A dementia-specific lever with no cardiometabolic analogue in the vault, and one of the largest single population-attributable contributors in the 2024 Lancet Commission. Distinctive because the exposure is correctable (hearing aids) and the correction is plausibly on the causal path — so the intervention question is answerable in principle, not just the association. The load-bearing move on this page is to keep those two questions apart: the exposure association is robust; the treatment benefit is not established to the same standard.
The exposure — hearing loss raises dementia risk [first-hand: Yu 2024]
- Pooled association (dichotomous yes/no). The largest meta-analysis to date — «the most extensive to date with fifty cohorts reporting on a total sample of 1,548,754 participants» — found hearing loss associated with incident dementia at HR 1.35 [1.26-1.45], k=30 cohorts. (Yu et al., 2024) The association held across the other cognitive outcomes: MCI 1.29 [1.11-1.50] (k=3), cognitive decline not specified as MCI/dementia 1.29 [1.17-1.42] (k=9). (Yu et al., 2024)
- Subtype specificity — the effect concentrates in Alzheimer’s, is null for vascular. Alzheimer’s disease dementia HR 1.56 [1.30-1.87] (k=4, the strongest single estimate); vascular dementia 1.30 [0.83-2.05] (k=3) was not statistically significant. (Yu et al., 2024) Yu discounts the VaD null on power, not absence of effect: only three studies looked at it, and two adjusted for cardiovascular factors that «may have diluted the association». (Yu et al., 2024) So the shape is outcome-specific — read the effect as Alzheimer’s-type, with vascular insufficient-evidence rather than no-effect.
- Publication bias was screened and not detected — Egger’s test non-significant for dementia (p=0.77, k=30) and for cognitive decline (p=0.12, k=9); funnel plots showed dispersion but no systematic asymmetry. (Yu et al., 2024)
De-secondhanding the Commission figure [F-refinement, NOT E-independent]
The page previously carried the exposure association from the 2024 Lancet Commission’s borrowed meta-analysis (HR 1.37 [1.00-1.87], I2=80%, 6 cohorts, n=666,370) (Livingston et al., 2024). Yu 2024 is the first-hand source for that relationship and supersedes the borrowed figure:
| Parameter | Livingston Commission MA | Yu 2024 MA | Same quantity? |
|---|---|---|---|
| Relationship | adult HL (yes/no) -> incident all-cause dementia | adult HL (yes/no) -> incident all-cause dementia | yes |
| Pooled effect | HR 1.37 [1.00-1.87] | HR 1.35 [1.26-1.45] | yes — near-identical point estimate |
| Evidence base | 6 dementia cohorts, n=666,370 | 30 dementia cohorts (50 total), N=1,548,754 | Yu is a superset / update |
| Precision | CI touches 1.00 (I2=80%) | CI 1.26-1.45, excludes 1.00 | Yu tightens materially |
This is a type-F claim-refinement: same relationship, a much larger and more precise pool, so the composite belief is firmer than either alone. It is NOT type-E independent corroboration — Yu shares authors with the held Commission (Gill Livingston is the Commission lead; Frank Lin is the ACHIEVE PI; Sergi Costafreda and Anne Schilder also overlap), so a shared-lineage re-pooling cannot raise confidence by independence. Yu itself frames the estimate as «overlapping but on the lower end of the confidence interval of the effect reported by the 2017 Lancet commission (1.9 [1.4-2.7])», noting the older estimates «relied on a substantially smaller sample of studies». (Yu et al., 2024)
Dose-response — graded, but the pooled slope rests on two studies
- Per 10 dB of hearing loss, dementia HR 1.16 [1.07-1.27] (pooled k=2: Deal 2017, Lin 2011). (Yu et al., 2024) This refines the Commission’s borrowed «4% … to 24% … per 10 dB» range (Livingston et al., 2024) into a single pooled estimate — but note the pooled slope draws on only two cohorts, so the dose-response is directional evidence, not a well-populated curve.
- By severity band, both mild (HR 1.27 [1.05-1.53]) and moderate-to-severe (HR 1.69 [1.29-2.22]) hearing loss raised dementia risk, but degree did not statistically moderate the association (p=0.09). (Yu et al., 2024) A monotone-looking severity gradient that does not reach significance as a moderator — consistent with a real graded effect, not proof of one.
