DSST indicates Digit Symbol Substitution Test; HCHS, Hispanic Community Health Study; and NHANES, National Health and Nutrition Examination Study.
aIn NHANES, only individuals aged 60 to 69 years had both audiometry and cognitive testing. Therefore, we were restricted to this age range for our analysis.
Linear regression was used for comparison. Multivariable models were adjusted for age, sex, educational level, and cardiovascular disease. The generalized additive model (GAM) regressions show that the association between hearing loss and cognition is not purely linear and that a stronger association between hearing loss and cognition may exist among individuals with normal hearing than those among those with HL. The SE CIs are in orange shading. HCHS indicates Hispanic Community Health Study; NHANES, National Health and Nutrition Examination Study; and SEVLT, Spanish-English Verbal Learning Test.
aStatistically significant for linear regression and GAM regression, except the multivariable models for the digit symbol substitute test in NHANES; coefficients and 95% CIs are included in Table 2.
All models were adjusted for age, sex, educational level, and cardiovascular disease. HCHS indicates Hispanic Community Health Study; NHANES, National Health and Nutrition Examination Study; and SEVLT, Spanish-English Verbal Learning Test.
aStatistically significant for a hearing stratum; coefficients and 95% CIs are included in Table 2.
bStatistically significant for a difference between the 2 hearing strata; coefficients and 95% CIs are included in eTable 3 in the Supplement.
eTable 1. Prevalence of Hearing Loss Categories
eTable 2. Linear Regression Models for z Scores of Cognitive Outcomes in Subjects With Normal Hearing
eTable 3. Interaction Term Coefficients for Separate Multivariable Linear Regression
eTable 4. Linear Regression Models for Cognitive Outcomes Based on Hearing Loss, Restricted to HCHS Subjects Ages 60-69 Years
eTable 5. Linear Regression Models for Digit Symbol Substitution Test Based on Hearing Loss in NHANES Using Sample Weighting
eFigure. Cognitive Performance vs Hearing in the National Health and Nutrition Examination Study Using Sample Weighting
Customize your JAMA Network experience by selecting one or more topics from the list below.
Identify all potential conflicts of interest that might be relevant to your comment.
Conflicts of interest comprise financial interests, activities, and relationships within the past 3 years including but not limited to employment, affiliation, grants or funding, consultancies, honoraria or payment, speaker's bureaus, stock ownership or options, expert testimony, royalties, donation of medical equipment, or patents planned, pending, or issued.
Err on the side of full disclosure.
If you have no conflicts of interest, check "No potential conflicts of interest" in the box below. The information will be posted with your response.
Not all submitted comments are published. Please see our commenting policy for details.
Golub JS, Brickman AM, Ciarleglio AJ, Schupf N, Luchsinger JA. Association of Subclinical Hearing Loss With Cognitive Performance. JAMA Otolaryngol Head Neck Surg. 2020;146(1):57–67. doi:10.1001/jamaoto.2019.3375
Is the association between hearing and cognition present among individuals who have classically defined normal hearing levels?
In this cross-sectional study of 6451 individuals, there was an inverse association between decreasing hearing and decreasing cognition among those classically defined as having normal hearing, after adjusting for confounders.
The findings suggest that the association between hearing loss and impaired cognition may be present at earlier levels of hearing loss than previously recognized; the current 25-dB threshold for defining adult hearing loss may be too high.
Age-related hearing loss (HL) is a common and treatable condition that has been associated with cognitive impairment. The level of hearing at which this association begins has not been studied to date.
To investigate whether the association between hearing and cognition is present among individuals traditionally classified as having normal hearing.
Design, Setting, and Participants
Cross-sectional study of 2 US epidemiologic studies (Hispanic Community Health Study [HCHS], 2008-2011, and National Health and Nutrition Examination Study [NHANES], 1999-2000, 2001-2002, and 2011-2012 cycles). The dates of analysis were November 2018 to August 2019. Multivariable generalized additive model (GAM) regression and linear regression were used to assess the association between HL (exposure) and cognition (outcome). Participants included 6451 individuals aged 50 years or older from the general Hispanic population (HCHS [n = 5190]) and the general civilian, noninstitutionalized US population (NHANES [n = 1261]).
