Audiometric Age-Related Hearing Loss and Cognition in the Hispanic Community Health Study

Abstract Studies associating age-related hearing loss (HL) with cognition have been limited by non-Hispanic cohorts, small samples, or limited confounding control. We overcome these limitations in the largest study of formal, audiometric HL and cognition to date using the multicentered Hispanic Community Health Study (n=5,277, mean age=58.4 [SD=6.2]). The main exposure was audiometric HL. The main outcome was neurocognitive performance. Adjusting for demographics, hearing aid use, and cardiovascular disease, a 20-dB increase (one-category worsening) in HL was cross-sectionally associated with worse performance in multiple neurocognitive measures: -1.53 (95% CI = -2.11, -0.94) raw score point difference on Digit Symbol Substitution Test, -0.86 (-1.23, -0.49) on Word Frequency Test, -0.76 (-1.04, -0.47) on Spanish-English Verbal Learning Test (SEVLT) 3 trials, -0.45 (-0.60, -0.29) on SELVT recall, -0.07 (-0.12, -0.02) on Six-Item Screener. Because HL is common and potentially treatable, it should be investigated as a modifiable risk factor for neurocognitive decline/dementia.

Inclusion of African Americans is uncommon (11) and inclusion of Hispanics is rare (3). A recent report on hearing health care from the U.S. National Academy of Sciences recommended studying diverse populations, which would allow generalizability (6). Second, many studies have been limited in sample size, which reduces the ability for sub-analyses. Third, studies have typically controlled for only hearing aid use as well as demographic and cardiovascular covariates (10). Fourth, many prior studies have used self-reported/ non-audiometric (3,12) or central auditory tests (13)(14)(15)(16), neither of which are the clinical standard for behavioral measurement of hearing thresholds (17,18).
We previously reported an association between observed (nonaudiometric) HL and incident dementia in an urban multiethnic elderly cohort (3). In the current study, we examine cross-sectional associations between audiometrically defined HL and cognitive performance in tests of psychomotor speed, attention, verbal functioning and memory, learning, and global cognition in more than 5,000 middle and late-age adults enrolled in the Hispanic Community Health Study (HCHS). We hypothesize that dose-dependent associations will exist, which will be robust to a variety of potential confounders. To the best of our knowledge, this is the largest study of formal, audiometric HL and cognition to date (10).

Cohort and Participants
The Hispanic Community Health Study/Study of Latinos (HCHS) is a longitudinal, multicenter community-based cohort study. It incorporates interviews, physical examinations, and specialized testing such as neuropsychological evaluations. All participants undergo audiometry. Assessments were conducted in English or Spanish, as preferred by the subject. Only the 2008-2011 wave of data is currently available; thus, cross-sectional analysis was performed.
Inclusion is illustrated in Figure 1. There were 14,155 total participants in the original data set. To restrict the analysis to people at risk for age-related HL, subjects under 50 years of age (n = 7,980) and those with self-reported early-onset HL (n = 212) were excluded.
Participants who were missing data for audiometry, any of the five primary outcomes, or covariate data were further excluded. In the final analysis 5,277 subjects remained.

Hearing
Hearing, measured using pure tone audiometry, was the exposure of interest. Hearing thresholds (in dB HL, decibel hearing level) were collected from frequencies 500-8,000 Hz. The four-frequency pure tone average was computed as the mean threshold (dB) at 500, 1000, 2000, and 4000 Hz. Hearing level was based on the pure tone average in the better hearing ear, consistent with prior epidemiologic studies of age-related HL (1,11). Hearing was primarily analyzed continuously. The severity of HL was categorized as follows: normal hearing (0-25 dB), mild (26-40 dB), moderate (41-55 dB), moderately-severe-or-worse (≥56 dB) (19).

