When a dementia study follows people for only a few years, its findings describe the years just before diagnosis — not the decades in which dementia actually develops. That distinction matters because the brain changes and the behaviour changes that precede a diagnosis can make a harmful exposure look protective, or make an innocent one look dangerous. Researchers call this reverse causation: the early disease changes the exposure, rather than the exposure changing the disease. The clearest way to see it is to take the same dataset, the same people and the same measurement, and simply vary how far in advance the measurement was taken.
Medical information disclaimer: This article is for general educational purposes only and does not provide medical advice, diagnosis, or treatment. Always consult a physician or other qualified health professional about symptoms, medications, tests, or treatment decisions.
Table of Contents
- The body weight example, where the effect flips sign
- Why the brain makes short windows unreliable
- Physical activity and frailty follow the same shape
- Treatment and prevention trials have the mirror-image problem
- How to read a headline about dementia risk
- What this does and does not change for families
- Frequently Asked Questions
The body weight example, where the effect flips sign
Body mass index (BMI, weight relative to height) offers the cleanest demonstration. In a pooled analysis of individual data from 1,349,857 dementia-free people across 39 cohorts, with 6,894 dementia cases, Kivimäki and colleagues in *Alzheimer's & Dementia* found that each 5-kg/m² higher BMI carried a hazard ratio of 0.71 (95% CI 0.66–0.77) when weight was measured less than 10 years before diagnosis. A hazard ratio below 1 means lower risk. Taken alone, that number says being heavier protects against dementia.
But the same people, measured 10–20 years out, gave 0.94 — essentially nothing. Measured more than 20 years out, the ratio was 1.16 (1.05–1.27): higher weight, higher risk. The authors read this as two processes stacked on top of each other. There is a real harmful effect of midlife obesity, and there is weight loss during the preclinical phase of dementia. People heading toward a diagnosis are heavier than average two decades out and lighter than average close to it, which is why the *Alzheimer's & Dementia* analysis treats the short-window result as an artefact rather than a finding.
Why the brain makes short windows unreliable
The reason is biological, not statistical. A Whitehall II analysis published in 2017 found that global cognitive decline is detectable at least 10 years before a clinical dementia diagnosis, and accelerates roughly 3 years before it. That means a study's "baseline" — the supposedly healthy starting measurement — is already capturing disease if it falls inside that window.
The participant has not yet been diagnosed, but the process is underway and is already shaping their weight, their activity, their sleep and their diet. So a short-follow-up study is not measuring what a healthy 55-year-old's habits do to their risk. It is measuring what early dementia does to a person's habits, run backwards through a statistical model that assumes the arrow points the other way.
Physical activity and frailty follow the same shape
Physical activity shows the pattern in its starkest form. In an individual-participant meta-analysis of 404,840 adults with a mean age of 45.5 and 2,044 incident dementia cases over 6.0 million person-years, Kivimäki and colleagues reported in the *BMJ* that physical inactivity predicted dementia when measured under 10 years before diagnosis (HR 1.40, 1.23–1.71) — and showed no association at all when measured more than 10 years before. UK Biobank narrowed the timeline further. The physical-activity association attenuated progressively across 2–6 years of follow-up, with no evidence of an association remaining after 5 years.
That finding, published in the *International Journal of Epidemiology*, is why analysts now apply a lag period — commonly 4 years — discarding the earliest cases rather than using all available follow-up. Frailty behaves the same way. Across four US and UK cohorts, Bayesian models found frailty trajectories accelerating 4–9 years before incident dementia, and frailty remained positively associated with dementia only among participants measured outside that prodromal span. Three different exposures, three different research groups, one shape.
Treatment and prevention trials have the mirror-image problem
Observational studies with short follow-up measure the wrong period. Trials with short follow-up measure too little of the right one. CLARITY AD randomised 1,795 people with early Alzheimer's disease for 18 months and, as Eisai reported at CTAD in November 2022, produced a 27% slowing of decline on the CDR-SB scale — a between-group difference of 0.45 points. The dispute over whether that difference is clinically meaningful is, at heart, a timing dispute. Eighteen months cannot show whether the gap between treated and untreated patients widens over time, holds steady, or closes.
Eisai extended the study into an open-label period of up to 48 months to test exactly that, which concedes the registration endpoint was a snapshot rather than a trajectory. Prevention trials are shorter still. FINGER ran 2 years in 1,260 at-risk older Finns and produced a between-group difference of 0.022 z-points per year on a neuropsychological test battery (95% CI 0.002–0.042, p=0.030), an effect size of about 0.13. The result was statistically positive. Two years is nowhere near long enough to say whether dementia incidence itself changes.
How to read a headline about dementia risk
The single most useful question about any dementia risk finding is: how long before diagnosis was the exposure measured? That is usually buried in the methods, but the study abstract often names the mean follow-up, which is a workable proxy. There is a real cost to the fix.
Lagging or excluding early cases trades bias for statistical power: it discards the cases closest to baseline and shrinks the number of events the analysis has to work with, which is why the *International Journal of Epidemiology* authors frame it as a trade-off rather than a correction. A well-lagged study with wide confidence intervals is not a failure; it is an honest one.
- **Follow-up under about 5 years** — treat as a description of the preclinical phase, not a risk factor. Useful for building early-detection tools; not a basis for changing habits.
- **A protective association that appears only in the first few years** — this is the specific pattern that should not be acted on. It is the signature of reverse causation, as the BMI reversal demonstrates.
- **Follow-up beyond 10–20 years, or an explicit lag period** — the finding is far more likely to reflect a causal effect.
- **Trials reporting cognitive-scale differences over 18–24 months** — these measure rate of decline over a short span, not whether dementia is prevented or ultimately delayed.
What this does and does not change for families
None of this undermines the case for midlife exercise, weight management and cardiovascular care. The long-window evidence — the 20-year-plus BMI data, in particular — points the same direction as the short-window evidence appeared to, for weight. It just points there for different reasons and on a different timescale. What it does change is how to interpret a relative's recent history.
If a parent lost weight or stopped walking in the two years before a diagnosis, that change is far more likely to be an early symptom than a cause. Reading it as a cause produces guilt that the evidence does not support. It also changes what to expect from a new drug's reported numbers. A 27% slowing measured over 18 months in early Alzheimer's is a statement about that 18 months, and the open-label extension running out to 48 months exists precisely because nobody yet knows what the curve does after the trial ends.
Frequently Asked Questions
What is a "lag period" in a dementia study?
The analyst deliberately ignores cases diagnosed in the first few years of follow-up and the exposure data closest to them. Four years is the common choice in UK Biobank analyses. It removes measurements likely contaminated by early disease, at the cost of fewer cases.
Does short follow-up make a study worthless?
No. It makes it a study of the preclinical phase rather than of causes. Findings about what changes in the years before diagnosis are genuinely useful for early detection — they are just not evidence about prevention.
How long before diagnosis do measurable changes start?
Global cognitive decline is detectable at least 10 years out and accelerates around 3 years out. Frailty trajectories accelerate 4–9 years before diagnosis. Weight begins diverging earlier still, given that the BMI association has already reversed by 10 years.
