Every row of the evidence table

The key point: the observational analysis showed light drinking looking protective, but the genetic (Mendelian randomisation) analysis did not, which the authors read as the protective appearance coming from people cutting down in early dementia.

This one reports light-to-moderate drinking as protective (HR 0.65) against non-drinking, and it flatly contradicts the Mendelian randomisation evidence. Infrequent drinkers were excluded, and non-drinkers include people who stopped for health reasons.

Sustained mild and moderate drinking carried lower risk than not drinking, sustained heavy drinking higher risk, and cutting down from heavy to moderate lower risk. Observational, with the same abstainer problem.

U-shaped: long-term abstainers had 47% higher risk than moderate drinkers, and above 14 units a week each extra 7 units raised risk by 17%. Neither end of the range looks like a benefit of drinking.

The exposure is cannabis use severe enough to need hospital care, not cannabis use in general, and the comparison is against other acute-care patients as well as the general population.

Direction is for QUITTING. Risk approached that of never-smokers and levelled off about seven years after quitting; weight gain after quitting narrowed the advantage. The authors flag residual confounding.

The two measures moved in opposite directions over 2015-2021: standardised incidence fell from 3.5% to 2.8% per year while prevalence rose from 10.5% to 11.8%; gaps between Black or Hispanic and White beneficiaries narrowed.

Direction is for later birth cohort: each successive cohort had a significantly lower incident dementia rate, independent of sex, education and age.

Direction is for later calendar epoch: incidence fell 44% from the first to the fourth epoch, but only among people with at least a high school diploma. The authors could not fully explain the decline.

Incidence was consistently but non-significantly lower in the 2000 subcohort (rate ratio 0.75, 95% CI 0.56-1.02, borderline). The authors read the pattern, together with lower mortality and larger brain volumes, as suggesting a decline, but the primary comparison did not reach significance.

Direction is for degenerative pathology profiles: more than 80% had mixed pathologies, and of five clustered profiles the LATE-NC/hippocampal sclerosis and the ADNC/amyloid angiopathy profiles showed the fastest cognitive decline.

Direction is for LATE-NC: found in 64% of brains, stages 1a, 2 and 3 were associated with dementia and lower MMSE, it was among the strongest independent predictors of dementia, and its attributable risk (24.2%) exceeded that of Alzheimer's neuropathologic change (18.6%); it usually occurred alongside other pathologies.

Direction is for carrying the risk factors. The Danish estimate (37.8% across 16 factors) is lower than the Lancet Commission's global 45%.

Direction is for carrying the risk factors: depression, diabetes, low education, smoking, obesity, physical inactivity and alcohol overconsumption each independently raised the hazard of early-onset dementia, and most associations were stronger for early-onset than late-onset dementia.

Direction is for carrying the risk factors. Which factor mattered most differed by APOE risk group, on 2,065 cases split across 11 factors and three genetic strata, so the subgroup estimates are imprecise.

Direction is for carrying hypertension, diabetes or smoking. The attributable share rose with the age at which the factors were measured (22% at 45-54, 44% at 65-74), which is what you would expect if some of it is reverse causation rather than cause.

Direction is for carrying the risk factors. 39 candidate factors were screened against 485 cases, so individual estimates are imprecise.

For the vascular risk factors: up to a quarter of the education-dementia association ran through hypertension, diabetes and related midlife factors, so carrying them accounted for part of the excess risk.

A lifetime-risk estimate (42% after age 55, higher in women, Black adults and APOE4 carriers) and a projection of new cases. It compares no exposure, so no direction is recorded.

Direction is for being African-American or American Indian/Alaska Native relative to Asian-American members: incidence was highest in those groups (26.6 and 22.2 per 1000 person-years) and lowest in Asian-Americans (15.2). The authors present this as a benchmark for reducing disparities, not as a biological effect of race.

Direction is for female sex: at higher amyloid levels women had higher plasma p-tau217 than men, greater tau deposition and accumulation at a given p-tau217 level, and faster cognitive decline in two of the cohorts.

Direction is for higher baseline plasma p-tau217 in cognitively unimpaired women: each SD increase more than doubled the hazard of MCI or dementia. Associations were larger in women over 70, APOE4 carriers and White women, and p-tau217 was not associated with incident MCI in Black women.

Direction is for the test's accuracy against the reference standard; an accuracy result, not a health outcome. The estimated age at plasma %p-tau217 positivity predicted age at symptom onset (adjusted R2 0.337-0.612) with a median absolute error of 3.0-3.7 years, which the authors call acceptable for trials but not yet for individual clinical use.

