The present study took data from the Chinese Longitudinal Healthy Longevity Survey (CLHLS), which is an ongoing longitudinal study that began in 1998, along with seven follow-up surveys, with the new participants being added to replace the deceased who passed away in 2000, 2002, 2005, 2008, 2011/2012, 2014, and 2018 (Yang and Meng, 2020). The CLHLS recruited a representative sample from 23 of the 31 provinces of China and is the largest database on the oldest-old in the world, with the survey areas covering 85% of the Chinese population (Zeng, 2012). More details about the sampling procedure and quality of data of this survey have been published elsewhere (Zeng, 2012). Ethics approval was obtained from the Research Ethics Committees of Peking University (IRB00001052-13074). All participants or their legal representatives signed written consent forms in the baseline and follow-up surveys.

The CLHLS consisted of questions regarding self-reported hearing difficulties only in waves 2011/2012 and 2014. Thus, this study employed two waves from the CLHLS longitudinal data harvested during 2011/2012 to 2014. A total of 10,890 participants were enrolled, including 10,744 aged 65 years or older, with the latest follow-up conducted in 2018. Accordingly, 9,674 respondents were interviewed in 2011/2012 wave, and 1,070 newly enrolled respondents were interviewed in 2014 wave, respectively. Data were incomplete for 1,703 participants, and the amount of missing data (key variables) ranged from 54 to 1,005. The longest follow-up period lasted for up to 8 years, and 47.5% (n = 5,099) of the participants died during the follow-up period until 2018. Details of the sampling method and calculation of weights have been published previously (Dawes et al., 2015). Characteristics of the raw dataset were shown in Supplementary Figure 1.

Hearing sensations were assessed in terms of self-reported measures, and all enrolled participants were required to attend a series of standardized training sessions prior to interviews. Self-reported hearing status data were based on the responses to the question: “Do you have any difficulty with your hearing? “YES” was coded as having HI, while “No” signified not having HI (Gao et al., 2020).” A systematic review compared the results obtained with self-report to a hearing question with those obtained by pure tone audiometry. They found that older adults with HI can be recommended for an epidemiologic study if audiometric measurements cannot be performed (Valete-Rosalino and Rozenfeld, 2005). Therefore, self-reported hearing status is a suitable option for large epidemiological studies (Valete-Rosalino and Rozenfeld, 2005; Deepthi and Kasthuri, 2012; Diao et al., 2014).

All participants were assessed for cognitive function by utilizing the Chinese version of the Mini-Mental State Exam (MMSE), a widely used cognitive test (Christensen et al., 2013). The test was tailored to the Chinese language based on the international standard of MMSE questionnaire, and had been proven to be reliable and valid in previous studies (Yi and Vaupel, 2010; Yuan et al., 2019; Duan et al., 2020; Zhang et al., 2021). The Chinese version of MMSE evaluates cognitive function in terms of 24 items, covering 7 sub-scales: orientation (4 points for time orientation and 1 point for place orientation); naming foods (naming as many kinds of food as possible in 1 min, 7 points); registration of 3 words (3 points); attention and calculation (mentally subtracting 3 iteratively from 20, 5 points); copy a figure (1 point); recall (delayed recall of the 3 words mentioned above, 3 points); and speech and language (2 points for naming objectives, 1 point for repeating a sentence, and 3 points for listening and following directions). The MMSE score ranges from 0 to 30. The higher the score, the better the cognitive function. The individuals who scored 25 or higher were considered to have normal cognitive function, and the summary accuracy at a cutoff value of 25 (10 studies) was sensitivity 0.87 and specificity 0.82 (Creavin et al., 2016). Cognitive impairment in this study was determined by its presence or absence according to this classification (Creavin et al., 2016).

Information on mortality was collected on the basis of death certificates provided by the local authorities. When such information was not available, relatives of the decedents were interviewed. Duration of follow-up was the time interval from the first interview date until the date of death. Participants who were alive at the last interview were regarded as being censored on the dates of their last interviews in 2018. The cause-specific mortality was not involved in this study because (1) a lot of the older adults died at home rather than in medical institutions where cause of mortality might be recorded, and (2) mortality surveillance systems are unsure in many survey fields (Duan et al., 2020).

A variety of variables were collected through a face-to-face interview against a standardized questionnaire, including sociodemographic features, lifestyles, health conditions and available medical services that were potentially associated with HI and cognitive impairment as suggested by previous studies (Dawes et al., 2015; Loughrey et al., 2018; Lv et al., 2019). Therefore, in this study, we assessed a range of these potential confounders by including the covariates age (continuous), gender (male or female), education background (no schooling or primary school or higher), occupation before retirement (manual or non-manual), ethnicity (Han or minority), residence (urban or rural), marital status (currently married and living with spouse, separated/divorced/never married, or widowed), tobacco smoking status (never or ever), alcohol drinking status (never or ever), regular leisure activities (yes or no), activity of daily living (ADL) (don’t need help or need help), having been diagnosed with hypertension (yes or no), having been diagnosed with diabetes (yes or no), and baseline cognitive function (impaired or not impaired).

First, we used multiple imputation to impute missing data for our raw dataset. The multiple imputation is based on chained equations and is commonly used for longitudinal studies. As multiple imputation uses information on baseline demographics and previous time points to predict missing values, the strategy assumes that data are missing at random, that is, that the missingness is related to observed data. We used some demographics such as gender, age, place of residence, as predictors, to impute the missing values of key variables. We established 10 imputed datasets and carried out the pooled statistical inference.

