiable risk factors each sex tends to carry, and those can be changed.

The check to run: Is this counting who has it, or who is getting it? See women and dementia risk.

6. The prodrome window: mistaking an early symptom for a cause

The trap. A study finds that people who take a drug — or have a symptom — go on to develop dementia more often, and it gets reported as the drug causing the dementia.

The real example. Disturbed sleep, anxiety, low mood and bladder trouble are all things dementia produces on its way in, years before anyone names it. So people prescribed sleeping tablets, sedatives and bladder drugs are disproportionately people already in the earliest phase of the illness.

There is a clean way to test for this: ask when the association appears. In a 2025 Canadian case-control study, chronic benzodiazepine use predicted dementia only within the four years before diagnosis. Go back further, before prodromal symptoms began, and the link faded. A genuine drug effect should not switch off the further back you look. Reverse causation should.

The same topic supplies a companion warning. The largest analysis of sleeping pills and Alzheimer's — 13 studies, 721,354 people — reported odds 29% higher (OR 1.29). The same paper, restricted to the studies that followed people forward in time, found no significant association (HR 1.17, 95% CI 0.87–1.58). Both numbers are in one abstract. Page one of the internet quotes the first.

The check to run: Could the outcome have caused the exposure? And: does this paper contain a second, quieter number? See do sleeping pills cause dementia?

7. The missing comparison group: compared with whom?

The trap. A drug's users are compared with the general population, when the honest comparison is with people who have the same condition.

The real example. Acid-reflux drugs. Pooled across nine prospective studies and 204,108 dementia cases, proton pump inhibitor users compared with non-users showed RR 1.16 (95% CI 1.00–1.35) — a small signal sitting right on the edge of significance.

The same researchers then compared PPI users with people taking a different acid-suppressing drug — people with broadly the same underlying complaint, the same reason to be medicated, the same tendency to visit doctors. The result: RR 1.03 (95% CI 0.66–1.62). Nothing.

Compare people with reflux to people with reflux and the signal disappears. Compare them to everyone else and it reappears. What was being measured was mostly the reason for the prescription, not the prescription.

The check to run: Compared with whom? An "active comparator" — people taking a different treatment for the same problem — is one of the most informative designs in this whole field, and one of the least reported. See does omeprazole cause dementia?

8. Relative versus absolute: the bigger percentage can be the smaller risk

The trap. A headline reports a percentage increase ("74% higher risk"). That is a relative figure. It says nothing about how many people are actually affected until you know how common the outcome was to begin with.

The real example. Spouses of people with dementia. A 2026 JAMA Network Open study of 955,105 married people in Taiwan compared spouses of people diagnosed with dementia with matched spouses of people who weren't. Women's risk was 1.74 times higher, and men's was 1.69 times higher. So far, women look slightly worse off (the two figures are close, and the authors call the link broadly similar in both sexes).

Now the absolute figures. Over five years, 6.47% of exposed women were diagnosed with dementia, against 3.72% of unexposed women: 2.75 extra cases per 100. For men it was 9.24% against 5.75%: 3.49 extra cases per 100. Men had the smaller relative increase and the larger absolute one, because they started from a higher baseline.

The same paper shows a starker reversal by age. For women under 55, the hazard ratio was 2.21 but the absolute increase was 0.33 extra cases per 100 over five years. For women aged 75 and over, the hazard ratio was a smaller 1.48, but the absolute increase was 5.00 extra cases per 100, fifteen times as many.

The check to run: Out of how many? Turn every percentage into "extra cases per 100 people, over how many years." If a paper doesn't let you do that, the percentage alone isn't enough to act on. See my spouse has dementia — does that raise my risk?

9. Read the interval, not just the headline number

The trap. A study reports a hazard ratio of, say, 1.58, and the headline says "58% higher risk." The number next to it in brackets, the 95% confidence interval, shows the range of values the data are compatible with. If that range runs almost down to 1.0 (no difference), the "58%" is a far shakier claim than it sounds.

The real example. Ultra-processed food. A 2026 American Journal of Public Health analysis of the US Health and Retirement Study compared older adults eating the most ultra-processed food with those eating the least. For dementia, the hazard ratio was 1.58, with a 95% CI of 1.003 to 2.48. The lower bound clears 1.0 by three thousandths. The test for trend (whether risk rose steadily as intake rose) gave p = .50, which is no sign of a steady gradient at all.

The same paper has a sturdier result for a broader outcome. For cognitive impairment without dementia, the hazard ratio was 1.46 (1.13–1.88) with a significant trend (p = .03). That's the finding to lead with. A headline built on the dementia number is leaning on the paper's weakest result.

The check to run: What's the bottom of the range? If the lower bound sits at or near 1.0, read the result as "possibly no difference." Then look for a dose-response trend: risk that climbs step by step with exposure is more convincing than one high group. See ultra-processed foods and dementia.