Audiometric HL (4-frequency pure-tone average).
Main Outcomes and Measures
Neurocognitive performance measured by the Digit Symbol Substitution Test (DSST) (score range, 0-113), Word Frequency Test (range, 0-49), Spanish-English Verbal Learning Test (SEVLT) 3 trials (range, 5-40), SEVLT recall (range, 0-15), and Six-Item Screener (range, 0-6); higher scores indicated better cognitive performance.
Among 6451 individuals, the mean (SD) age was 59.4 (6.1) years, and 3841 (59.5%) were women. The GAM regression showed a significant inverse association between hearing and cognition across the entire spectrum of hearing after adjusting for demographics and cardiovascular disease. In separate multivariable linear regressions stratified by the classic binary definition of HL, decreased hearing was independently associated with decreased cognition in adults with normal hearing (pure-tone average ≤25 dB) across all cognitive tests in the HCHS. For example in this group, a 10-dB decrease in hearing was associated with a clinically meaningful 1.97-point (95% CI, 1.18-2.75) decrease in score on the DSST. When using a stricter HL cut point (15 dB), an association was also present in NHANES. The associations between hearing and cognition were stronger or equivalent in individuals with normal hearing than among those with HL. For example, there was a 2.28-point (95% CI, 1.56-3.00) combined cohort DSST score decrease per 10-dB decrease among individuals with normal hearing vs a 0.97-point (95% CI, 0.20-1.75) decrease among those with HL, with a significant interaction term between continuous and binary hearing.
Conclusions and Relevance
An independent association was observed between cognition and subclinical HL. The association between hearing and cognition may be present earlier in HL than previously understood. Studies investigating whether treating HL can prevent impaired cognition and dementia should consider a lower threshold for defining HL than the current 25-dB threshold.
Age-related hearing loss (HL) is one of the most prevalent disorders of aging, affecting two-thirds of individuals 70 years or older.1 However, treatment is rare. Only 14% of adults with HL in the United States wear hearing aids.2 Hearing aids are uncommon even in countries where they are covered by national health care, such as the United Kingdom.3 Cognitive impairment and dementia, among the most pressing worldwide public health concerns, have recently been found to be associated with HL.4-6
A recent review article estimated that completely preventing or treating HL may be associated with a 9.1% reduction in new dementia cases, the largest reduction among all known risk factors.7 This conclusion is supported by longitudinal studies4,8-10 that have reported that HL is associated with later cognitive impairment (whereas cognitive impairment is not associated with later HL9) and by early uncontrolled treatment studies.11,12 These findings have had major policy implications, including recently passed US legislation that allows hearing aids for mild to moderate HL to be available without a prescription.13
The current understanding of the association between hearing and cognitive impairment is that the association is first seen with mild HL. At this and higher levels, the risk of cognitive impairment and dementia increases in a dose-dependent fashion.8,10 To our knowledge, the association between cognition and HL in the category of normal hearing, defined widely in adults by a pure-tone average (PTA) of 25 dB or less,14 has not been studied. Because associations have been noted between mild HL and dementia,10,15 the association between HL and cognitive impairment may be present at earlier stages (ie, when HL is subclinical [less than mild]). This hypothesis is important because the 25-dB threshold is arbitrary, with some researchers suggesting that it is too lenient.16,17 Furthermore, it is unclear when to begin treatment for HL, and guidelines are lacking or are not evidence based.
In this study, we examined the association between the continuum of hearing performance, particularly in individuals considered to have normal hearing (PTA ≤25 dB), and cognition. We examined 2 large US cross-sectional cohorts, assessed multiple different cognitive outcome measures, and used different types of regression techniques to explore this association. We hypothesized that the association between hearing and cognition is already apparent among individuals with subclinical HL. We also hypothesized that the association between hearing and cognition is similar between individuals with subclinical HL and those with traditionally defined HL.
Participants from both the Hispanic Community Health Study (HCHS) and the National Health and Nutrition Examination Study (NHANES) were included in this cross-sectional study. The flow of participant inclusion and exclusion is shown in Figure 1. The Columbia University Institutional Research Board deemed this secondary analysis of anonymized, publicly available data to be nonhuman research. Informed consent was waived.
