Gender Disparities in the Relationship between Education and Self-Reported Health Across Cohorts in China

Background: Variation in the relationship between education and health has been studied intensely over the past few decades. Although there is abundant research on gender disparity and cohort variations in the relationship between education and health, based on samples from the U.S. and Europe, research about China is limited. Given the specific social changes in China, the study is designed to analyze the gender difference and cohort variations in the education-health relationship. Method: Longitudinal, nationwide data from the Chinese Family Panel Studies from the years 2010 to 2016 are statistically analyzed. Self-reported health is measured by respondents’ subjective assessment of their health. The highest level of education earned operationalizes the education measure. Each cohort is defined by a distinct period of social change in China. The age-vector model is used to analyze gender and cohort variations in the association between education and self-rated health. Results: Men report better health than women, but the relationship between education and health for women is stronger than for men. Educational gaps in self-rated health do not change significantly for cohorts before 1955 and cohorts after 1977, but the gaps become stronger for cohorts between 1956 and 1976. There is gender disparity within the cohort variations in the education-health relationship. For women, the education-health relationship in the 1956-1960, 1967-1976 and 1977-1983 cohorts is significantly stronger than for the 1908-1938 cohort. While not as strong, the education-health relationship remains consistent across all cohorts for men. Conclusions: The study findings support the resource substitution hypothesis and not the rising importance hypothesis in China. Considering the findings on gender disparity and difference in cohort effects, we discuss the potential influences of the unique social transformation and educational expansion in China.

graduates in China in 1999, but four years later, the number of graduates rose to 1.88 million. By 2017, there were around 7.36 million college graduates in China [27]. The likelihood of female high school students getting into colleges has been reported to be the similar to that of male high school students [28]. However, the significant educational advancement in China resulted in a market devaluation of educational credentials, and the influx of college credentials also made the labor market more competitive [29].
These social upheavals and policy changes in China occurred at different time points, and therefore their influences may vary for birth cohorts who came of age in different historical periods. This study aims to examine the education-health relationship from both gender and cohort perspectives in the Chinese context. Data from four waves (2010,2012,2014,2016) of a nationwide survey of adults aged 22-years and older are used to answer the following research questions: (1) Is the education gap in health larger for men or women? (2) Is there any inter-cohort variation in the association between education and health? (3) Is there any gender difference in the relationship between education and health in inter-cohort variations?

Methods Sample
We use four waves of data from the Chinese Family Panel Studies (CFPS) from 2010-2016. All four years of CFPS are nationwide surveys of the Chinese population aged 18 years and older. The surveys were administered by the Institute of Social Science Survey at the Peking University of China and were designed to study the historical change of society, economy, population, education, and health in China. The survey collects panel data at the individual, household, and community levels. Respondent selection was guided by implicit stratification and multi-stage probability proportional to size (PPS) sampling. The panel design provides an opportunity for cohort analysis of social and economic change over time. We identified the sample of adults aged 22