Direction is for higher baseline plasma p-tau217 in cognitively unimpaired adults: each SD raised the hazard of progression by 38%, and the very-high group had a 38% absolute 5-year risk of progression.

Direction is for higher plasma p-tau217: each SD raised the hazard of incident cognitive impairment by 41%, more steeply in APOE-ε4 carriers (76%) than non-carriers (26%), with survival curves separating 3-4 years after the blood test.

Direction is for the test's accuracy against the reference standard; an accuracy result, not a health outcome. Combined plasma %p-tau217 and Aβ42/40 detected abnormal CSF amyloid status (AUC 0.949) and each predicted rising amyloid PET over time in people who started below the positivity threshold, replicated in two further cohorts; p-tau231 and GFAP added no clear independent value.

The findings point different ways: lower eGFR raised every blood biomarker except Aβ42/40 (a test-interpretation caveat), yet impaired kidney function did not itself raise dementia risk (HR 0.93), while the association between high NfL and dementia was about twice as strong in people with impaired kidneys.

Direction is for the blood test's prognostic performance: plasma p-tau217 predicted cognitive decline and progression to MCI as well as tau PET (R2 0.33 vs 0.34; HR 1.57 vs 1.61), and using the blood test first cut the number needed to screen for a trial by 94%. A prognostic-performance result, not a health outcome.

Direction is for abnormal baseline plasma p-tau217: people positive for both plasma amyloid and tau progressed to MCI at 7.8 times the rate of those negative for both, and tau-positive/amyloid-negative people at 4.3 times. CSF and PET gave similar results; adding plasma Aβ42/40 did not improve prediction.

The two measures behaved differently: baseline plasma p-tau217 predicted later cognitive decline about as well as entorhinal tau PET (ρ -0.47 vs -0.55), but over time p-tau217 rose then decelerated and tracked concurrent cognitive change only weakly (ρ -0.24) or not at all in the PET substudy, whereas regional tau PET change correlated strongly with concurrent decline.

Direction is for higher baseline plasma GFAP in cognitively unimpaired adults: it predicted faster cognitive decline, CDR progression, cortical atrophy and conversion to amyloid positivity, with stronger effects in females.

Direction is for higher plasma pTau217: it was associated with global and domain-specific cognitive decline within both racial groups and discriminated 3-year dementia well (AUC 0.81 in Black and 0.77 in White adults); biomarker levels were similar across races and the association with decline was similar, though pTau217 tracked faster global and semantic memory decline in Black adults.

Direction is for adverse biomarker levels: only the Alzheimer-specific markers measured in midlife (lower Aβ42:Aβ40, higher p-tau181) predicted dementia two decades later, while every marker measured in late life did, NfL most strongly (HR 1.92 per SD). Biomarkers changed faster in APOE ε4 carriers and with midlife hypertension or diabetes.

Direction is for high serum NfL or GFAP: the top tertile of either raised incident dementia risk by 38-49% and roughly tripled dementia-specific mortality; being in the top tertile of both doubled dementia risk and raised dementia mortality nine-fold. NfL was independently associated with faster cognitive decline.

Direction is for astrocyte reactivity: amyloid was associated with higher plasma p-tau only in people positive for astrocyte reactivity, and an AD-like pattern of tau-PET accumulation as a function of amyloid appeared only in those individuals, cross-sectionally and longitudinally.

Direction is for a rising serum NfL: its rate of change separated mutation carriers from non-carriers 16.2 years before expected symptom onset (versus 6.8 years for a single level), peaked around symptom conversion, and predicted cortical thinning and cognitive decline.

The two comparisons point different ways: overall utilization and payments were similar between amyloid-positive (A4) and amyloid-negative (LEARN) participants, but among A4 participants claims indicators of progression carried 45-103% higher payments.

Direction is for the AD-associated complement pattern: age-related falls in C4, C4b, Factor I and Properdin and rises in Factor D were more pronounced in people who later developed AD, detectable in the preclinical phase, and the same profile was specific to AD among dementia subtypes.

The association was strongest at moderate intake rather than the highest.

882 cases across five quintiles and several interaction strata, so the subgroup estimates rest on small numbers.

The association is reported in people who already have elevated Alzheimer's blood markers, which is the point of the study.

Independent of overall diet quality. Substituting olive oil for margarine or mayonnaise was also associated with lower risk; substituting for butter or other vegetable oils was not.

Significant on one adherence score (MEDAS, HR 0.77) but not the other (PYRAMID, HR 0.86, 95% CI 0.73-1.02).