The HCHS is a multicenter US community-based prospective study. Almost all older adult participants underwent both audiometry and cognitive testing. Testing was conducted in English or Spanish per individual preference. Only the 2008-2011 wave of data was available; therefore, a cross-sectional analysis was performed. There were originally 14 155 participants. To restrict the analysis to individuals at risk for age-related HL, those younger than 50 years (n = 7980) and individuals with early-onset HL (n = 205) were excluded. Participants with missing audiometry, cognitive testing, or covariate data were further excluded, leaving 5190 participants for analysis.
The NHANES is a biannual cross-sectional study designed to be representative of the US population. This second cohort was chosen to confirm findings noted in HCHS and to extend generalizability to a non-Hispanic population. Only the 1999-2000, 2001-2002, and 2011-2012 cycles contained both audiometry and cognitive testing. These cycles were merged for analysis, resulting in 30 760 total participants. However, only those aged 60 to 69 years underwent both audiometry and cognitive testing; therefore, 28 294 individuals outside this age range were excluded, leaving 1392 participants. Participants without the Digit Symbol Substitution Test (DSST) or covariate data were further excluded, leaving 1261 participants for analysis. Of these, 308 of 1261 (24.4%) came from the 1999-2000 cycle, 282 of 1261 (22.4%) came from the 2001-2002 cycle, and 671 of 1261 (53.2%) came from the 2011-2102 cycle.
The HCHS and NHANES contained the same hearing and covariate data, as well as DSST scores. A combined cohort of 6451 individuals aged 50 years or older was thus created for additional analyses with greater power.
In both HCHS and NHANES, hearing was assessed with pure-tone audiometry. Unaided ear-specific hearing thresholds were measured from 500 to 8000 Hz. Higher decibel thresholds indicated worse hearing. The primary exposure was HL, as defined by the 4-frequency PTA in the better ear, representing the mean threshold at 500, 1000, 2000, and 4000 Hz. The choice of the better ear PTA was consistent with prior epidemiologic studies8,10 of age-related HL.
Hearing loss was defined by the globally used clinical and research cutoff of a PTA exceeding 25 dB.14 Traditionally, normal hearing is defined as a PTA of 25 dB or less. In a sensitivity analysis, we defined strict HL as a PTA exceeding 15 dB. In this report, we defined subclinical HL as a PTA of 1 to 25 dB. In another sensitivity analysis herein, we defined strict subclinical HL as a PTA of 1 to 15 dB.
The following 5 cognitive outcomes designed to assess a wide range of neurocognitive function across a variety of demographic backgrounds were studied: DSST, Word Frequency Test, Spanish-English Verbal Learning Test (SEVLT [3 trials and recall]), and Six-Item Screener. Details of these tests have been described.18 The HCHS included all measures, whereas NHANES only included the DSST. All measures were analyzed continuously, with higher scores indicating better cognitive performance. Score ranges are listed in Table 1.
Speed and attention were assessed with the DSST (from the Wechsler Adult Intelligence Scale–Revised [WAIS-R]) and the Word Frequency Test (letter fluency test). The DSST, a widely used measured of cognitive function in epidemiologic studies, assesses psychomotor speed and attention.19 Participants filled a series of symbols corresponding to specific digits within 90 seconds (HCHS) or 120 seconds (NHANES).20 To allow comparison of DSST scores across studies, the HCHS score was scaled by multiplying by 1.33. The Word Frequency Test is a measure of verbal fluency. Participants were asked to say words starting with certain letters as fast as possible during two 60-second trials.21
Verbal memory and learning were measured with the SEVLT. The participant was read a set of common words and then asked to recall them. Three separate trials were conducted with distinct word sets.22 The sum of the number of words across the 3 trials was recorded (SEVLT 3 trials), with a possible score ranging from 0 (the worst score) to 45 (the best score). A distracting word set was then read and repeated back. Immediately afterward, the participant attempted to recall the initial word set (SEVLT recall).
Global cognitive function was assessed with the Six-Item Screener.23 This assessment tool is a short screen for cognitive impairment containing 6 questions.
