The NHANES, conducted by the National Center for Health Statistics (NCHS) since the early 1960s, is an ongoing series of cross-sectional surveys designed to assess health and nutritional status in the civilian non-institutionalised US population.

In NHANES 1999–2004, half of the subjects aged 20–69 years were randomly selected to participate in the Audiometry Examination Component. Subjects were excluded if they used hearing aids that were not able to be removed for testing or had sufficient ear pain at the time of the examination that they could not tolerate headphones.22 The eligible sample size was 5742 participants: 1807 in 1999–2000, 2046 in 2001–2002 and 1889 in 2003–2004.

Audiometry examination was performed in a sound-isolated room in the mobile examination centre by health technicians trained by a National Institute for Occupational Safety and Health-certified audiologist. Instrumentation for the Audiometry Component included an Interacoustics Model AD226 audiometer with standard TDH-39 headphones and EtymoticEar Tone 3A insert earphones.22 Pure-tone air conduction hearing thresholds were obtained on both ears at frequencies of 0.5, 1, 2, 3, 4, 6 and 8 kHz across an intensity range of −10 to 120 dB. Participants with missing value at one or more frequencies were excluded from analysis. Procedural details in collecting the audiometric data have been described elsewhere.22 23

We computed hearing thresholds (decibels) at speech frequencies as a pure-tone average (PTA) of 0.5, 1, 2 and 4 kHz and the hearing thresholds at high frequencies as a pure-tone average (high PTA) of 3, 4 and 6 kHz.24 According to classification by the WHO, hearing loss is defined as PTA greater than 25 dB in either ear.24 25 Noise notch was defined as the hearing threshold at 3, 4 and/or 6 kHz at least 10 dB greater than at 1 or 2 kHz and at least 10 dB greater than at 6 or 8 kHz.11 12

Of the initial sample of 5742 participants eligible for inclusion in the audiometry examination, 324 (6%) were excluded from the analysis because a test was not performed at all and 152 (3%) were excluded because tests have missing values at one or more frequencies. As an additional measure of reliability of participant responses, all audiograms tested the 1 kHz frequency twice in each ear.22 Three participants were excluded because there was >10 dB difference between the 1 kHz test–retest thresholds.24 An additional 452 participants were excluded because of unilateral hearing loss defined as >10 dB difference between the PTAs of left and right ears. Therefore, audiometric results for 4811 participants were included for analysis in the present study.

Participant occupation information was obtained from the occupation questionnaire, which contains personal interview data on employment and variables relating to the work environment, such as workplace noise exposure history.26 We used the participant's longest job as a measure of past noise exposure. The longest job information was obtained by asking the question directly: ‘What kind of work were you doing the longest?’.27 Occupation data were then coded by trained coders using the US Census Bureau's Census 1990 or 2000 indexes of occupation, which are the three-digit NCHS Occupational Classification Source Codes.28 In the publicly available NHANES data, these codes were collapsed into 41 occupation categories.27 Of the 4811 participants with available audiometric measurements, 4547 had information available on their longest job—2498 participants who answered their longest job plus an additional 2049 participants who answered their longest job were the same as their current job and for whom the current job code was assigned as their longest job. To link the O*NET noise estimates to the 41 occupation categories, we grouped the 801 O*NET SOCs available into the corresponding NHANES occupation categories and computed the averages of the O*NET noise scores in each category. Because military occupations were not included in the O*NET survey, we could not generate an O*NET noise score for the military occupation group and so resulted in 40 occupation groups for analysis. After linking O*NET noise estimates to occupation categories, we assigned an average noise score to each participant. Of the 4547 participants with available longest job information, we excluded 66 participants whose longest jobs were not coded to any of the 41 occupation groups and 59 participants who had military occupation, yielding 4422 participants available for statistical analyses.

We also computed the weighted averages of O*NET noise scores for the 40 NHANES occupation groups, accounting for the SEs of the noise scores in each O*NET SOC group and which reflect the precision of the O*NET survey data (see supplemental table 1). Because SEs were not available for 58 of the 801 SOCs, we used the unweighted averages as our primary index of occupational noise and examined the weighted ones as a sensitivity analysis. The unweighted and weighted average O*NET scores for 40 occupation categories are presented in table 1.

