A dementia research association disappears after adjustment when the original link was carried by something else — usually other illnesses, income and education, or early-life ability — rather than by the exposure being studied. Once researchers statistically account for those background factors, the apparent effect shrinks toward "no difference," which means the first number was measuring the company the exposure keeps, not the exposure itself. Adjustment is the standard tool for this. It asks what the risk would look like if two groups matched on the listed factors, and the honest answer is often that the headline finding does not survive the comparison.
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
- What "adjustment" actually does to a risk number
- When the real cause sits earlier in life
- Reverse causation — the disease arriving before the diagnosis
- Adjustment can also delete a real effect
- Surviving adjustment is not proof either
- Reading a dementia risk headline without being misled
- Frequently Asked Questions
What "adjustment" actually does to a risk number
Observational studies report risk as a hazard ratio: 1.40 means the exposed group developed dementia 40% faster over follow-up, and 1.00 means no difference. Each ratio comes with a 95% confidence interval, the range of values compatible with the data. When that interval crosses 1.00, the study cannot distinguish the finding from chance. A 2026 population-based cohort published on PubMed shows the pattern in full.
Metabolic dysfunction-associated steatotic liver disease — fat build-up in the liver tied to diabetes, obesity and blood pressure — was linked to dementia at HR 1.40 (95% CI 1.17–1.67) in the crude analysis. After adjusting for comorbidities and socioeconomic factors, it fell to HR 1.12 (95% CI 0.94–1.34), an interval that includes no effect at all. Nothing about the patients changed. What changed is that the comparison now holds other illness and social position constant, and most of the original signal turns out to have lived there.
When the real cause sits earlier in life
The most instructive disappearances involve factors that precede the exposure by decades. In a Norwegian cohort of more than 207,000 men reported in PNAS, more education looked protective against dementia — until researchers controlled for cognitive test scores measured in young adulthood. The association then vanished entirely (HR 1.08, 95% CI 0.91–1.28, p=0.40).
Read carefully, that result says the schooling signal largely reflected ability that was already present before the schooling. People who tested higher in their twenties both stayed in education longer and developed dementia later, and the second fact was not produced by the first. This matters for anyone making decisions on the basis of risk lists. "More years of study lowers your dementia risk" and "people who study longer tend to have brains that age differently anyway" carry very different practical implications.
Reverse causation — the disease arriving before the diagnosis
Dementia changes behaviour and health years before anyone receives a diagnosis. Someone in the early silent phase may exercise less, drink differently, lose weight, or stop managing their blood pressure well. A study that records those changes at baseline and counts dementia later will read the consequence as the cause.
The 2024 Lancet standing Commission on dementia attributes part of the instability in published associations to exactly this, and asks researchers to do two specific things: report follow-up time separately for cases and non-cases, and test what happens when dementia diagnosed within 5–10 years of baseline is excluded. If an association survives that exclusion, it is harder to explain as the disease announcing itself early. When you read a study, look for whether it did this. A cohort with three years of follow-up and no lag analysis is measuring a much shorter window than it appears to.
Adjustment can also delete a real effect
The mirror-image error is less discussed and just as consequential. Adjusting for a mediator — a step on the causal path between exposure and outcome — removes part of the very effect the study set out to measure. Work by Schisterman and colleagues in *Epidemiology*, available through PMC, shows this biases the estimate toward the null, and that if the mediator has other causes shared with the outcome, conditioning on it opens a collider path that can increase, reduce, or even reverse the apparent effect.
Concretely: if exercise protects the brain partly by lowering blood pressure, then adjusting for blood pressure subtracts that route and understates exercise. The resulting "no association" is an artefact of the model, not a finding about exercise. This is not confined to individual papers. A meta-research scoping review in the International Journal of Epidemiology examined overadjustment bias specifically within systematic reviews of socioeconomic health inequalities — so pooled summary estimates, the kind usually treated as the strongest evidence, can carry the problem forward too.
Surviving adjustment is not proof either
A risk factor that holds up under adjustment has cleared one hurdle, not all of them. Residual confounding — the influence of things that were never measured — remains. A three-cohort analysis of modifiable dementia risk factors across HRS, CHARLS and ELSA, published in Frontiers in Public Health in January 2025, states plainly that unmeasured confounders including APOE ε4 genotype and socioeconomic status persist after adjustment. Genetic methods often disagree with cohort results.
A systematic review of Mendelian randomisation studies by Desai and colleagues in the European Journal of Neurology found adiposity, alcohol, blood pressure and physical activity produced non-concordant directions of effect compared with observational findings. A factor can survive adjustment in a cohort and still fail a genetic causal test. Attenuation patterns can themselves be diagnostic. A late-life statin and LDL analysis of 6,977 people, reported on PubMed, found dementia risk elevated at initiation (HR 1.66, 95% CI 1.16–2.38) but attenuating under lipid adjustment and lagged analyses — which the authors read as residual confounding, treatment selection and reverse causation rather than a genuine hazard from the drug.
Reading a dementia risk headline without being misled
The practical question is rarely "is this study wrong" but "how much weight should this carry in my decisions." A few checks separate a durable finding from a fragile one: None of this argues against the standard advice on hearing aids, blood pressure, activity and social contact — those carry benefits well beyond dementia. It argues against treating a single attenuating hazard ratio as a reason to start or stop anything, and against reading a disappeared association as proof that the exposure is harmless.
- Find the adjusted estimate, not the crude one. If only one number appears in the coverage, the adjusted figure is the one the researchers stood behind.
- Check whether the confidence interval crosses 1.00. It it does, the study did not detect an effect, whatever the point estimate suggests.
- Look for a lag analysis excluding cases diagnosed within 5–10 years of baseline, as the Lancet Commission recommends.
- Ask what was adjusted for. Adjusting for a plausible mediator (blood pressure in an exercise study) can hide a real effect rather than reveal a false one.
- Treat prevention percentages as conditional. The Lancet Commission's widely quoted "45% of cases preventable" figure explicitly assumes every one of its 14 associations is causal, and the Commission states most of that evidence comes from observational studies from which causation cannot be inferred.
Frequently Asked Questions
Does a disappearing association mean the first study was badly done?
Usually not. Crude and adjusted estimates are both reported on purpose, and the gap between them is information — it shows how much of the raw signal came from comorbidities, income, education or prior ability rather than the exposure.
Why do genetic studies contradict cohort studies so often?
Mendelian randomisation uses inherited variants, which are fixed at conception and so cannot be confounded by later lifestyle. Desai and colleagues found adiposity, alcohol, blood pressure and physical activity gave non-concordant directions of effect versus observational work, which points to unmeasured confounding in the cohorts.
Should I ignore the "45% of cases preventable" figure?
Treat it as a ceiling under a stated assumption rather than a forecast. The Lancet Commission is explicit that the figure assumes each of its 14 risk-factor associations is causal, and that most of the underlying evidence is observational.