Variables that might confound the association between HL and cognitive performance were included in multivariable models. For both HCHS and NHANES, the same covariates with the same coding were used, including age, sex, educational level, and cardiovascular disease. Educational level was coded as a 4-level categorical variable (less than ninth grade, high school, trade school or some college, or college degree or higher). Cardiovascular disease is a potential confounder because it could contribute to both HL and impaired cognition. A composite cardiovascular disease variable aggregating several risk factors was generated to avoid multicollinearity.24 One point was assigned for each of 3 present risk factors, including coronary artery disease, hypertension, and/or history of transient ischemic attack or stroke. In addition, 1 point was added for impaired glucose tolerance and 2 points for diabetes. The score range was 0 to 5.
Descriptive analyses were performed on the separate and combined data sets. Means (SDs) were used to describe continuous variables. Frequencies and proportions were used for categorical variables. Three regression modeling strategies were used to flexibly examine associations between exposure (hearing [continuous]) and each outcome (cognition [continuous]).
The first strategy is linear regression. This approach assumes that the average score on the cognitive outcome assessment changes linearly with hearing. The univariable model is represented by cognitive score = β0 + β1hearing + ε, where β0 is the intercept, β1 is the change in average cognitive score for a 1-dB increase in PTA, and ε is an error term. This model is simple and easy to interpret; however, it does not capture nonlinear associations.
The second strategy is generalized additive model (GAM) regression,25 assuming a smooth, possibly nonlinear association between hearing and cognition. This technique produces an association between hearing and cognition that is determined by the data rather than the modeler. In other words, GAM regression does not assume a specific a priori functional association (eg, linearity) between hearing and cognition. Therefore, it may allow a better fit than models assuming a strict linear association. The univariable GAM regression is represented by cognitive score = β0 + g(Hearing) + ε, where β0 and ε are as described previously and g(hearing) is a smooth, nonlinear function that is estimated using penalized regression splines.26 A smoothing parameter controls the smoothness of the estimate for g(Hearing), which is selected to minimize the generalized cross-validation score, a measure of model fit.26 Unlike linear regression models, GAM regressions do not produce point estimates. Instead, the meaning of the modeled association comes only from the visual plot of the estimate of g(Hearing).
Separate linear regressions for individuals with and those without HL are the third strategy. Linear regression models are fit such that the linear association between hearing and cognition depends on whether a person has normal hearing or HL. The model is represented by cognitive score = β0 + β1hearing + β2group + β3hearing × group + ε, where group is 1 if an individual has HL or 0 if he or she has normal hearing. The other components are similar to those specified in the first model. The threshold between normal hearing and HL was the widely used 25-dB cutoff,14 supported a posteriori from the GAM regression. An alternative, stricter HL cut point of 15 dB was also used.27 Testing β3 = 0 allows us to assess whether the linear association between hearing and cognition changes depending on whether the individual has or does not have HL.
For each of the 3 regression strategies, we fit univariable models with hearing as the sole associated risk factor. We then fit similar multivariable models, adjusting for age, sex, educational level, and cardiovascular disease score. All models were fit separately on the HCHS and NHANES samples. Models with the DSST as an outcome were also fit for a combined sample.
In a sensitivity analysis, regressions were restricted to those participants who did not wear hearing aids. We also restricted the HCHS age range to 60 to 69 years to match that of NHANES. Finally, sample weighting was used for NHANES regressions to account for the complex sampling design.28
We computed z scores to express score differences in SDs. All hypothesis tests were 2 sided. Significance was defined at the α = .05 level. Analyses were performed from November 2018 to August 2019 using R, version 3.5.1 (R Foundation for Statistical Computing) with RStudio, version 1.2.1181 (RStudio Inc). Packages included mgcv (GAM regression), car (regression diagnostics), and survey (sample weighting).
Baseline participant characteristics are listed in Table 1. Of 6451 participants, 3841 (59.5%) were women, and the mean (SD) age was 58.3 (6.2) years in the HCHS participants and 63.9 (2.8) years in the NHANES participants. Hearing aid use was only 0.9% (46 of 5190) in HCHS participants. Among NHANES participants who had hearing aid data recorded (978 of 1261), use was 4.7% (46 of 978). Given the low prevalence of hearing aid use, it was difficult to specifically analyze data among hearing aid users. The distribution of HL, including categorization across cohorts, is summarized in eTable 1 in the Supplement.
