Educational level as a predictor of the incidences of non-communicable diseases among middle-aged Japanese: a hazards-model analysis

It is well known that health outcomes are closely related to educational level [1]. Specifically, lower levels of education have been found to be related to higher incidence and prevalence of non-communicable diseases (NCDs) [2, 3]—including cardiovascular and cerebrovascular diseases [4‐7], cancers [8, 9], diabetes [10‐12], hypertension [13], and chronic respiratory diseases [14]—in addition to higher levels of cause-specific and all-cause mortality [15‐17]. However, most of the preceding studies on educational inequalities in NCD incidence have been prospective cohort analyses, which focused on the prevalence or cumulative incidence over the follow-up period among respondents with no baseline disease (e.g. [5, 11, 14]), or (repeated) cross-sectional analyses, which compared the prevalence of the diseases among respondents with different levels of educational attainment (e.g. [2, 4, 6, 18]). In comparison, hazards model analyses, which aimed to relate the time passage before the incidence of the disease to educational level, have been relatively scarce [12].

A related issue that should be further addressed is to what extent educational inequalities in NCD incidence are mediated by socioeconomic status (SES), other than educational level, and health behaviour. Low income and unstable job statuses, which are likely linked to low educational level, are reasonably predicted to raise the risk of NCD incidence, as documented by previous studies [2, 3]. Similarly, some types of health behaviours, such as smoking, heavy alcohol drinking, physical inactivity, and unhealthy diet, are expected to mediate the impact of educational level on NCD incidence [1]. To be sure, many studies have assessed the importance of these factors in explaining the educational inequalities in mortality [19‐22], worsening frailty [23], and mental health [24]. However, the meditating effects of these factors on educational inequalities in NCD incidence have been largely understudied. If the mediating effects are limited, policy interventions to improve SES or promote healthy lifestyle after the completion of education would not be expected to be effective in moderating educational inequalities in health during later life. Indeed, some studies have provided evidence for the limited mediating effects of SES and health behaviour on worsening frailty [23] and NCD prevalence [18].

It is also interesting to explore gender differences in the association between educational inequalities and NCD incidence. Some studies have found steeper educational gradients in the incidence of some types of NCD among women than among men [2, 5]. Gender differences may also depend on the type of NCD and there may be also gender differences in contributions of SES and health behaviour in explaining educational inequalities [25].

Considering the obtained knowledge and limitations of the existing literature, the current study conducted a hazards model analysis to examine educational inequalities in NCD incidence, using the 12-wave longitudinal data obtained from nationwide population-based surveys in Japan. The hazards models were estimated for six types of NCDs, and educational inequalities were compared across them. The results are expected to help evaluate the relevance of the observations obtained by preceding prospective cohort and cross-sectional studies. In addition, the extent to which education inequalities in NCD incidence were mediated by SES, other than educational level, and health behaviour was examined. Specifically, household-size adjusted income and job status as well as four types of health behaviour (smoking, heavy alcohol drinking, physical inactivity, and unhealthy diet) at baseline were examined as potential mediators. Furthermore, gender differences in education inequalities were explored.

Japan faces a growing burden of NCDs, while its rate of longevity is one of the highest in the world; NCDs accounted for 54.5% of total causes of death in 2017, and deaths caused by cancers and heart diseases, the top two causes of death among NCDs, increased from 11.6 and 8.7 per million, respectively, in 1970 to 29.9 and 16.4 per million in 2017 [26]. The increasing burden of NCDs has enhanced the need for health promotion policies to prevent and control NCDs, especially under extended longevity [27, 28]. The current study focused on the risks of NCDs among middle-aged Japanese, who were 50–59 years old at baseline, given that the prevalence of key NCDs has been shown to increase substantially during the middle age years [26].

The key features of the study sample are summarised in Table 1, along with the gender difference and its statistical difference for each variable. Low, middle, and high educational levels comprised 19.4, 55.2, and 25.4%, respectively, for men, and 18.4, 74.9, and 6.7%, respectively, for women, indicating higher educational attainment among men compared to women. The proportion of stable job status was much higher among men (84.2%) than among women (25.4%). Women displayed healthier lifestyle than men in all four aspects (smoking, heavy alcohol drinking, physical inactivity, and unhealthy diet).