Significant on one adherence score (MEDAS, HR 0.77) but not the other (PYRAMID, HR 0.86, 95% CI 0.73-1.02).

Risk ratios close to one across every intake strategy, including substituting fish for meat. This is the one study here that emulates a trial rather than comparing eaters with non-eaters, and it disagrees with the pooled cohorts.

The dementia-only estimate barely cleared significance (HR 1.58, 95% CI 1.003-2.48) and its test for trend did not (P=0.50).

518 cases. Replacing 10% of ultra-processed food with minimally processed food was associated with 19% lower risk.

Dose-response by distance for dementia; no association with Parkinson's disease or multiple sclerosis.

Concentrating activity into one or two days worked about as well as spreading it across the week.

Direction is for 12 weeks of cycle-ergometer exercise: putative glymphatic influx in the putamen and meningeal lymphatic vessel size and flow increased, and inflammation-related plasma proteins fell in step with the imaging changes. A mechanistic imaging result, not a cognitive outcome.

866 cases over a short follow-up for a disease with a long prodrome, so some of the association may be people already declining moving less.

The largest effect for physical activity anywhere in this table by a wide margin. On 678 cases over 6.3 years, in a disease with a long prodrome, reverse causation is the more likely explanation for an effect that size.

Explicitly finds no neuroprotective effect, and attributes the lower risk seen in other cohorts to reverse causation: activity fell in the nine years before diagnosis. This contradicts the pooled cohort evidence.

Direction is for APOE genotype after 90: ε2 carriers had lower dementia risk, while the ε4 association was not statistically significant. Women and Black participants had higher risk than men and Asian participants. Age-standardised incidence was 116.82 per 1000 person-years.

Direction is for carrying APOE ε4: 10-year dementia risk ratios of 1.59 (Filipino, confidence interval crossing 1) to 1.96 (Japanese), similar across the four groups despite carrier prevalence ranging from 14% to 25%. Not modifiable.

Direction is for each additional ε4 allele: it did not change amyloid-related p-tau217 secretion, but for any given p-tau217 level tau spread faster across connected regions, and the p-tau217 threshold marking early tau-PET increases was 26-76% lower than in non-carriers.

Direction is for APOE ε4 carriage, which strengthened the association between amyloid and plasma p-tau; plasma GFAP strengthened the p-tau-tau PET link mainly in non-carriers, and high GFAP went with faster longitudinal tau accumulation.

Direction is for carrying two APOE4 copies: near-complete penetrance of Alzheimer's biology and earlier symptom onset.

Direction is for carrying APOE2, which is associated with longer life irrespective of Alzheimer's status. The lifespan figures come from gene-targeted mice, not from the human cohort.

Direction is for midlife hypertension. Genetic risk for Alzheimer's did not modify it, which the authors read as everyone with hypertension standing to gain from treatment.

Direction is for a higher polygenic score: each standard deviation raised AD odds 1.14-1.52-fold across European, African American, Caribbean Hispanic and Native American Hispanic groups (1.21-1.65 in replication cohorts), and the score tracked worse cognition, more AD neuropathology, smaller hippocampi and abnormal CSF biomarkers.

Direction is for carrying the risk factors. Which factor mattered most differed by APOE risk group, on 2,065 cases split across 11 factors and three genetic strata, so the subgroup estimates are imprecise.

Direction is for a higher integrative polygenic score: dementia hazard rose 1.15-fold per SD in older community-dwelling people independently of APOE, generalising to Europeans, East Asians and memory-clinic patients, and reaching five-fold risk when combined with APOE ε4 in memory-clinic patients. Not modifiable.

Direction is for a higher polygenic score: consistently associated with AD risk, age at onset and CSF biomarkers independently of APOE in Europeans and with AD risk in many other ancestries; a cross-ancestry score with the APOE region improved the association, and the signal weakened as the diagnosis was broadened.

Direction is for a higher polygenic risk score: each SD raised the dementia hazard 1.35- to 1.42-fold after adjustment including APOE ε4, and the new scores modestly improved prediction (AUC 74.7% to 75.7-76.1%) while still explaining less risk than APOE. Not modifiable.

Direction is for one copy of the APOE3 Christchurch variant: median onset of cognitive impairment at 52 versus 47 years in matched PSEN1 E280A carriers without it, with preserved brain metabolism and fewer vascular amyloid features in the few imaged or autopsied. Not modifiable.

Among 37 heterozygous carriers (ages 56.6-82.1) none had developed AD or MCI by the March 2022 data freeze and apolipoprotein B was lower than in non-carriers, but the authors say continued follow-up is necessary and no homozygotes were found, so protection could not be tested. No incidence comparison is reported, so no direction is recorded.

