In July 2024, The Lancet published the third update of the standing Commission on dementia prevention, intervention, and care. The Commission, led by Professor Gill Livingston at University College London, identifies fourteen modifiable risk factors for dementia across the life course — less education in early life; hearing loss, high LDL cholesterol, depression, traumatic brain injury, physical inactivity, diabetes, smoking, hypertension, obesity, and excessive alcohol use in midlife; untreated vision loss, social isolation, and air pollution in later life. Two of these factors, high LDL cholesterol and untreated vision loss, were added in 2024 to the twelve identified in the 2020 update. The Commission’s summary finding — that ‘around 45 per cent of cases of dementia are potentially preventable by addressing 14 modifiable risk factors at different stages during the life course’ — has since circulated widely in health reporting, public policy discussion, and consumer longevity commentary.
The Commission’s work is serious and its methodology is disclosed. The forty-five per cent figure is not wrong. But it is not the kind of number it is now being read as. It is a theoretical upper bound derived by a specific statistical procedure, not a policy target, and not a forecast. Treating it as either — which a large share of the subsequent coverage does — takes a careful piece of population epidemiology and turns it into something the evidence cannot support.
The case of this article is narrow and precise. The Commission’s list of risk factors is useful. The evidence for the individual risk factors is mixed in quality but, for most of them, real. The forty-five per cent figure, read correctly, is a measure of what can in principle be attributed to those risk factors under a maximally favourable reading. Read as a forecast of what can be done, it overstates what public-health intervention can achieve — and understates a question the Commission does not, and cannot, answer: what fraction of the remaining fifty-five per cent is inherited, stochastic, or otherwise beyond the reach of the interventions the Commission proposes.
WHAT THE COMMISSION ACTUALLY DID
The Commission’s core analytic move is the calculation of population attributable fractions. For each risk factor, the Commission pools the best available evidence — meta-analyses, cohort studies, and where available randomised trials — to estimate a relative risk. It then pairs that relative risk with an estimated prevalence of the risk factor in the population. The product, with adjustment for the overlap between risk factors, is the population attributable fraction: the share of dementia cases that would not have occurred, in that population, if that risk factor had been entirely absent. The Commission sums adjusted PAFs across the fourteen factors and reports the total.
The arithmetic is careful. The framing is not pitched as a claim about achievable prevention. The Commission’s own text, read closely, is consistently qualified: ‘potentially preventable’, ‘if all fourteen risk factors are eliminated’, ‘a theoretical estimate of maximum impact’. The forty-five per cent is what happens arithmetically if every person in the world has secondary education, no hearing loss, no vision loss, no hypertension, no diabetes, no obesity, no smoking, no excessive drinking, no depression, no traumatic brain injury, adequate physical activity, normal LDL cholesterol, meaningful social contact in later life, and exposure to air below current particulate-pollution guidelines. It is not what happens under any plausible public-health programme.
WHAT A POPULATION ATTRIBUTABLE FRACTION IS — AND IS NOT
This distinction is where the popular reading of the number departs from the evidence. A population attributable fraction describes a counterfactual: the share of a health outcome that is ‘attributable’, statistically, to an exposure. It does not describe what an intervention can achieve. The two are different quantities, and the difference is structural rather than technical.
Three separate gaps intervene. First, the PAF assumes that the statistical associations in the underlying studies are causal, and causal in the direction the intervention implies. Some of the fourteen factors have strong causal evidence from randomised trials — smoking, hypertension, and hearing loss among them. Others have largely observational evidence, where residual confounding cannot be ruled out and where the direction of effect is plausible but not established. Treating all fourteen as equally causally demonstrated collapses a real distinction in the underlying literature.
Second, the PAF assumes that eliminating a risk factor produces the full counterfactual effect estimated from the epidemiology. Real interventions do not eliminate risk factors; they reduce them. A population-level anti-smoking programme that halves smoking prevalence captures, in principle, roughly half of the PAF attributable to smoking — if it captures the benefit proportionally. It often does not, because residual smokers tend to be those for whom cessation is hardest, and the residual risk is not linearly related to prevalence. The population-level realisation rate for most risk-factor interventions is significantly below the theoretical ceiling.
Third, the PAF does not account for overlap that is not already modelled, for competing risks, or for the fact that many of the factors co-occur. A person who has less education, lower income, worse housing, greater air-pollution exposure, worse access to hearing care, and higher rates of untreated hypertension is not someone for whom fourteen parallel interventions are independently available. The statistical model handles overlap arithmetically; the underlying social structure that produces the co-occurrence does not yield to the same arithmetic.
A population attributable fraction describes a counterfactual. It does not describe what an intervention can achieve. The two are different quantities, and the difference is structural rather than technical.