Outcomes
We examined as outcomes the five core cognitive measures collected in the HCHS. These measures included tests of global cognitive function, verbal memory, and frontal-executive abilities. For all tests, higher scores indicate better cognitive performance. All tests were analyzed continuously.
Attention and speed were measured with the Digit Symbol Substitution Test (DSST) of the Wechsler Adult Intelligence Scale-Revised and the Word Frequency Test (letter fluency test). The DSST measures psychomotor speed and attention. Subjects are asked to fill in a series of symbols corresponding to specific digits within 90 seconds (20). It is a widely used measured of cognitive function in epidemiologic studies (21). The observed score range was 0 (worst) to 80 (best). The Word Frequency Test is a measure of verbal fluency. Subjects are asked to state words that begin with certain letters as quickly as possible during two 60-second trials (22). The observed score range was 0 (worst) to 49 (best).
Verbal learning and memory were measured with the Spanish-English Verbal Learning Test (SEVLT). The subject is read a list of common words over three separate trials and asked to recall the list after each learning trial (23). The sum of the number of words across the three trials is recorded (SEVLT 3 trials), with a possible score ranging from 0 (worst) to 45 (best). A distracting word list is then presented and repeated back. Immediately after, the subject is asked to recall the initial word list (SEVLT recall), with a possible score ranging from 0 (worst) to 15 (best).
Global cognitive function was measured by the Six-Item Screener (24). This is a brief, six-item screen for cognitive impairment. The score ranges from 0 (worst) to 6 (best).

Covariates
Covariates that might confound the relationship between HL and cognition were included in the multivariable model. Demographics covariates included age, gender, and years of education. Gender was assessed by the examiner asking the participant what his/ her gender was. Hearing aid use was coded as yes/no depending on whether subjects self-reported any use in the past 12 months. Cardiovascular disease was treated as a confounder because it may cause both HL and cognitive impairment. A composite cardiovascular disease variable aggregating multiple risk factors was created to avoid multicollinearity. Multicollinearity, which occurs if related individual predictors are independently included in a model, can result in overadjustment (25) and erratic model behavior (26). A point was assigned for each of three risk factors that were present: coronary artery disease, hypertension, and/or self-reported stroke/ transient ischemic attack. Additional points were added for diabetes based on fasting serum glucose and, if available, oral glucose tolerance test, and HbA1C (1 point for impaired glucose tolerance, 2 points of diabetes) (27). Depressive symptoms, which could act as both a confounder and/or a mediator (28), were assessed with the Center for Epidemiologic Studies Depression Scale-10 (CESD-10) (29). Anxiety was assessed with the State-Trait Anxiety Inventory-10 (30,31). Cognitive testing was performed in English or Spanish based on the participant's preference. Bilingualism was assessed by whether the subject spoke English, Spanish, or both at home.

Statistical Analysis
One-way analysis of variance was used to compare means of continuous variables of interest between subjects across four levels of hearing categories (normal, mild, moderate, moderately-severe-orworse). The dependent variable was the variable of interest (age, education, etc.). The chi-squared and Fisher's Exact Tests were used to compare categorical variables across hearing categories.
Regression was used to examine associations between hearing performance examined continuously and the cognitive outcomes. Univariable regression was performed first, followed by multivariable regression adjusting for possible confounders. Linear regression was used as outcomes were continuous (DSST, SEVLT, and word frequency, Six-Item Screener). We additionally looked for interaction between hearing × gender and hearing × age To allow comparisons across different cognitive tests, z-scores were computed. Ordinal logistic regression (proportional odds model) was used for the Six-Item Screener in a sensitivity analysis where it was treated as an ordinal categorical variable (since the score only ranged from 0 to 6). In another sensitivity analysis, a quadratic term for age was added to the main multivariable models to attempt to better control for the confounding effect of age.
For the interpretation of models, we report HL changes in 20-dB increments. This is because 20 dB is approximately equivalent to a 1-category worsening (i.e. across standard clinical categories of normal, mild, moderate, moderately-severe, severe, and profound). (Some studies in the literature have used 10-dB increments. Because we used linear regression, our effect sizes and confidence intervals for 20-dB increments can be converted to those of 10-dB increments simply by dividing by 2.) We estimated the approximate years of aging corresponding to the cognitive score reduction seen with a 20-dB increase in HL. This was obtained by dividing the cognitive score reduction per 20-dB HL in the fully adjusted model (

Baseline Characteristics
Baseline respondent characteristics, categorized by hearing level, are reported in Table 2. The mean age was 58.4 years (SD = 6.2). Of 5,277 subjects, 2,023 were men (38.3%). Across hearing categories, there was a significant difference for age, gender, education, hearing aid use, cardiovascular disease score, and all cognitive tests.

Univariable Analysis
Scatterplots showing the unadjusted relationships between hearing (pure tone average, dB HL) and the cognitive outcomes appears in Figure 2. There was a significant inverse relationship between HL and all outcomes (p <.001, r s range −0.21 to −0.10). The corresponding regressions for these univariable analyses are discussed below. The univariable relationships between demographic variables (age, gender, education) and cognitive outcomes appear in Supplementary Table 1.

