How do chronic diseases affect personal and household income? A double debiased machine learning analysis of the China health and retirement longitudinal study (CHARLS) in older adults

Abstract

Objectives

This study aimed to investigate the impacts of chronic diseases such as hypertension, dyslipidaemia and diabetes on personal and household income among ageing Chinese adults. The primary hypothesis was that these chronic diseases have differential effects on the socioeconomic status of individuals and households, with gender and age influencing these relationships.

Design

Prospective cohort study using double/debiased machine learning (DDML) techniques to analyse data from the China Health and Retirement Longitudinal Study (CHARLS).

Setting

Nationally representative sample of ageing Chinese adults, with data collected from multiple regions across China. The sample represents a variety of both urban and rural settings.

Participants

A total of 69 457 participants entered the study, with 69 457 completing it. The sample included both male and female participants, with the majority being of Han Chinese ethnicity. Participants were selected based on the presence of hypertension, dyslipidaemia and diabetes, and exclusion criteria included: no information on age (n=4307), no information on gender (n=12), no information on medical insurance (n=177).

Primary outcome measures

The primary outcome measures, as outlined in the study protocol, were the associations between three chronic diseases (hypertension, dyslipidaemia and diabetes) and personal income (LPI) as well as household income (LHI). These associations were measured using the DDML method, which provided both overall measurements and gender-specific subgroup analyses. There were no significant deviations between the planned and actual outcome measures, and all outcomes were assessed as originally intended.

Results

Dyslipidaemia was positively associated with LPI (coefficient=0.078, 95% CI 0.052 to 0.105) but negatively associated with LHI (coefficient=−0.049, 95% CI −0.084 to –0.015). Diabetes showed stronger positive effects on LPI (coefficient=0.093, 95% CI 0.052 to 0.135) and negative effects on LHI (coefficient=−0.094, 95% CI −0.147 to –0.041). Gender-specific analyses revealed that dyslipidaemia had a stronger association with LPI in males (95% CI 0.080 to 0.163) compared with females (95% CI 0.007 to 0.075). For diabetes, males experienced larger increases in LPI (95% CI 0.053 to 0.190) compared with females (95% CI 0.015 to 0.117). Additionally, reductions in LHI were more pronounced in females with diabetes (95% CI −0.187 to –0.043).

Conclusions

Chronic diseases, particularly dyslipidaemia and diabetes, significantly affect the socioeconomic status of ageing Chinese adults, with distinct gender-specific impacts. These findings highlight the importance of targeted interventions to address the income disparities linked to chronic diseases. Further research is needed to explore the long-term effects of disease management on socioeconomic outcomes.

Trial registration number

Prospective, observational, community-based cohort study using 2011–2018 CHARLS data from 28 provinces in mainland China, with the registration number IRB00001052-11015, following ethical approval from the Biomedical Ethics Committee of Peking University.

Introduction

The global population is rapidly ageing. It is estimated that the number of adults ≥65 years will double to 1.5 billion by 2050.¹ Meanwhile, the population aged 80 and above is expected to triple between 2019 and 2050, reaching 426 million.² Chronic diseases—including hypertension, dyslipidaemia and diabetes—pose significant global public health challenges by imposing considerable burdens on healthcare systems and diminishing patients’ quality of life.3,6 The prevalence of the three diseases is driven not only by biological factors but also by a range of socioeconomic and demographic determinants. According to the 2021 Global Burden of Disease study, the global mortality rates for hypertension, dyslipidaemia and diabetes were 0.138%, 0.047% and 0.068%, respectively. In China, these rates were 0.204%, 0.059% and 0.068%, higher than global rates (http://ghdx.healthdata.org/gbd-results-tool). Moreover, disparities in health insurance coverage may exacerbate these trends by limiting access to preventive care and timely interventions. Understanding these complex relationships is essential for healthcare personnel, who play a pivotal role in patient care, education and chronic disease management.⁷

Income and economic well-being among ageing populations are also pressing concerns in China. Chronic diseases are increasingly common among older adults, imposing significant financial burdens. While much of the existing literature has focused on the influence of income levels on health outcomes,8,13 few studies have explored how chronic diseases exert influence on individual income and household income, particularly regarding gender differences. This disease-income relationship is crucial to understand, as chronic conditions often result in increasing healthcare expenditures,^{14 15} reduced labour force participation^{16 17} and diminished earning potential.^{18 19} In China, individuals and families affected by chronic diseases face heightened economic insecurity.²⁰ For example, it was reported that ischaemic heart diseases were associated with higher medical costs, suggesting a direct link between chronic conditions and financial difficulties.²¹

