Data from the Chinese Longitudinal Healthy Longevity Survey (CLHLS) [23] were used. The CLHLS is a nationwide population-based longitudinal survey conducted in a sample of randomly selected counties and cities in 23 of the 31 provinces in China. More details about the survey design and data quality are available elsewhere [17, 23]. We used data collected during 2011–2018, which included 3 waves: 2011–2012, 2014, and 2017–2018. The baseline wave (2011–2012) comprised 9,679 participants aged 65 years and over. Individuals without any follow-up after baseline (n = 791), with a negative follow-up time (n = 180) or a missing disability status (n = 499), or without complete data concerning smoking status, fruit/vegetable intake, physical examination, and stress status (n = 25) were excluded from analyses (accounting for a total of 15.45%). Thus, the final analytical sample included 8,184 participants. The CLHLS was approved by the research ethics committees of Duke University and Peking University (IRB00001052–13074).

Economic status, educational attainment were used to characterize SES. Economic status was measured by the following question: “how do you rate your economic status compared with that of other local people?” and categorized as high (very rich or rich), intermediate (average), or low (poor or very poor). Educational attainment was categorized into low (no schooling), intermediate (1–6 years), and high (7 or more years) by the number of schooling years.

Occupational position was another common indicator being used to assess SES. Occupational position was divided into high (government, institutional and managerial personnel; professional and technical personnel; and military personnel), intermediate (clerks, service industry employees, and manual workers), and low (farmers and the unemployed) according to individuals’ major occupation before age 60. However, about 90% of females in the low occupation and about 81% of males in the low occupation. Occupation may be not much variability in this indicator to pick up exposure to occupational hazards and the like which is what one wants to assess when linking occupation with disability. Therefore, we only displayed the results and discussed the results in discussion, not draw conclusions.

We divided health state into three states: disability-free, disability, and mortality. Disability was defined as needing assistance in at least one of the activities of daily living (ADL): bathing, dressing, going to the toilet, indoor transferring, continence and feeding. DFLE was defined as expected life years without disability.

Demographic characteristics and impact factors were assessed at every follow-up wave. Individual characteristics included gender (male/female), age, and region (urban/rural). Risk factors included smoking, inadequate fruit/vegetable intake, not undergoing physical examinations, and feeling stress. Smoking status was categorized as smoking and never smoked, with the former category including both currently and formerly smoking. Fruit/vegetable intake was measured with the following question: “Do you eat fresh fruits or vegetables?” It was categorized as daily/almost daily, occasionally, and rarely/no. Inadequate fruit/vegetable intake included “occasionally” and “rarely/no” intake categories. Not undergoing physical examinations was defined as participants not undergoing a regular physical examination once every year. Stress status was assessed with the following question: “Do you often feel fearful or anxious?” It was categorized as feeling stress (always or often) or no stress (sometimes, seldom, or never).

First, logistic regression models were used to estimate the association of SES with risk factors. A multistate Markov model (MSM) [24] was fitted to estimate the hazard ratios (HRs) for the association of health transitions with SES and risk factors. We fitted 6 MSMs that allowed 4 transitions: from disability-free to disability (disability incidence), from disability-free to death, from disability to disability-free (recovery from disability), and from disability to death. The first model (model 1) adjusted for age, gender, and region. Subsequently, smoking status, fruit/vegetable intake, physical examination status and stress status were entered into model 1 separately as time-dependent covariates (models 2–5) and then simultaneously into model 1 (model 6). In these models, sex-age-region-specific transition probabilities were estimated.

Then, to calculate DFLE₆₅and LE₆₅, we used microsimulation [25] to simulate a cohort of 100,000 persons at age 65. In this 65-year-old cohort, the distribution of region and gender was sourced from the 2010 census data of China [26], and the gender-region-specific distribution of SES and disability was based on the observed prevalence of the CLHLS’s 2011–2012 wave. From age 65 to death, the health and survival trajectories of each individual in this cohort were governed by the transition probabilities output from the MSM model. The 95% CIs (from the 2.5th and 97.5th percentiles) were estimated by bootstrapping with 1,000 independent replications.

Finally, we set the prevalence of the risk factors to zero in all the SES groups (elimination scenario). We compared the results of the elimination scenario to the results of the current situation to quantify the effect of the risk factors on the socioeconomic disparities in DFLE and LE.

We used economic status and educational attainment as measures of SES to conduct the above analysis separately. These analyses were mainly performed using R software (version 3.5.1) and SAS version 9.4.

The baseline characteristics of the participants are shown in Table 1. The sample consisted of 3667 males and 4557 females. Of the participants, 19.72%, 26.64%, 28.98% and 24.66% were ages 65–74, 75–84, 84–95, and ≥95 years, respectively. More than half of the participants had low educational attainment (58.82%), only 10.51% had high educational attainment; 15.82% and 17.36% reported low and high economic status, respectively. There were 4349 deaths during 2011–2018. More than a quarter of the older adults were disabled at baseline (26.56%).

Socioeconomic disparities in DFLE₆₅ and LE₆₅ are displayed in Table 2. A positive socioeconomic gradient was seen in LE₆₅ as well as DFLE₆₅ when using economic status and educational attainment as measures of SES. LE₆₅ between the high and low economic status groups differed by 2.20 (95% CI, 1.10–3.41) years for males and 2.04 (1.01–3.29) years for females. The DFLE₆₅ disparity in economic status was 1.51 (0.52–2.59) and 1.29 (0.19–2.52) years for males and females, respectively. Similarly, the LE₆₅ disparity between the high- and low-education groups was 2.28 (1.11–3.46) years for males and 1.79 (0.66–3.07) years for females. For DFLE₆₅, the education disparity was 1.88 (0.74–2.93) years for males and 1.32 (0.13–2.60) years for females. Overall, the proportion of DFLE₆₅ was lower in the high-SES group than in the low-SES group using economic status and educational attainment as markers of SES, although the difference in education was not statistically significant.