Multivariable Analysis
The summary for linear regression models appears in Table 1. In the univariable model (model 1), HL was significantly associated with a lower score on all five cognitive outcomes. Covariates were added to the models to adjust for potential confounding. Adding demographic factors (model 2; age, gender, education) attenuated the relationship between hearing and all cognitive outcomes. Adding hearing aids (model 3) had a small effect. Finally, adding cardiovascular risk factors (model 4) also had a small effect. Scatterplots of this fully adjusted model appear in Figure 2. For all models, p < .001.
A 20-dB increase in HL (controlling for confounders) was equivalent to an aging effect (i.e. the model-predicted pure effect of aging) of 2.2 years for the DSST, 9.8 years for the Word Frequency Test, 3.9 years for the SEVLT 3 trials, 4.8 years for the SEVLT recall, and 0.7 years for the Six-Item Screener.
We additionally computed z-scores to compare the association across cognitive outcomes. The strength of association between hearing and cognition showed similar results across outcomes.
(Supplementary Table 2) The fully adjusted model was then used to calculate the difference in cognitive outcomes for each HL category (the median withincategory value was chosen) compared to normal (0 dB) hearing ( Finally, we looked for effect modification (interaction) by gender and age. Significant effect modification was noted in only a few instances, including the DSST (hearing × gender, interaction term coefficient p < .01 and hearing × age, p < .05) and the Six-Item Screener (hearing × gender, p < .05). In the DSST, scores differed by −1.  Table 4.
In a sensitivity analysis, depressive symptoms assessed by the CESD-10 was added as a covariate because depression may be either a confounder or a mediator of the relationship between hearing and cognition (28,(32)(33)(34). When added to the fully adjusted model, the relationship between hearing and all cognitive outcomes was modestly attenuated but remained significant. (Supplementary Table 3) Similar findings were noted if anxiety assessed by the State-Trait Anxiety Inventory was added as a covariate. (Supplementary Table3) We also added preferred language during cognitive testing as well as bilingualism as covariates. When added to the fully adjusted model, the relationship between hearing and all cognitive outcomes was minimally attenuated. (Supplementary Table 3) The fully adjusted model was also restricted to only subjects with no or mild HL or, separately, to those who did not use hearing aids. In both scenarios, the hearing-cognitive outcome relationships were not attenuated. (Supplementary Table 3).   Notes: Linear regression was used for all cognitive outcomes. Results are for the fully adjusted model, which adjusts for demographics, hearing aid use, and cardiovascular risk factors. HL, hearing loss; SEVLT = Spanish-English Verbal Learning Test. *A 20-dB increase in HL is approximately equivalent to a 1-category worsening (categories are: normal, mild, moderate, moderately-severe, severe, profound).  Adding a quadratic term for age to the fully adjusted linear models to attempt to better control for the confounding effect of age did not change the relationship between hearing and cognition (Supplementary Table 3). In addition, it did not appreciably decrease the Akaike information criteria, indicating no improvement in model fit (<0.02% change for all models).
We also used ordinal logistic regression to analyze the Six-Item Screener since the score only ranges from 0 to 6. This modeling technique provides the odds of a lower score category (i.e. lower whole number score) versus a higher score category. In the fully adjusted model (model 4), the odds being in a lower score category increased 1.14 times (95% CI, 1.02, 1.27; p < .05) for every 20-dB increase in HL.

