Abstract
Background
Research on public health insurance (PHI) and human capital has focused on direct pathways through which increased medical utilization improves human health capital. However, a possible indirect pathway may also exist: PHI may expand the budget for nonmedical expenditures, some of which are important human capital inputs. This study uses the New Rural Cooperative Medical Scheme (NRCMS) to examine whether two empirical associations are consistent with a potential indirect pathway: NRCMS participation is associated with higher individual caloric intake, and caloric intake is associated with subsequent labor productivity.
Methods
Using unbalanced panel data from the 2009 and 2011 waves of the China Health and Nutrition Survey, we examined two hypotheses: NRCMS participation is associated with higher individual caloric intake, and higher caloric intake is associated with higher productivity, as measured by earnings per working day. We used propensity score matching for the first hypothesis and inverse probability weighting regression for the second hypothesis.
Results
NRCMS participation was associated with approximately 294 kcal/day higher caloric intake. A caloric difference of this magnitude was associated with 2.97%–3.91% higher labor productivity in rural China in the late 2000s.
Conclusions
These findings suggest that nonmedical consumption may be a relevant pathway linking public health insurance to human capital in low-income rural settings. While causal interpretation remains limited, they underscore the importance of examining the broader human capital implications of health insurance beyond health outcomes.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1186/s13561-026-00833-w.
Keywords: Public health insurance, Human capital, Caloric intake, Labor productivity, Rural China
Introduction
Poor people remain trapped in poverty because limited budgets prevent sufficient investment in human capital [1]. Health-related financial burdens constrain such investments through multiple channels. High medical costs directly restrict access to medical care, weakening the labor force’s human health capital [2, 3]. At the same time, such costs indirectly reduce nonmedical consumption [4, 5], with items such as food being crucial for human capital [6, 7]. Many governments have introduced public health insurance (PHI) programs to reduce such financial burdens.
Previous studies have documented the overall effects of PHI and shown that it reduces poverty and increases income among disadvantaged groups [8–10]. In addition to these impact assessments, researchers have explored the mechanisms of human capital. The direct pathway is widely recognized to operate through increased medical care, which boosts health status [11]. Multiple studies, which have examined diverse PHI schemes, have consistently indicated that PHI increases the utilization of outpatient, inpatient, and preventive medical services [12–14] and improves health [15, 16]. However, Lei and Lin [17] suggested that this direct pathway is limited in some contexts, leaving conclusions about the direct pathway obscure.
In addition to the direct pathway, PHI may also generate an indirect pathway for human capital. By reducing medical expenses [18, 19] and the necessity for precautionary savings [20, 21], PHI may increase the budget for nonmedical expenditures. Some categories of nonmedical consumption may function as human capital inputs. For manual laborers in many developing countries, food is an important example, as calories are a critical resource for sustaining labor productivity [6, 7, 22, 23].1 This suggests a potential indirect pathway: PHI may increase caloric intake and thereby support productivity.
However, this indirect pathway remains ambiguous. First, some studies have indicated that PHI increases overall food consumption [25, 26], but this does not necessarily imply an increase in caloric intake. Food expenditure is an important indicator of improved welfare, but it may also partly reflect factors that are not directly related to human capital, such as taste-oriented food choices, rather than increased energy intake. Caloric intake, by contrast, can be interpreted more directly as a nutritional input relevant to labor productivity for manual laborers in low-income, labor-intensive settings [6, 7, 22, 23]. Examining caloric intake therefore allows us to assess whether PHI is associated with nutrition-related human capital. Second, even if PHI is associated with higher caloric intake, the productivity implications of additional calories remain uncertain, because the role of calories in supporting work capacity may differ across levels of development and population groups. Consequently, it is necessary to examine both the calorie and productivity components to assess the empirical relevance of this potential indirect pathway.
This study examines a potential indirect pathway through China’s New Rural Cooperative Medical Scheme (NRCMS), a major PHI program introduced in 2003 for rural residents. Financed jointly by individuals, rural communities, and government subsidies, it features low premiums and voluntary participation. The NRCMS aims to increase access to medical care and has been shown to boost nonmedical consumption [27]. In the 2000s, caloric intake in rural China remained insufficient [28, 29], making it plausible that NRCMS participation may be linked to higher caloric intake and subsequent productivity.
The study examines two hypotheses: whether NRCMS participation is associated with individuals’ calorie consumption in the year of participation, and whether calorie consumption is associated with productivity in the following year (productivity was measured as earnings per working day).
The China Health and Nutrition Survey (CHNS), an unbalanced panel survey that collects individual data biannually and quadrennially from 1989 to 2015, enabled us to examine these two hypotheses. Among its survey waves, the 2009 and 2011 waves are particularly suitable for this purpose because they allow us to link NRCMS participation and caloric intake in 2009 to labor productivity in 2010. For the first hypothesis, we used the variables of program participation and calorie consumption from the 2009 survey wave. For the second hypothesis, we used the “total income for 2010” divided by the “number of working days in 2010” recorded in the 2011 survey wave. Leveraging the panel data structure, we used the caloric intake of a particular person in 2009 to explain the productivity of the same person in 2010, which helps mitigate concerns about reverse causality from higher productivity to more calorie consumption.2
We used two empirical methods appropriate for the nature of each data structure. For the first hypothesis (NRCMS participation and calorie intake), we applied propensity score matching (PSM) to adjust for observed differences related to self-selection in participation. For the second hypothesis (calorie intake and productivity), we used multiple regression with inverse probability weighting (IPW) to control for covariates and mitigate potential attrition bias possibly stemming from the unbalanced panel. The estimates indicate that NRCMS participation was associated with approximately 294 kcal/day higher caloric intake. A caloric difference of this magnitude was associated with approximately 2.97%–3.91% higher labor productivity, even after controlling for other possible determinants. These two associations are consistent with a potential indirect pathway involving NRCMS participation, caloric intake, and subsequent productivity.
This study contributes to the existing literature in two ways. First, it contributes to the public health and health insurance literature. Previous studies have shown that PHI can raise income or reduce poverty among vulnerable groups [8–10], but the human capital channels underlying these broader economic outcomes need further examination. Beyond the direct pathway through medical care and health capital, this study documents two related associations: NRCMS participation is associated with higher caloric intake, and higher caloric intake is associated with higher labor productivity. These patterns are consistent with a potential indirect pathway in which PHI may relax budget constraints in ways that support nutrition-related human capital in low-income rural settings.
Second, this study contributes to development and labor economics by highlighting calories as a component of productivity-related human capital, a classic hypothesis proposed by Strauss [6]. We argue that calories are crucial in a human capital framework in which physical labor dominates subsistence income.
The remainder of this paper is organized as follows. "Background" section introduces the background of the NRCMS, and "Hypotheses" section provides the hypotheses. "Methods" section presents the data and empirical strategy. "Results" section reports the results and robustness checks. "Discussion" section discusses the study’s limitations, and "Conclusions" section concludes the study.