Assumptions of linear regression were assessed and met. In simple, univariable linear regression models, cognitive performance consistently decreased as hearing decreased (Figure 2A and B and Table 2). For example, for every 10-dB decrease in hearing, the DSST score decreased by 3.12 (95% CI, 2.72-3.52) points in the combined cohort. This association held across cohort studies (HCHS, NHANES, and the combined cohort for the DSST). It also held across different cognitive tests (Word Frequency Test, SEVLT 3 trials, SEVLT recall, and Six-Item Screener in HCHS).
Associations remained, albeit attenuated, in multivariable linear regression models after adjusting for confounders (Figure 2C and D and Table 2). For example, for every 10-dB decrease in hearing, the DSST score decreased by 1.10 (95% CI, 0.74-1.46) points in the combined cohort, adjusting for confounders. All associations were statistically significant except for the multivariable model for the DSST in NHANES.
The GAM regression was used to explore nonlinear associations between hearing and cognition. The estimated smooth effect curves demonstrating the associations between HL and the DSST are shown in Figure 2. As hearing decreased, the DSST score declined (Figure 2A and C). Qualitatively similar visual patterns occurred for participants in HCHS and NHANES. These same associations were found in the univariable analysis and multivariable analysis after adjusting for age, sex, educational level, and cardiovascular disease. In the multivariable analysis, the association was attenuated. Unexpectedly, the association appeared to be consistently stronger (ie, the decrease in cognitive performance was steeper) among individuals with normal hearing than among those with HL.
The GAM regression analysis was repeated in individuals from the HCHS for the 4 other cognitive tests, including the Word Frequency Test, SEVLT 3 trials, SEVLT recall, and Six-Item Screener (Figure 2B and D). In all cases, the association between hearing and cognition appeared to be stronger among those with normal hearing than among those with HL. For some cognitive outcomes (ie, DSST and Word Frequency Test), an inflection point was seen between normal hearing and HL.
Almost all associations were statistically significant at the α = .05 level (ie, the smooth effect curve is nonconstant). The exception was the multivariable model for the DSST in NHANES.
Because a stronger association was apparent between hearing and cognition among those with normal hearing in GAM regression, 2 separate groups of multivariable linear regression models were created, one for those with normal hearing (PTA ≤25 dB) and another for those with HL (>25 dB). The adjusted association between hearing and DSST score was significant for individuals with normal hearing and for those with HL in HCHS and the combined cohorts (Figure 3A and Table 2). In adults with normal hearing, a 10-dB decrease in hearing was associated with a clinically meaningful 1.97 (95% CI, 1.18-2.75) decrease in score on the DSST in the HCHS or a 2.28 (95% CI, 1.56-3.00) decrease in the combined cohort. In adults with hearing loss, a 10-dB decrease in hearing was associated with a 1.01 (95% CI, 0.19-1.84) decrease in score on the DSST in the HCHS or a 0.97 (95% CI, 0.20-1.75) decrease in the combined cohort.
We then repeated this analysis in HCHS participants for the 4 other cognitive tests (Figure 3B and Table 2). The adjusted association between hearing and cognition was statistically significant among participants with normal hearing but did not reach statistical significance among those with HL. For example, in adults with normal hearing, a 10-dB decrease in hearing was associated with a 0.63 (95% CI, 0.26-0.99) decrease in score on the Word Frequency Test. In adults with HL, however, there was no significant association.
In addition, we computed z scores to express cognitive outcome score differences in more familiar units of SD. Among participants with normal hearing in HCHS, the cognitive outcome decrease associated with a 10-dB decrease in hearing ranged from SDs of 0.08 (95% CI, 0.03-0.14 [Six-Item Screener]) to 0.13 (95% CI, 0.08-0.18 [SEVLT recall]) after adjusting for confounders (eTable 2 in the Supplement). Put in context, the cognitive outcome decrease associated with a 25-dB decrease in hearing (from 0-dB perfect hearing to the 25-dB threshold for HL) ranged from SDs of 0.20 (95% CI, 0.08-0.35 [Six-Item Screener]) to 0.33 (95% CI, 0.02-0.45 [SEVLT recall]) after adjusting for confounders.
