Can treating hearing loss slow cognitive decline?

Hearing intervention may help some older adults at greater risk of cognitive decline. Universal cognitive benefit and dementia prevention remain unestablished. A 2026 analysis of retesting provides a useful sensitivity check; our previous attempt to exclude measurement explanations by assigning two auditory tests a 20% weight was invalid.

Correction, September 20, 2026: We withdraw the claims that measurement bias cannot explain the ARIC signal and that the intervention delays decline by about 1.4 years. We correct score normalization, the consequences of a persistent gain beginning after baseline, and an invented three-test multiplicity adjustment. We add the actual statistical analysis plan, the 2025 supplementary tables and the 2026 retest analysis, retaining unresolved source discrepancies.

01What the randomized comparison establishes

ACHIEVE enrolled adults aged 70–84 with untreated hearing loss and no substantial cognitive impairment. It assigned 490 to hearing intervention and 487 to health education. Intervention comprised hearing aids, audiological counselling and follow-up. Of the participants, 238 came from the longstanding ARIC cohort and 739 were newly recruited community volunteers. The primary outcome was three-year change in a global cognitive factor score, not dementia or neuropathology.s1

The recruitment-source × assignment × time interaction had p=0.010. Comparing one significant subgroup with one nonsignificant subgroup would not establish this difference. Reconstructing standard errors from the published intervals and assuming independent stratum estimates gives a contrast difference of 0.252 SD, approximate 95% CI 0.061–0.443 and p≈0.0098. This does not refit the mixed model, couple clustering, degrees-of-freedom correction or multiple imputation.

ARIC controls declined by 0.402 SD; 0.191/0.402=47.5%, explaining the rounded 48% headline. This is not a 48% reduction in dementia. Multiplying the ratio by three produces about 1.4 years only under a fixed linear-rate conversion; it does not identify disease postponement. The combined upper limit of 0.081 is about 40% of the combined control decline of 0.202. Calling this a “narrow null” that excludes meaningful relative benefit is too broad.

Sample-size and precision weighting of subgroup point estimates gives approximately 0.0004 and −0.006. These illustrate cancellation; neither equals nor reproduces the adjusted combined estimate of 0.002. Recruitment source does not identify why effects differ. Slower decline, test familiarity and control participants obtaining hearing aids may contribute. Such control-group uptake was 7.8% in ARIC and 19.4% in de novo participants.s1

02Prespecification and the actual multiplicity plan

The main paper describes recruitment-source stratification as prespecified. Public SAP version 5.2 also lists it in section 7.4(f), and section 7.6(e) proposes excluding the two auditory-only tests.s1s5

Our previous article incorrectly said that domain outcomes lacked multiplicity adjustment, then multiplied the global-cognition p-value of 0.027 by three. That mixes different testing families. The paper specifies alpha 0.05 for the primary outcome, Hochberg adjustment across four secondary outcomes, post hoc application to stratified analyses, and alpha 0.10 for interactions.

There is also a source discrepancy: SAP 5.2's revision history and section 7.7 describe five outcomes, including global cognition, whereas the main paper describes four secondary outcomes. We did not obtain the full main-paper appendix or an explanation of that change. We therefore retain reported p-values and distinguish the analysis levels; we neither invent an adjustment nor certify complete reproduction of the multiplicity procedure. The ARIC language-domain estimate, 0.229 SD (0.050–0.408; p=0.012), is supportive but cannot establish a mechanism.s1s5

03What the 2025 risk analysis adds

Pike and colleagues trained their prediction model in 2,692 ARIC members who did not participate in ACHIEVE. However, the highest-quartile cut point was chosen after visualizing the risk × treatment × time spline in the trial. External prediction training does not make treatment-effect subgroup selection externally validated.s2

The authors report 61.6% slower decline in the highest quartile (95% CI 33.7%–94.1%). This is another analysis of the same trial. Table 2's 0.208 SD is the difference between high- and lower-risk treatment contrasts, not the treatment contrast within the high-risk group. Adding the lower-risk contrast, −0.047, gives a high-risk point contrast of 0.161 SD. Its interval requires coefficient covariance. We quote the percentage and its interval as author-reported, without claiming independent reproduction.