Confined to people who said the aid actually improved their hearing (HR 0.86); those reporting poor improvement showed nothing (HR 0.98).

Direction is for hearing loss. Risk was lower among those using hearing aids (aHR 1.06) than those not (aHR 1.20). The authors note their estimates are smaller than earlier studies'.

Direction is for having cataracts or myopia. A Mendelian randomisation analysis supported the cataract-vascular dementia link.

Direction is for midlife hypertension. Genetic risk for Alzheimer's did not modify it, which the authors read as everyone with hypertension standing to gain from treatment.

Regression discontinuity across a birth-date eligibility threshold, which makes confounding by health-seeking behaviour less likely than in a plain cohort comparison. Eligibility did not change uptake of other preventive services or other chronic-disease diagnoses, which is the check that makes the design convincing.

Regression discontinuity across the 2 September 1933 eligibility threshold; stronger in women than men.

Direction is for a hospital-treated infection: the raised dementia hazard (1.54) was barely attenuated by adjustment for frailty (1.49) and persisted among non-frail participants (1.34), so pre-existing frailty and comorbidity did not explain it.

Direction is for delirium on admission: it raised the risk of incident dementia across the whole spectrum of long-term conditions, with the largest relative risk (3.38) in people without multiple long-term conditions.

One of the largest effect sizes in this table (subdistribution HR 3.00), rising about 20% with each further episode. The authors argue for a causal reading.

Read with care: the authors' own reading is that a fall flags undetected dementia rather than causing it, and they recommend cognitive screening after a fall. A one-year window cannot separate cause from early sign.

The authors state the data are compatible with anything from a 39% decrease to a 5% increase in risk, against DPP-4 inhibitors rather than no treatment.

For GLP-1 receptor agonists specifically: HR 0.95 (95% CI 0.87-1.04) against DPP4 inhibitors in the intention-to-treat analysis. Only the continuous-use analysis reached significance (HR 0.79).

Seven separate comparisons against seven other antidiabetic drugs, so the comparator is always another drug rather than no treatment.

Designed around new diabetes diagnoses with a landmark analysis and a one-year lag, which removes the prevalent-user advantage that inflates metformin's apparent benefit in the pooled cohorts.

Comparator is other oral antidiabetic drugs, not no treatment.

The exposure actually studied is STOPPING metformin early, which carried 21% higher dementia incidence than staying on it. Read as indirect support for metformin, not as a trial of it.

For SGLT2 inhibitors specifically: HR 0.86 (95% CI 0.79-0.94) against DPP4 inhibitors, and HR 0.70 in the as-treated analysis.

After the first year the data rule out more than a 3% change either way. Diagnoses were 46% higher in the first year, which the authors attribute to people being seen more often just after starting a statin. This sits against the Chicago cohort, which found benefit in APOE4 carriers.

Benefit was confined to APOE4 carriers (HR 0.60) and absent in non-carriers (HR 0.96). Observational, and it disagrees with the larger target-trial emulation that found nothing.

The authors read these symptoms as early markers of neurodegeneration rather than causes — which is also how an earlier analysis of this same cohort read its own null result for midlife depression.

Direction is for having depression (51% higher risk). Treated depression carried lower risk than untreated (HR 0.7), which is the more interesting half but is an observational comparison of people who did and did not get treatment.

Midlife depressive symptoms did not raise dementia risk (HR 1.21, 95% CI 0.95-1.54), even when chronic. The association closer to diagnosis is read by the authors as depression being a prodromal feature, not a cause.

Direction is for MORE education. Up to a quarter of the effect ran through midlife vascular risk factors, and the authors say treating those factors would not close the educational gap.

Distress rose, and rose most in APOE4 homozygotes. The authors nonetheless conclude disclosure by a trained provider was generally safe and well tolerated in this psychologically pre-screened group.

Direction is for LOSING slow-wave sleep. 52 cases in 346 people, so the estimate is imprecise, and slow-wave sleep also falls with Alzheimer's genetic risk, which muddies cause and effect.

Direction is for SHORT sleep (6 hours or less): higher dementia risk than a 7-hour sleep duration.

Isolation (being alone) carried the risk (HR 1.26); loneliness (feeling alone) did not once depressive symptoms were accounted for (HR 1.04, 95% CI 0.94-1.16). That split matters, and the loneliness meta-analysis here reaches the opposite conclusion about loneliness.

Significant only for contact measured at age 60, not at 50 or 70. The authors note that keeping up contact may itself be a marker of cognitive reserve rather than a cause of it.