Statistically significant differences in the adjusted association of hearing with cognition between stratified groups (ie, normal vs HL) were then tested by creating a single multivariable regression model with an interaction term between continuous and binary hearing. Statistically significant differences were observed for the DSST in the combined cohort. The association between hearing and the DSST score was significantly stronger among those with normal hearing (2.28; 95% CI, 1.56-3.00 point DSST score decrease per 10-dB decrease in hearing) than among those with HL (0.97; 95% CI, 0.20-1.75), after adjusting for confounders (interaction term coefficient, 1.29; 95% CI, 0.22-2.35) (Figure 3A and B and eTable 3 in the Supplement).
In an additional analysis, the cut point between normal hearing and HL was changed to a stricter 15-dB cutoff.27 The adjusted association between hearing and cognition was significant for all models in both hearing strata except for the DSST scores in the NHANES group among individuals with strict HL and the Six-Item Screener (Figure 3C and D and Table 2). Statistically significant differences between the adjusted association of hearing with cognition between stratified groups (ie, strict normal ≤15 dB vs strict HL >15 dB) were observed for all tests except the SEVLT 3 trials and the Six-Item Screener (Figure 3C and D and eTable 3 in the Supplement). The association between hearing and cognition for 3 of the 5 tests and for the DSST in HCHS and NHANES participants was significantly stronger among individuals with strict normal hearing than among those with strict HL after adjusting for confounders. For example, the association between hearing and the DSST score was significantly stronger among those with strict normal hearing (3.94; 95% CI, 2.43-5.46 point DSST score decrease per 10-dB decrease in hearing) than among those with strict HL (0.63; 95% CI, 0.12-1.14), after adjusting for confounders (interaction term coefficient, 3.32; 95% CI, 1.78-4.86).
In a sensitivity analysis, regressions in HCHS participants were restricted to those who did not wear hearing aids. Alternatively, hearing aid use was added as a model covariate. No clinically meaningful changes were noted. In a second sensitivity analysis, the HCHS group was restricted to individuals aged 60 to 69 years to match those in NHANES group. A significant association remained between hearing and the DSST among individuals with normal hearing in the HCHS group and the combined cohort. The associations between hearing and the other cognitive measures among participants with normal hearing in the HCHS group were variably attenuated and nonsignificant (of note, the HCHS sample size was reduced to 1636 from 5190) (eTable 4 in the Supplement) Finally, sample weighting was applied to NHANES regressions. The association between hearing and the DSST was significantly stronger among participants with normal hearing than among those with HL (eTable 5 and eFigure in the Supplement).
An inverse association between decreasing hearing and decreasing cognition was observed among people classically considered to have normal hearing. This association remained despite controlling for confounders, including age, sex, educational level, cardiovascular disease, or hearing aid use. The association between hearing and cognition was stronger among those with classically defined normal hearing compared with those with HL. The decrease in cognition per unit decrease in hearing among individuals with normal hearing was clinically meaningful.18
These findings are contrary to the prevailing understanding of HL as a risk factor for cognitive impairment. Most researchers consider that the association between hearing and cognition does not begin unless individuals have HL defined by the widely used 25-dB cutoff.14,29 This is based on the assumption that hearing better than this threshold is asymptomatic and without clinical consequence. We hypothesized that an equivalent association would exist among individuals with normal hearing and individuals with HL. Instead, the association not only existed among individuals with normal hearing, but also in 1 case was considerably stronger than that in individuals with HL. This was initially and unexpectedly observed in GAM regressions, which did not assume any particular associations (eg, linearity) between hearing and cognition.
Stricter definitions of normal hearing exist. Commonly, a 20-dB threshold is used in children and, rarely, 15 dB in adults.27,29 We reanalyzed the data using the stricter 15-dB cutoff, which made the previously observed patterns more pronounced. Across 2 unrelated large, national cross-sectional studies and across 5 different cognitive tests, there was an association between hearing and cognition among those with a strict definition of normal hearing.