Measures
The self-reported health variable measures respondents' subjective assessment of their health. Respondents were specifically asked, 'How good is your health in general?' The Likert scale in 2010 includes the response options of 'very bad,' 'bad,' 'a little bad,' 'fair,' and 'good. ' We coded the first of these four items as '0' and "good" health as '1.' The Likert scale from 2012 to 2016 includes the response options of 'bad,' 'fair,' 'little good,' 'good,' and 'very good.' The first two items were coded as '0,' and the last three answers were coded as '1.' The self-reported health measure is regarded as a valid and reliable measure of health as it encompasses the subjective experience of fatal and nonfatal diseases and the general feeling of well-being [30][31]. The selfreported health variable is highly correlated with objective measures of health, such as mortality, morbidity, or diagnosis from a clinical exam. The self-rated health measure is a salient predictor of morbidity and mortality [32], and an even stronger predictor of physical health, mortality, and chronic diseases [14].
Education was measured by asking, 'What is the highest degree you have completed?' Respondents could select one of the following answer categories: 'not received education,' 'primary school,' 'junior high school/professional high school,' 'senior high school,' 'junior college,' 'college,' and 'undergraduate.' For individuals who were still attending school, they were asked which year they attended at the time of the survey. The answers range from 0 years to 22 years. Given that adults aged 24 years or younger may not have completed their educational careers by the time they were surveyed, we selected samples with respondents aged 25 years or older in an attempt to avoid assessing effects of education on health prematurely [21,33].
Additionally, employment status, family income, and frequency of physical exercise are important factors affecting health [35][36], so we use them as control variables. For employment status, we used the 'not employed' answer as the reference category. We used average family annual income to measure economic background. The frequency of physical exercise was measured by asking, 'how often do you exercise in the last month?' The answer options range from 'never,' 'one time a month,' 'two or three times a month,' 'two or three times a week, and 'almost every day,' and they were coded as 0 to 4 respectively.

Models
The longitudinal design in CFPS allows researchers to study between-person and withinperson changes over the four survey waves, whereas a repeated cross-sectional design would confound age effect with period and cohort effect due to the collinearity between age, period, and cohort [37]. Hierarchical growth curve models tested for cohort variations in age trajectories [38][39]. We use an age vector model to adjust for age and cohort effect over time because a traditional growth curve model would not accurately map the age sequence of self-rated health for short-term panel data -a 6-year period is too short to estimate the trend for an entire life course. The age vector model measures the between-person differences in age and cohort effects, and estimates within-person effects using the passage of time (i.e., survey waves) rather than the sequence of age [39]. An age vector model is also the preferred method for testing a cohort-level hypothesis with short-term panel data. This model relaxes the linear assumption of age imposed by the traditional growth curve model, allowing for us to model higher polynomials of age between individuals [39].
Since self-reported health is a dichotomous variable in our study, we used a two-level hierarchical binary logistic regression model to estimate the probability of being in good health, adjust for within-person changes at the first level, and adjust for between-person differences at the second level. This model estimates the gender differences for both age trajectories and cohort variations in the association between education and health over the life course. We formulate a series of hierarchical linear models using HLM 7.1 software. The full model (Model 4) that controls for all interaction terms and covariates is described as follows: The level-1 model characterizes within-individual change across survey waves after controlling for a series of variables at level-2.
The level-2 model estimates the between-individual change in health with age, and assesses whether there are patterns in the association between education and health in the age trajectory for gender across different cohorts. Level-2 consists of an intercept component that measures fixed effects for all individuals and a slope equation (i.e., linear growth rate) that measures the changes in fixed effect over time.
The linear growth rate of the period . 1% equation is expressed as the following: . The linear growth rate component includes all of the corresponding variables from the intercept component, except for 89:_=> % because we assume the rate of change in age effect is the same across all survey waves. We also do not control for random effect for the slope equation, because we lack the statistical power and we assume the slope coefficients are fixed for all respondents. Lastly, we examine the full model separately for each gender, allowing us to compare the significance of association and general direction between the female-only model (Model 5) and the male-only model (Model 6).

Results
Gender difference: Results for the resource substitution hypothesis Table 1 shows the descriptive statistics for all the variables in the analytic models. As expected, men on average reported better health status, higher educational attainment, higher income, and higher physical exercise rates than their female counterparts. The sample size for each cohort by gender is shown in Table 2.  In Table 3, we present the results from a series of age vector models. Based on the odds ratio in Model 2 for males, men reported better health than women. This model tests the resource substitution hypothesis, which supposes that there is a substantial educational difference in the slope of gender (Model 1). For male respondents, a one-year increase in education yields a change in log odds of 0.07 or an odds ratio of 1.073. Thus, a female with one additional year of education is on average 1.012 times more likely than a male with one additional year of education to report a status of healthy in the survey holding all else constant. In other words, the association between education and health is weaker among men than among women. The gender difference becomes statistically significant in the slope model when some variables are added, but it is not statistically significant in the intercept model. This pattern supports the resource substitution hypothesis and is consistent with the results of related studies in the United States [9].