Discussion
We found that age-related HL, defined audiometrically, was associated with lower cognitive performance in all tests and domains, including tests that can be taken independent of hearing ability. A dose-response relationship was observed. Furthermore, this association was present even at mild levels of hearing impairment.
We found robust and significant associations between HL and all five cognitive outcomes, spanning domains of psychomotor speed, attention, verbal functioning, learning, verbal memory, and global cognition. In the fully adjusted models, HL was associated with decreased scores in the DSST, Word Frequency Test, SEVLT3 trials, SEVLT recall, and Six-Item Screener. The strength of association between hearing and cognition was similar and statistically equivalent across these outcomes. These models controlled for a wide variety of confounders including demographics (age, gender, and education), hearing aid use, and cardiovascular risk factors.
Our data extend findings of prior primarily educated/white (7-10) cohorts to Hispanics, the fastest growing ethnicity in the United States. Although it is difficult to compare studies because of heterogeneity of outcomes and covariates, the hearing-cognition relationship apparently holds across numerous racial/ethnic groups (3,11). Interestingly, controlling for bilingualism or testing language had no effect.
There is growing medical and policy interest in the association between age-related HL and cognition (1)(2)(3). This stems from the high burden of untreatable dementia (4), high prevalence of HL (80% of those over 80 years (35)), and low prevalence of hearing aid use (under 25% of those over 80 years (36)).
The observed association between HL and cognition is clinically meaningful. To put the findings in context, the reduction in cognition associated with a modest 1-category hearing drop in the fully adjusted models was equivalent to an additional 0.7 (Six-Item Screener) to 10 (Word Frequency Test) years of aging.
Even with mild HL, there was a measurable difference in cognitive outcomes compared to normal hearing individuals. This finding is meaningful because mild HL is very common and begins around age 50 (35). Large differences were noted for the most severely hearing-impaired category.
We examined whether the relationship between HL and cognition varied by gender or by age. Despite a few instances of significant interaction, there was no meaningful pattern observed.
The association between hearing and cognition remained significant after adjusting for depressive symptoms. Depression could act as a confounder, causing poorer hearing test and cognitive test performance due to lack of motivation (i.e. measurement error of both the exposure and outcome.) Depression could also act as a mediator, whereby HL causes depression, which, through additional steps such as social withdrawal, results in poorer cognition. Although there is inadequate evidence to state that HL causes depression, they have been associated in cross-sectional and longitudinal studies (4,28,(32)(33)(34). Prior studies of audiometric (9,37) or non-audiometric (38) HL have rarely taken comorbid depressive symptoms into account.
Cognitive dysfunction has many different risk factors. As a result, only a small fraction of the overall variability in cognition is explained by HL. However, any relationship between HL and cognition is noteworthy because of the high prevalence of HL (80% in people over 80 years) and poor treatment levels (under 20% use hearing aids or cochlear implants). Thus, a relatively high proportion of cognitive dysfunction could be attributable to HL. A recent review predicted that preventing or treating HL could result in the single largest percentage reduction in dementia of any known risk factor (4).
A methodologic concern is that HL could result in lower performance on cognitive tests because subjects could not hear the instructions or the oral presentation of test stimuli. In this scenario, hearing-impaired subjects would be misclassified as having worse cognition, creating a false-association. This seems unlikely because neuropsychological testing in HCHS was conducted in a quiet, welllit setting with a single trained examiner (39). In this environment, HL should not interfere with clearly spoken instructions except when profound. Fewer than 0.1% of subjects had profound HL and fewer than 0.4% had severe or profound HL. Furthermore, we found a strong inverse association between HL and cognition even when we restricted the analysis to subjects with normal hearing or mild HL, or to only subjects who did not use hearing aids. Finally, strong associations were seen in the DSST, which is a non-verbal test. Although initial instructions are spoken, the subject repeats non-scored trials until he/she clearly understands the task. After this point, the entire test is performed in silence with pen and pencil.
This study had limitations. Our analyses were cross-sectional and limited to the only available wave of HCHS. Thus, our inferences about temporality and causality are limited at this time. Although findings remained significant after controlling for numerous confounders, residual confounding remains a possibility. For example, as-yet unidentified factors could cause both HL and impaired cognition. We used linear regression, which is easy to interpret and apply to clinical scenarios. Future studies should also explore non-linear regression techniques.
However, this study has important strengths. With more than 5,000 included subjects this is the largest study to date examining the association of formally tested audiometric hearing and cognition. The large sample size provided power to examine for interaction. We also performed unique sensitivity analyses, including adding bilingualism or testing language as a covariate, adding depressive symptoms as a covariate, restricting to subjects with no more than mild HL, and restricting to subjects without hearing aids. These sensitivity analyses had no or little effect on the hearing-cognition relationship.
Finally, prior studies, including some with larger sample sizes (16), have relied on self-reported (3,12) or central auditory tests (14)(15)(16) instead of audiometrically measured peripheral hearing tests. Self-reported HL can be subject to bias (17) since it does not always correlate to true, audiometric hearing levels. Central auditory tests, while often affected by peripheral auditory function, do not ideally measure age-related changes of the peripheral auditory system. Neurodegeneration underlying cognitive impairment and dementia may cause central auditory dysfunction, but should not cause peripheral age-related HL (40,41).