Despite existing insights, significant gaps remain in understanding how hypertension, dyslipidaemia and diabetes impact income among ageing Chinese adults. To address this gap, we analyse data from the China Health and Retirement Longitudinal Study (CHARLS) using Double/Debiased Machine Learning (DDML) to examine how various chronic diseases affect the elderly income—both at the individual and household levels—and to explore the underlying reasons for these effects. This method relaxes rigid assumptions, enabling direct extraction of variable relationships.²² As an unbiased estimator based on machine learning, DDML effectively captures both causal linear and fully non-linear effects,²³ enhancing its versatility and robustness. By integrating causality through machine learning techniques, DDML facilitates the direct discovery of variable relationships, mitigates confounding effects and strengthens the validity of causal relationships.^{24 25}

In healthcare policy, identifying vulnerable populations to income loss due to chronic diseases can facilitate the design of targeted financial and healthcare support programmes. From a nursing and preventive care perspective, a better understanding of these economic impacts can assist healthcare personnel in reducing the long-term financial burdens on patients and their families. By shifting the focus from “income as a determinant of health” to “health as a determinant of income”, our research offers a novel perspective on the economic consequences of ageing-related chronic diseases, thereby supporting the development of more effective, context-specific health interventions.

Materials and methods

We conducted a prospective, quasi-experimental cohort study using DDML design, embedded within the nationally representative CHARLS (CHARLS, 2011–2018). The study included community-dwelling adults aged ≥45 years from 28 mainland provinces who were first diagnosed by a physician with stroke, heart disease, diabetes or cancer between waves. These individuals were matched to never-diagnosed controls based on pre-event mental health trends and baseline covariates. No randomisation or blinding was performed. The survey covers both urban and rural areas across the north-east, north, east, south-central, south-west and north-west regions of China and is publicly available.

Study sample

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• Used nationally representative longitudinal data, enhancing the generalisability of the findings, and employed a large sample size, improving statistical power.

• Applied Double/Debiased Machine Learning for robust statistical analysis.

• Included gender-specific subgroup analyses, providing detailed insights.

• Chronic diseases were measured using a dichotomous variable, and reliance on self-reported diagnoses may have introduced recall bias or misclassification.

• The study focused solely on single chronic diseases (eg, hypertension) and did not consider the effects of multiple chronic diseases (eg, comorbid diabetes or dyslipidaemia) on income.

This research used longitudinal data from the CHARLS across four waves (2011, 2013, 2015 and 2018), with a sample of 73 953. CHARLS is a biennial survey conducted by the National School of Development at Peking University, aiming to be representative of the Chinese residents aged over 45 years and their family members at the national level, both in urban and rural areas.²⁶ Then, CHARLS adopted a multistage, stratified, probability proportional to size random sampling method to extract 450 communities (including villages) from 150 counties/districts across 28 provinces of China. CHARLS collects extensive information at the individual and household levels.²⁷ All data were collected by face-to-face computer-aided personal interviews. The survey includes information on demographics, family, health status, healthcare and medical insurance, employment and household economy (income, consumption and wealth). CHARLS is unique in that few other surveys in China that collect data on health and disability using such a variety of measures.

Participants with missing critical demographic or socioeconomic information were excluded: 4307 samples with missing age data, 12 with missing gender information and 177 with incomplete medical insurance records, resulting in 69 457 samples. For disease-specific analyses, exclusion criteria were applied sequentially: (a) Hypertension cohort: after excluding 20 250 samples with missing hypertension status, 49 207 samples remained, comprising 13 571 hypertensive and 35 636 non-hypertensive samples, (b) Dyslipidaemia cohort: from the baseline cohort, 4387 participants without dyslipidaemia data were excluded, leaving 65 070 samples (6441 dyslipidaemia and 58 629 non-dyslipidaemia) and (c) Diabetes cohort: a total of 19 684 samples with incomplete diabetes records were removed, resulting in 49 773 samples, including 3723 diabetic and 46 050 non-diabetics.