Background
In the late 1950s, the Chinese government established the Cooperative Medical System (CMS) to improve rural access to medical care. The system operates mainly on rural community funding, with contributions from individuals and government subsidies to build health centers and provide medicines, enabling affordable care [30]. By the 1970s, the CMS covered 90% of China’s rural population and was considered a successful model for addressing healthcare in developing countries [31]. However, by the 1980s, rural economic system reforms replaced collective agricultural production with private production, causing the CMS to lose its primary funding source and collapse [32]. By 1998, rural insurance coverage fell to 9.5% [32].
Against this backdrop, the financial burden of health issues became the leading cause of poverty in rural China in the 1990s [33]. This poverty is not only a short-term financial loss due to out-of-pocket health expenditures but also a long-term poverty trap hindering human capital accumulation. First, rural populations without the CMS struggled to access adequate medical care, leaving their health unprotected [34]. Second, food intake in rural China was limited and the absence of health insurance constrained household food consumption [35]. According to a report by the United States Department of Agriculture on China, in 2003, rural households in the middle-income quintile had a per capita annual income of approximately US$275 (including the imputed value of own-produced food consumed at home), and they devoted nearly 45% of their total expenditure on food, underscoring tight food budgets [36]. With physically demanding work common among rural residents, food and calorie shortfalls help explain low productivity [29].
To mitigate this issue, the Chinese government launched the NRCMS in 2003 to increase the utilization of medical care and nonmedical consumption [27]. The NRCMS is jointly funded by individuals, rural communities, and government subsidies, offering below-market-price health insurance to rural populations through voluntary participation. By 2008, the NRCMS had expanded nationwide.
Hypotheses
Figure 1 presents a conceptual diagram illustrating potential pathways linking the NRCMS to human capital and productivity through two channels: medical care and food consumption. First, medical care is a critical investment in health capital [37]. The direct pathway of the NRCMS may increase medical care use, thereby improving health (the blue lines in Fig. 1). Second, nonmedical consumption, such as food, is also important for human capital. While food is related to health, its caloric content is emphasized in labor economics as a classic determinant of productivity for manual laborers [6, 7, 22, 23].3 Insufficient food intake remained a concern in rural China during the 2000s ("Background" section), making this mechanism particularly relevant. Therefore, a potential indirect pathway may enhance productivity by increasing caloric intake.
Fig. 1.
Potential pathways linking PHI to human capital
To examine the empirical relevance of this potential indirect pathway, this study evaluated the following hypotheses:
Hypothesis 1: NRCMS participation is positively associated with individuals’ caloric intake in the year of participation (corresponding to “ [1]” in Fig. 1).
Hypothesis 2: Caloric intake in the year of participation is positively associated with productivity in the following year (corresponding to two routes in Fig. 1: “ [2]”, the independent (unmediated) calorie–productivity link; and “ [2]′”, the health-mediated pathway whereby calories may first improve health and then enhance productivity).4
Our main quantity of interest for Hypothesis 2 is the overall association (“ [2]” + “ [2]′”).
To reduce observable confounding, we controlled for covariates. In Fig. 1, 
denotes the covariates that jointly influence NRCMS participation and caloric intake (Hypothesis 1), and 
denotes the covariates that jointly influence caloric intake and subsequent productivity (Hypothesis 2). We used PSM to adjust for observed differences in 
and weighted multiple regression to control for 
. "Methods" section provides the exact variable lists and method details.
Figure 1 raises an econometric issue for examining Hypothesis 2 (the overall association “ [2]” + “ [2]′”): whether we should control for health. On the one hand, if we include health to account for this channel, what we can estimate as the coefficient of calories is the independent association alone (≈“ [2]”), which can be an underestimation of the overall calorie–productivity association (“ [2]” + “ [2]′”), if “ [2]′” exists. On the other hand, if we do not include health as an intermediary variable, the coefficient of calories can be upwardly biased (≈“ [2]” + “ [2]′” + “ [3]”). Consequently, given the model structure and available variables, neither specification captures the overall association (“ [2]” + “ [2]′”); instead, the association can be bracketed between the estimate with health controls (≈“ [2]”) and without health controls (≈“ [2]” + “ [2]′” + “ [3]”). "Results" section reports the estimates with and without health controls and provides discussion based on the range suggested by the regression results.
Methods
Data
We used data from the CHNS [38]. The CHNS was established by the Carolina Population Center at the University of North Carolina in Chapel Hill and the National Institute for Nutrition and Health at the Chinese Center for Disease Control and Prevention. The survey covered 15 provinces through multistage cluster sampling. It included the following variables used in this study: individual characteristics, NRCMS participation, caloric intake, income, and working days. The database included biannual and quadrennial surveys since 1989, with nutritional data updated to 2011. The CHNS followed up on each sample in every wave. Sample attrition across waves included individuals who were unwilling to participate and those who migrated outside the survey areas. To compensate for this attrition, the CHNS added new interviewees to maintain the sample size.
We screened the data using three criteria. First, we retained individuals with an agricultural household registration, which was the NRCMS’s target population. Second, we included individuals aged 18–60 years, the main income-generating group. In China, 18 years is the legal age of adulthood, and 60 years is the threshold for old age. Finally, we excluded abnormal samples with a daily caloric intake of over 8,000 kcal, weight under 30 kg, or height under 80 cm.
Accordingly, we used the 2009 wave data to test Hypothesis 1. Then, we used income in 2010 divided by working days in 2010, both recorded in the 2011 wave, as a proxy for productivity in 2010. Previous empirical studies have also used income per unit of time as a practical measure of labor productivity [7, 39, 40]. In this study, the income variable comes from the CHNS constructed individual income file and captures all income components that can be measured at the individual level, including wage employment and agricultural production, but excluding components such as government subsidies, gifts, and rental income.5 Therefore, our productivity measure should be interpreted as broader individual earnings capacity per working day, rather than as a pure measure of physical productivity. Because this measure may be affected by determinants other than caloric intake, the analysis for Hypothesis 2 controls for other factors in various specifications by including factors such as education, health conditions, previous year’s productivity, regional dummies, and respondents’ main occupation recorded in 2009.
Productivity in 2010 was matched with calorie data from the 2009 wave to examine Hypothesis 2. Although the sample data were collected over a decade ago, this study remains relevant for drawing insights applicable to low-income countries that have suffered from insufficient calorie intake in rural areas, at similar stages as China.
Empirical strategy
Propensity score matching
Hypothesis 1 examined whether NRCMS participation was positively associated with calorie consumption in 2009. As participation in the NRCMS is voluntary, self-selection occurs. We adopted the PSM approach by referencing the methods used by Trujillo et al. [41] and Sun and Lyu [42] to analyze health insurance policies; the latter assessed the NRCMS.