For comparison, Chi-square test was used for categorical variables and the analysis of variance was employed for continuous variables. The Kaplan–Meier method was employed to plot the survival curves in terms of baseline hearing status and gender. Multivariable-adjusted hazard ratios (HRs) and 95% confidence intervals (CIs) of all-cause mortality by hearing status were calculated by using five dependent Cox proportional hazards models: Model 1: no variables adjusted; Model 2: additionally adjusted for gender, age, education background, residence, and marital status based on model 1; Model 3: additionally adjusted for smoking status, drinking status, regular leisure activities, and ADL based on model 2; Model 4: additionally adjusted for two kinds of diseases (hypertension and diabetes) based on model 3; and Model 5: additionally adjusted for baseline cognitive function based on model 4. Our test ascertained that the proportional hazard assumption was not been violated.

Next, in order to assess disparities across different populations, we conducted subgroup analyses in terms of baseline cognitive function (MMSE score ≥25 versus <25 points), age (65–79 versus ≥80 years), and gender (female versus male), respectively. In addition, we also examined whether the association of HI with mortality differed by baseline cognitive function and gender by separately adding an interaction term to the fully adjusted model. Moreover, we further examined the association of HI with all-cause mortality by seven cognitive subdomains among Chinese older adults.

In the end, to resolve the problem with the loss to follow-up, we performed a sensitivity analysis by removing incomplete cases. Additionally, we also conducted another sensitivity analysis by eliminating those who died within the half of the year after the baseline survey to account for the possibility that the pre-mortality dropped in hearing function and/or disease status could have influenced our results.

A two-tailed p-value of less than 0.05 was considered statistically significant. All analyses were performed by using R software package (version 4.1.1, the R Foundation for Statistical Computing, Vienna, Austria).

Table 1 presents the descriptive baseline characteristics in terms of different hearing states. A total of 10,744 enrolled participants aged 65 or older participated in the baseline survey in waves during 2011/2012 and 2014, and they were followed up for at least one wave. All participants were aged 86 years on average (range: 65–114 years) and more than half of them (55.5%) were female. Participants with HI more likely to be older, female, lower-educated, rural residents, widowed, separated, divorced or never married, no regular leisure activities, need help in ADL, having had hypertension and diabetes, and worse cognitive function. However, participants with HI were less likely to smoke and drink, but were more likely to have disease conditions (Table 1).

Kaplan–Meier survival curve revealed a conspicuous probability of survival relationship between hearing status and all-cause mortality for total population (log-rank test for trend: p < 0.001) (Figure 1). The median survival time of none-HI and HI participants were 4.7 and 2.7 years, respectively, among the total population.

Table 2 shows the multivariable-adjusted HR and 95% CI of all-cause mortality by hearing status among Chinese older adults aged ≥65 years. The five multivariate models displayed consistent patterns between HI and all-cause mortality, and the association attenuated with more covariates included in the models. The unadjusted model (Model 1) showed that HI was significantly associated with the all-cause mortality. Compared to participants without HI, those who reported HI during the follow-up period were 2.29 times more likely to have all-cause mortality (95% CI: 2.16, 2.42). After adjustment for gender, age, education background, residence, and marital status in Model 2, the association between HI and mortality diminished significantly but remained comparable, with HR = 1.28 (95% CI: 1.20, 1.36). Model 3 (HR = 1.14, 95% CI: 1.07, 1.21) showed that HR decreased from 1.28 to 1.14 (i.e., 14% hazard ratio reduction) after adjustment for smoking, drinking, regular leisure activities, and ADL based on Model 2. Model 4 was additionally adjusted for two kinds of diseases (hypertension and diabetes) based on Model 3, and the results were consistent with Model 3. Model 5 was additionally adjusted for baseline cognitive function based on model 4, and the results suggested that HI was associated with a 7% [HR = 1.07, (95% CI: 1.00, 1.14)] increase in the risk of all-cause mortality compared with individuals without HI.

Figure 2 shows the HR of all-cause mortality by HI for different subgroups, conducted as separate models for each subgroup with full adjustment as in Model 5. The relationship and the effect sizes between HI and all-cause mortality were consistent across all subgroups in terms of cognitive function, age and gender. To examine whether the association between HI and all-cause mortality is modified by different cognitive states, we further tested the interaction between hearing status and cognitive function. The results showed that there was a significant interaction between HI and cognitive impairment (p < 0.001). HI participants with cognitive impairment showed higher mortality risk (HR = 2.31, 95% CI: 2.13, 2.51) than other groups (Figure 3). Similar associations were observed in the interaction analysis based on hearing status and gender. We also found that males had a higher risk of mortality (HR = 1.55, 95% CI: 1.37, 1.76) than their female counterparts, when they had the same level of HI (Figure 3).

We examined the seven cognitive subdomains separately for the association between HI and all-cause mortality. Impairment in orientation, naming foods, registration, attention and calculation, copy figure, delayed recall, and speech and language was independently associated with elevated mortality risk in the participants with HD, with HR ranging from 1.28 (copy figure) to 1.46 (speech and language) (Table 3).

Among the cohorts, there was almost no change in the association between HI and all-cause mortality after excluding participants lost to follow-up or with survival time less than half of the year. The association was still robust after further adjustment for potential confounders (Supplementary Tables 1, 2).