For continuous control variables (age, household expenditure, education level), missing values were imputed using the mean of the observed data. This method preserved the sample size while minimising potential bias from complete case exclusion.

We recognise that in the CHARLS dataset, the panel is unbalanced (eg, some individuals exit in later waves, some new individuals enter), which raises potential concerns of loss to follow-up and sample selection. To address this, we adopted a DDML approach which (1) In the first stage uses machine learning to model the high-dimensional individual and time-varying covariates including participation/exit indicators, (2) Includes the exit/entry mechanism into the nuisance function estimation so as to account for non-random attrition and (3) In the second stage uses orthogonalised scores for causal effect estimation, thereby reducing bias due to the selection/attrition mechanism.^{28 29}

All statistical analyses were performed on the final cleaned datasets. Final sample sizes and subgroup distributions are summarised in figure 1.

Figure 1. Flowchart for study population selection. CHARLS, China Health and Retirement Longitudinal Study.

All data were openly available from the official CHARLS website (http://charls.pku.edu.cn/). Ethical approval for all the CHARLS waves was granted from the Institutional Review Board at Peking University (IRB approval numbers: IRB00001052-11014 and IRB00001052-11015). During the fieldwork, each respondent who agreed to participate in the survey was asked to sign two informed consent forms.

Measurement of main variables

Outcome variable: In our research, the primary outcome variable comprises total income over the past year, consisting of two principal components: total personal income (Outcome Variable 1) and total household income (Outcome Variable 2). Personal income is operationalised through two constituent elements: pension income and after-tax self-earned income (SEI). Pension income was measured through respondents’ self-reported receipts from three distinct sources: (a) Occupational pensions administered by governmental entities, organisational bodies, social insurance institutions or commercial insurance providers, (b) Supplementary corporate pension insurance programmes and (c) Commercial pension insurance schemes. After-tax SEI was calculated based on respondents’ declarations of their net salary and bonus earnings from the preceding year, which is a widely used approach in prior research.^{18 30} Household income was aggregated from five constituent categories in the survey instrument: wage income; personal transfer income (including pensions and old-age allowances); household agricultural income; household self-employment income; and household transfer income (encompassing subsidies from ecological conservation programmes such as farmland reforestation and agricultural subsidies). This multi-component measurement framework aligns with established methodologies in the field.³¹

Independent variables: the primary independent variables are binary indicators of chronic conditions. Specifically, Independent Variable 1 denotes whether a participant has hypertension, Independent Variable 2 denotes whether a participant has dyslipidaemia, and Independent Variable 3 denotes whether a participant has diabetes. These conditions are emblematic of chronic diseases due to their high epidemiological prevalence, pathophysiological interrelations and modifiability.32,34 New-onset cases of hypertension, dyslipidaemia and diabetes were defined based on participants’ self-reported doctor diagnoses and reported treatment. The other group was defined as no diagnosis or report of the disease. For new respondents, or for those who had not been previously diagnosed with any chronic disease at the time of the last survey, the following questions were asked sequentially: “Have you been diagnosed with hypertension by a doctor?”, “Have you been diagnosed with dyslipidaemia (defined as elevated low-density lipoprotein, triglycerides and total cholesterol levels, or a low level of high-density lipoprotein) by a doctor?”, “Have you ever been informed by a doctor that you have diabetes or elevated blood glucose (including impaired glucose tolerance and elevated fasting glucose)?”.

Control variables: our analysis controls for several covariates that may influence income, including gender,³⁵ age,³⁶ medical insurance status,³⁷ the year of data collection and total household expenditures over the past year. Macroeconomic shocks (eg, recessions, inflation), policy changes (eg, minimum wage laws) and technological shifts can systematically affect income trends. By controlling for the survey year, we account for time-specific unobservable factors. Household expenditure serves as a proxy for wealth, consumption patterns and financial obligations—factors that are often correlated with income. Including household expenditure as a control helps mitigate omitted variable bias (eg, wealthier households typically exhibit higher income and spending). The estimated effects were adjusted for four to five covariates. In our study, we chose not to modify or exclude zero income values. The presence of zero values is meaningful in the context of rural China, where some households may report zero income due to a lack of earnings, but may rely on savings, remittances or other non-monetary resources for their livelihood. These zero values reflect a real aspect of the economic landscape in these areas and should not be ignored. Given the significance of these zero values, we decided to include them as they are in our analysis, without transformation. We believe that removing or altering them would distort the representation of households with no current income but still with substantial economic activity or support from alternative sources. We log-transformed total household income, total household expenditure and total personal income because these variables are highly right-skewed and contain outliers. The log transformation helps reduce the influence of extremely large values and stabilises the variance, making comparisons across households more reliable and reducing the impact of a few extreme observations on the analysis. This approach is widely used in studies related to income and expenditure.38,40