We followed the standard PSM approach [43]. First, we generated propensity scores with a probit model of NRCMS participation using the following individual characteristics (
in Fig. 1): prior year income, education, age, gender, marital status, height, weight, and regional dummies.6 Second, we used single nearest-neighbor matching without replacement to match individuals in the treatment (NRCMS participants) and control (NRCMS nonparticipants) groups. Third, we checked whether the matching met two key PSM assumptions: balancing property and common support. Finally, we calculated the average treatment effect on the treated (ATT) as follows:
 |
1 |
where calorie represents an individual’s caloric intake, 
denotes the individual characteristics condensed into a propensity score, P, and the NRCMS indicates the program participation (1 for participation and 0 for nonparticipation). The first term on the right-hand side of the equation represents the average caloric intake of the treatment group (NRCMS = 1) after matching, and the second term represents the counterfactual caloric intake of the treatment group.
Similarly, the average treatment effect on the untreated (ATU) was calculated as follows:
 |
2 |
Combined with the ratio of people in the two groups p, the average treatment effect (ATE) was calculated as
 |
3 |
Inverse probability weighting regression
To test Hypothesis 2, we used caloric intake data from 2009 and productivity data from 2010 (reported in the 2011 wave) based on an unbalanced panel dataset. We controlled for the covariates 
in Fig. 1 and health using a multiple regression approach. As the attrition between the two datasets was likely to be nonrandom, the IPW method was applied [44–46]. In the first stage, we used a logit model to estimate the likelihood that individuals observed in 2009 would also appear in 2011. The dependent variable was continued observation (1 if observed in both years and 0 if observed only in 2009), and the independent variables were 2009 individual-level determinants of attrition, such as income, education, age, gender, marital status, height, weight, and region dummies. The weights w were generated based on the estimated probability of continued observation. Finally, the association between caloric intake and productivity was estimated using the following formula:
 |
4 |
where i represents the individual and k denotes the suffix for the regression coefficient 
. Productivity represents the labor productivity in 2010, Calorie indicates the caloric intake in 2009, and 
comprises a set of control variables. We included the 
covariates from Fig. 1 such as age, gender, education, marital status, height, weight, and regional dummy variables. We reported the results without controlling for health and then added controls for health status and medical care utilization. We measured health by the number of individuals’ sick days 28 days before the survey interview, following previous studies [17]. We evaluated medical care utilization through responses to the question “What did you do when you felt ill?,” which we coded as 1 if the respondent sought formal medical care and 0 otherwise. The coefficient of interest (
) represents the percentage difference in productivity associated with a unit increase in caloric intake, given that the dependent variable is the log of productivity.
As productivity had a persistent component, we reported a column that included 2008 productivity as a lagged outcome control following Banerjee et al. [47]. This mitigated the possibility that individuals who were already more productive in 2008 consumed more calories in 2009 and remained more productive in 2010. It was calculated as income in 2008 divided by working days based on the 2009 CHNS wave. In addition, we reported a specification that includes dummy variables for respondents’ main occupation recorded in 2009. This check helps address the concern that the estimated association between caloric intake and productivity may partly reflect occupational differences in work conditions or physical work demands.
Results
Descriptive statistics
Table 1 presents the definitions and descriptive statistics of the variables from the 2009 wave dataset used in Hypothesis 1. The outcome variable was calorie, with an average value of 2,268 kcal. Gender differences in caloric intake (not presented in the table) indicated that men consumed 2,451 kcal per day, whereas women consumed 2,066 kcal per day. According to the dietary reference intake standards of the Chinese Nutrition Society [48], the calorie requirements for moderate physical laborers, such as waiters and salespersons, are 2,250 kcal for men and 1,800 kcal for women. The requirements for heavy physical laborers, such as farmers and workers, are 3,000 kcal for men and 2,350 kcal for women.7 In 2009, the rural population in China had a relatively low nutritional intake, barely sufficient for moderate physical labor and insufficient for heavy physical labor. Therefore, if the rural population had a higher budget for nonmedical consumption, they might have considered increasing their caloric intake; higher caloric intake may support physical labor productivity, particularly in the case of heavy physical labor.
Table 1.
Definition and descriptive statistics of variables from the 2009 wave data set for Hypothesis 1
| Variables | Definition | Mean | SD | Min | Max |
|---|---|---|---|---|---|
| Calorie | Average caloric intake of the respondent in a single day (100 kcal) | 22.68 | 6.56 | 5.59 | 69.41 |
| Income | Annual total income before the survey year (1000 yuan) | 13.98 | 25.07 | -121.60 | 600.00 |
| Education | Years of education | 7.15 | 3.30 | 0 | 16 |
| Age | Age of the respondent | 43.25 | 10.45 | 18 | 60 |
| Gender | 1 for male, 0 for female | 0.52 | 0.50 | 0 | 1 |
| Marital status | 1 for married, 0 for other | 0.90 | 0.30 | 0 | 1 |
| Height | Height of the respondent (cm) | 161.93 | 8.18 | 138.40 | 186.50 |
| Weight | Weight of the respondent (kg) | 60.95 | 10.75 | 34.00 | 103.70 |
Data source: CHNS 2009 survey wave
Some income data are negative owing to losses from own land farming or livestock production. Sample size: n = 2810
Moreover, Table 1 includes individual characteristics. The rural population’s total income in the year before the 2009 survey averaged approximately 14,000 yuan. This amount was less than half of the 2008 urban per capita income of 29,229 yuan, indicating a deficiency [49]. Additionally, the average number of years of education was only 7.15 years.
As NRCMS participation is voluntary rather than randomly assigned, participants and nonparticipants exhibited significant differences in their characteristics (Table 2). Our t-test for mean differences indicated that participants had significantly lower income and education levels than nonparticipants, suggesting that disadvantaged groups were more inclined to join the NRCMS. The participants were significantly older, which may be because the financial burden of health issues increases with age. Furthermore, the participants had significantly lower height and weight, which suggests food insecurity, as height usually reflects food intake in early life [50] and weight is related to past calorie accumulation [51]. Overall, the differences between participants and nonparticipants (Table 2) highlight the rationale for using PSM.
Table 2.
Differences in characteristics between the NRCMS participants and nonparticipants in 2009
| Variables | Participants | Nonparticipants | Mean difference | ||
|---|---|---|---|---|---|
| Mean | Std. dev. | Mean | Std. dev. | p-value | |
| Income | 13.66 | 24.95 | 20.70 | 26.70 | 0.00 |
| Education | 7.03 | 3.23 | 9.71 | 3.60 | 0.00 |
| Age | 43.44 | 10.36 | 39.29 | 11.47 | 0.00 |
| Gender | 0.52 | 0.50 | 0.59 | 0.49 | 0.13 |
| Marital status | 0.91 | 0.29 | 0.82 | 0.38 | 0.00 |
| Height | 161.81 | 8.19 | 164.30 | 7.76 | 0.00 |
| Weight | 60.80 | 10.69 | 64.15 | 11.47 | 0.00 |
| Observations | 2681 | 129 | |||
Data source: CHNS 2009 survey wave
The p-value indicates the result of a t-test comparing the means of characteristics between participants and nonparticipants. Sample size: n = 2810
Table 3 presents the definitions and descriptive statistics for Hypothesis 2. Owing to sample attrition, the sample size decreased from 2,810 to 1,611 participants, with an attrition rate of approximately 40%. Although the means are comparable, attrition of this magnitude can affect the estimation. We applied IPW to adjust for potential attrition bias.
Table 3.
Definitions and descriptive statistics for Hypothesis 2 (2009 data combined with 2010 productivity)
| Variables | Definition | Mean | SD | Min | Max |
|---|---|---|---|---|---|
| Productivity | Total income divided by working days in 2010 (yuan/day) | 80.41 | 125.12 | -187.21 | 2115.87 |
| Calorie | Average caloric intake of the respondent in a single day (100 kcal) | 23.05 | 6.61 | 6.14 | 53.22 |
| Education | Years of education | 7.06 | 3.22 | 0 | 16 |
| Age | Age of the respondent | 44.01 | 8.76 | 18 | 58 |
| Gender | 1 for male, 0 for female | 0.52 | 0.50 | 0 | 1 |
| Marital status | 1 for married, 0 for other | 0.94 | 0.23 | 0 | 1 |
| Height | Height of the respondent (cm) | 161.86 | 8.27 | 138.60 | 186.50 |
| Weight | Weight of the respondent (kg) | 61.40 | 11.08 | 34.00 | 103.70 |
Data source: CHNS 2009 and 2011 survey waves
Some income data are negative owing to losses from own land farming or livestock production, which in turn causes negative productivity. Sample size: n = 1611
NRCMS participation and caloric intake
Main PSM results
PSM was used to estimate the probit model of participation in the NRCMS, and Table A.1 in Additional file 1 presents the results. Then, using the matching method explained in "Propensity score matching" section, we checked whether the matching met the requirements for common support and balancing property. The solid and dashed lines in Fig. 2 represent the propensity score distributions of NRCMS participants and nonparticipants, respectively. Their overlap represents the common support area. Among the 129 nonparticipants, 123 observations were matched, while 124 observations were matched among the 2,681 participants. Table 4 compares the characteristics of the two groups after matching. No significant differences were observed, indicating that the balancing property was satisfied. Therefore, all prerequisites for PSM were met.
Fig. 2.
Common support between the participants and non-participants
Table 4.
Balancing property after matching
| Variable | Mean (participant) | Mean (nonparticipant) | p-value (t-test) |
|---|---|---|---|
| Income | 22.39 | 19.68 | 0.54 |
| Education | 9.48 | 9.48 | 0.99 |
| Age | 38.55 | 39.65 | 0.45 |
| Gender | 0.62 | 0.58 | 0.49 |
| Marital status | 0.82 | 0.85 | 0.63 |
| Height | 165.53 | 163.98 | 0.12 |
| Weight | 65.45 | 63.84 | 0.27 |
See Table A.1 for propensity score estimation. For matching, a single nearest-neighbor matching without replacement is used. Among 129 non-participants, 123 samples are matched; 124 samples are matched among 2681 participants
Table 5 presents the results of ATT, ATU, and ATE based on PSM. We used bootstrapping to compute standard errors for robust estimates. In the ATT row, the mean represents the average daily caloric intake of the matched NRCMS participants (2,466 kcal). The counterfactual estimate of the participants’ average daily caloric intake in the case of nonparticipation in the NRCMS was 2,057 kcal. Hence, the ATT estimate for caloric intake was 409 kcal, significant at the 1% level. Similarly, the ATU estimate was 179 kcal, significant at the 5% level. Finally, using Eq. (3), we calculated the ATE estimate as 294 kcal, significant at the 1% level.
Table 5.
Estimation results of the impact of the NRCMS on caloric intake by PSM
| Mean | Counterfactual | Coefficient | Bootstrap std. err. | p-value | |
|---|---|---|---|---|---|
| ATT | 24.66 | 20.57 | 4.09 | 0.91 | 0.00*** |
| ATU | 20.60 | 22.38 | 1.79 | 0.81 | 0.03** |
| ATE | 2.94 | 0.71 | 0.00*** |
Bootstrap standard errors are calculated to represent sampling variability. ** and *** indicate statistical significance at the 5% and 1% levels, respectively. The unit of calorie is 100 kcal. See Table A.1 for propensity score estimation. For matching, single nearest-neighbor matching without replacement is used
The difference between the mean and counterfactual should be equal to the coefficient. However, owing to rounding, a minor discrepancy may exist that is not an error
ATT Average Treatment Effect on the Treated, ATU Average Treatment Effect on the Untreated, ATE Average Treatment Effect
Robustness checks for hypothesis 1
The PSM estimation in this study raises two concerns. First, nonparticipants might differ from participants in unobserved factors, especially when the number of nonparticipants was small. This concern is related to the ignorability assumption underlying PSM: after conditioning on observed covariates, treatment assignment should be independent of potential caloric intake outcomes. Second, because PSM relies on matched observations, many participants were not used in the matched-sample comparison, which may limit the representativeness of the estimates.
To address the first concern, we conducted sensitivity analyses using Rosenbaum bounds [52]. This method introduces a gamma parameter to simulate the possible impact of unobserved covariates on propensity scores. If the results remain significant at higher gamma values, the PSM estimates are less sensitive. A gamma value of 2 indicates absolute insensitivity to hidden bias, values between 1.5 and 2 indicate relative insensitivity, and values below 1.5 suggest a lack of robustness [53]. Table A.2 in Additional file 1 presents the results of the sensitivity analyses. When gamma was 1.8, the results remained significant at the 5% level; when gamma was 2, the results remained significant at the 10% level. Therefore, the PSM result for Hypothesis 1 remained relatively insensitive to hidden bias.
To address the second concern, we estimated IPW and doubly robust augmented IPW (AIPW) specifications using the full sample. IPW adjusts for observed confounders by weighting observations with inverse propensity scores [54]. Unlike PSM, which uses only matched observations and may bias results by ignoring unmatched ones, IPW includes the full sample and weights observations to balance covariates between participants and nonparticipants. We further implemented doubly robust AIPW [55, 56]. Details of the AIPW specification are provided in Additional file 1. Table A.3 in Additional file 1 presents the IPW and AIPW results. The IPW estimate was 286 kcal/day, and the AIPW estimate was 312 kcal/day; both were significant at the 1% level and close to the PSM estimate of 294 kcal/day. These results support the robustness of the positive association between NRCMS participation and caloric intake under alternative observable-selection estimators.
Within the constraints of the available data, PSM, Rosenbaum bounds, IPW, and AIPW provide feasible ways to address the main empirical concerns as far as possible. However, these checks do not eliminate all limitations of the comparison. Rosenbaum bounds indicate that the PSM result is relatively insensitive to possible hidden bias, but they are based on the matched sample and do not address the potential representativeness issue caused by excluding unmatched participants. IPW and AIPW use the full sample and therefore reduce dependence on the matched subset; however, they still rely on observed covariates and cannot assess sensitivity to hidden bias in the same way as Rosenbaum bounds. Thus, no single method fully addresses both concerns. Nevertheless, the estimates are consistently positive and statistically significant across PSM, Rosenbaum bounds, IPW, and AIPW. We therefore interpret the results as evidence of a positive association between NRCMS participation and caloric intake, while acknowledging that causal interpretation remains limited.
Caloric intake and labor productivity
Main IPW regression results
For the IPW regression, we used a logit model to predict whether individuals observed in 2009 were also observed in 2011 (Table A.4 in Additional file 1).8 Table 6 presents the regression results with IPW for the log of productivity. We excluded 12 observations (10 negative values and 2 zeros, of 1,611) owing to log transformation; we verified that the calorie coefficient was significant when estimated without the log specification, indicating that the results in Table 6 were not driven by the exclusion of negative and zero productivity values.9 In Column (1), the coefficient on caloric intake was significant at the 1% level, with a coefficient of 0.0133. Moreover, education was significantly positive; however, the coefficient was relatively small. Column (2) includes additional control variables, with the coefficient of caloric intake remaining stable at 0.0127. Age had a positive coefficient, whereas age squared was negative, implying that productivity increases with age, peaks at approximately 39 years, and then declines in an inverted U-shape. As Columns (1) and (2) do not control for health, the calorie coefficient captures “ [2]” + “ [2]′” + “ [3]” in Fig. 1, which is plausibly upward-biased relative to the overall association (“ [2]” + “ [2]′”).
Table 6.
Inverse probability weighted regression results explaining the log of earnings per working day
| (1) | (2) | (3) | (4) | (5) | |
|---|---|---|---|---|---|
| Calorie | 0.0133*** | 0.0127*** | 0.0112** | 0.0101** | 0.0108** |
| (0.00448) | (0.00452) | (0.00449) | (0.00448) | (0.00460) | |
| Gender | 0.0164 | 0.0287 | 0.0431 | 0.0394 | -0.00743 |
| (0.0582) | (0.0802) | (0.0807) | (0.0802) | (0.0806) | |
| Education | 0.0183** | 0.0132 | 0.0119 | 0.00670 | 0.00254 |
| (0.00825) | (0.00881) | (0.00877) | (0.00867) | (0.00900) | |
| Age | 0.0577** | 0.0520** | 0.0525** | 0.0702*** | |
| (0.0257) | (0.0259) | (0.0263) | (0.0266) | ||
| Age squared | -0.000746** | -0.000672** | -0.000698** | -0.000873*** | |
| (0.000304) | (0.000306) | (0.000309) | (0.000313) | ||
| Marital status | -0.165 | -0.131 | -0.192* | -0.188* | |
| (0.110) | (0.111) | (0.106) | (0.106) | ||
| Height | 0.00233 | 0.00189 | 0.00175 | 0.00145 | |
| (0.00540) | (0.00546) | (0.00546) | (0.00552) | ||
| Weight | -0.00342 | -0.00336 | -0.00460 | -0.00405 | |
| (0.00335) | (0.00338) | (0.00339) | (0.00346) | ||
| Sick days | -0.0305 | -0.0303 | -0.0301 | ||
| (0.0203) | (0.0197) | (0.0199) | |||
| Medical care | 0.0469 | 0.0350 | 0.0183 | ||
| (0.0889) | (0.0884) | (0.0898) | |||
| Productivity_2008 | 0.00131*** | 0.00120*** | |||
| (0.000442) | (0.000421) | ||||
| Region dummy | √ | √ | √ | √ | √ |
| Main occupation dummy | √ | ||||
| N | 1599 | 1599 | 1577 | 1552 | 1551 |
The numbers in parentheses represent the standard errors. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. See Table A.4 for the computation of inverse probability weights
Using the same specification as Column (2), we also conducted additional heterogeneity checks for Hypothesis 2 to examine whether the association between caloric intake and the log of productivity differs across subgroups. We estimated subgroup regressions based on 2009 information on age, income, gender, and physical labor intensity classified according to respondents’ main occupation, and then compared subgroup coefficients using Wald tests (Table A.6 in Additional file 1). The clearest heterogeneity appears by age: the coefficient is larger among individuals younger than 45, and the subgroup difference is statistically significant. The coefficients are also larger in the lower-income group and in the medium/high physical-intensity group than in their counterpart groups, but the subgroup differences are not statistically significant. Taken together, these patterns are suggestive that additional calories may be more productivity-relevant for individuals with greater physical work demands or tighter nutritional constraints.
Column (3) adds health status and medical care; the caloric intake coefficient declines slightly to 0.0112, likely because conditioning on health captures only “ [2]” in Fig. 1. This coefficient is smaller than the overall association (“ [2]” + “ [2]′”); however, the difference between Columns (2) and (3) is small, suggesting that the magnitude of “ [2]′” + “ [3]” is limited (0.0127 − 0.0112 = 0.0015). Sun and Lyu [42] found that the NRCMS’s impact on adult health is less, and Lei and Lin [17] revealed that the impact is statistically nonsignificant, limiting the bias from the NRCMS—health channel (“ [3]”). In addition, the health-mediated calorie channel (“ [2]′”) was limited in our data, leading to the close values between “ [2]” and “ [2]” + “ [2]′” + “ [3]”. Columns (4) and (5) further add 2008 productivity as a lagged outcome covariate and dummy variables for respondents’ main occupation recorded in 2009, respectively; the coefficient changes little and remains significant, supporting Hypothesis 2.
Across Columns (1)–(5), each 100-kcal increase is associated with 1.01%–1.33% higher productivity. As the sample mean of calorie was approximately 2,300 kcal (Table 3), the estimated elasticity was approximately 0.23–0.30, consistent with Strauss [6] and Kumar et al. [23].10 Given that NRCMS participation was associated with 294 kcal higher caloric intake ("NRCMS participation and caloric intake" section), a caloric difference of this magnitude corresponds to approximately 2.97%–3.91% higher labor productivity.