The characteristics between participants with and without the longest job information were similar in terms of prevalence of hearing loss, age, body mass index (BMI) and the status of hypertension and diabetes. Compared to included participants, excluded participants were less likely to be male, non-Hispanic white and smokers, less educated, and less exposed to occupational noise (see supplemental table 2).

Other demographic and hearing-related variables were obtained from the NHANES questionnaires. BMI was calculated as weight in kilograms/height in metres squared (missing for 49 participants). Use of ototoxic medication was counted when participants reported medications of aminoglycoside, loop diuretics, antineoplastic drugs or non-steroidal anti-inflammatory drugs (missing for five participants). Smoking pack-years were computed and grouped into non-smokers, smokers who smoked <20 pack-years or smokers who smoked >20 pack-years (missing for 392 participants). Hypertension was defined as self-reported physician diagnosis, use of antihypertensive medication, systolic blood pressure ≥140 mm Hg or diastolic blood pressure ≥90 mm Hg at the time of examination (missing for 189 participants). Diabetes mellitus was defined as those who self-reported either previous physician diagnosis or use of antihyperglycemic medication (missing for two participants). Non-occupational noise exposures were determined by audiometry questionnaires asking if the subject had ever been exposed outside of work to the noise of a firearm for a mean of at least once a month for 1 year (missing for five participants) and if the subject had ever been exposed outside of work to loud noise (eg, power tools or loud music) for a mean of at least once a month for 1 year (missing for six participants). Socioeconomic status, such as education attainment, may be a potential confounder, but we did not consider it because education is highly correlated with job title used for our O*NET noise estimates, and controlling for such a socioeconomic status may result in over adjustment for our exposure effect.

Our study sample was limited to adults who had complete information on these important covariates, and therefore, a total of 3828 participants were available for data analyses.

All statistical analyses were performed using SAS survey procedures (SAS V.9.2) and the Rsurvey package (R V.2.9.1) to account for the complex survey design of the NHANES.1 29 Sample weights for the combined 6-year sample were used as per NCHS recommendations in order to provide annual national estimates, which accounted for the unequal probabilities of selection due to over sampling and non-response.24

Linear regressions were used for continuous hearing thresholds in each frequency and PTA. Hearing threshold outcomes were log-transformed to normalise distributions. Eighty subjects (2.1%) with better-than-normal hearing had zero or negative hearing thresholds. We excluded these subjects to better interpret regression results of log-transformed thresholds in our primary linear regression analyses. Linear regressions including all available subjects were considered in sensitivity analyses to evaluate the effect of the exclusion. In sensitivity analyses, a constant 6 was added before log-transformation to make all hearing threshold values positive and the resulting data as normal as possible. We examined the O*NET noise score as a continuous variable and in quintiles. For the latter, we tested for linear trend across quintiles using ordinal terms. For dichotomous hearing loss and noise notch outcomes, we determined the OR and 95% CI using logistic regression models. In multiple regression analyses, we identified a priori those covariates that needed to be controlled for, based on biological consideration and the current state of the literature: age (years), sex, race/ethnicity, BMI (kilograms per square metre), cigarette smoking (pack-years), ototoxic medication, hypertension, diabetes mellitus, firearm noise exposure and recreation noise exposure.30 We fit age and age-squared to capture non-linear effects of age. To identify influence of potential confounders, we developed sequential models: (1) age, sex and race/ethnicity adjusted; (2) additionally adjusted for BMI, ototoxic medication, cigarette smoking, current diagnosis of hypertension and current diagnosis of diabetes and (3) additionally adjusted for recreational noise and firearm noise exposures. We examined subgroup analyses by age, sex and race/ethnicity to identify the most susceptible subpopulations.