We obtained the previously unread supplement. These are risk × treatment × time interactions, not within-group effects:

Directions show some consistency, but precision and cut-point robustness are mixed. Interactions using cognition-only scores or scores excluding hearing and cognition also cross zero. Table S3 has an internal discrepancy: a normal approximation from 0.158 (0.068–0.249) gives p≈0.00062, while the table prints <0.0001. We retain its interval and document the mismatch rather than replace the authors' inference with an approximation.s6

Comparison with non-trial ARIC participants does not add a randomized control: selection, score standardization and follow-up calculations differ. A significant comparison in one arm and a nonsignificant one in the other do not establish a treatment effect. A 2026 biomarker study explores selecting faster-declining participants, but its power calculations assume a 33% intervention effect. It informs trial design, not independent verification of hearing-treatment efficacy.s2s9

04Why the measurement model cannot rule out an alternative

Two of ten tests use exclusively auditory stimuli. The other eight are not necessarily unaffected by hearing, spoken instructions or cognitive load. The global score comes from a latent-variable model, not a simple average of ten standardized tests. Language-domain findings are informative but are not a negative control guaranteed to have zero auditory influence. Pike's discussion explicitly considers reduced cognitive load during language tests. Pre-assessment speech-understanding checks mitigate problems without guaranteeing their absence.s1s2s5s10

Two hypothetical examples expose the missing assumptions:

• Scale. For ten equally weighted, unit-variance tests with common correlation ρ, the mean has SD √[(1+9ρ)/10]. If two tests each shift by δ test SD, the standardized composite changes by 0.2δ/√[(1+9ρ)/10]. A 0.191 composite-SD difference requires δ≈0.30, 0.51 or 0.71 when ρ=0, 0.2 or 0.5. It does not always require 0.96. These correlations are analyst-selected, not ACHIEVE estimates or empirical bias bounds.
• Time. Suppose test performance improves immediately by 0.191 SD after intervention and the gain persists, while latent decline proceeds at exactly the same rate in both groups. Baseline-to-year-three change still differs by 0.191. A shift present equally at baseline and follow-up cancels; a persistent shift beginning after baseline need not. A sustained level gain and slower decline can produce identical final contrasts.

These examples do not show that bias caused the result or estimate its probability. They show that the previous exclusion argument was invalid. Discrimination needs actual scoring parameters, administration records, intermediate visits, less auditory-dependent outcomes and pathological measurements.

05A more direct test from the 2026 retest analysis

Wang and colleagues used 2,571 in-person assessments from the same 977 participants, principally at baseline and years one and three; 554 had all three assessments. Their primary global score excluded Digit Span Backward and Logical Memory. A baseline-zero/follow-up-one indicator estimated a persistent retest-associated gain of 0.11 SD (0.07–0.15).s7

Yet eTable 4 shows very little change in intervention coefficients after adjustment. The ARIC coefficient excluding those tests changed from 0.066 to 0.065; including them, it changed from 0.079 to 0.078. The figure retains the table's coefficient scale and does not recast these as the 2023 three-year contrasts. Its ARIC/de novo header counts, 239/738, disagree with the main text's 238/739; we do not use those counts in calculations.s8

This weakens an explanation based entirely on the modeled retest component. It cannot rule out every auditory, expectation, administration or missingness mechanism. The analysis is post hoc, assumes missingness at random and a time model, and does not perform multiple imputation. Its retest indicator cannot completely separate visit-specific practice from genuine change. It supports the subgroup signal without establishing slower neurodegeneration.

Imaging evidence is not wholly absent: a conference-supplement abstract on 445 participants from the same trial reports nominally protective cortical-thickness findings. Abstract-level information is insufficient to audit all regional tests and analysis choices; it cannot upgrade the evidence to established disease modification.s11

06Events, safety and the remaining boundary

The incident-impairment HR was 0.90 (0.61–1.33) overall, 0.94 (0.54–1.64) in ARIC and 0.89 (0.48–1.67) in de novo participants. This composite includes adjudicated dementia, MCI or a subsequently confirmed MMSE decline of at least three points or telephone equivalent. It is not dementia alone. The intervals detect no difference while retaining potentially meaningful benefit and harm.s1

The trial reported no unexpected adverse events judged related to participation, not an absence of all adverse events. One hundred participants did not complete the year-three visit: 24 lost to follow-up, 26 withdrawn, 34 deceased and 16 without completed cognitive assessment. The primary analysis combined 862 year-three in-person assessments, nine eligible pre-death assessments and 106 imputed scores. Those counts describe different things.s1

The defensible conclusion is an undetected overall effect and a prespecified ARIC signal for slower cognitive-score decline. Risk enrichment and retest analyses help delimit it but do not constitute independent trials or establish dementia prevention. This article does not derive hearing-aid prescriptions or screening rules.

No individual data were obtained or trial models refitted. The reproducibility package contains located summary inputs, model code, four original figures and numerical validation. Full source documents are not redistributed. The unavailable main appendix, SAP testing-family discrepancy and supplementary p-value/count discrepancies remain explicit. Evidence was checked through September 20, 2026. Codex performed the revision, self-review and editorial sign-off as one agent, not independent human or clinical review.