Robustness test results for Hypothesis 2
Although Table 6 addresses several potential biases, an additional concern is the appropriateness of using a linear model to test Hypothesis 2. Therefore, we applied a nonparametric kernel regression to examine whether the calorie–productivity association was sensitive to functional form and compared it with the original linear estimates [57]. Bekpulatov et al. [58] used this method as a robustness check.
Table A.5 in Additional file 1 presents the nonparametric kernel regression results for the average marginal effect of calories on the log of productivity. A 100-kcal increase in caloric intake was associated with approximately 1.5% higher productivity, significant at the 1% level. This estimate is close to and slightly higher than the result in "Main IPW regression results" section, suggesting the robustness of the original linear specification. Nevertheless, given the available data structure, the estimated calorie–productivity relationship should be interpreted as an association after controlling for observed covariates, rather than causal evidence.
Discussion
Taken together, the results are consistent with a plausible indirect pathway in which NRCMS may support labor productivity by increasing nutrition-related nonmedical consumption in the late 2000s. In particular, NRCMS participation was associated with higher caloric intake, and higher caloric intake was associated with higher earnings per working day. This pattern suggests that, in low-income rural settings where caloric intake remains insufficient and physical labor is common, health insurance may affect human capital not only through medical care, but also by relaxing financial constraints on productivity-related consumption such as food.
This study has two important limitations. First, although we conducted several robustness checks, these analyses cannot fully eliminate endogeneity concerns, as discussed above. Therefore, causal interpretation remains limited, and the results should be interpreted as associations rather than definitive causal effects.
Second, the findings are primarily derived from a low-income rural context in which caloric intake remained insufficient and physical labor was common. The observed associations may therefore not be generalizable beyond such settings. In middle- and high-income contexts, additional calories may have limited productivity value or may even be harmful when excessive caloric intake and obesity-related risks become more relevant concerns [24, 59, 60]. This boundary is also important for interpreting rural China as development proceeds. Recent research on rural China has begun to examine concerns such as overweight and unhealthy dietary changes, rather than calorie insufficiency alone [61]. Therefore, the findings should be interpreted as most relevant to settings where calorie constraints and manual labor remain important.
Conclusions
Overall, the findings suggest that public health insurance may relax financial constraints in ways that support nutrition-related human capital in low-income settings. This implication should not be directly generalized to middle- or high-income settings. Future research could proceed in two directions. First, in low-income settings where caloric intake remains insufficient and manual labor is common, such as parts of sub-Saharan Africa and South Asia, it would be valuable to examine whether public health insurance supports labor productivity by increasing caloric intake. Second, recent studies have increasingly examined factors that improve dietary quality [62–66], as well as the potential links between dietary quality and labor-market outcomes [67, 68]. A possible direction for future research is therefore to examine whether a policy–diet quality–productivity pathway exists alongside the calorie pathway considered in this study.
Supplementary Information
Additional file 1: Appendix A. Supplementary regression and robustness results.
Acknowledgements
The authors would like to express their gratitude to the faculty members of Kyoto University for their valuable suggestions on this paper. The authors also thank the participants of the 2025 Annual Conference of the Agricultural Economics Society of Japan and the Kyoto Development Economics Research Student Workshop for their helpful comments and feedback.
Abbreviations
- ATE
Average Treatment Effect
- ATT
Average Treatment Effect on the Treated
- ATU
Average Treatment Effect on the Untreated
- CHNS
China Health and Nutrition Survey
- CMS
Cooperative Medical System
- IPW
Inverse probability weighting
- NRCMS
New Rural Cooperative Medical Scheme
- PHI
Public health insurance
- PSM
Propensity score matching
Authors’ contributions
Y.H. conceptualized the study, conducted the formal analysis, and drafted the manuscript. K.K. supervised the study and reviewed and edited the manuscript. Both authors read and approved the final manuscript.
Funding
The authors gratefully acknowledge the financial support from the Japan Science and Technology Agency (JST) SPRING Program [grant number JPMJSP2110]. The funder had no role in the conceptualization, study design, data collection, data analysis, decision to publish, or preparation of the manuscript.
Data availability
The datasets analysed during the current study are available in the China Health and Nutrition Survey (CHNS) website, https://chns.cpc.unc.edu/data/datasets/.
Declarations
Ethics approval and consent to participate
This study used secondary data from the China Health and Nutrition Survey and did not collect primary data.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
In developed contexts, extra calories have no productivity effect or even negative effects since low-income groups often have excessive caloric intake [24]. This issue is beyond the scope of this study, which focuses on settings with insufficient caloric intake in poor areas.
Although we exploit the panel’s timing by linking 2009 intake to 2010 productivity for the same individual, our estimation remains cross-sectional in levels; accordingly, this specification does not estimate within-individual fixed effects.
As productivity is measured in the subsequent period in our study ("Introduction" section), this timing helps mitigate concerns about reverse causality from productivity to caloric intake.
Given documented food scarcity and the absence of evidence of excessive high-calorie intake in rural China during our study period, we expect the health-mediated channel (“ [2]′”) to be positive or at least nonnegative.
Some income sources, such as agricultural production, are generated through joint household production. For these sources, CHNS allocates household net income to individual members according to their reported time spent in the relevant activity [38]. Although this constructed variable may still partly reflect joint household production and local market conditions, it is the closest available measure of productivity in our setting, to our knowledge.
The regional dummies were based on China’s seven geographic regions. Among the nine provinces in the 2009 wave, Jiangsu and Shandong were in East China; Henan, Hubei, and Hunan were in Central China; Guangxi was in South China; Guizhou was in Southwest China; and Liaoning and Heilongjiang were in Northeast China.
In nutrition science, the quantitative method for measuring the degree of physical exertion is called the physical activity level. According to the Chinese Nutrition Society [48], China classifies physical jobs into light, moderate, and heavy levels.
Income, age, marriage, and weight significantly and positively contributed to participation in the 2011 survey. This result is consistent with the design of the CHNS database. Individuals who moved to nonsurvey areas were not tracked ("Data" section). Individuals who had a higher income, were married, or were older were less likely to move. This result underscores the necessity of using weighted regression to explicate varying selection probabilities.
Log transformation is standard in labor economics, as it reduces skewness and facilitates interpretation in percentage terms.