The prevalence of risk factors according to SES is presented in Figure 1. Economic status and educational attainment were negatively associated with the prevalence of inadequate fruit/vegetable intake (p < 0.05), whereas only economic status was negatively related to the prevalence of feeling stress (p < 0.001). Lower economic status and educational attainment were associated with higher prevalence of not undergoing physical examinations (p < 0.05). Overall, there was little socioeconomic difference in the prevalence of smoking.

Smoking was the risk factor for disability incidence (HR = 1.20, 95% CI = 1.01–1.43) (Table 3). For mortality from disability-free, not undergoing physical examination was the risk factor (HR = 1.25, 95% CI = 1.05–1.53). For mortality from disability, feeling stress, not undergoing physical examinations, the occasional and rare/no intake of fruit/vegetable were the risk factors, and the HRs (95% CI) were 1.27 (1.08–1.49), 1.11 (1.01–1.22), 1.15 (1.03–1.28), and 1.28 (1.16–1.43), respectively.

The association of transitions with SES and the risk factors are shown in Table 3. For mortality from disability-free, the HR (95% CI) for low economic status (high as reference) was 1.52 (1.06–2.17) in the model adjusted for age, gender, and region. After adjusting smoking, fruit/vegetable intake, physical examination, and stress, the HR was attenuated to 1.40 (0.97–2.01). Similarly, for mortality from disability, after adjusting these factors, the HR for low economic status decreased from 1.28 (1.10–1.49) to 1.16 (0.99–1.36), and the HR for intermediate economic status slightly decreased from 1.16 (1.03–1.32) to 1.13 (1.00–1.28). The association of economic status with disability incidence and recovery had no statistical significance (p > 0.05).

Low education was only associated with disability incidence (HR = 1.36, 95% CI = 1.04–1.78). After adjusting for the above factors, the HRs changed little.

The DFLE₆₅ and LE₆₅ disparities in economic status were substantially reduced by eliminating not undergoing physical examinations or inadequate fruit/vegetable intake and somewhat reduced by eliminating feeling stress, but these disparities changed little by eliminating smoking (Figure 2). For example, the DFLE₆₅ disparity between males with high and low economic status was reduced from 1.51 (0.52–2.59) years to 1.08 (0.04–2.34) years by eliminating inadequate fruit/vegetable intake. That is, 28.48% of the DFLE₆₅ disparity in economic status for males was attributed to inadequate fruit/vegetable intake. Inadequate fruit/vegetable intake, not undergoing physical examinations, and stress contributed in total to 35.10% of the DFLE₆₅ disparity in economic status for males and 57.36% for females. The contribution from these factors to the LE₆₅ disparity in economic status was 26.36% and 42.65% for males and females, respectively. However, the contribution of these risk factors to education disparities in DFLE₆₅ and LE₆₅ was small (approximately 10%) (Table 4).

Our study provided updated results on the socioeconomic disparities in DFLE and LE among older adults using different measures of SES in the 2010s in an LMIC. Moreover, the SES and risk factors may consequently change with time. Therefore, we used repeated measures of SES and risk factors as time-dependent variables to estimate the effect of risk factors on this disparity. A study has suggested that the effect of risk factors will be underestimated if only risk factors are assessed at the first follow-up [33].

However, there are some limitations. First, risk factors were self-reported and categorized broadly on two or three levels, which potentially underestimated their effects. Second, we did not use income or wealth to assess SES due to its potential multiple sources (e.g., cumulative income throughout life, monetary assistance from family members and various sideline economic activities) and the difficulty of obtaining accurate accounts. Second, economic status was self-perceived and subjective. However, subjective SES, such as self-perceived economic status and adequacy of income, has also been used to assess the association of SES with health in previous studies [37–39]. Subjective SES has been thought to capture differences in wealth [40]. Furthermore, self-perceived economic status may have been more related to relative SES within the respondents’ region. Some of the high-SES people from rural areas, for instance, may have had lower absolute SES than intermediate-SES people from the cities. Local governments may find it effective to make scientific decisions targeted to relatively poor older adults within their regions. Third, losses to follow-up and from nonresponse may have biased our results. The excluded participants tended to live in rural areas and have lower economic status than the eligible participants, but the difference was small (Supplementary Table S1). Fourth, confounding factors that were unknown or not included in the analysis, such as moderate or heavy drinking, may have contributed to an overestimation of outcomes. However, it was difficult to obtain alcohol consumption in our data. Finally, the MSM model we fitted based on the Markov process assumed that the transitions between health statuses depended only on the present status, not on the sequence of events that preceded it and thus did not account for individual heterogeneity in disability status history.

In conclusion, DFLE₆₅ and LE₆₅ disparities existed in economic status and education. Moreover, the LE₆₅ disparity was greater than the DFLE₆₅ disparity. The mechanisms of the DFLE₆₅ and LE₆₅ disparities in economic status and education differed. Physical examinations, fruit/vegetable intake and stress had great effects on DFLE₆₅ and LE₆₅ disparities in economic status but not in education. Future policies should pay more attention to older adults’ economic status and education and use different interventions for different groups.