No effect-modifier was found — the moderator-null [first-hand: Yu 2024]
Across 37 HR-reporting studies, «None of the factors investigated moderated the relationship between hearing loss and any type of incident cognitive impairment» — not baseline age, hearing-assessment type, follow-up length, nor adjustment for baseline cognition or vascular factors; meta-regression on female proportion, baseline age, and follow-up years was likewise null. (Yu et al., 2024) Decision consequence: the observational evidence supplies no route-(b) effect-modifier for the exposure association — no identified stratum in which hearing loss carries a relatively larger dementia risk. Keep two objects distinct: Yu’s null is about the exposure relative effect, whereas ACHIEVE’s pre-specified interaction (pinteraction=0.010, below) is about the treatment relative effect — a positive route-(b)-form signal, but hypothesis-generating (one subgroup, lenient alpha, a contamination confound). So the established reading of the ARIC subgroup stays baseline-risk (route-a, absolute-benefit — robust even if the interaction is a false positive), not a confirmed effect-modification.
Reverse-causation guard [first-hand: Yu 2024]
Yu’s causal read is guided by Bradford Hill criteria and defends temporality directly: it excluded studies with baseline dementia and those with under-two-year follow-up. But it does not over-claim — «dementia has a long prodrome of several years so reverse causality cannot be completely excluded». (Yu et al., 2024) As a probe, Yu tested whether follow-up length moved the association and found no significant effect (longer latency did not shrink the effect, which argues against pure reverse causation). (Yu et al., 2024) The Commission’s complementary note stands: a shared cardiovascular pathology has «not been reported» to account for the association (Livingston et al., 2024). Kept as a medium-confidence causal read, not high — an incompletely-excludable prodrome is exactly why.
The treatment lever — hearing aids [intervention arm; confidence LOW]
The intervention question — does correcting hearing loss lower dementia risk? — is a different and weaker evidence state than the exposure association above.
- Observational (Yeo 2023, first-hand). An 8-study pool (n=126,903, follow-up 2-25 years) found «significantly lower hazards of any cognitive decline among hearing aid users compared with participants with uncorrected hearing loss (HR, 0.81; 95% CI, 0.76-0.87; I2 = 0%)» — a 19% lower hazard, with the incident-dementia subgroup at HR 0.83 [0.77-0.90] (k=4). (Yeo et al., 2023) A short-term arm (11 studies, n=568) found a 3% test-score improvement (ratio of means 1.03 [1.02-1.04]). (Yeo et al., 2023) These are the figures the 2024 Commission borrowed (its «19% (0.76-0.87)» and «17% (0.77-0.90)» (Livingston et al., 2024)); Yeo is the first-hand source.
- Three first-hand caveats the borrowed version dropped, all load-bearing:
- The pooled benefit is observational, and self-selected. Yeo is by its own description a «multiadjusted observational meta-analysis of 31 observational studies» (Yeo et al., 2023); the 8-study longitudinal pool contains no RCT. Hearing-aid users self-select on health, motivation and resources — a classic healthy-user route to a spurious protective HR that adjustment does not remove -> The Observational-Trial Discordance, Upgrading Observational Evidence.
- Overall GRADE quality was low. «Overall quality of evidence was low when assessed using GRADE». (Yeo et al., 2023)
- The search predates the definitive trial. Yeo searched only to July 2021, so it cannot include ACHIEVE (2023) — the pooled 19% is a pre-RCT observational estimate. The short-term test-score arm carries its own artifact, which Yeo names: «The supposed improvement in cognitive test scores is confounded by the fact that participants can simply hear the instructions of the test better after hearing restoration». (Yeo et al., 2023)
- RCT (ACHIEVE 2023) — null overall on the primary outcome, now held first-hand. The first hearing-aid RCT (N=977, aged 70-84, untreated hearing loss; hearing aids + audiological support vs a health-education control; 3-year change in a global-cognition factor score) found no between-group difference in the total cohort: difference 0.002 SD [95% CI -0.077 to 0.081], p=0.96 (intervention -0.200 [-0.256 to -0.144] vs control -0.202 [-0.258 to -0.145]). (Lin et al., 2023) This first-hand result de-secondhands the borrowed account the page previously carried via the Commission and Yu; the overall null on the continuous cognition slope is well-powered (the trial was sized on it) and clean.