Strauss [6] estimated an elasticity of 0.34 using data on farmers in Sierra Leone, whereas Kumar et al. [23] found an elasticity of 0.24 for workers in India. Behrman et al. [22] estimated an elasticity of 0.61 for farmers in Pakistan, and Jha et al. [7] reported an elasticity of approximately 0.9 for female farmers in India. The magnitude of the elasticity is largely influenced by the degree of poverty.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Ghatak M. Theories of poverty traps and anti-poverty policies. World Bank Econ Rev. 2015;29(Suppl 1):S77–105. [Google Scholar]
- 2.Peters DH, Garg A, Bloom G, Walker DG, Brieger WR, Rahman MH. Poverty and access to health care in developing countries. Ann N Y Acad Sci. 2008;1136:161–71. [DOI] [PubMed] [Google Scholar]
- 3.Wagstaff A, Lindelow M, Gao J, Xu L, Qian J. Extending health insurance to the rural population: an impact evaluation of China’s new cooperative medical scheme. J Health Econ. 2009;28:1–19. [DOI] [PubMed] [Google Scholar]
- 4.Asfaw A, von Braun J. Is consumption insured against illness? Evidence on vulnerability of households to health shocks in rural Ethiopia. Econ Dev Cult Change. 2004;53:115–29. [Google Scholar]
- 5.Wagstaff A. The economic consequences of health shocks: evidence from Vietnam. J Health Econ. 2007;26:82–100. [DOI] [PubMed] [Google Scholar]
- 6.Strauss J. Does better nutrition raise farm productivity? J Polit Econ. 1986;94:297–320. [Google Scholar]
- 7.Jha R, Gaiha R, Sharma A. Calorie and micronutrient deprivation and poverty nutrition traps in rural India. World Dev. 2009;37:982–91. [Google Scholar]
- 8.Korenman SD, Remler DK. Including health insurance in poverty measurement: the impact of Massachusetts health reform on poverty. J Health Econ. 2016;50:27–35. [DOI] [PubMed] [Google Scholar]
- 9.Chen S, Li J, Lu S, Xiong B. Escaping from poverty trap: a choice between government transfer payments and public services. Glob Health Res Policy. 2017;2:15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Remler DK, Korenman SD, Hyson RT. Estimating the effects of health insurance and other social programs on poverty under the Affordable Care Act. Health Aff (Millwood). 2017;36:1828–37. [DOI] [PubMed] [Google Scholar]
- 11.Pan J, Lei X, Liu GG. Health insurance and health status: exploring the causal effect from a policy intervention. Health Econ. 2016;25:1389–402. [DOI] [PubMed] [Google Scholar]
- 12.Miller G, Pinto D, Vera-Hernández M. Risk protection, service use, and health outcomes under Colombia’s health insurance program for the poor. Am Econ J Appl Econ. 2013;5:61–91. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Sparrow R, Suryahadi A, Widyanti W. Social health insurance for the poor: targeting and impact of Indonesia’s Askeskin programme. Soc Sci Med. 2013;96:264–71. [DOI] [PubMed] [Google Scholar]
- 14.Ghislandi S, Manachotphong W, Perego VM. The impact of universal health coverage on health care consumption and risky behaviours: evidence from Thailand. Health Econ Policy Law. 2015;10:251–66. [DOI] [PubMed] [Google Scholar]
- 15.Quimbo SA, Peabody JW, Shimkhada R, Florentino J, Solon O. Evidence of a causal link between health outcomes, insurance coverage, and a policy to expand access: experimental data from children in the Philippines. Health Econ. 2011;20:620–30. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Sood N, Bendavid E, Mukherji A, Wagner Z, Nagpal S, Mullen P. Government health insurance for people below poverty line in India: quasi-experimental evaluation of insurance and health outcomes. BMJ. 2014;349:g5114. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Lei X, Lin W. The new cooperative medical scheme in rural China: does more coverage mean more service and better health? Health Econ. 2009;18(Suppl 2):S25–46. [DOI] [PubMed] [Google Scholar]
- 18.Babiarz KS, Miller G, Yi H, Zhang L, Rozelle S. New evidence on the impact of China’s new rural cooperative medical scheme and its implications for rural primary healthcare: multivariate difference-in-difference analysis. BMJ. 2010;341:c5617. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Nguyen KT, Khuat OT, Ma S, Pham DC, Khuat GT, Ruger JP. Impact of health insurance on health care treatment and cost in Vietnam: a health capability approach to financial protection. Am J Public Health. 2012;102:1450–61. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Gruber J, Yelowitz A. Public health insurance and private savings. J Polit Econ. 1999;107:1249–74. [Google Scholar]
- 21.Cheung D, Padieu Y. Heterogeneity of the effects of health insurance on household savings: evidence from rural China. World Dev. 2015;66:84–103. [Google Scholar]
- 22.Behrman JR, Foster AD, Rosenzweig MR. The dynamics of agricultural production and the calorie-income relationship: evidence from Pakistan. J Econom. 1997;77:187–207. [Google Scholar]
- 23.Kumar K, Gautam S, Singh A, Ram F, Singh A. Does consuming more calories make you earn more? Nutrition-productivity link revisited in India [working paper]. SSRN. 2015. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2610452. Accessed 19 May 2026.
- 24.Dall TM, Fulgoni VL III, Zhang Y, Reimers KJ, Packard PT, Astwood JD. Predicted national productivity implications of calorie and sodium reductions in the American diet. Am J Health Promot. 2009;23:423–30. [DOI] [PubMed] [Google Scholar]
- 25.Hamid SA, Roberts J, Mosley P. Can micro health insurance reduce poverty? Evidence from Bangladesh. J Risk Insur. 2011;78:57–82. [Google Scholar]
- 26.Asfaw DM, Shifaw SM, Belete AA, Aychiluhm SB. The impact of community-based health insurance on household’s welfare in Chilga District, Amhara Regional State, Ethiopia. Front Public Health. 2022;10:868274. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Bai CE, Wu B. Health insurance and consumption: evidence from China’s new cooperative medical scheme. J Comp Econ. 2014;42:450–69. [Google Scholar]
- 28.Peng X, Conley D. The implication of health insurance for child development and maternal nutrition: evidence from China. Eur J Health Econ. 2016;17:521–34. [DOI] [PubMed] [Google Scholar]
- 29.You J, Wang S, Roope L. Intertemporal deprivation in rural China: income and nutrition. J Econ Inequal. 2018;16:61–101. [Google Scholar]
- 30.Feng X, Tang S, Bloom G, Segall M, Gu X. Cooperative medical schemes in contemporary rural China. Soc Sci Med. 1995;41:1111–18. [DOI] [PubMed] [Google Scholar]
- 31.Zhang D, Unschuld PU. China’s barefoot doctor: past, present, and future. Lancet. 2008;372:1865–7. [DOI] [PubMed] [Google Scholar]
- 32.Liu Y. Development of the rural health insurance system in China. Health Policy Plan. 2004;19:159–65. [DOI] [PubMed] [Google Scholar]
- 33.Zhan P, Shen Y, Li S. Multidimensional poverty and poverty reduction policies in rural China during 1995–2013. China Econ. 2019;14:25–38. [Google Scholar]
- 34.Liu Y, Hsiao WC, Li Q, Liu X, Ren M. Transformation of China’s rural health care financing. Soc Sci Med. 1995;41:1085–93. [DOI] [PubMed] [Google Scholar]
- 35.You X, Kobayashi Y. The new cooperative medical scheme in China. Health Policy. 2009;91:1–9. [DOI] [PubMed] [Google Scholar]
- 36.Gale F, Huang K. Demand for food quantity and quality in China. Economic Research Report No. 32. Washington, DC: United States Department of Agriculture, Economic Research Service; 2007. https://www.ers.usda.gov/publications/45708. Accessed 19 May 2026.