Biologically, HL is a continuum. Hearing is most commonly measured by the quietest intensity of sound necessary to hear a pure tone 50% of the time. The value of a 0-dB hearing level is artificially set based on normative data from standardizing organizations (eg, American National Standards Institute).29 Technically, any value higher than this level (>0 dB) is worse than normal. In practice, more liberal cutoffs are used. For research purposes, we recommend the term subclinical HL, defined as imperfect hearing (>0 dB) but within the currently accepted normal range. For clinical purposes, more study is needed before recommending a change to the historical HL categorization. One possibility is formally introducing a lay-friendly term, such as borderline HL, for 16 to 25 dB. This term is sometimes used today, but definitions vary.
To our knowledge, no prior study has examined whether the association between hearing and cognition (excluding dementia) is present in subclinical HL. Studies30-33 that categorize HL (into mild, moderate, severe, profound, or present) have only used the entire category of 25 dB or less as a reference for normal. In doing this, an a priori assumption is made that the association between hearing and cognition begins in individuals with hearing of 25 dB or worse. Our data suggest that this assumption is incorrect.
To explore associations in HL of 25 dB or less, hearing must be defined as a continuous variable (or hypercategorized within that range). Several studies15,31,34,35 have shown an inverse linear association between cognition and the entire spectrum of continuous HL, beginning at 0 dB. However, these studies did not specifically examine the association in individuals with HL of 25 dB or less. For example, it is possible (if not commonly assumed) that the regression is primarily influenced by those with HL exceeding 25 dB and the fit is poor among those with HL of 25 dB or less. The GAM regression and stratification of individuals by HL category that we performed suggest that this may not be the case.
Several studies have examined the association between dementia and continuous HL, including HL of 25 dB or less specifically. No associations were seen among those with hearing of 25 dB or less. These findings may be explained by dementia being a discrete diagnosis that is met by reaching thresholds of severe cognitive and functional impairment.36 It seems unlikely that subclinical HL would contribute sufficiently to dementia risk to cross this threshold after adjusting for other risk factors.8,10
Although additional studies are needed, the data imply that the cut point for HL should be closer to 0 dB. Even individuals with hearing of 15 dB or less were at risk for reduced cognition. When considering HL as a continuum, 0 dB may be considered perfect hearing, and any threshold higher than this might be considered a loss. Examination of normative data support this recommendation given that most healthy young people have a PTA lower than 25 dB and individuals with a PTA of 25 dB or less may consult clinicians for HL.16,17
The discussion to create a stricter definition of HL contrasts with the already poor insight, late diagnosis, and low treatment level of age-related HL as it is currently defined. Among US adults with mild HL (25-40 dB), only about 3% wear hearing aids,2 a statistic similar to what we observed. Even among adults who perceive difficulty in hearing, less than half report seeing a hearing specialist within 5 years.37
The association between HL and dementia has been observed to begin with mild HL.8,10 It seems logical that cognitive impairment, a precursor to dementia, would be associated with subclinical HL, the precursor to mild HL. In some models, the association between hearing and cognition was stronger among individuals with subclinical HL than among those with outright HL. If HL and cognition are associated, a possible explanation is that the first stages of impaired communication carry the greatest functional consequences. This is speculative because, although mild HL has been associated with consequences of impaired communication, such as loneliness,38 the social associations of subclinical HL have not been studied to our knowledge. This represents a future avenue for research.
Another explanation for an association between subclinical HL and cognition may be inadequately sensitive tests for HL. Hearing testing is typically performed in quiet settings. Yet, social exchanges in the real world often occur in settings with background noise. It is possible that some individuals with subclinical HL and impaired cognition had more impaired word understanding in noisy settings than individuals with subclinical HL and normal cognition. This explanation could be tested by introducing word-in-noise hearing tests into epidemiologic studies. In addition, if stimulating social discourse can enhance or maintain cognition, then any biological advantage to facilitate this process, including superior hearing, may benefit cognition. Muscle mass exceeding what is considered normal may produce a better runner and more physically fit individual. Likewise, better hearing beyond what we define as normal may produce a better listener and a more cognitively fit individual. This may be particularly true in challenging listening environments, such as a noisy classroom, where those with supernormal hearing may be at a competitive advantage.
In a sensitivity analysis, hearing aid users were removed from the multivariable models for HCHS (hearing aid use was not consistently recorded in NHANES). There was no meaningful change in results. This was not surprising because hearing aid use was less than 1%.
