The inequality of health-income effect in employed workers in China: a longitudinal study from China Family Panel Studies

The data are from the China Family Panel Studies (CFPS). The CFPS is a high-quality biennial household longitudinal survey, which is collected by the China Social Sciences Research Centre (ISSS) at Peking University. The survey focuses on the individual economic and non-economic welfare of Chinese residents, including individual’s employment status, educational achievement, health-related behaviours, family relationships, migration status and etc. Considering the great regional differences in Chinese society, CFPS adopts the proportional probability sampling (PPS) with implicit stratification, multi-stage, multi-level and population proportionality. The cross-sectional response rates of surveys in 2012 and 2014 were 74.1 and 72.8% respectively, and the cross-survey tracking rates were 80.6 and 83.8% respectively. As for the missing data, CFPS team has adopted a variety of processing methods to reduce the missing rate. For example, if personal income is missing, replace it with the average of the income range. If it is still missing or less than 100 yuan, it shall be replaced by the summation of the sub items related to income(see [31], for more details about CFPS). CFPS has so far conducted six surveys from 2010 to 2018, and each survey covers 25 provinces and cities with a total sample of 16,000 households and approximately 30,000 family members in these households.

In this study, we use data from waves of 2012, 2014, 2016, and 2018. After dropping observations with missing information on health indicators and other covariates, the final sample of our study includes 19,540 person-year observations aged between 18 and 70 years.

The key independent variable in our analysis is self-reported health. Compared with other health indicators such as healthcare use, disability, illness history and biomarkers [6, 32‐34], self-reported health is less costly and easy to administer in large social surveys such as CFPS. In addition, self-reported health has been shown a good predictor of both mortality and morbidity and large use in literature [35, 36]. Besides, we use lagged self-reported health as a proxy of current health to deal with the endogeneity problem caused by the reverse causality between health and income.

Self-reported health are extracted from the survey questionnaire and measured on a Likert scale with levels of excellent, very good, good, average and poor. We merge excellent, very good and good into one category due to the rather limited sample size of individuals whose reported health are excellent and very good health.¹

Table 1 provides the distribution of self-reported health in each wave. First, we observe that most people reported their health as good or above, accounting for 71% of observations. 24.4% reported their health as average and 4.6% as poor. Second, we can also detect a time change. In the year of 2010, only 2.6% of workers reported poor health, against 7.6% reporting poor in the year of 2016. In contrast, 78.1% of workers reported good health in 2010, but the proportion decreased to 64.6% in 2016. Table 1 suggests that the health status of workers was getting worse over the period of 2010 and 2016.

The outcome variable is the self-reported annual income from the individual’s current jobs, calculated as the sum of salary, various bonuses and allowances of workers. We choose annual income instead of monthly wages as the dependent variable because the former measure is less subject to monthly variation which might be common for certain type of jobs. It is worth mentioning that we exclude farmers or self-employed workers since they do not get paid by employers. The annual income is calculated in RMB Yuan in 2018 value. Table 2 below shows the average annual income of individuals with different self-reported health status in each wave and in all four waves.

The overall average annual income for workers who reported good, average, and poor health are 33,283 Yuan, 28,018 Yuan, and 23,308 Yuan, respectively. It suggests that workers with better self-reported health have higher earnings. Table 2 also shows that workers’ annual income increased continuously over the years from 2012 to 2018, but the income gap between different health cohorts became larger overtime. This tells us better health not only brings a higher earing for workers in China but also causes the income gap to widen overtime. Though the descriptive analysis in Table 2 strongly suggests a correlation between health and income, this correlation can be a spurious relationship without further control on other potential confounding factors.

Another main independent variable is worker’s occupation type. We divide workers into two groups, formally and informally employed workers. The former includes workers employed by the government, public and private enterprises, and large-scale organizations with labour contract. The latter includes workers who do not have a formal labour contract or a long-term contractual relationship with their employers. This distinction is important since workers of unstable employment were more vulnerable to shocks such as health deterioration. A dummy indicator is used to represent an individual’s occupation, with one for workers who were formally employed and zero otherwise.

In regressions, we control a rich set of characteristics of workers including educational attainment, age and its squared term, sex, nationality, registered residence as urban and rural, marital status as well as residential area² as East China, Central China, Northeast China and West China.

We apply a Mincer [37] equation in which the log of annual income is modelled as a function of education and experiment in human capital. The general form of the Mincer Equation as follows:

Where ln(y) denotes the log of income, Sch and Exper denote the schooling and working years. X represents the other independent variables than education and experience, including sex, marital, living areas. Other unobservable factors are contained in the error term ε,which satisfied E(ε| X) = 0.