Statistical analysis

Analyses were performed using Stata MP V.17.1 (Stata, College Station, Texas, USA). This study used independent-samples t-tests to compare differences between continuous variables that were normally distributed and expressed as means±SD. The χ² test—used when expected cell counts were <5—was employed to examine associations between categorical variables, which are presented as frequencies and percentages. Two-sided p values <0.05 were considered statistically significant.

This study conducted a sensitivity analysis by grouping the data based on gender.

DDML is a framework designed to provide more accurate estimates in statistical models, particularly when handling high-dimensional data. As outlined in the literature,^{23 29} DDML corrects for the bias that arises from overfitting when estimating causal effects models in high-dimensional settings. This method has been widely applied to estimate health intervention effects and policy impacts. For example, one study leveraged DDML to estimate weighted cumulative treatment effects in time-to-event outcomes, demonstrating its utility in observational studies where randomised controlled trials are not feasible.⁴¹ Similarly, another study employed DDML in mobile health intervention policies, highlighting its potential in personalised medicine and health policy evaluation.⁴²

To implement this framework, we specified a Partially Linear Model using the ddml package. The process can be divided into two key stages:²⁹

1. Nuisance Parameter Estimation: this stage focuses on estimating confounding variables that are not of primary interest but may influence the outcome. We employed Gradient Boosting Regressors (via the ‘pystacked’ command) to estimate the conditional expectations of the outcome E(Y|X) and the treatment E(D|X). To avoid overfitting, we used fivefold cross-fitting.

2. Primary Parameter Estimation: this stage focuses on estimating the effect of treatment variable, adjusted for the nuisance parameters. Using the residuals from the first stage, the final causal effect was estimated by regressing the residualised outcome on the residualised treatment variable.

We employ DDML as our primary analytical tool. DDML not only facilitates the direct extraction of relationships between input and output variables from datasets, but also effectively mitigates confounding effects, thereby enhancing the validity of the causal relationships.^{24 25} A salient feature of these models is their ability to maintain interpretability, enabling researchers to rigorously analyse the impacts of various causal factors or policies. This evolution represents a significant advancement in computational data analysis, bridging the gap between empirical data and causal understanding. First, DDML generally outperforms traditional statistical methods (eg, difference-in-differences). Second, DDML relaxes the strict assumptions of traditional models by learning relationships between variables directly from the data.²² Notably, DDML is structured around machine learning algorithms and is classified as an unbiased estimator. In our study, we set a random seed (the default seed=42) for reproducibility which is aligned with prior research. We used gradient boosting to perform the fitting process, leveraging their capacity to capture complex, nonlinear relationships. In addition to capturing causal relationships, DDML is also capable of modelling fully nonlinear effects.²³

Our research used the DDML method to investigate the relationship between chronic conditions and income. This approach effectively eliminates canonical bias by employing DDML in the estimation of auxiliary equations, thereby overcoming the challenge of dimensionality associated with an excessive number of control variables in traditional linear regression. Moreover, it improves the accuracy of estimation by using non-parametric machine learning models to capture nonlinear relationships without pre-specifying the functional form between variables.²⁹

Patient and public involvement

None. This study employed secondary data analysis using the CHARLS dataset. As the research used pre-collected, de-identified data, patient and public involvement in the study design, implementation or dissemination phases was not applicable.

Result

The samples of hypertension, dyslipidaemia and diabetes were 49 207, 65 070, and 49 773, respectively. Significant gender differences were observed across all conditions (hypertension: p<0.001; dyslipidaemia: p<0.05; diabetes: p<0.001). Males constituted 48.27% of the non-hypertension group and 46.11% of the hypertension group, while females showed higher proportions in all disease groups (eg, 53.89% in hypertension). Age was significantly higher in disease groups compared with non-disease groups (p<0.001). Log-transformed household expenditure (LHE) and income (LHI) exhibited significant differences across groups, with lower LHI in hypertension (p<0.001) and higher LHE in dyslipidaemia (p<0.001). Medical insurance coverage exceeded 93% in all groups, with minor but significant variations (p<0.001) (table 1).

Table 1. Baseline characteristics and intergroup comparisons of samples with hypertension, dyslipidaemia and diabetes.

Variable	Total n=49 207		χ2/t	Total n=65 070		χ2/t	Total n=49 773		χ2/t
	Non-hypertension n=35 636	Hypertension n=13 571		Non-dyslipidaemia n=58 629	Dyslipidaemia n=6441		Non-diabetes n=46 050	Diabetes n=3723
Gender			18.432***			5.06*			12.330***
Male	17 201 (48.27%)	6257 (46.11%)		28 163 (48.04%)	2999 (46.56%)		22 082 (47.95%)	1674 (44.96%)
Female	18 435 (51.73%)	7314 (53.89%)		30 466 (51.96%)	3442 (53.44%)		23 968 (52.05%)	2049 (55.04%)
Age	59.26±9.95	63.21±9.98	−39.270***	60.02±10.48	61.07±9.28	−7.729***	60.34±10.26	62.59±9.28	−12.980***
LHE	10.16±1.03	9.96±0.98	19.464***	10.09±0.98	10.26±0.98	−13.395***	10.16±1.03	10.18±0.96	−0.807
LHI	8.86±1.55	8.58±1.48	17.776***	8.66±1.51	8.66±1.47	−0.163***	8.85±1.56	8.63±1.43	8.421***
LPI	8.89±1.30	8.73±1.24	11.434***	8.82±1.23	8.96±1.25	−9.138***	8.82±1.32	8.89±1.27	−2.746**
HLEA	6.04±4.07	5.83±3.70	5.153***	5.84±3.61	6.61±3.69	−16.292***	5.94±4.09	6.48±3.82	−7.848***
Medical insurance			2.757			29.136***			0.051
No	2032 (5.70%)	827 (6.09%)		3774 (6.44%)	304 (4.72%)		2520 (5.47%)	207 (5.56%)
Yes	33 604 (94.30%)	12 744 (93.91%)		54 855 (93.56%)	6137 (95.28%)		43 530 (94.53%)	3516 (94.44%)

Significance levels are denoted as follows: *p<0.05, **p<0.01, ***p<0.001.

HLEA, Highest Level of Education Attained; LHE, Logarithm of past year’s total household expenditure; LHI, Logarithm of past year’s total household income; LPI, Logarithm of past year’s total personal income.

In our research, distinct associations were observed between health conditions and income variables. Hypertension showed no significant association with LPI (coefficient=0.012, 95% CI −0.012 to 0.036) or LHI (coefficient=−0.016, 95% CI −0.049 to 0.016). Dyslipidaemia was positively associated with LPI (coefficient=0.078, 95% CI 0.052 to 0.105) but inversely associated with LHI (coefficient=−0.049, 95% CI −0.084 to –0.015). Diabetes demonstrated stronger positive effects on LPI (coefficient=0.093, 95% CI 0.052 to 0.135) and negative effects on LHI (coefficient=−0.094, 95% CI −0.147 to –0.041) (table 2).

Table 2. Associations between chronic conditions and income-related variables.

	Coef.	SD	Z	95% CI
Hypertension
LPI	0.012	0.012	0.990	−0.012	0.036
LHI	−0.016	0.017	−0.990	−0.049	0.016
Dyslipidaemia
LPI	0.078	0.014	5.750	0.052	0.105
LHI	−0.049	0.018	−2.780	−0.084	−0.015
Diabetes
LPI	0.093	0.021	4.380	0.052	0.135
LHI	−0.094	0.027	−3.510	−0.147	−0.041

Horizontal lines are 95% CI.

Description of each variable in the formula: Y (outcome variable 1: LPI, outcome variable 2: LHI), D (treatment variable: hypertension or not; dyslipidaemia; presence or absence of diabetes), X (control variables: gender, age, LHE, HLEA are controlled).

HLEA, Highest Level of Education Attained; LHE, Logarithm of past year’s total household expenditure; LHI, Logarithm of past year’s total household income; LPI, Logarithm of past year’s total personal income.