WHAT THE COHORT REPLICATIONS SUGGEST
Work published since the Commission makes the point in a more empirical way. In late 2025, researchers at the University of Wisconsin–Madison attempted to replicate the Commission’s model in a single longitudinal cohort — the Wisconsin Longitudinal Study, which has followed more than 5,500 participants with seventy years of prospectively collected data. When the Lancet risk-factor model was applied to a single, well-characterised population rather than to the pooled meta-analytic evidence, the effect estimates attenuated. Individual risk factors retained statistically meaningful associations with dementia, but the total fraction attributable to them was smaller than the headline figure. The replication’s authors concluded, with characteristic academic reserve, that ‘effect estimates derived from meta-analytic approaches can be influenced by publication bias, inconsistently defined constructs/covariates, neglecting genetic interactions, and largely based on cross-sectional findings’.
This is not a refutation of the Commission. Meta-analysis and single-cohort replication are different exercises, and disagreement between them is not scandalous; it is expected. But it underlines what the forty-five per cent figure is. It is a pooled, assumption-heavy estimate of an upper bound. The evidence from single cohorts, which is the evidence closer to what a real public-health programme would encounter, gives a lower number.
THE ONE PLACE THE EVIDENCE HAS BECOME SHARPER
If the argument of this piece is that the forty-five per cent figure is over-read, a companion observation is that one of its components has recently become notably sharper. The ACHIEVE trial, published in The Lancet in 2023, randomised 977 older adults with untreated hearing loss to hearing-aid provision versus a successful-ageing health-education control. In the full cohort, cognitive decline over three years did not differ significantly between arms. In the prespecified subgroup of participants at elevated risk of cognitive decline — those recruited from the longstanding ARIC cardiovascular cohort — the intervention reduced global cognitive decline by approximately forty-eight per cent over three years.
ACHIEVE is not definitive. Trials of any intervention against cognitive decline are difficult; three years is short; the subgroup effect requires replication. But it is the first randomised evidence of a hearing-loss intervention moving a cognitive outcome in a population at elevated risk. It is also the cleanest demonstration of something the Commission’s framing tends to obscure: the individual risk factors have very different evidentiary structures. Some have trial evidence; most do not. The forty-five per cent figure averages across that distinction.
WHAT THE HONEST READING LOOKS LIKE
The honest reading of the 2024 Lancet Commission is that it represents a careful synthesis of an uneven evidence base. Its list of modifiable risk factors is the best available list, and its call for investment in population-level prevention is defensible on public-health grounds. The forty-five per cent figure is useful as a rhetorical anchor, in the way that any upper-bound estimate is useful: it tells the reader that prevention is worth taking seriously, and that the share is not trivial.
What the figure is not, and cannot be, is a target. It is not a policy outcome any programme can aim to deliver. It is not a forecast of what prevention will achieve in the coming decades. It is not, despite the most frequent misreading, a statement that forty-five per cent of the people who currently develop dementia would have avoided it under better circumstances. It is a statistical summary of what would arithmetically follow from the elimination of a set of risk factors under the Commission’s specific assumptions — a calculation that the Commission itself describes as theoretical.
The wider implication is a distinction that the longevity discussion generally has trouble holding. There is a difference between the evidence that a risk factor exists and the evidence that addressing the risk factor is achievable at scale, affordable, and durable. The Commission is, with care, about the first. The forty-five per cent figure is, on the best reading, an invitation to take the second question seriously. What it is not is the answer to that question. The answer depends on what public health, policy, clinical practice, and structural reform can actually do — and the answer is emphatically smaller than forty-five per cent. Understanding that is not pessimism. It is the precondition for doing the work that remains.
PRIMARY SOURCES
— Livingston G, Huntley J, Liu KY, et al. Dementia prevention, intervention, and care: 2024 report of the Lancet standing Commission. The Lancet. 2024 Aug 10;404(10452):572–628. https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(24)01296-0/fulltext
— Lin FR, Pike JR, Albert MS, et al. Hearing intervention versus health education control to reduce cognitive decline in older adults (ACHIEVE): a multicentre, randomised controlled trial. The Lancet. 2023;402(10404):786–797. https://pubmed.ncbi.nlm.nih.gov/37478886/
— Williams VJ, Trane R, Sicinski K, et al. Life Course Modifiable Risk Factors of Dementia: Replicating the 2024 Lancet Commission Model in a Single Longitudinal Cohort. Alzheimer’s & Dementia. 2025. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12726475/
— Ngandu T, Lehtisalo J, Solomon A, et al. A 2 year multidomain intervention of diet, exercise, cognitive training, and vascular risk monitoring versus control to prevent cognitive decline in at-risk elderly people (FINGER): a randomised controlled trial. The Lancet. 2015;385(9984):2255–2263. https://pubmed.ncbi.nlm.nih.gov/25771249/