Cohort variations in education and health: Results for the rising importance hypothesis
In order to understand the cohort variations in self-reported health, we added the cohort variable in Model 2. The odds ratios for each cohort in Model 2 show that the younger cohort reports as healthier than the oldest cohort, and the difference in self-reported health between the oldest cohort and the 1961-1966 cohort is statistically significant.
To test the rising importance hypothesis, which supposes that the association between education and health becomes stronger over time for a given population, we included interaction effects between education and cohort in Model 3. Model 3 reports education-by-cohort interaction effects on health. We found that the association between education and health does not change significantly for cohorts before 1955, and the association becomes stronger for younger cohorts. For the 1956-1960 cohort, a one-unit increase in education yields a change in the log odds of 1.099. The odds ratio for the 1908-1938 cohort is 1.056. The odds ratio of a one-unit education gain for cohort 1956-1960 over the odds ratio of a one-unit education gain for the 1908-1938 cohort is 1.041. In other words, respondents from cohort 1956-1960 with one additional year of education are on average 1.041 times more likely than respondents from cohort 1908-1938 with one additional year of education to report healthy in the survey holding all else constant. With one additional year of education, respondents from the 1967-1976 cohort and the 1977-1983 cohort are on average 1.059 times and 1.064 times respectively more likely to report as healthy compared with respondents from the 1908-1938 cohort. We also note that the rising trend disappears in the youngest cohort. Based on these findings, the full models do not support the hypothesis of rising importance.

Gender difference in education and health across cohorts
We also established full models separately for men and women. Model 4 shows that the odds ratios of the interaction between education and cohort are significantly more than 1 among women for the 1956-1960, 1967-1976 and 1977-1983 cohorts. The results in Model 4 suggest that education's positive effects on health increased for the 1956-1960, 1967-1976 and 1977-1983 cohorts. Female respondents from the 1956-1960, 1967-1976 and 1977-1983 cohorts with one additional year of education are on average 1.063 times, 1.088 times, and 1.102 times more likely than respondents from the 1908-1938 cohort (P<0.05) to report better health, respectively. The pattern for the relationship between education and health for the female subsample is the same as that of the full sample. As with the rising importance hypothesis, the rising trend also disappears in the youngest cohort. In Model 5, education is not positively related to good health for men (P>0.1). The interactional effects between education and cohort are not statistically significant. In other words, the education effects are the same across all eight cohorts among men, but there is a gender difference in education and health across cohorts.