One methodologic concern is whether HL lowered performance on cognitive tests because individuals could not hear instructions. This would misclassify hearing-impaired individuals as having worse cognition, creating a false association. This is unlikely because neuropsychological testing was conducted in a quiet, well-lit setting with trained examiners.39 In this setting, HL should not influence clearly spoken instructions unless profound. In addition, associations were seen for the DSST, which is a nonverbal test. Initial instructions are oral; however, the participant repeats nonscored trials until he or she clearly understands what to do. The remainder of the test is performed in silence.
This study has strengths. We used 2 large national studies and multiple different cognitive outcomes. Numerous ethnic groups and centers were included, which is important for generalizability.4 Results were robust to 2 different cut points for HL. Also, our conclusions are supported by various regression methods. This includes GAM regression, which (unlike linear regression) does not have any assumptions about particular functional associations between hearing and cognition. This study represents the largest investigation of formal, audiometrically measured HL and cognition to date.
This study has limitations. We used cross-sectional studies, which preclude causal inference. Although other studies have shown that HL predicts later cognitive impairment,4,8-10 we cannot address whether subclinical HL and cognitive impairment are causally related. Early declines in both hearing and cognitive performance may be associated with common aging-related processes. We adjusted for major potential confounders, including age, sex, educational level, and cardiovascular disease. However, it is possible that unknown factors confound the association between hearing and cognition. Because HCHS and NHANES are population-based studies, few individuals had severe to profound HL, which increases the uncertainty in regressions among these individuals. This uncertainty could be addressed in future cohort studies by oversampling individuals with severe to profound HL. Also, HCHS and NHANES exclude institutionalized individuals, limiting generalizability.
In this study, decreasing hearing was independently associated with decreasing cognition among adults with subclinical HL. The current 25-dB threshold for defining adult HL may be too high. Studies investigating whether treating HL can prevent impaired cognition and dementia should consider a lower threshold for defining HL.
Accepted for Publication: September 16, 2019.
Corresponding Author: Justin S. Golub, MD, MS, Department of Otolaryngology–Head and Neck Surgery, Vagelos College of Physicians and Surgeons, New York-Presbyterian/Columbia University Irving Medical Center, Columbia University, 180 Ft Washington Ave, Harkness Pavilion, Floor 8, New York, NY 10032 (firstname.lastname@example.org).
Published Online: November 14, 2019. doi:10.1001/jamaoto.2019.3375
Author Contributions: Dr Golub had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Concept and design: Golub, Brickman, Luchsinger.
Acquisition, analysis, or interpretation of data: Golub, Ciarleglio, Schupf, Luchsinger.
Drafting of the manuscript: Golub, Ciarleglio.
Critical revision of the manuscript for important intellectual content: All authors.
Statistical analysis: Golub, Ciarleglio.
Obtained funding: Golub, Schupf, Luchsinger.
Administrative, technical, or material support: Golub, Schupf.
Supervision: Golub, Luchsinger.
Conflict of Interest Disclosures: Dr Golub reported receiving travel expenses for industry-sponsored events from Cochlear, Advanced Bionics, and Oticon Medical; consulting fees from Oticon Medical, Auditory Insight, Optinose, and Decibel Therapeutics; and honoraria from Abbott; and reported that his department received unrestricted educational grants from Storz, Stryker, Acclarent, 3NT, and Decibel Therapeutics. Dr Luchsinger reported being editor in chief of Alzheimer Disease & Associated Disorders and receiving a stipend from its publisher and reported being a paid consultant for vTv Therapeutics. No other disclosures were reported.
Funding/Support: This study was funded by the National Institute on Aging of the National Institutes of Health (grants K23AG057832 and L30AG060513 to Dr Golub and grant K24AG045334 to Dr Luchsinger).
Role of the Funder/Sponsor: The funding source had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.
Additional Contributions: Nicholas S. Reed, AuD (The Johns Hopkins University), Jay T. Rubinstein, MD, PhD (University of Washington), Ward R. Drennan, PhD (University of Washington), Ilana Cellum, AuD (Columbia University), Jessica Galatioto, AuD (Columbia University), and Megan Kuhlmey, AuD (Columbia University), provided insight on the definition of the 0-dB hearing level. They were not compensated for their contributions.
Create a personal account or sign in to: