Data source and selection of variables
Data were obtained mainly from the projects innovating payment systems and improving health benefits, which were jointly carried out by Harvard University, Oxford University, Fudan University, and Ningxia Medical University(2009, 2011, and 2012). The data from 2015、2019 and 2022 were extracted from the National Natural Science Foundation of China(from a follow-up study of the project) (21).
A multi-stage stratified cluster randomized design was used to obtain a representative sample from each county. We selected a total of four counties in Ningxia Province, including two project counties(Haiyuan and Yanchi) and two control counties(Pengyang and Xiji). In each county, all the villages were divided into three economic levels, 40%of the sample villages were selected. Then, using the household head roster, 33 households(20 households in the control counties) in each village were selected by systematic sampling. Members of the sample households who had been living there for more than 6 months were selected as respondents. Data from 2 follow-up surveys in 2019 and 2022 were selected for this study. This study was conducted on residents who went out to work and had an age requirement for respondents, with limited earning power for those below 18 and above 60 years of age, thus removing these two categories of samples (22); and removing samples where key variables were missing in the merging process of different sub-databases. The final data was obtained for 26,738.
Explained variable:
Self-rated health(SRH) was used as the explained variable, an indicator of overall health status that has been utilized in many social surveys (23–25), which came from the question in the questionnaire of “What do you think about your health status. ”The choices of self-rated health status are as follows: (1) Very good; (2) Good; (3) General; (4) Not good; (5) Very bad. In the empirical study, three category variables are generated.
SRH is a widely used indicator of public health because of its easy availability in large-scale surveys (26)(39) and its importance in predicting mortality (27)(40).To avoid the potential bias caused by sparse data (28)(41), the original five-point Likert scale data were dichotomized into good SRH(i. e. “very good” and“good”), fair(i.e. “fair”) and poor SRH(i. e. “bad”, “very bad”). In our model, good SRH was coded as 2, fair was coded as 1 and poor SRH was coded as 0.
Explanatory variables:
In this study, the concept of "outworking" is defined as whether you went out to work for income in the past year. In the empirical study, the categorical variable is generated. The answer“no” is assigned 0, and“yes” is assigned 1.
Control variables:
In this study, the control variables were referred to Grossman's (29) health needs model, and three main types of control variables were chosen: one was individual characteristics, household characteristics, and health service characteristics variables. The control variables in this study were individual characteristics, household characteristics, and health service characteristics of the respondents, such as gender, age, marital status, education level, annual per capita household income grouping, household size, drinking water, type of housing, type of toilet, chronic diseases, and two-week visit rate were controlled. Where the annual household income per capita is categorized into I, II, III, IV, and V, from lowest to highest, according to the internationally recognized income quintile method (30).
Measurement model
Outworking is not a random occurrence; it is also the result of a combination of individual and household characteristics and health service utilization. Direct comparisons of the health status of rural residents before and after outworking are therefore prone to sample selection bias and endogeneity problems, making it impossible to accurately assess the net effect of outworking on the health of rural residents. To address these issues, the propensity score matching method (PSM) can be used to calculate propensity scores to determine the experimental and control groups for analysis. The advantage of propensity score matching is that it "overcomes the problem of bias in observable variables by matching experimental and control group samples on a range of observable characteristics" (31). However, the endogeneity problem cannot be addressed fundamentally because of the unreasonable assumption that "unobserved characteristics do not affect health status" in the matching process, which can be corrected by the difference-in-differences method (DID), which can improve the deficiencies of propensity score matching (32). Therefore, a combination of these two methods, the PSM-DID method, was used to analyze the net effect of outworking on the health of rural residents. Then, the measurement model can be set as follows:
First, the propensity score was calculated. First, covariates were selected, which included relevant variables affecting the health status of rural residents; secondly, the conditional probability of whether the sample went out to work or not, i.e. the propensity score, was estimated through the covariates using a probit model as follows:
$$P\left(Xi\right)=Pr(D=1|Xi)$$
Where \(\text{D}=1\) refers to residents outworking and \(\text{X}\text{i}\) refers to the observable covariates.