The prevalence of each NCD over the sample period was compared by educational level in Table 2. Without controlling for any other factor, an inverse educational gradient was observed for diabetes and stroke among both genders, while such a gradient was observed for heart disease and hypertension only among women. Hyperlipidaemia was positively associated with educational level among both men and women, and cancer was not related to it among either men or women.

The results of the analysis of variance, which tested the null hypothesis that there was no difference in prevalence of each NCD across regional areas, are presented in Table 3. Judging by the p-values, six out of twelve gender-NCD combinations revealed regional heterogeneity. This result indicated the need to control for regional areas in regression analysis.

The estimated RIIs of educational level for baseline SES and health behaviour are summarised in Table 4. To assess the RII of educational level, the estimated OR of each outcome responding to an increase in the ridit score of educational level from 0 (highest) to 1 (lowest) was calculated. Judging by the estimated ORs, lower educational levels were positively associated with low income and unstable job status. The association between lower educational levels and no job was positive for men but negative for women; the latter result probably reflected the existence of highly-educated, dependent wives with no paid job. All four types of unhealthy behaviour were also found to be positively associated with lower educational level; these educational gradients were consistent with the argument that lower SES and unhealthy behaviour can explain educational inequalities in NCD incidence.

Hazards model analysis began by graphically illustrating the Kaplan-Meier survival estimates by educational level in the case of diabetes (without controlling for any other variable). As shown in Fig. 1, the differences in cumulative incidences of diabetes tended to widen as time passed particularly among women. Educational inequalities in the incidence of diabetes were confirmed more formally by Cox proportional hazards models. The estimation results of Models 1 and 2 for diabetes are presented in Table 5. The HRs of the incidence of diabetes in response to an increase in the ridit score of educational level from 0 (highest) to 1 (lowest) (in both Models 1 and 2) as well as to lower SES and unhealthy behaviour (in Model 2) were calculated. The HR for the ridit score indicates the RII by educational level. The estimated HRs were well above one in all model specifications, confirming the inverse educational gradient in the incidence of diabetes, and the gradient was steeper for women than for men.

It was also found that among both men and women, the HRs for the ridit score of educational level was not much attenuated and remained highly significant in Model 2, which controlled for baseline SES and health behaviour. Meanwhile, the results of SES and health behaviour in Model 2 were mixed; the incidence was positively associated with smoking, heavy alcohol drinking, and physical inactivity among men and only with no job among women, while no association was observed for any other variable.

Similar Cox proportional hazards models were estimated for other NCDs as well, and their estimated HRs for educational level are summarised in Table 6. Even after controlling for baseline SES and health behaviour, educational inequalities were observed for diabetes and stroke among both men and women and for hypertension only among women. Heart disease and cancer were not associated with educational level, while hyperlipidaemia had a positive association. Only small to moderate parts of the HR were attenuated by baseline SES and health behaviour for NCDs that had inverse educational gradients. When the degree of attenuation was calculated by (HR in Model 1 – HR in Model 2) / (HR in Model 1–1) [22], it was in the range between 0.0% (for hypertension among women) and 32.7% (for stroke among men). It was also observed that attenuation was more limited among women than among men for diabetes and stroke.

Another noticeable finding was that, as already observed for diabetes in Fig. 1 and Table 5, the estimated HRs for the ridit score of educational level tended to be higher among women than among men for NCDs whose incidences were negatively associated with educational level. To examine gender differences more formally, Table 7 presents the estimated HRs for the ridit score, women, and their intersection terms, obtained by Model 2 using the entire sample. The estimated HR for the intersection term was significantly higher than one for diabetes and hypertension, indicating sharper inverse educational gradients for these diseases among women than among men. The HR for the intersection term was greater than 1 for heart disease, stroke, and hyperlipidaemia, but non-significant.

This study examined educational inequalities in NCD incidence using the data obtained from a 12-wave longitudinal nationwide survey of middle-aged individuals in Japan. Unlike most of the preceding studies, this study applied a hazards model analysis to the data over a long follow-up period (11 years) to capture the association between educational level and NCD incidence. In addition, the potential mediating effects of baseline SES and health behaviour as well as gender differences were explicitly considered. Key findings and their implications are summarised as follows.

First, educational inequalities were confirmed for selected types of NCDs. Even after controlling for baseline SES and health behaviour, Cox proportional hazards model regressions showed that lower educational level was positively associated with the incidences of diabetes and stroke among both men and women, and with hypertension only among women. The observations of the inverse educational gradients were largely consistent with those of preceding studies (for diabetes [2, 10‐12, 18], stroke [2, 5], and hypertension [2, 13]).