- The subgroup benefit is a pre-specified interaction — hypothesis-generating, not established effect-modification. A pre-specified sensitivity analysis stratified by recruitment source showed the effect differed between the higher-risk ARIC cohort and the healthy de-novo volunteers (pinteraction=0.010, tested at a lenient prespecified alpha<0.10). In ARIC (n=238) the intervention gave a 48% smaller 3-year cognitive decline: difference 0.191 [0.022 to 0.360], p=0.027; in de novo (n=739) the difference was -0.061 [-0.151 to 0.028], p=0.18 — near-zero, if anything favouring control. (Lin et al., 2023) Read this as one pre-specified subgroup of one trial at a liberal alpha — the route-(b) false-positive generator — not an established relative-effect modifier. Two facts keep it hypothesis-generating: (i) the harder incident-impairment endpoint was null in every stratum (below), and (ii) a differential-contamination confound is present — control drop-in (obtaining aids outside the study) was 19.4% in de novo vs 7.8% in ARIC, which biases the de-novo comparison toward null on its own, mimicking effect-modification by baseline risk. (Lin et al., 2023) The ARIC protective estimate did grow under per-protocol and CACE (compliance-adjusted) analyses vs intention-to-treat. (Lin et al., 2023) Yu, whose authors include the ACHIEVE PI, records the same status directly: it «was a pre-planned but secondary analysis, and we therefore need to see if further RCTs to replicate this effect in people at higher risk for dementia». (Yu et al., 2024)
- On the incidence quantity Yeo estimated, ACHIEVE is null but statistically compatible — softening the headline discordance. ACHIEVE’s secondary incident-cognitive-impairment outcome (adjudicated dementia/MCI or a persistent MMSE drop) showed no reduced hazard in any stratum: total HR 0.90 [0.61-1.33] (p=0.59), ARIC 0.94 [0.54-1.64], de novo 0.89 [0.48-1.67]. (Lin et al., 2023) These hazard ratios are the quantity commensurable with Yeo’s observational cognitive-decline HR 0.81 [0.76-0.87] and dementia HR 0.83 [0.77-0.90] — and ACHIEVE’s CI includes Yeo’s point estimate, so on the same quantity the RCT is underpowered (wide CI), not contradicting. The clean overall null sits on a different quantity — the continuous 3-year cognition slope — which Yeo never estimated. The discordance is therefore an average-effect erasure plus a quantity mismatch, not a head-on reversal.
Exposure is not intervention — the distinction that governs the decision
A robust risk factor does not certify a treatment target. Hearing loss raising dementia risk (Yu, 50 cohorts, tight CI) does not establish that fitting hearing aids lowers it — that inference needs intervention evidence, and the intervention evidence is weaker on every axis:
| Exposure association | Intervention benefit | |
|---|---|---|
| First-hand source | Yu 2024 (50 cohorts) | Yeo 2023 (observational) + ACHIEVE RCT |
| Best design | large cohorts, screened for pub-bias | 1 RCT (null overall) + observational pool |
| Direction/size | HR 1.35 [1.26-1.45] | observational 0.81 [0.76-0.87]; RCT null overall (0.002 SD, p=0.96), 48% in high-risk ARIC subgroup (pre-specified, hypothesis-generating) |
| Confounding risk | reverse causation (prodrome) | healthy-user self-selection (device users) |
| Confidence | medium | low |
The Yeo-observational vs ACHIEVE-RCT gap is a textbook instance of The Observational-Trial Discordance: a large, consistent, I2=0% observational signal beside an overall-null RCT. But before crowning the trial, match the quantities — the parameter table shows the apparent clash is partly non-commensurable, and the commensurable part is not a reversal:
| Parameter | Yeo 2023 (observational) | ACHIEVE 2023 (RCT) | Same quantity? |
|---|---|---|---|
| Design | pooled observational cohorts, no RCT | randomized, health-education control | no |
| Exposure contrast | aid users vs uncorrected HL (self-selected) | randomized intervention vs control | no — self-selected vs randomized |
| Continuous cognition | not estimated | 3-yr global-cognition slope diff 0.002 SD [-0.077 to 0.081], p=0.96 | no — Yeo has no continuous-slope estimate |
| Incident decline/dementia (HR) | cognitive-decline HR 0.81 [0.76-0.87]; dementia 0.83 [0.77-0.90] | incident cognitive impairment HR 0.90 [0.61-1.33] | yes — both time-to-event hazards |
| Verdict on the HR row | — | ACHIEVE 0.90 CI includes Yeo 0.81/0.83 -> underpowered, not contradicting | commensurable; no clash |
So the discordance is resolved by two moves, not one: (1) the clean overall null is on a continuous quantity Yeo never estimated, and on the quantity Yeo did estimate the RCT is merely imprecise; (2) where the RCT is powered (the continuous primary), healthy-user self-selection in the device-user arm inflates the observational HR, randomization removes it, and a signal survives only where absolute baseline risk is high enough to show one -> the ARIC subgroup. A confounded-observational + baseline-risk + non-commensurable-endpoint resolution.