For hypertension, gender exhibited a null association with LPI. Dyslipidaemia was strongly associated with LPI in males (95% CI 0.080 to 0.163) but weakly in females (95% CI 0.007 to 0.075). Diabetes displayed gender-specific effects, with larger LPI increases in males (95% CI 0.053 to 0.190) compared with females (95% CI 0.015 to 0.117). LHI reductions were more pronounced in females with diabetes (95% CI −0.187 to –0.043) (table 3).

Table 3. Gender analysis of chronic conditions and income-related variables.

	Gender	Coef.	SD	Z	95% CI
Hypertension	LPI
	Male	0.035	0.019	1.800	−0.003	0.072
	Female	−0.006	0.015	−0.400	−0.036	0.024
	LHI
	Male	0.001	0.024	0.040	−0.046	0.048
	Female	−0.030	0.023	−1.320	−0.075	0.015
Dyslipidaemia	LPI
	Male	0.122	0.021	5.680	0.080	0.163
	Female	0.041	0.017	2.370	0.007	0.075
	LHI
	Male	−0.061	0.026	−2.350	−0.113	−0.010
	Female	−0.044	0.024	−1.820	−0.092	0.003
Diabetes	LPI
	Male	0.121	0.035	3.480	0.053	0.190
	Female	0.066	0.026	2.530	0.015	0.117
	LHI
	Male	−0.068	0.039	−1.710	−0.145	0.010
	Female	−0.115	0.037	−3.150	−0.187	−0.043

Horizontal lines are 95% CI.

Description of each variable in the formula: Y (outcome variable 1: LPI, outcome variable 2: LHI), D (treatment variable: hypertension or not; dyslipidaemia; presence or absence of diabetes.), X (control variables: age, LHE, HLEA are controlled).

HLEA, Highest Level of Education Attained; LHE, Logarithm of past year’s total household expenditure; LHI, Logarithm of past year’s total household income; LPI, Logarithm of past year’s total personal income.

Discussion

Our research investigated the associations between chronic health conditions (hypertension, dyslipidaemia and diabetes) and income. The results also revealed notable gender differences: females exhibited a higher prevalence of these conditions, and older age emerged as a consistent risk factor across all disease groups. Income disparities were also apparent; individuals in the hypertension group tended to have lower household incomes, whereas those with dyslipidaemia reported higher household expenditures. Furthermore, gender analyses uncovered distinct patterns. For instance, dyslipidaemia showed a stronger association with personal income among males, while females with diabetes experienced more pronounced reductions in household income, a finding that may be linked to traditional caregiving roles. Overall, these findings underscore the complex relationships among chronic diseases, income and gender, thereby providing a robust foundation for the development of targeted interventions and policy reforms.

Our results indicate that females exhibit a higher prevalence of chronic conditions—for example, a 53.89% prevalence of hypertension—which aligns with global trends indicating that women often bear the dual burdens of biological vulnerability and caregiving responsibilities.^{43 44} The results also identified significant gender differences in the prevalence of hypertension, dyslipidaemia and diabetes, with women demonstrating higher rates of hypertension and diabetes. Furthermore, postmenopausal hormonal changes may exacerbate hypertension risk.^{45 46} Nevertheless, our results may also be inconsistent with those of some studies. For instance, it was found that men are more likely to develop hypertension and diabetes at an earlier age than women, particularly in Western populations.⁴⁷ These discrepancies may indicate that age distribution, lifestyle, regional, cultural and healthcare accessibility factors may exert influence on the gender-specific prevalence of chronic diseases.

Hypertension exhibits a null association with income in our adjusted analyses, diverging from previous studies that linked hypertension to productivity loss.⁴⁸ Although descriptive comparisons revealed that individuals with hypertension had lower household expenditure, personal income and household income compared with those with normal blood pressure, this discrepancy may be attributed to our cohort’s high medical insurance coverage (>93%). Such coverage might mitigate financial strain despite underlying economic stress,¹² limited healthcare accessibility, poor dietary habits¹¹ and reduced physical activity observed in lower-income groups.⁴⁹ In contrast, dyslipidaemia was positively associated with personal income yet inversely related to household income. One possible explanation is that individuals with dyslipidaemia may benefit from a high-calorie diet¹⁰ and enjoy enhanced workplace benefits.⁹ However, these individuals may also face substantial out-of-pocket medical expenses for managing complications that fall outside standard insurance coverage, which could ultimately reduce overall household income.