Table 2 shows descriptive characteristics of study participants. Overall, 456 subjects (11.9%) had a mild or greater hearing loss. After accounting for sampling weights, cluster and strata of the NHANES complex design, the mean of O*NET noise scores in the entire population was 3.06 (SE=0.02), and subjects with hearing loss had a significantly higher noise score than those without hearing loss (3.26 (SE=0.04) vs 3.04 (SE=0.02)). The means of O*NET scores in the entire population and for subjects with and without hearing loss which were not considered sampling weights were 3.09 (SD=0.60), 3.26 (SD=0.65) and 3.07 (SD=0.59), respectively. Distributions of the O*NET noise scores in the NHANES participants are shown in supplemental figure 1. Subjects with hearing loss were older (54.8 vs 40.4 years), more likely to be male (66% vs 46%), non-Hispanic white (80% vs 71%) and ever smoker (59% vs 46%), used ototoxic medication (24% vs 15%), less educated (percentage of greater than high school diploma 43% vs 60%), more likely to be exposed to occupational noise (45% vs 32%) and firearm noise (13% vs 7%), and more likely to have hypertension (43% vs 21%) and diabetes (12% vs 3%) than those without hearing loss.

Table 3 presents the distributions of participant characteristics by quintiles of the O*NET noise scores. Subjects with higher occupational noise were more likely to be male and of race/ethnicity other than non-Hispanic white, more likely to have higher hearing thresholds, less educated and more likely to be exposed to firearm and recreational noises than those with lower occupational noise.

Table 4 shows per cent changes in PTA in associations with the O*NET score as a continuous variable and in quintiles in various covariate-adjusted models. An increase in occupational noise exposure either as a continuous variable or in quintiles was significantly associated with higher (poorer) hearing thresholds. In the fully adjusted model (model C), subjects in the highest noise quintile had 22.5% (95% CI 11.0% to 35.2%) higher hearing thresholds than those in the lowest quintile. A unit increase in the O*NET noise score was associated with a 15.4% (95% CI 9.7% to 21.5%) increase in hearing thresholds in the fully adjusted model. The same trends were also observed in high PTA (see supplemental table 3) and all individual frequencies (see supplemental figure 2). Table 4 was designed to show results from the linear regression of log-transformed PTA with the O*NET score, which excludes the subjects with zero and negative hearing thresholds (2.09%) for better interpretation of log-transformation. Supplemental table 4, panel A versus B., compares results in subjects with only positive hearing thresholds with results in all available subjects. Panel A was designed to show results from the linear regression of log-transformed (PTA+6) with the O*NET score in all available subjects, whereas panel B was designed to show results from the linear regression of log-transformed (PTA+6) with the O*NET score in subjects with only positive hearing thresholds, the same subjects as table 4. From the comparison of supplemental table 4, panel A versus B, linear regression in subjects with only positive hearing thresholds showed consistent patterns to those of linear regression in all available subjects.

Table 5, panel A, shows the association between O*NET noise score and the risk of hearing loss in different covariate-adjusted models. There were significant dose-dependent relationships in all models, and further adjustment for potential confounders including ototoxic medication, cigarette smoking, hypertension, diabetes, recreation noise and firearm noise did not change the result. The fully adjusted OR for hearing loss comparing the highest versus the lowest quintiles was 2.07 (95% CI 1.18 to 3.63).

Table 5, panel B, shows OR for risk of the noise notch by quintiles of O*NET noise score. The association showed a statistically significant dose-dependent relationship, and the increase pattern in the risk of noise notch in association with O*NET noise score explains better the dose-dependent relationship between occupational noise and hearing loss than that in the OR of the risk of hearing loss by the O*NET noise score in panel A. The fully adjusted OR for noise notch comparing the highest versus the lowest quintiles was 1.51 (95% CI 1.09 to 2.09).

As a sensitivity analysis, we also examined associations with the weighted O*NET noise scores. Overall, associations were similar to those with unweighted scores (supplemental tables 5 and 6).

We examined the association of the O*NET noise score, modelled as a continuous variable, with PTA in the subgroups by important demographic characteristics such as sex, age and race/ethnicity. The association was strongest among the age group of 40–59 years, whereas it was less significant among the age group over 60 years. Men or non-Hispanic white also had higher O*NET noise-associated PTAs than did women or other race/ethnic groups (see supplemental table 7).