Decision relevance
- Default the stratification to route-(a) baseline-risk, even though ACHIEVE supplies route-(b)-form evidence. ACHIEVE’s pre-specified interaction (pinteraction=0.010) is effect-modification of the treatment effect in form, but it is hypothesis-generating — one subgroup of one trial at a lenient alpha, with a differential-contamination confound (de-novo control drop-in 19.4% vs ARIC 7.8%) that alone can manufacture the split. So the decision rests on the route-(a) reading: absolute benefit is largest where baseline dementia risk is highest (the ARIC arm), which recommends targeting high-risk groups without needing the subgroup claim to be true, and is robust if the interaction is a false positive -> Baseline Risk and the Relative-Absolute Split. The aided-benefit claim still rests mainly on consistent-but-confounded observational evidence plus a well-powered overall-null RCT, not on RCT proof of benefit.
- The lever is cheap and low-harm, and correction exists for most hearing loss — so even under low confidence in the population-average benefit, the expected-value case for high-baseline-risk older adults is favourable; the gap is uptake, not availability. Reducing harmful noise exposure is the upstream (primary-prevention) arm.
- Open loop / next evidence. The decisive missing piece is a replication RCT powered in high-risk strata; ACHIEVE’s own investigators say so. Until then, treat hearing loss to prevent dementia is a reasonable bet at high baseline risk, not an established general-population recommendation.
Self-critique [run 2026-09-04, before commit — ACHIEVE de-secondhand]
- ACHIEVE first-hand did not launder the subgroup into an established effect. The 48% ARIC reduction is stated with its CI, its lenient interaction alpha, the every-stratum incidence-HR null, and the drop-in contamination confound — kept explicitly hypothesis-generating, and the decision defaulted to route-(a) baseline-risk so it does not depend on the subgroup being real. No “aids cut dementia 48%” headline.
- Obs-vs-RCT filed as a DISTINCTION, not a fresh tension (fake-tension check). The parameter table
shows the overall null (continuous 0.002 SD slope) and Yeo’s HR 0.81 are different quantities
(not-joined check (ii): different unit/outcome); on the commensurable incidence-HR quantity ACHIEVE 0.90
[0.61-1.33] includes Yeo’s estimate, so there is no head-on contradiction to file. Correctly homed as an
instance of The Observational-Trial Discordance, not a new
[[tension]]page. - Independence honest — no E claimed. Lin (ACHIEVE PI) co-authored held Yu 2024, and ARIC investigators (Coresh/Mosley/Knopman) recur across the cluster, so ACHIEVE is not type-E-independent of the hearing/Livingston lineage; its value is the RCT design (obs-vs-RCT), added as F-refinement, not independent corroboration.
- Exposure/intervention confidence kept separate and honest. The page-spine (exposure) is medium; the intervention claim is explicitly low, labelled at point of use, and the RCT null-overall leads the aid section rather than the observational 19%. No false “treat-to-prevent” headline.
- De-secondhanding did not launder independence. Yu is marked F-refinement, explicitly NOT E-independent, with the shared-author list named — so the tighter CI is not sold as independent corroboration of the Commission. Yeo is genuinely independent of Yu/Livingston (separate Singapore group), but it supplies the observational arm at low GRADE, not an independent RCT, so it raises no confidence on causation.
- Subtype-null handled as insufficient, not no-effect. Vascular-dementia null is reported with Yu’s own power/dilution caveat, not asserted as absence of effect.
- Multi-source dementia reference page —
confidence: medium(exposure spine). Orbits Dementia Prevention and Modifiable Risk Factors; not a competing nucleus.