Discussion
This study assesses the resource substitution hypothesis and rising importance hypothesis in the Chinese context. Results reveal that the resource substitution hypothesis is supported in the Chinese context, which is consistent with related U.S. studies. However, compared to previous findings about these two hypotheses from the United States, there are two notable differences in the Chinese context. First, the effect of education on health has not increased from the oldest cohort to the youngest cohort, and the gaps in health remained stable for some cohorts. Second, for the rising importance hypothesis, there is a gender difference in the educational effects on health across cohorts in the Chinese context.
We think potential explanations for the U.S.-China differences lie in the role of sociocultural and policy change. First, the association between education and health-related resources in China may be one of the reasons for the complex trend across cohorts. Eating dinner and drinking wine with friends, colleagues, and superiors is a crucial way to maintain a social network and to obtain resources in Chinese society but can be a negative effect for health outcomes [40]. Before the Reform and Opening-Up, individuals had limited access to food and clothing under the socialist economy. After decades of Opening-Up, more people began to enjoy abundant material prosperity [41]. Individuals with higher education and more purchasing power are more likely to participate in social gatherings, indulge in unhealthy diets, and consume more alcohol [39]. Furthermore, these behaviors are empirically found to be more pronounced among men than women [41], which could possibly be supported by a strong positive relationship between education and drinking in the male sample in our study.
As for the cohort patterns, we argue that the drastic educational landscape changes in China affect the relationship between education and health-related resources. The rapid increase in social and economic benefits from education for the 1967-1976 cohort may be due to the recovery of the college entrance exam in 1977. Although there were less than 1.4 million students who graduated from college prior to 2003, credentials became more critical for finding jobs after the Opening-Up and Reform periods in 1978. Members of the 1967-1976 and 1977-1983 cohorts who graduated from college could acquire well-paying jobs compared with cohort members who did not receive a college degree. However, with educational expansion starting in 1999 in China, the number of people receiving a college education has been increasing rapidly. According to the Chinese Educational Statistics Report (2018), the number of people receiving a college education rose from 1.08 million in 1998 to 7.62 million in 2017 [27]. Compared with older cohorts, the children of Later Opening-Ups (1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994) can receive an education more easily, but the valuation of the same academic degree has decreased. Educational expansion resulted in the phenomenon of overeducation, which lessens the income benefits from educational attainment due to the mass of individuals who have similar credentials vying for similar, limited opportunities. Members of the youngest cohort face more competition in the labor market than the members of the Early Opening-Ups, for whom the adverse effects from educational expansion had not caught up. Educational expansion weakens the role of education in improving access to both socioeconomic and healthrelated resources [29,43]. The rate of returns for education even declined in 2009 [24]. As a result, the rising trend disappears for the 1984-1994 cohort.
The gender difference in education and health across cohorts can be attributed to women experiencing greater difficulty with gaining access to educational and health-related resources. Results show that the links between education and health are stronger among women than among men in China, which support the 'resource substitution hypothesis' from the perspective of cohorts. Women have been in a socially disadvantaged position for decades [44], so they have fewer resources to rely on. On the contrary, men have more resources presently and historically, so education is less important for men than for women. Another possible reason for this gender disparity is females' disadvantages in the labor market. Due to governmental deregulation of the market, enterprises began to employ more men than women, and discrimination against women in the labor market has increased since the "reform and open-up" [45][46]. Moreover, education is more important for women in securing jobs than for men. Hence, education is more important for women than for men across cohorts.
Although self-reported health is a valid measure of health status [30], it is susceptible to effects based on individual characteristics and cultural contexts. A study found that people with higher education and income tend to report their health optimistically in China [47]. Given its vulnerability to individual and social influences, the association between education and selfreported health may be receiving too much attention in the literature. The vignette method could be incorporated into studies in the future to provide more qualitative information if related items are included in the design of data collection instruments. The relationship between education and health is complex or potentially reciprocal -people with higher education can improve their health, but people may drop out of high school due to severe health problems. The primary goal of our study is to illustrate the demographic and cohort patterns for the education and health association. As such, we cannot establish a causal relationship between education and health, however different methods and data may allow for the conclusion of causal inferences in future studies.

Conclusions
In conclusion, the study investigated how the categories of gender and cohort moderate the returns to education on health, as well as how the gender-cohort interactions condition the education-health associations over time. The results did not support the rising importance hypothesis, especially among men. This different pattern suggests that broad contextual factors such as gender and cohort can significantly shape the education-health patterns across cohorts in China. Our findings show that the gender difference in the association between education and health is significant, but China's unique history of educational and health development, as well as distinctive dining culture, may have also influenced these education-health patterns. The rapid social change and the Chinese authoritarian state require various theoretical frameworks to explain the gender disparity in educational benefits across cohorts.

CFPS: Chinese Family Panel Survey
Ethics approval and consent to participate Not applicable

Availability of data and materials
The datasets generated and /or analyzed during the current study are available in the Chinese Family Panel Survey, http://www.isss.pku.edu.cn/cfps/index.htm.