Subgroup disparities in the mediating roles of income and social support between education level and depressive symptoms among Chinese middle-aged and older adults: an analysis based on the China health and retirement longitudinal study

Abstract

Objectives

This study aimed to identify potential subgroups and explore how the mediating effects of income and social support in the relationship between education level and depressive symptoms vary across these groups among middle-aged and older Chinese adults.

Design

This study used statistical learning techniques to analyse data from the China Health and Retirement Longitudinal Study (CHARLS), adopting a cross-sectional approach.

Setting

The study analysed data from the 2011 baseline survey of CHARLS. Depressive symptoms were evaluated based on the Centre for Epidemiologic Studies Depression Scale form. Education, social support (informal social support, community support and public support), lifestyle choices (smoke, alcohol and social activity), health conditions (hypertension, dyslipidaemia, diabetes, lung disease, heart problem, kidney disease, digestive disease, arthritis, disability, pain and activity limitation), demographic background (urban residence, sex, age, marital status, body mass index and retirement status) were included in the analysis.

Participants

8161 adults over the age of 45 were included.

Primary outcome measures

The primary outcome measure was the association between educational level and depressive symptoms mediated by income and social support. These associations may vary across subgroups and were measured using a heterogeneous structural equation modelling method with automatic subgroup detection.

Results

Three subgroups were identified and their distinct mediation pathways were revealed. Subgroup 1 had lower depressive symptoms, higher income, better community support and more urban residents. In this subgroup, education was positively associated with income and community support. Higher income and higher community support were associated with fewer depressive symptoms. Subgroup 2 had more severe depressive symptoms, more rural residents and higher rates of chronic conditions and body pain. In this subgroup, neither the direct nor the indirect effects of education were statistically significant. Subgroup 3 had the lowest depressive symptom severity, more rural residents and lower income. In this subgroup, education had a direct protective effect and the indirect pathway through community support was also significant.

Conclusions

Income and community support mediate education-depression associations with subgroup heterogeneity. Policies enhancing education, income and community resources could mitigate depressive symptoms.

Trial registration number

Not applicable (secondary analysis); CHARLS ethical approval: IRB00001052-11015 (Peking University Biomedical Ethics Committee).

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This study used a nationally representative sample of middle-aged and older adults from the China Health and Retirement Longitudinal Study.

• A heterogeneous structural equation model that simultaneously detects subgroups and selects key mediators was employed, which estimates mediation effects in a completely data-driven manner without imposing assumptions on the causes of heterogeneity.

• Informal social support was assessed using only dichotomous variables, which may not fully reflect the true degree of support received.

• The generalisability of these findings is limited by the Chinese middle-aged and older adult sample and the cross-sectional design, which prohibits causal conclusions.

Introduction

Depressive symptoms among older adults are becoming more common globally. WHO estimates that between 10% and 15% of those aged over 65 experienced depressive symptoms, with rates as high as 45% in certain regions.¹ Depression in later life exacerbates the outcomes of numerous medical conditions, increases healthcare costs and places additional strain on both family support systems and healthcare resources. In China, about 22.7% of Chinese seniors aged 60 and above are affected by depression, impacting not just their health but also straining families, communities and the healthcare system.² To develop effective interventions for managing depression among the elderly, it is vital to explore the relationship between depressive symptoms and potential risk factors. This involves not just recognising factors that have a significant impact on depressive symptoms but also understanding the pathway through which the risk factors affect depression and how they are inter-related.³

Recent studies have extensively examined the effect of education level on depressive symptoms and studied the underlying mechanisms involved. Bjelland et al conducted a study to determine the relationship between education level and depression among Norwegian adults and to examine if this relationship is mediated by other factors.⁴ Their longitudinal analysis showed that lower levels of education were significantly linked to both anxiety and depression. Of the factors mediating this relationship, economic hardship was found to substantially reduce the observed effect of education on depression. Another study investigated the effect of education level on depressive symptoms in Chinese older adults.⁵ It was found that education level influenced depressive symptoms through parallel mediating effects of economic security level and subjective memory ability. Other studies have examined socioeconomic status (eg, education level, household income and occupational status) as a moderator and explored the mediating role of social support in depressive symptoms. Zhou et al revealed that the relationship was partially mediated by social support, indicating that higher social support was negatively associated with depressive symptoms.⁶ Lian et al found that social frailty including infrequent social activity and infrequent contact was associated with incidence of depressive symptoms.⁷ However, as pointed out by Schaan, while education, work and income are commonly used as interchangeable measures of socioeconomic status, education is the foundational factor influencing employment opportunities, income level and wealth accumulation.⁸ In turn, higher occupational status and income levels are positively correlated with improved mental health outcomes. Therefore in this study, we chose education level as the predictor variable, considered income, social support as mediators and depressive symptoms as the outcome variable.