10. Check what was actually measured

The trap. "Dementia risk" in a headline, or even in a scientific citation, can stand for a much broader outcome than dementia itself: mild cognitive impairment, a memory-test score, or a risk score calculated from other factors. Each step of retelling tends to round the outcome up to "dementia."

The real example. The 2026 Taiwanese spouse study (capsule 8) compares its result with an earlier Korean couples study, describing it as reporting "an odds ratio of 1.74 for dementia." But the Korean paper (KLOSCAD, JAMA Network Open, 2021) defines "cognitive disorder", in both partners, as mild cognitive impairment or dementia: OR 1.74 (95% CI 1.12–2.69). Its authors add that they could not analyse dementia and mild cognitive impairment separately because the sample was too small. The number is right, but the outcome isn't the same one, and this happened in a peer-reviewed citation, one step from the source.

The check to run: Dementia, or something broader? Find the outcome in the paper's methods or abstract, not in the headline or in another paper's summary of it. Mild cognitive impairment, test-score change and "risk scores" are each worth knowing about, but none of them is a dementia diagnosis.

11. A retracted paper doesn't disappear from the internet

The trap. When a journal retracts a paper, the record is updated, but news stories, blog posts and AI-generated summaries that cited it usually aren't. The claim keeps circulating with a respectable-looking journal name attached.

The real example. Ultra-processed food again. A 2023 paper in Alzheimer's & Dementia reported "long-term results from the Framingham Offspring Study" on ultra-processed food and dementia. It was retracted. The journal's notice says the authors did not have the appropriate approvals from the US National Heart, Lung, and Blood Institute for use of the data. That's a data-permission failure, not a finding that the result was wrong, but a retracted paper can't be relied on either way.

To make it harder, a different research group, Framingham Heart Study investigators working under an approved protocol, later published an analysis of the same question in the same Offspring cohort (Journal of Prevention of Alzheimer's Disease, 2025). Both papers have "ultra-processed food", "dementia" and "Framingham" in the title. Pages that cite "the Framingham study" on this topic may be citing either one.

The check to run: Is it still standing? Search the paper's title on PubMed. A retraction is flagged at the top of the record. When two studies share a dataset name, check the authors and year, not just the cohort.

The short version, as a checklist

When you meet a frightening or exciting dementia number, ask:

1. How common is this condition in people like me? (base rate)
2. What else happened to this group — including dying? (competing mortality)
3. Was it randomised, and who chooses to do this? (healthy-user bias)
4. Is this about a population or about me? (attributable fractions)
5. Is it counting who has it or who is getting it? (prevalence vs incidence)
6. Could the outcome have caused the exposure? (the prodrome window)
7. Compared with whom? (the missing comparator)
8. Out of how many — extra cases per 100, over how many years? (relative vs absolute)
9. What’s the bottom of the confidence interval — and is there a trend? (read the interval)
10. Is the outcome dementia, or something broader? (check what was measured)
11. Has the paper been retracted or corrected? (retractions)

And one more that catches a surprising amount: does the same paper contain a second number that points the other way? In our experience it often does, and it is usually the more honest one.

A bonus check: an evidence grade is not a danger rating

The trap. A review says the evidence for something is "weak," and it gets read as "the thing is harmless." Or a review calls it "highly probable," and that gets read as "the effect is large."

The real example. Air pollution and dementia, graded twice in the same year by two serious teams. A 2026 umbrella review of 54 systematic reviews found air pollution to have the strongest evidence of any urban-environment exposure for cognitive health — "highly probable." A 2025 analysis using the deliberately conservative Burden of Proof framework rated the same link two stars out of five: "weak and/or inconsistent evidence."

Neither is wrong. The first is grading consistency of direction across many studies. The second is grading magnitude and agreement about magnitude, and penalising unexplained disagreement between studies. A factor can be one of the best-supported items on a list and one of the least precisely measured at the same time.

The check to run: Is this grade about whether the effect is real, or about how big it is? Those two questions have different answers more often than you would expect. See air pollution and dementia risk.

What this means for what we publish

We apply these to ourselves, because the same traps are available to anyone selling a health product — including us.

It is why we do not give anyone a personal "your dementia risk is X%" figure. A number like that would need to survive every trap on this page, and it cannot: it would mix population attributable fractions with individual odds, ignore competing mortality, and imply a precision that no risk model has for an individual.

What the evidence does support is more modest and more useful: showing which of the modifiable factors you currently carry, which ones the research supports acting on, and letting you work on them consistently. That is what our free profile does, and where the evidence is weak we would rather say so than round it up.

You can read how we research and review our content — and the independent physician audit of the app, with all nine findings published, including the ones that were critical.

How we research and review our content →

An independent physician audited the Solenna app. We published all nine findings →

Educational content only. This page explains how to read research; it does not diagnose any condition and is not a substitute for professional medical advice. Discuss anything concerning your own health, medication or test results with your clinician.