Study population
The CHARLS, a nationally representative survey of China’s middle-aged and elderly population, provides a high-quality public micro-database with social, economic, and health information. Samples were selected by using multistage probability sampling taking into consideration regional and socioeconomic disparities. The probability-proportional-to-size (PPS) sampling technique was used to select 150 county-level units from all counties of China except Tibet. PPS sampling was also applied to select three primary sampling units (PSUs) from each county-level unit. PSUs represent the lowest level of government organization and consist of administrative villages (cun) in rural areas and neighborhoods (shequ or juweihui) in urban areas. Eighty or more households with age-eligible members were selected within each PSU, and one age-eligible member was randomly selected from qualified households. If the chosen person was willing to participate, this person and his or her spouse were interviewed. All stages of sampling were conducted by a computer to avoid potential biases arising from human manipulation.
CHARLS also has good cross-study comparability of results because the survey instrument was developed on the basis of the best international practices and harmonized with over 25 leading international research studies in the Health and Retirement Study model. [24]
The data used in our study were obtained from the CHARLS dataset collected in 2015, which included a total of 20967 individuals. Information on demographic characteristics, health-related behaviors and lifestyles, and health status were collected through face-to-face interviews using the questionnaire. Most of the interviewers were recruited from local colleges and universities and had received 9 days of rigorous training on the questionnaire content, interview techniques, security and quality control, and face-to-face interview practices. The surveyed data were recorded by using a computer-assisted personal interview (CAPI) system, which could help substantially improve the quality of the surveyed dataset by detecting errors in data inputs. When the interviewer enters an input item with a logical error or abnormal value, the system will pop out a message to alert the interviewer to this dubious entry. In this way, the participants no longer need to read the questionnaire, and the interviewers would ask them face to face and help them understand the questions and corresponding options. Anthropometric and physical measurements were provided, and blood samples were collected by trained nurses from township hospitals or China’s Center for Disease Prevention and Control (CDC) according to the standard protocol.
The exclusion criteria of the study were as follows: (1) demographic data were not recorded; (2) aged <45 years old; and (3) creatinine or hearing status data were not recorded. A total of 12508 participants were included in our final analysis. The participant selection process is shown in Figure 1.
Hearing loss
In our study, hearing loss was identified through self-reporting. Objective measurement, such as audiometry, was not provided in CHARLS, but several studies have demonstrated the reliability of self-reported hearing loss [25, 26]. We used hearing-related questions in the CHARLS survey to determine whether a patient had hearing loss. The hearing-related CHARLS survey questions are listed in Table 1. Participants were asked these questions by trained investigators through face-to-face interviews. A participant was defined as having hearing loss if he or she met one of the following three criteria: 1) had a hearing problem; 2) wore a hearing aid; and 3) had a poor hearing status.
Kidney function decline
The estimated glomerular filtration rate (eGFR) is considered the best overall index of kidney function in health and disease [27]. In this study, we determined eGFR using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation [28] as follows:
[Please see the supplementary files section to view the equation.] (1)
where Scr represents serum creatinine measured in units of mg/dL. κ is 0.7 for females and 0.9 for males, α is −0.329 for females and −0.411 for males, min refers to the minimum value of Scr/κ and 1, and max refers to the maximum value of Scr/κ and 1.
At the beginning of this study, we attempted to classify the eGFR into five groups: ≥90, 60–89, 30–59, 15–29, and <15 mL/min/1.73 m2, according to the different stages of CKD defined by the Kidney Disease: Improving Global Outcomes (KDIGO) guideline [29]. We calculated the numbers and percentages of participants in different eGFR groups, and found that only 74 (0.59%) participants with eGFR of 15–29 mL/min/1.73 m2 and 8 (0.06%) participants with eGFR <15 mL/min/1.73 m2, which may lead to limited statistical power to examine the association between eGFR and hearing loss. This is a study based on general population, and thus there are less people in lower eGFR groups. Therefore, we finally used only three eGFR groups of ≥90, 60–89, and <60 mL/min/1.73 m2) for analysis.
Other predictor variables
We chose predictor variables by referring to hearing loss risk factors reported in the existent literatures [5, 30] and their availabilities in the CHARLS dataset. Besides eGFR, predictor variables used in the regression models included demographic characteristics (i.e., age, gender, education, area of residence), health-related behaviors (i.e., smoking and drinking status) and cardiovascular risk factors (i.e., body mass index [BMI], central obesity, hypertension, diabetes, stroke, high-density lipoprotein [HDL] cholesterol, and low-density lipoprotein [LDL] cholesterol). Demographics and health-related behaviors data were obtained from the questionnaire. BMI was defined as weight in kilograms divided by the square of height in meters, and acquired through physical measurements. Participants were categorized as underweight (<18.5 kg/m2), normal weight (18.5 to 24.9 kg/m2), overweight (25.0 to 29.9 kg/m2), and obese (≥30 kg/m2) according to their BMI. In our study, central obesity was defined as a waist circumference of ≥80 cm for females and ≥102 cm for males. Hypertension was defined as mean systolic blood pressure (SBP) ≥140 mmHg, mean diastolic blood pressure (DBP) ≥90 mmHg, or a self-report of hypertension. Diabetes was defined as fasting plasma glucose ≥126 mg/dL, HbA1c concentration ≥6.5%, or a self-reported doctor diagnosis. Stroke was defined as a self-reported history of doctor diagnosis. HDL and LDL cholesterol levels were directly obtained from the results of laboratory tests.
Statistical analysis
Descriptive statistics for continuous variables are presented by using means and standard deviations (SD), while frequencies and percentages are used to describe categorical variable characteristics. Student’s t-test was used to compare the mean values of continuous variables between participants with and without hearing loss, and differences in hearing loss prevalence among different categorical variable groups were tested by using Pearson’s test. The association between reduced eGFR and hearing loss was modeled by using a logistic regression function, and odds ratios (ORs) with 95% confidence intervals (CIs) of hearing loss for different eGFR categories were calculated. Multivariable logistic regression models were constructed to adjust for potential confounding variables (i.e., age [45-54 years, 55-64 years, ≥65 years], gender [male, female], education [illiterate, literate, primary school, middle school, high school and above], area of residence [urban or rural], smoking [never, current, past], drinking [never, current, past], BMI [underweight, normal weight, overweight, obese], central obesity [yes or no], hypertension [yes or no], diabetes [yes or no], stroke [yes or no], HDL cholesterol [continuous value], and LDL cholesterol [continuous value]). To build the multivariate logistic regression models, we used a separate unknown category to represent missing data for BMI, central obesity, and stroke, which had missing rates greater than 1%. Other variables included in the models that had missing rates lower than 1% were not preprocessed as cases containing missing data were automatically deleted during logistic regression analysis.
All analyses were performed by using STATA software (version 14.0). All p values were based on two-sided tests with a significance level of 0.05.