Meanwhile, among both men and women, the incidence of heart disease or cancer had no educational inequalities and the incidence of hyperlipidaemia was positively associated with educational level. Notably, the observed positive association between educational level and hyperlipidaemia was somewhat surprising, but it was consistent with the observation among Korean male adults [41]. One possible reason to explain this result is that higher educational level may lead to a dietary style linked to higher risks of hyperlipidaemia. Indeed, total energy and fat intakes have been found to be positively associated with household expenditure, which is closely linked to educational level, among both male and female adults in Japan [42], possibly because higher-educated, higher-SES individuals have more chances to dine out [41].

Although identifying the reasons to account for the differences across NCDs is beyond the scope of this study, the findings generally underscore the importance of educational level as a key predictor of NCD incidence. To be sure, it cannot be concluded from the results of this study that educational level is a key determinant of NCDs. However, it can be reasonably argued that educational level can be a reliable signal of the risk of the incidences of selected NCDs in later life.

Second, the results revealed that the mediating effects of SES, other than educational attainment, and health behaviour on NCD incidence are relatively limited. Lower income and job status as well as unhealthy behaviour have been found to explain income inequalities in health outcomes [19‐22, 24]. Actually, the results showed that lower educational level tended to increase risks of low income, unstable job status, smoking, heavy alcohol drinking, physical inactivity, and unhealthy diet, and that some of these factors modestly raised the risk of NCD incidence. However, inclusion of these factors led to just low-moderate attenuation in educational inequalities of NCD incidence, a result consistent with the observations in some preceding studies [18, 23]. This finding suggests that educational attainment as a predictor of NCD incidence is mostly independent of these factors. However, there can be other potential mediators—such as obesity [19], health literacy [43], social participation [23], and accessibility to health care service—that were not considered in this study. Nevertheless, the results point to limited effectiveness of policy interventions to improve SES or promote healthy lifestyle in reducing the educational inequalities in NCD incidence.

Finally, the results showed significant gender differences in educational inequalities in the incidences of diabetes and hypertension. Similar observations have been obtained in preceding studies for diabetes [2, 18] and hypertension [2]. The exact reason for these differences and also the reason for no gender differences observed for other diseases were not addressed in this study. However, it might be possible that gender differences in social roles, exposures to stressors, and coping strategies may confound gender differences in educational inequalities in NCD incidence [44, 45].

It should be noted that this study had several limitations, besides the observations being limited to middle-aged Japanese people and the reliability of self-reported NCD diagnosis. First, SES and health behaviour of the sample are likely to change across the periods possibly due to health reasons, even though educational attainment can be considered mostly fixed in this study sample. Hence, focusing exclusively on these factors at the baseline may not precisely capture their mediating effects. Second, the results are likely to have been affected by the definitions of health behaviours and other variables. For instance, while physical inactivity was defined as engaging in no leisure-time physical activity in this study, people can be very active in their mainstream life without engaging in leisure-time physical activity. Third, as mentioned above, there are potentially other mediators than those of SES and health behaviour considered in this study for educational inequalities in NCD incidence. Adding these factors in the analysis, the attenuation of estimated educational inequalities from Model 1 to Model 2 may be reduced, possibly affecting the interpretation regarding the degree of independent impact of educational level on NCD incidence. Fourth, respondents who dropped out without incidence were censored in each regression model, as is usually the case of hazards model analysis. Although the average attrition rate was as low as 4% in each wave, neglecting the possibility of the attrition due to the incidence may have led to an underestimation of the risk of incidence.

This study underscores educational inequalities in the incidences of selected NCDs, especially among women. Inverse educational gradient in NCD incidence was not much attenuated by adjusting for mediating effects of baseline SES and health behaviour. Further analysis is needed to identify the mechanism linking educational level and incidences of each specific NCD. Still, the observation that educational level was associated with NCD incidence in a relatively non-mediated manner has important implications for public health. Educational level can be utilized as a reliable signal of the risk of NCD incidence in general, suggesting that the policy measures to reduce such a risk should be targeted to individuals with lower educational level. It can be also argued that policy support to help socioeconomically disadvantaged children to enhance educational attainment is recommended to reduce inequality in health in later life.