Also, the effect of health on productive time and thereby income is hypothesized in Grossman model [3], which add new a perspective of the human capital research that health is a component that affects income. Following the expansion of Mincer Equation in previous literature, we model the log of annual income as a function of health status, age, sex, education, marital status, and other of covariates. Endogeneity may arise if unobserved individual characteristics are correlated with health variable. In panel data, by assuming that these unobserved individual characteristics as time-constant, we can use fixed effect models. The model presents as in Eq. (2):

Where the subscripts i and t denote individual and survey year, respectively. β and γ are estimated parameters. lnincome_it denotes the log of the annual income of individual i measured in RMB Yuan in time t, H_it denotes health status of individual i in time t. X_it is a series of dummy indicators of an individual’s characteristics as covariates. α_i represents the individual specific effect, which is assumed time-irrelevant in models of fixed effects. ε_it is a random error term and is often assumed to be iid normally distributed.

The majority of current analyses using classical regression models show that, on average, health has a significant impact on income. However, in the case of the widening income gap, estimating the average effect of health on income would underestimate the impact for poor workers and underrate the inequality of the health-income effect. To better understand the heterogeneous treatment effects of health on income for individuals with varying income levels, we resort to quantile regression.

Quantile regression was first introduced by Koenker and Bassett in 1978 [38], as an extension of the classical regression model and the panel quantile regression is quantile regression using panel data. Considered a linear model of the τ th quantile.

\( {y}_i={x}_i^T{\beta}_{\tau }+{e}_i,\mathrm{i}=1,\dots, \mathrm{n}, \),

Where the τ th quantile of e_i is zero. The estimator of β_τ can be derived from solving:

The quantile regressions are modelled as the conditional quantiles of the dependent variable and solved by minimizing the sum of weighted residual absolute values of residuals instead of minimizes the sum of squared residuals in OLS estimates. Compared with linear models, the quantile models are often less biased to skewed data and could capture varying effects. Therefore, quantile regressions fit non-normal distributed data and result in more robust estimates.

In addition, quantile regressions are well-known to be more robust to the outliers and require much weaker assumptions for the distribution of the error term compared with classical linear regressions.

The panel data quantile regression is the extension of quantile regression in panel data by Koenker [39]. Considered Eq. (2) where the τ th quantile of the conditional distribution of lnincome_it given explanatory variable vector X is specified as: Where Q_τ(lnwage_it | X) denotes the quantile τ of log monthly wage conditional on the vector of regressors.

The panel quantile model combines the advantages of panel data and quantile regression, in the sense that it effectively controls unobserved individual heterogeneity while being more robust to functional misspecification. Using panel quantile regression, we are able to recover the relationship between the dependent variable and other covariant based on the distribution of dependent variable at different quantiles [40]. The results of panel quantile regression show the different effects for specific quantile with robustness, which is suitable for our research.

In addition, we are also interested in the heterogeneity of the health-income effect among different demographic and socio-economic groups measured as sex, register residence and residential area. As male and female workers are endowed with different physical and mental capacity, they are expected to have different levels of health and income [41, 42]. According to the existing literature, there are vast differences between rural and urban residents in China, with the effect of health on wages is more pronounced for rural residents [43, 44]. Moreover, we test the health impact based on residential area.

Finally, we analyse the influence mechanism of health-income effect, and discuss the influence path of health on income by adding the interaction terms between health and other variables.

Finally, we analyse the influence mechanism of health-income effect, and discuss the influence path of health on income by adding the interaction terms between health and other variables.

Concerned with the problem of multicollinearity in regressions, we report bootstrap errors (bootstrap of 1000 times) rather than the asymptotic standard error to guarantee the efficiency of the regression model. All statistical and regression analysis is performed using R 3.5.1.

The descriptive statistics for both the dependent variable and independent variables are presented in Table 3.

The overall average annual income is 31,538 Yuan, with 34,226 Yuan for males and 27,607 Yuan for females. 69.7% males reported good health, and the proportion for female is 72.8%. On average, female workers are 2 years younger than male workers. More than two thirds of males (70.3%) and females (73.7%) have completed education lower than senior high school, while 29.7% of males and 26.3% females have completed higher education. Approximately 68.8% of males and 70.4% of females live in rural areas and about 45.1% males and 39.5% females have formally employed works.

The first column of Table 4 shows the regression result using the fixed-effect model, (2) to (5) columns show the panel quantile regression results at 25th, 50th, 75th, and 90th quantile of the income distribution.

The fixed-effect model suggests that, the workers with average and good health have an average annual income 15.8 and 11.3% higher than ones of poor health, after controlling for age, marital status, registered residence, and occupation. Due to the specification of the fixed-effect model, time-irrelevant variables such as sex, education, residential areas are dropped.

Apart from this mean effect recovered by linear fixed effect models, panel quantile regressions indicate that workers reporting their health as good and average earned 21.9 and 17.1% higher than those who report poor health at lower income levels (e.g. the 25th quantile). The gaps narrow down to 16.9 and 12.3% for the median income group and further to 16.4 and 8.2% at higher income levels (e.g. the 75th quantile). These estimates suggest that workers distributed at the 25th and 50th percentile of income are affected more by health and health status may have an insignificant effect on workers at the upper tail of the income distribution. Other variables have signs as expected. For example, education is statistically significant at conventional levels at all quantiles, which is consistent with existing empirical evidence on the return of schooling.

Furthermore, we have conducted the falsification test which include the forward treatment variable as additional controls – a way commonly used as the post estimation test in panel data regressions – to test the model assumptions of our analysis. In the test regressions, we have found that when the health in prior period is controlled, the current health is only marginally significant in fixed effect model, and mostly insignificant in panel quantile regression, which supports our identification assumption. The test results are shown in the Table 11 in Appendix.

Figure 1 depicts effects of health on annual income on the full range of income quantiles from zero to one. The dashed lines represent the coefficient estimates for have the average and good health against bad health using the panel quantile model and the solid line and dotted line represent the coefficient estimates using the fixed-effect model. It is evident that the lower- and middle-income workers are more affected by health status.

From Fig. 1 we can conclude that, first, better health brings higher income in the sense that good or average earn more than those reporting poor. Second, the effect of health on income dampens with the increase of the income distribution, suggesting a more crucial health-income effect for low- income workers. Third, the fixed-effect model estimates have roughly the same magnitude as panel quantile estimates at 75th quantile, suggesting that the classical estimation may underestimates the health-income effect for low-income workers but overestimates the effect for high-income workers. On the other hand, using panel quantile regression, we are able to recover heterogeneous returns of health on income regarding different income groups while controlling of endogeneity.

In this section, we discuss the relationship between health and income, using alternative health indicators to verify the reliability of our previous finding. Following Cai et al. [45], we generate a new health index called ‘health score’, which combines two questions ‘Have you suffered from a chronic or acute disease?’ and ‘How severs was the illness or injury?’ in the questionnaire. Based on responses to the two questions, health score is defined as an ordered response variable with four different levels: no chronic disease (s1), have a chronic illness but not severe (s2), have chronic diseases somewhat severe (s3) and quite severe (s4). We treat health score as a binary variable with four levels.

Table 5 shows the estimation results of the panel quantile regression model using health score as the health indicator, which suggests better health measured as health score dummies significantly implies better earnings for those whose income are distributed at 25th and 50th quantile, after controlling for covariates. Nevertheless, the effect is insignificant for workers whose income are distributed at the upper ends.

Compared with panel quantile regression using lagged self-rated health, the results in column (2) to (5) in Table 4 and Table 5 are similar, which indicates that our regression results are robust to particular health measures used.

Table 4 shows that male workers’ incomes are higher than females at all quantiles after controlling for other factors, but the pattern is not clear. To test the potential differences in the effect of health on income for sex heterogeneity, we discuss the health-income effect by sex.

Table 6 shows the health-income effects by sex using the self-reported health index. Estimations reveal that health is more important for female workers than males, since the health-income effect is more significant for females. Male workers with good and average health earned slightly higher annual income than male workers with poor health, but it is not statistically significant for most male workers. As the distribution annual income increased, the health-income effect shows a downward trend. For instance, annual income for males with average health is 13.4% higher than those with poor health at 25th quantile, but only 7.5% higher and insignificant at 75th quantile.

For female workers, the health-income effects are larger for workers placed at lower and middle range of the income distribution than those placed at the higher end. The similar pattern is observable for males even though with smaller effects. For example, females with good health earn 41.3 and 29.1% higher than those with poor health at the 25th and 50th percentiles, but only 12.1% higher at the 90th quantile and no longer significant.

Figure 2. shows the variation of panel quantile estimates for workers with different sex and health status under full income distribution. The dashed line represents male and dotted line represents female. We find that female workers are more affected by health than males for two health statuses. With the increase of income distribution, the health-income effect for females decreasing faster than males, which means health is more important for female workers than males, especially for low-income workers.

Next, we test the heterogeneity of the registered residence for workers. In general, rural workers earned 39.9% less than urban workers on average in the fixed-effect model, but in the panel quantile regression, the income gap is relatively small. Even for rural workers that distributed at 25th quantile, the income gap is 29%, and the difference is narrowing along with the annual income distribution. So, we are interested in how did health affect income for workers with the different registered residence.

From the regression results in Table 7, the effect of rural workers’ health on income is higher than that of urban workers. For urban workers, the impact of health on income is not significant for all income quantiles. For rural workers, health has a considerable effect on income for most workers, and the health-income effect is decreasing of the income distribution, from 25.5 to 17.5%, but is insignificant at 90th income distribution. Generally speaking, health is vital for rural workers, maintain good health contribute to higher income.

The last heterogeneity test in this paper is the health-income effect for workers from different residential areas. Since we will assess whether the achievements of economic growth were distributed across the country, we divide these 25 provinces and cities into four economic regions as East China, Central China, Northeast China, and West China. Among these four economic regions, the East has the most powerful economic strength, followed by the Central and Northeast, and West is the weakest.

We present the quantile regression results grouped by four economic regions in Table 8 below to show the variance effect of health on income across the country. Table 8 shows four economic regions have the different health-income effect, and the effect is more significant for the relatively developed area in east China.

From previous regressions, we find that the health-income effects are robust even when we used different health indicators and consider the heterogeneity of workers. The observed health-income effect weakens with the increase of income of workers and are more relevant to low- and middle-income workers. Our hypothesis for the results is that health not only directly affects workers’ income but also affects income through workers’ occupation choices. Given the segmented job market in China, we have recognized the financial implication of health change would be quite different for workers with more secured jobs like those belonging to public sector and workers with less secured jobs. We hypothesized that the secured jobs would help to buffer the negative effect of health deterioration on income reduction. To verify our hypothesis, we have included an analysis in which occupation is dichotomized as stable and unstable jobs and interacted with health. Thus, we divide occupations in the questionnaire into two categories based on job characters. The first category, termed as formal employment, consists of jobs with long-term contracts with employers such as enterprise and government and regular working hours and salaries but with requiring higher education or job skills. The other category that we term as informal employment include workers who engage on manual or repetitive work with lower requirements of education or professional knowledge, who are temporarily employed or paid by days and easily replaced.

The income distributions of these two types of workers are illustrated in Table 9. We find that with the increase of the income quantile, the proportion of workers engaged in formal employment was higher, while low-income workers were more involved in informal employment. The percentage of the formal employment increased from 8.7% at the lower tail of income to 68.8% at the upper.

We add the interaction of health and occupation in the fixed-effect model and the panel quantile regression to explain how health affects income. We also apply mean-centering method [46] to reduce nonessential collinearity problem after including interaction terms. The results are presented in Table 10.

In both the fixed-effect model and the panel quantile model, the interaction between health and occupation is significantly and negatively correlated with income, that the health-income effect may only hold for workers with unstable employment. For those with stable employment, the employment serves as the safety net which could buffer the shock of health deterioration on income decrease. These results are highly suggestive that health not only affects income directly but also affects income through workers’ occupation.

To the best of our knowledge, no previous study has explored inequality of health-income effect for workers distributed on a wide income spectrum in China using panel quantile regression. In addition to the discussion of the health-income effect, we have investigated its heterogeneity for different demographic and socio-economic groups and possible mechanism to explain the effect. Our main finding that deterioration of health most affect workers located at the lower tail of income distribution, suggests that public policies should be tailored particularly to alleviate the negative impact of health deterioration on income for lower and middle-income groups.

This study has some limitations. First, the lagged self-reported health, which is used as the primary health measure in our study to deal with the endogeneity problem, may suffer from the measurement error in the sense that it is self-reported and lagged. To reduce the self-report bias, we have examined an alternative measure which is based on chronic illness history and hopefully less subject to self-report bias. Using lagged measure of health might be problematic given that health is a continuous and changeable indicator. Generally speaking, lagged health is often better than current health. When we use lagged health as an exposure variable to analyse the impact of health on income, the effect can be overestimated. Nevertheless, in our analysis, we have found that the income gap exists not only between workers who are placed at the two ends of health spectrum – the ones with good health and with bad health, but also exists between the ones with poor health and with average health. This speaks that despite the possible overestimation, the true effect of health-income gradient is unlikely to be neglectable. Second, to avoid the temporal variation we measure income of workers by their annual income, which, however, may be subject to year-to-year variation. Nevertheless, it is well-known that measurement errors for the outcome variable can only result in less efficiency of parameter estimates, but not bias. Finally, we have only examined occupation as the possible mediator accounting for the health-income effect. Even that we have found that occupation is indeed one important channel, there might be other important channels, which we have to leave for future research.