- 37.Leibowitz AA. The demand for health and health concerns after 30 years. J Health Econ. 2004;23:663–71. [DOI] [PubMed] [Google Scholar]
- 38.China Health and Nutrition Survey. China Health and Nutrition Survey data sets. https://chns.cpc.unc.edu/data/datasets/. Accessed 19 May 2026.
- 39.Feldstein M. Did wages reflect growth in productivity? J Policy Model. 2008;30:591–4. [Google Scholar]
- 40.Sahn DE, Alderman H. The effects of human capital on wages, and the determinants of labor supply in a developing country. J Dev Econ. 1988;29:157–83. [Google Scholar]
- 41.Trujillo AJ, Portillo JE, Vernon JA. The impact of subsidized health insurance for the poor: evaluating the Colombian experience using propensity score matching. Int J Health Care Finance Econ. 2005;5:211–39. [DOI] [PubMed] [Google Scholar]
- 42.Sun J, Lyu S. Does health insurance lead to improvement of health status among Chinese rural adults? Evidence from the China family panel studies. Int J Health Serv. 2020;50:350–9. [DOI] [PubMed] [Google Scholar]
- 43.Caliendo M, Kopeinig S. Some practical guidance for the implementation of propensity score matching. J Econ Surv. 2008;22:31–72. [Google Scholar]
- 44.Wooldridge JM. Inverse probability weighted estimation for general missing data problems. J Econom. 2007;141:1281–301. [Google Scholar]
- 45.Matsouaka RA, Atem FD. Regression with a right-censored predictor using inverse probability weighting methods. Stat Med. 2020;39:4001–15. [DOI] [PubMed] [Google Scholar]
- 46.Carry PM, Vanderlinden LA, Dong F, Buckner T, Litkowski E, Vigers T, et al. Inverse probability weighting is an effective method to address selection bias during the analysis of high dimensional data. Genet Epidemiol. 2021;45:593–603. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Banerjee AV, Cole S, Duflo E, Linden L. Remedying education: evidence from two randomized experiments in India. Q J Econ. 2007;122:1235–64. [Google Scholar]
- 48.Chinese Nutrition Society. Chinese dietary reference intakes. Beijing: China Science Publishing & Media Ltd.; 2013. [Google Scholar]
- 49.National Bureau of Statistics of China. The average wage of urban employees in 2008 was 29,229 yuan [in Chinese]. 2009. https://www.stats.gov.cn/sj/zxfb/202303/t20230301_1919501.html. Accessed 19 May 2026.
- 50.Deaton A. Height, health, and development. Proc Natl Acad Sci U S A. 2007;104:13232–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 51.Saris WHM. Very-low-calorie diets and sustained weight loss. Obes Res. 2001;9(Suppl 4):S295–301. [DOI] [PubMed] [Google Scholar]
- 52.Rosenbaum PR. Overt bias in observational studies. Observational studies. 2nd ed. New York: Springer; 2002. pp. 71–104. [Google Scholar]
- 53.Xu X, Zheng Y, Fang J, Huang J, Yang X, Zhu X, et al. Associations between regular physical exercise and physical, emotional, and cognitive health of older adults in China: an 8-year longitudinal study with propensity score matching. Front Public Health. 2024;12:1301067. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 54.Lunceford JK, Davidian M. Stratification and weighting via the propensity score in estimation of causal treatment effects: a comparative study. Stat Med. 2004;23:2937–60. [DOI] [PubMed] [Google Scholar]
- 55.Bang H, Robins JM. Doubly robust estimation in missing data and causal inference models. Biometrics. 2005;61:962–73. [DOI] [PubMed] [Google Scholar]
- 56.Funk MJ, Westreich D, Wiesen C, Stürmer T, Brookhart MA, Davidian M. Doubly robust estimation of causal effects. Am J Epidemiol. 2011;173:761–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 57.Rios-Avila F. Smooth varying-coefficient models in Stata. Stata J. 2020;20:647–79. [Google Scholar]
- 58.Bekpulatov F, Younis I, Cheng L. The impact of natural resource management, innovation, and tourism development on environmental sustainability in low-income countries. Resour Policy. 2023;86:104088. [Google Scholar]
- 59.Goettler A, Grosse A, Sonntag D. Productivity loss due to overweight and obesity: a systematic review of indirect costs. BMJ Open. 2017;7:e014632. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 60.Averett SL. Obesity and labor market outcomes. IZA World Labor. 2014;32:1–32. 10.15185/izawol.32. [DOI]
- 61.Huang L, Ren Y, Zamani O, Loy JP. The curse of television: evidence from nutritional outcomes in rural China. World Dev. 2025;192:107014. [Google Scholar]
- 62.Feng H. Using clean fuels promotes diverse diets and health in Chinese households. Food Secur. 2025;17:1207–22. [Google Scholar]
- 63.Zheng X, Lin Q, Xie G, Yu X, Liu C. The road to healthier diet: benefits of the Five Vertical and Seven Horizontal National Trunk Highway System in China. J Health Econ. 2026;107:103139. [DOI] [PubMed] [Google Scholar]
- 64.Olsho LEW, Klerman JA, Wilde PE, Bartlett S. Financial incentives increase fruit and vegetable intake among Supplemental Nutrition Assistance Program participants: a randomized controlled trial of the USDA Healthy Incentives Pilot. Am J Clin Nutr. 2016;104:423–35. [DOI] [PubMed] [Google Scholar]
- 65.Huang W, Xiang K, Yu X, Zou H. Polluted air, healthier diets: household food consumption patterns in response to air quality in China. J Health Econ. 2025;104:103068. [DOI] [PubMed] [Google Scholar]
- 66.Yan Z, Feng H, Li Z, Lin W. From policy to plate: the impact of livestock and poultry regulation on meat consumption in rural China. Agric Econ. 2025;56:574–86. [Google Scholar]
- 67.Feng H, Yan Z, Wang HH. Free from the kitchen fueled by clean energy: women’s employment and nutritional benefits in rural China. Agric Econ. 2026;57:e70075. [Google Scholar]
- 68.Custodio E, Jiménez S, Ramos MP, Sartori M, Ferrari E. Diet quality rather than caloric intake associated with labour wages in Kenya. Food Secur. 2025;17:345–61. [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Additional file 1: Appendix A. Supplementary regression and robustness results.
Data Availability Statement
The datasets analysed during the current study are available in the China Health and Nutrition Survey (CHNS) website, https://chns.cpc.unc.edu/data/datasets/.