Understanding heterogeneity within a population is also an important topic in social science research.⁹ As such, it is of considerable interest to investigate the heterogeneity of the mediating pathway among education level, social support, income and depressive symptoms. You et al found that there was a huge gap in the health-related quality of life between urban and rural middle-aged and elderly Chinese. Income, education, occupation and frequency of contact with children play positive mediating roles in the difference of health-related quality between urban and rural older residents.¹⁰ Yang et al revealed that the association between social engagement and depression symptoms differed in people with different sociodemographic characteristics and health conditions.¹¹ The subgroups identified, who were middle-aged and elderly residents living in rural areas and had <6 hours of sleep, as well as those residing in urban areas, may gain greater benefits from social engagement compared with the entire population. Another study examined the temporal trend and subgroup variations in the prevalence and treatment of depression among middle-aged and older adults in China.¹² The study identified disparities in prevalence based on age, gender and province.

In many practical studies, the groups are not explicitly defined in advance of analysis. Therefore, an automatic subgroup detection method is essential for specifying heterogeneous mediating structures and each individual’s subgroup membership. To the best of our knowledge, no study has yet integrated mediation analysis, subgroup detection and pathway identification to comprehensively investigate the relationship between education level, depressive symptoms and other influencing factors. This analysis framework is essential for the development of targeted public health strategies and precise interventions. To address this goal, we used a heterogeneous mediation analysis approach proposed by Xue et al.¹³ For the first time, this method is implemented to analyse the mechanism underlying depression in the elderly population. We visualised the subgroup characteristics and differences, and provided a more comprehensive understanding of the population heterogeneity. Moreover, the bootstrap method was incorporated to assess the statistical significance of the identified pathways.

Methods

Data source

This study used data from the China Health and Retirement Longitudinal Study (CHARLS) study, which focuses on China’s middle-aged and elderly populations, drawing from a sample of households with members aged 45 years or above.¹⁴ We used the 2011 baseline survey data for subgroup detection and disparity analysis because it captures participants’ initial health, socioeconomic and demographic characteristics prior to subsequent changes, such as the COVID-19 pandemic in 2019. By anchoring our analysis in the baseline wave, we minimise biases introduced by later attrition, including selective dropout due to mortality, illness or non-response, which may overrepresent healthier or more stable individuals in follow-up waves. The 2011 data thus provides the most complete and representative cohort for identifying meaningful subgroups before longitudinal attrition affects sample composition. The sample exclusion criteria of this study were as follows: (1) Individuals with a high rate (more than 30%) of missing data in critical modules,^{15 16} where imputation would be inaccurate, (2) Individuals aged below 45 and (3) Individuals who had three or more missing items on the Centre for Epidemiologic Studies Depression Scale (CES-D) form. After these exclusions, the final sample size was 8161. Figure 1 described the detailed information of the study sample screening flowchart. Table 1 summarises the definitions and measurement details for all study variables.

Figure 1. Flow chart for the inclusion and exclusion criteria. CES-D, Centre for Epidemiologic Studies Depression Scale.

Table 1. Description of the independent variable, mediators and confounders.

Variable	Definition
Education	Education level (1: elementary school or above, 0: otherwise)
Live with children	Live with all children (1: yes, 0: no)
Weekly contact	Have weekly contact with children not living together (1: yes, 0: no)
Transfer	Any transfer from non-coresident family members (1: yes, 0: no)
Income	Household per capita income (in Yuan)
Pension	Receiving any pension type, excluding wages and basic firm pensions (1: yes, 0: no)
Health insurance	Possessing any health insurance type, including private and government plans (1: yes, 0: no)
Community activity 1	Whether the village/community has dancing team or other exercise organisations (1: yes, 0: no)
Community activity 2	Whether the village/community has room for card and chess games (1: yes, 0: no)
Community activity 3	Whether the village/community has outside exercising facilities (1: yes, 0: no)
Community activity 4	Whether the village/community has elderly care centre (1: yes, 0: no)
Community activity 5	Whether the village/community has elderly association (1: yes, 0: no)
Community activity 6	Whether the village/community has activity centre for the elderly (1: yes, 0: no)
Community activity 7	Whether the village/community has organisations for helping the elderly and the handicapped (1: yes, 0: no)
Smoke	Smoking or not (1: yes, 0: no)
Alcohol	Drinking any alcoholic beverages last year (1: drinking more than once a month, 0: for less than once a month or none)
Social activity	Attain any social activity (1: engaged in group activities; 0: none)
Hypertension	Have been diagnosed with hypertension (1: yes, 0: no)
Dyslipidaemia	Have been diagnosed with dyslipidaemia (1: yes, 0: no)
Diabetes	Have been diagnosed with diabetes or high blood sugar (1: yes, 0: no)
Lung disease	Have been diagnosed with chronic lung disease (1: yes, 0: no)
Heart problem	Have been diagnosed with heart problems (1: yes, 0: no)
Kidney disease	Have been diagnosed with kidney disease (1: yes, 0: no)
Digestive disease	Have been diagnosed with stomach or other digestive disease (1: yes, 0: no)
Arthritis	Have been diagnosed with arthritis or rheumatism (1: yes, 0: no)
Disability	Any physical disabilities/brain damage/mental retardation/vision problem/hearing problem/speech impediment (1: yes, 0: no)
Pain	Currently feel any body pains (1: yes, 0: no)
Activity limitation	Any limitation of activities of daily living (1: yes, 0: no)
Urban residence	Living area (1: urban community, 0: rural village)
Sex	1: female, 0: male
Age	In years
Marital status	Marital status (1: married with spouse present, 0: others)
BMI	Body mass index
Retirement status	Processed retirement status (1: yes, 0: no)

To identify outliers in the continuous variables, we applied the median absolute deviation (MAD) rule, defining outliers as values falling outside the range of median±5 × MAD. Rather than excluding these cases, which would reduce statistical power and potentially bias the sample, we treated them as missing values. They were subsequently imputed together with other missing data using the multivariate imputation by chained equations method, implemented with the R package mice. This package creates multiple imputations for missing data using the fully conditional specification approach, where each incomplete variable is imputed with its own tailored model. It can handle various data types and provides diagnostic tools to assess imputation quality. For the imputation of each incomplete variable, all variables listed in table 1 were incorporated into the imputation model.

Outcome variable

Depressive symptoms were evaluated based on the CES-D form, which is a widely used self-report scale designed to measure the level of depressive symptomatology in the general population. CHARLS participants were asked the frequency of experiencing the following items to evaluate how they felt and behaved over the previous week: (1) I was bothered by things that don’t usually bother me, (2) I had trouble keeping my mind on what I was doing, (3) I felt depressed, (4) I felt everything I did was an effort, (5) I felt hopeful about the future, (6) I felt fearful, (7) My sleep was restless, (8) I was happy, (9) I felt lonely and (10) I could not get “going”. Each item was measured using a 4-point Likert scale from 0 (rarely or none of the time) to 3 (most or all of the time). Items 5 and 8 were reversed-scored. The total score ranges from 0 to 30 with higher scores indicating greater severity of depressive symptoms.

Independent variable and mediators

Education was the independent variable in this study and it was measured by the highest level of education attained. The mediators considered in our study were income and social support, with the latter comprising three components: informal social support, community support and public support. Previous studies have treated social support as a multidimensional construct that includes informal social support, community support and public support.^{6 17} Informal social support refers to the assistance and emotional comfort an individual receives from their family. Community support refers to the resources and services provided to individuals within the local village or community. Public support (also referred to as formal support) refers to assistance and benefits provided by government agencies. Informal social support was measured using three indicators: whether participants lived with all children, whether they had weekly contact with children not living together and whether they had received any transfer from non-coresident family members. This measurement approach is consistent with previous research, where informal social support was operationalised based on two key aspects: the frequency of contact with children and the receipt of financial transfers from family members living outside the household.18,20 Previous studies also used binary measures of contact with children621,23 and financial support from family members.^{24 25} Prior research indicates that individuals living with their children are generally assumed to have regular contact with them.²¹ Therefore, living with all children and having weekly contact with non-coresident children both reflect parent-child contact and we include both measures to capture the dimension of emotional support within informal social support. Furthermore, transfer from non-coresident children reflects economic support from children. We summed these three indicator variables to create a composite measure representing the overall level of informal social support. Community support was evaluated through the presence of external exercise facilities, dance or exercise groups, organisations aiding the elderly and handicapped, activity centres for the elderly and elderly associations in the community or village. These variables were summed to represent the comprehensive level of community support. Public support was measured by two indicators related to formal social security and welfare for older adults: whether the participant had a pension and whether the participant had medical insurance or welfare. These two indicators were summed to measure overall public support, following the operationalisation established in prior research.^{6 17}

Confounding variables

In addition to the independent variable and mediators, we also incorporated various confounding variables, including lifestyle choices (smoking, alcohol consumption and social activity), health conditions (hypertension, dyslipidaemia, diabetes, lung disease, heart problem, kidney disease, digestive disease, arthritis, disability, pain and activity limitation) and demographic backgrounds of the participants (urban residence, sex, age, marital status, body mass index (BMI) and retirement status).

Statistical analysis

The heterogeneous structural equation modelling method with automatic subgroup detection¹³ was implemented. Let $X_i$ represent the education level of subject $i$ and $Y_i$ represent the depressive symptom score. Denote ${\mathit{M}}_i$ as the potential mediators: income, informal social support, community support and public support. Let ${\mathit{Z}}_i$ denote the confounders. Suppose that the population can be divided into $H$ subgroups, each possessing its own mediating pathway. The value of $H$ is not known in advance and we can determine the optimal number of subgroups based on some model selection criteria. The structural equation model can be written as

$${\displaystyle {\displaystyle \begin{array}{c}{\mathbf{M}}_i={\mathbf{b}}_h{X}_i+\mathrm{\Gamma }{\mathbf{Z}}_i+{\delta }_i,\\ Y_i={\beta }_h{X}_i+{\theta }_h^T{\mathbf{M}}_i+{\gamma }^T{\mathbf{Z}}_i+{\epsilon }_i,\end{array}}}$$ (1)

where ${\beta }_h$ is the direct effect of education level within subgroup $h$, and ${\theta }_h^T{b}_h$ represents the overall mediation effect of education level transmitted via the mediators. $\Gamma$ and $\gamma$ are the effects of the confounders on the mediators and depressive symptoms, respectively.

We detected the subgroups, identified important mediating pathways and estimated the effects simultaneously by minimising the objective function in (2). The first term handles subgroup assignment by evaluating each participant’s fit to different subgroup models. Participants are more likely to belong to subgroups whose models fit their data best. The second term merges similar groups into shared model forms, reducing redundant subgroups. The third component identifies key mediators within subgroups, while the final component selects relevant confounders.

$${\displaystyle {\displaystyle {\displaystyle \begin{array}{rl}L(\theta )=& \sum _{i=1}^N\underset{1\le h\le H}{min}\{\mathrm{trace}[({\mathbf{M}}_i-{\mathbf{b}}_h{X}_i-\mathbf{\Gamma }{\mathbf{Z}}_i{)}^T({\mathbf{M}}_i-{\mathbf{b}}_h{X}_i-\mathbf{\Gamma }{\mathbf{Z}}_i)]\\ +{\left(Y_i-{\beta }_h{X}_i-{\theta }_h^T{\mathbf{M}}_i-{\gamma }^T{\mathbf{Z}}_i\right)}^2\}\\ +{\lambda }_0\sum _{1\le h_1,h_2\le H}\left(|{\beta }_{h_1}-{\beta }_{h_2}|+\Vert {\theta }_{h_1}-{\theta }_{h_2}{\Vert }_1+\Vert {\mathbf{b}}_{h_1}-{\mathbf{b}}_{h_2}{\Vert }_1\right)\\ +{\lambda }_1\sum _{j=1}^P\left\{1-\frac{1}{(1+c_0|b_{h,j}|)(1+c_0|{\theta }_{h,j}|)}\right\}\\ +{\lambda }_2\left\{\Vert \gamma {\Vert }_1+\sum _{j=1}^p\Vert {\mathbf{\Gamma }}_j{\Vert }_1\right\}\end{array}}}}$$ (2)

A criterion similar to the Bayesian information criterion (BIC) was used to determine the optimal values of the tuning parameters and the number of subgroups. Prior to analysis, all continuous mediators, confounders and the outcome variable were standardised. This was achieved by subtracting the sample mean from each value and dividing by the sample SD. Binary variables were not standardised and were kept in their original coding.

To quantify uncertainty and assess the stability of variable selection, we conducted a bootstrap analysis by resampling the data 100 times. Each time we drew a bootstrap sample of the same size as the original dataset with replacement (bootstrap sample size=8161). We then fit the model in (1) and obtained subgroup membership and coefficient estimates for each bootstrap sample. A permutation issue may arise when identifying subgroups across bootstrap samples: a subgroup labelled as “1” in one sample may correspond to subgroup “2” in another. This occurs because subgroup numbers are merely nominal labels without practical meanings. To solve this problem and ensure consistent subgroup interpretation across bootstrap samples, we aligned the subgroup labels in each bootstrap sample with those derived from the original dataset. This alignment was achieved by permuting the subgroup labels to maximise the overlap of individuals within each group between the bootstrap sample and the original dataset. After aligning the subgroups, we calculated a selection frequency for each variable in the outcome model (the second model in Equation (1)) to evaluate the variable selection stability. This frequency represents the percentage of bootstrap samples in which the variable was selected as important (ie, with a non-zero effect), with higher values indicating more stable selection.

To determine the statistical significance of the direct effects, indirect effects and confounder effects, we then computed their 95% CIs using the percentile method. The 2.5th and 97.5th percentiles of the 100 bootstrap estimates were taken as lower endpoint and upper endpoint of the interval. This percentile method provides an empirical CI without assuming a specific distribution of the data. Statistical significance was determined by checking whether zero fell outside this interval; if the interval excluded zero, we concluded that the effect was statistically significant at the 5% level.

Patient and public involvement

None.

Results

Detected subgroups

The optimal number of subgroups as determined by the minimisation of the BIC criterion was three. The first subgroup consisted of 2439 individuals, the second subgroup comprised 2445 individuals and the third subgroup included 3277 individuals. To better comprehend the characteristics of the subgroups, we illustrate the distribution variations of some key factors across all three groups, as presented in figure 2.

Figure 2. Comparison of depressive score, income, BMI, informal social support, community support, public support, education level, married status, sex, urban residence, arthritis and pain in the three groups. (In all subfigures, the left, middle and right bars represent groups 1, 2 and 3, respectively. For subfigures (a–c), the y-axis ranges from a lower level to a higher level. In subfigures (d–f), the y-axis shows the cumulative percentage, and the coloured segments from top to bottom correspond to increasing levels of the variable. In subfigures (g–l), green denotes a value of zero, while orange denotes a value of one). BMI, body mass index.

Subfigure (a) indicates that the peaks of the distributions of depressive symptom scores for both group 1 and group 3 were <12. Given that the threshold for defining a depressive episode is 12 points in the CES-D score,¹⁰ this suggests that the majority of individuals in these two groups were not suffering from depression. Moreover, a greater proportion of individuals in subgroup 3 exhibited low scores of depressive symptoms. In contrast, group 2 displays a bimodal distribution with two peaks. The second peak significantly exceeds the threshold of 12 points in the CES-D score. This implies that the severity of depressive symptoms was notably higher in group 2.

Subfigure (b) shows that, while the peaks of the income distributions of the three subgroups were in approximately the same position, the spreads in the distributions were significantly different. The distribution for group 1 is more dispersed, with a larger number of individuals with higher income. The distributions for groups 2 and 3 were more concentrated, with more individuals receiving lower income.

Subfigures (e) and (j) reveal that group 1 enjoyed higher levels of community support and had the highest proportion of urban residents compared with the other groups. Regarding chronic conditions and body pain, subfigures (k) and (l) suggest that group 2 has the highest proportion of individuals experiencing arthritis and body pain. As for the remaining figures, the distributions of education level, informal social support, public support, BMI, marital status and sex were similar among all subgroups.

Identified mediation pathway

Our primary objective is to examine variations in depressive symptom scores attributable to education level, while accounting for mediators (eg, income and social support components) and potential confounders. Hence the unstandardised coefficients which correspond to the original scale of the data are more suitable for practical interpretations. Therefore, we transformed the estimated effects back to their original scale for reporting in this section. The estimated effects of education and confounding factors on both income and social support components are presented in table 2. The estimated effects of education, income, social support components and confounding factors on depressive symptoms are shown in tables3 4. To provide a clear visualisation of the mediation pathway, we further summarise these results and present the estimated coefficients and the identified pathway for the three subgroups in figure 3.

Table 2. Heterogeneous effects of the independent variable and homogeneous effects of confounders on the mediators.

	Income	Informal social support	Community support	Public support
Group 1, education	8283.8289	0	2.3994	0
Group 2, education	1317.5390	0	−0.7818	0
Group 3, education	−4326.1199	0	−0.3801	−0.0284
Smoke	−99.2357	−0.0128	−0.2400	0
Alcohol	0	0	−0.1084	0
Social activity	62.4997	0	0	0.0125
Hypertension	−206.479	0	0	0
Dyslipidaemia	0	0	0	0
Diabetes	0	0	0	0
Lung disease	0	0	0	0
Heart problem	0	0	0	0
Kidney disease	0	0	0	0
Digestive disease	0	0	0	0
Arthritis	−240.9617	−0.0011	−0.0804	0
Disability	−938.7070	0	−0.1209	−0.0023
Pain	−1003.2334	0	−0.2473	0
Activity limitation	0	0	0	0
Urban residence	1199.1188	0	1.6414	−0.0306
Sex	0	0.0402	−0.1274	−0.0015
Age	−106.5156	0.0172	0.0010	0.0044
Marital status	−535.4398	−0.0026	−0.1660	0.0022
BMI	92.0997	0	0.0141	0.0031
Retirement status	3727.6835	−0.0260	0.0462	0.1044

BMI, body mass index.

Table 3. Heterogeneous direct effects of the independent variable and heterogeneous mediator effects on the outcome.

	Depressive symptoms
	Group 1	Group 2	Group 3
Education	0	−1.2902	−1.1343
Income	−3.1194e-06	−3.6654e-04	6.6151e-05
Informal social support	0	0	0
Community support	−0.7835	−3.0878	1.8775
Public support	0	0	0

Table 4. Homogeneous confounder effects on the outcome.

Depressive symptoms
Smoke	0	Alcohol	0	Social activity	−0.3817	Hypertension	0
Dyslipidaemia	0	Diabetes	0	Lung disease	0	Heart problem	0
Kidney disease	0	Digestive disease	0.4125	Arthritis	0.8320	Disability	0.410
Pain	3.0946	Activity limitation	0	Urban residence	0	Sex	0.4328
Age	0	Marital status	−1.0814	BMI	−0.0582	Retirement status	0

BMI, body mass index.

Figure 3. The identified mediation pathway. Subfigure (a) is the mediation pathway for subgroup 1. Subfigure (b) is the mediation pathway for subgroup 2. Subfigure (c) is the mediation pathway for subgroup 3.

Figure 3A corresponds to subgroup 1, characterised by low depressive symptoms, higher proportions of individuals with greater annual income and community support, and the greatest concentration of urban residents. We can see that there was no mediating pathway through informal social support and public support. And the education level did not have a direct impact on depressive symptoms. Specifically, the mediator model showed that there was a positive association between education level and household per capita income. In the mediator model (the first model in Equation (1)), the independent variable education had an estimated effect of 8283.82 on the mediator income for subgroup 1. This indicates that individuals with an education level of elementary school or above earned an additional ¥8283.82 annually compared with those with a lower education level. A higher education level was also positively associated with the amount of community support an individual could receive. In the same model, the effect of education on the mediator, community support, was estimated to be 2.39. This suggests that people with an educational level of elementary school or higher received two to three more instances of community support. The outcome model showed that both income and the level of community support were negatively associated with the severity of depressive symptoms. In the outcome model (the second model in Equation (1)), the mediator income had an estimated effect of −3.11e-06 on the outcome variable, depressive symptoms. On average, an increase of ¥8283.82 in income was associated with a decrease of 0.025 points in the depressive score. In the same model, the estimated effect of community support on depressive symptoms was −0.78. Receiving two more components of community support was associated with a decrease of 1.56 points in the depressive score. Considering that the mode of depressive score in group 1 was approximately 9, the influence of education level via income and community support was actually noticeable. Therefore, policies aimed at enhancing education and income levels, as well as increasing community support, could potentially contribute to further prevention of depression.

Figure 3B corresponds to subgroup 2, which is characterised by more severe depressive symptoms, a higher proportion of rural residents and greater prevalence of chronic conditions and body pain. It is shown that both the mediating effect of education through income and the direct effect of education contributed to lower depressive symptoms in this subgroup. In the mediator model, the effect of the independent variable education on the mediator income was estimated to be 1317.53. This indicates that individuals with at least an elementary school education earned an additional ¥1317.53 annually. In the outcome model, the effect of the mediator income on the outcome variable depressive symptoms was estimated to be −3.66e-04. Accordingly, an increase of ¥1317.53 in income was associated with a 0.48-point decrease in the depressive score. In the outcome model, the direct effect of education on depressive symptoms was estimated to be −1.29, suggesting that individuals with at least an elementary school education had depressive scores 1.29 points lower than those with a lower education level. In the mediator model, the effect of education on community support was estimated to be −0.78. This indicates that individuals with an education level of elementary school or higher were likely to live in communities with slightly fewer supportive resources. In the outcome model, the effect of community support on depressive symptoms was estimated to be −3.08. This suggests that lower community support was associated with higher depressive symptoms.

Figure 3C represents subgroup 3, the group with the lowest severity of depressive symptoms, a greater proportion of rural residents and lower income levels, but a lower prevalence of chronic conditions and body pain. The direct effect of education suggests that a higher level of education was associated with lower scores of depressive symptoms. However, the estimated indirect effects through income and community support were contrary to expectations. In the mediator model, individuals who attained at least elementary school were likely to earn less, with the estimated effect of education on income being −4326.11. They were also likely to live in communities with slightly fewer support facilities, with the estimated effect of education on community support being −0.38. In the outcome model, lower income was associated with fewer depressive symptoms, with the estimated effect of income on depressive symptoms being 6.61e-05. Lower community support was also associated with fewer depressive symptoms, with the estimated effect of community support on depressive symptoms being 1.87. This may be a distinct characteristic of this subpopulation, which, despite having lower income, exhibits healthier conditions.

Significance of the effects

Using the bootstrap approach described in the Method section, we calculated the frequency with which each variable was selected as an important predictor for depressive symptoms and the 95% CIs of the effects. The results are summarised in tables 5 and 6. Some CIs in table 6 have zero as the upper endpoint; these results should be interpreted alongside the selection frequencies reported in table 5. For example, the CI for the direct effect of education on depressive symptoms in group 3 is (−3.9795, 0). This suggests a potential protective effect of education. While there is some uncertainty about the presence of an effect, table 5 indicates that education received a non-zero coefficient in 86% of bootstrap samples, suggesting that the effect is frequently estimated to be protective. Table 6 shows that education had significant indirect effects through income and community support in subgroup 1. Education also demonstrated a significant direct effect and a significant indirect effect via community support in subgroup 3. For the confounders, we can conclude that being female, having chronic conditions such as stomach or other digestive diseases, arthritis or rheumatism, possessing a physical disability and experiencing body pain were all significantly associated with more severe depressive symptoms. Conversely, being married and having a higher BMI were significantly associated with lower depressive symptom scores.

Table 5. Percentage of bootstrap resamples showing nonzero effects on depressive symptoms in the second model in Equation (1).

	Education (direct effect)	Income	Informal social support	Community support	Public support
Group 1	66%	94%	54%	79%	6%
Group 2	83%	92%	83%	87%	15%
Group 3	86%	68%	56%	92%	2%
	Smoke	Alcohol	Social activity	Hypertension	Dyslipidaemia
	16%	4%	93%	4%	0%
	Diabetes	Lung disease	Heart problem	Kidney disease	Digestive disease
	0%	46%	54%	3%	100%
	Arthritis	Disability	Pain	Activity limitation	Urban residence
	100%	100%	100%	74%	36%
	Sex	Age	Marital status	BMI	Retirement status
	100%	66%	100%	100%	0%

BMI, body mass index.

Table 6. 95% CIs for significant effects.

Direct effect	Group 3, education → depressive symptoms	(−3.9795, 0)
Indirect	Group 1, education → income → depressive symptoms	(−2.5478, 0)
Effect	Group 1, education → community support → depressive symptoms	(−2.3923, 0)
	Group 3, education → community support → depressive symptoms	(−2.4582, 0)
Confounder	Digestive disease → depressive symptoms	(0.2102, 0.9659)
Effect	Arthritis → depressive symptoms	(0.4392, 1.0383)
	Disability → depressive symptoms	(0.2781, 1.0858)
	Pain → depressive symptoms	(2.8692, 3.7877)
	Sex → depressive symptoms	(0.2141, 0.8832)
	Marital status → depressive symptoms	(−1.1610, −0.3385)
	BMI → depressive symptoms	(−0.1304, −0.0240)

BMI, body mass index.

Discussion

We used baseline survey data from CHARLS to examine subgroup disparities in the mediating pathways between educational attainment and the severity of depressive symptoms. Three distinct subgroups among the elderly Chinese population were identified. These subgroups demonstrate distinct patterns in the relationships between educational attainment, income, community support and depressive symptom severity, reflecting heterogeneous pathways in mental health disparities.

Subgroup 1, demonstrating lower depressive symptom severity, was characterised by significantly higher annual incomes, better community support and greater urban residency prevalence compared with other subgroups. Our findings based on this subgroup revealed a positive association between education level and household per capita income, as well as the amount of community support an individual receives. These factors, in turn, were negatively associated with the severity of depressive symptoms, suggesting that increased education and income, coupled with substantial community support, can mitigate depressive symptoms among older urban residents.