Interestingly, our results indicate that women with diabetes experience more significant reductions in household income. This marked decline is likely to be attributed to employment interruptions related to caregiving responsibilities, aligned with prior research indicating that women often sacrifice employment opportunities to manage chronic illnesses.⁵⁰ These socioeconomic penalties exacerbate health inequities and restrict access to essential care. In light of these insights, nursing advocacy should prioritise establishing workplace partnerships⁵¹ that promote flexible work arrangements⁵² and employer-sponsored health screenings,⁵³ in addition to developing community-driven respite care programmes to alleviate caregiving burdens.⁵²

Diabetes indicated pronounced gender-specific effects. Among women, diabetes was strongly associated with a reduction in household income, underscoring the economic penalties arising from traditional caregiving responsibilities⁵⁰ and employment interruptions.⁵⁴ This reduction in household income among women may also reflect regional factors, such as limited social services and disparities in healthcare access, which further exacerbate income losses, in contrast to high-income countries where robust social security systems mitigate these impacts.^{55 56} Conversely, among men, diabetes was more strongly linked to increases in personal income. This positive association may be driven by occupational dynamics (eg, labor-intensive jobs) that inadvertently raise personal earnings despite increasing health risks. Especially in China, where traditional norms often associate men with career-related stress and women with family and caregiving roles, these divergent income effects highlight the complex association between chronic disease and gender-specific socioeconomic roles.⁸ Moreover, the overall higher prevalence of hypertension, dyslipidaemia and diabetes among women-potentially due to inherent biological factors, entrenched social roles and delayed access to preventive care^{57 58}—underscores the urgent need for gender-sensitive screening programmes and educational interventions focused on self-management strategies, such as dietary modifications and stress reduction. To address these disparities, nursing interventions should be tailored to the unique challenges faced by each gender. For women, establishing workplace partnerships that offer flexible work arrangements and on-site health screenings, along with community-based respite care programmes, could help mitigate the economic impact of caregiving responsibilities. For men, occupational health nurses should advocate for safer work conditions and regular health assessments in high-risk industries. These targeted strategies are critical for reducing gender-based income disparities associated with chronic diseases and for promoting equitable health outcomes.

Limitations

This study has several methodological limitations that should be considered when interpreting the results. First, chronic diseases were measured using a binary variable, specifically whether a participant had ever been diagnosed by a doctor. While this approach is efficient, it may introduce bias, as it relies on self-reported data that are susceptible to recall bias or misclassification. To reduce these potential biases, future studies could incorporate more precise biochemical indicators, such as blood pressure, blood lipids and blood glucose levels, to measure chronic diseases more accurately.

Second, this study focused solely on the relationship between a single chronic disease (eg, hypertension) and income, without considering the combined effects of multiple chronic diseases (eg, comorbid diabetes or dyslipidaemia). Chronic diseases often occur in combination, and their interactions could have more complex effects on income that this study did not capture. Future research should explore the effects of multiple chronic conditions simultaneously to gain a more nuanced understanding of how these diseases interact and influence income.

Another limitation is the reliance on self-reported health data, which may affect the accuracy and reliability of the findings. Participants may not always recall or report their health status accurately, especially regarding chronic disease diagnoses. To enhance data quality, future research could combine self-reported data with objective health indicators or clinical assessments.

Finally, the study used the gradient boosting method for DDML fitting, but did not employ other robustness checks, such as random forests or nnet. While the gradient boosting method is robust, the lack of additional checks may limit the model’s robustness and the generalisability of the results. Future studies could apply a variety of fitting methods and conduct robustness tests to ensure the reliability and consistency of the findings.

Conclusion

This research highlights the essential role of nursing in tackling the interconnected challenges of chronic disease, income inequality and gender disparities. By advocating for policy reforms and implementing targeted interventions, healthcare professionals can empower patients to break the cycle of health inequity. Future efforts should focus on intersectional approaches that address the distinct needs of diverse populations, ensuring equitable healthcare access for all. The findings emphasise the importance of interventions that address both the health and economic challenges faced by ageing Chinese adults with chronic diseases.

The data employed in this study were sourced from the CHARLS.⁵⁹

Data availability statement

Data are available in a public, open access repository.