Subgroup 2, characterised by a higher severity of depressive symptoms, was predominantly composed of rural residents with lower income levels and a higher prevalence of chronic conditions and body pain. In this subgroup, we observed that both the indirect effects of education through income and the direct effects of education contributed to alleviating depressive symptoms. Higher education levels were associated with increased income, which subsequently corresponded to a decrease in depressive scores. Furthermore, the direct effect of education on depressive scores was negative, indicating that education itself can serve as a protective factor against depression. Interestingly, the indirect effects of education through community support suggested that individuals with an education level of elementary school or higher were likely to reside in communities that offer slightly fewer supportive resources. This lower level of community support was associated with higher levels of depressive symptoms, highlighting the importance of community resources in managing depressive symptoms among the rural elderly population. Pei et al investigated the longitudinal trajectories of depressive symptoms and found that individuals residing in rural areas, those with primary school education or less, non-participation in social activities and those suffering from multiple chronic diseases were more likely to exhibit higher and increasing levels of depressive symptoms.²⁶ Their identified subgroup shares some similarities with our subgroup 2 in terms of key characteristics.

Subgroup 3, with the lowest severity of depressive symptoms, a larger proportion of rural residents and lower income and a lower prevalence of chronic conditions and body pain, presented a complex scenario. A higher level of education was associated with lower depressive symptoms, yet the indirect effects through income and community support were somewhat counterintuitive. It was found that individuals with at least elementary school education were likely to earn less and reside in communities with fewer support facilities. Surprisingly, lower income and a lower level of community support were associated with fewer depressive symptoms. Despite these findings, it is important not to overlook the positive direct effect of education on reducing depressive symptoms. Policies that promote lifelong learning and provide accessible educational opportunities for the elderly could be beneficial. In addition, policies could focus on improving the quality of community support rather than just the quantity to ensure that it effectively meets the needs of the elderly.

Chen and Wang investigated depression prevalence trends and subgroup variations across China, finding significant geographic disparities.¹² These regional differences were attributed to several factors: (1) Health behaviours (alcohol use, physical activity), (2) Social support (trust networks, healthcare access), (3) Multimorbidity rates and (4) Economic conditions (income levels, welfare services, vulnerable populations like left-behind children and empty-nest elderly). The conclusions drawn from their research align partially with the findings of the present study, particularly in highlighting the significant roles of social support and income in subgroup identification. Tian et al found that participants with moderate, increasing or high depressive symptom trajectories had lower education levels, higher rural residence rates and greater comorbidity and disability burdens compared with those with low depressive symptoms.²⁷ Other subgroup analysis revealed that social engagement may be more effective in preventing depressive symptoms for people who live in the urban area.¹¹ These findings enhance our understanding of the characteristics of the subgroups identified through our analysis.

Our research makes several contributions to this field. First, to the best of our knowledge, this was the first study to explore the subgroup disparities in the mediating roles of income and community support in the association between education and depressive symptoms. Second, existing research on subgroup analysis typically employed a two-stage approach, first classifying subgroups based on specific features (eg, levels of depressive symptoms) and then analysing the association between subgroup membership and other risk factors. Unlike traditional two-stage approaches, the method we used in this study can simultaneously identify subgroups and analyse the mediation pathways, reducing potential bias from sequential modelling. Moreover, by integrating subgroup detection and mediation analysis within a unified modelling framework, our approach captures critical interdependencies that stepwise methods may overlook, thereby yielding more robust and interpretable results. Consequently, our approach facilitates clearer interpretation of how mediation pathways vary across subgroups. Third, our research underscores the importance of tailoring interventions based on the unique characteristics and needs of different subgroups and highlights the potential role of education in mitigating depressive symptoms and improving the mental health of older individuals.

Nevertheless, our study has several limitations that should be acknowledged. First of all, our analysis revealed that some of the proposed mediators were not significant. This null finding may reflect either a true absence of mediation effects in this context or methodological limitations in detecting subtle pathways. Future research could employ more advanced statistical approaches, such as post-selection inference. The null finding may also be due to the measurement of the mediator. For example, the variables used in this study for informal social support only captured the presence of financial transfers and interaction, rather than their amount or frequency. As a result, they may not have fully reflected the intensity of informal social support. This measurement limitation could partially account for the non-significant mediating effect observed. Second, while CHARLS provides longitudinal data, our analysis was restricted to a single wave due to the complexity of the modelling approach. This limits our ability to infer temporal or causal relationships. Nevertheless, the use of a single wave for population-based analysis is well established in the literature using CHARLS data.¹⁹ A potential extension of this work could involve a longitudinal mediation analysis with automatic subgroup detection that takes advantage of data from all visits. Lastly, participants with substantial missing data were removed, and our study included only individuals from the Chinese population. These factors may limit the generalisability of the findings.

Data availability statement

Data are available in a public, open access repository. The datasets used and analysed in the current study are available upon reasonable request from the official CHARLS website (https://charls.charlsdata.com/pages/data/111/en.html).