Second, propensity score matching was used to split the experimental and control groups. The propensity score matching-difference-in-differences model was used to analyze the effect of outworking on the health status of rural residents. A double-difference analysis was first conducted on the successfully matched sample to overcome the shortcomings of the PSM method, and the difference-in-differences model (DID) was as follows:
$${Y}_{it}=\alpha +\gamma {D}_{t}+\beta {x}_{it}+{\mu }_{i}+{\lambda }_{t}+{\epsilon }_{it}$$
where \(\text{Y}\) refers to health status, \(\text{D}\) refers to whether residents go out to work, \({\mu }\) refers to unobservable individual fixed effects and \({\lambda }\) refers to time-fixed effects. The specific model is as follows:
$$AT{T}_{PSM-DID}=E[{Y}_{1}^{T}-{Y}_{0}T\mid P(X),D=1]-E[{Y}_{1}C-{Y}_{0}C\mid P(X),D=0]$$
Where \({Y}_{1}^{T}\) and \({ Y}_{0}\text{T}\) refers to the health status of the experimental group after and before the outworking occurred, respectively, and \({Y}_{1}\text{C}\) and \({ Y}_{0}\text{C}\) refers to the health status of the control group after and before the outworking. In this study, 2 periods of panel data were selected for 2019 and 2022. Residents who did not go out to work in 2019 but did so in 2022 were defined as the experimental group and those who did not go out to work in either year were defined as the control group.
Table 1
Variable definitions and descriptive statistics.
Variables | Definition | 2019年 | 2022年 |
Mean | Standard error | Mean | Standard error |
Explained variable (Y) | | | | | |
Self-rated health | Bad = 0, Fair = 1, Good = 2 | 1.487 | 0.739 | 1.246 | 0.776 |
Explanatory variable (X) | | | | | |
Outworking | Yes = 1 and No = 0 | 0.220 | 0.414 | 0.160 | 0.367 |
Control variable (C) | | | | | |
Gender | Male = 1 and Female = 0 | 0.525 | 0.499 | 0.524 | 0.499 |
Age | 18–29 years old = 1,30–44 years old = 2,45–60 years old = 3 | 2.023 | 0.839 | 2.114 | 0.848 |
Marital status | 1 = unmarried, 2 = married, 3 = divorced/widowed | 1.776 | 0.472 | 1.769 | 0.481 |
Education level | 1 = no schooling, 2 = primary school, 3 = junior high school, 4 = senior high school or above | 2.543 | 1.055 | 2.612 | 1.067 |
Annual per capita household income | I group = 1, II group = 2, III group = 3, IV group = 4, V group = 5 | 3.000 | 1.414 | 3.000 | 1.414 |
Family size | 1–3 persons = 1,4–5 persons = 2,≥6 persons = 3 | 2.046 | 0.723 | 1.957 | 0.748 |
Type of drinking water | 1 = tap water, 2 = cellar water, 3 = else | 1.858 | 0.454 | 1.990 | 0.216 |
Housing type | 1 = brick soil concrete, 2 = brick wood,3 = full brick,4 = else | 2.366 | 0.878 | 2.377 | 0.747 |
Toilet type | 1 = dry toilet,2 = water flushing type, 3 = else | 1.212 | 0.536 | 1.219 | 0.550 |
chronic disease | Yes = 1 and No = 0 | 0.188 | 0.391 | 0.181 | 0.385 |
Two-week visit rate | Yes = 1 and No = 0 | 0.060 | 0.238 | 0.032 | 0.176 |
Analysis of propensity score matching results
The nearest neighbor match (1:1 ratio) was used to calculate the average treatment effect (ATT) for the treatment group. Before calculating the ATT, a balance test and a co-support test were performed (33). As depicted in Fig. 1 for the balance test results, the covariate standardization bias (% bias) was considerably lower for the matched post-treatment and control groups (both < 10%). The results of the common support test are shown in Fig. 2, where most observations fall within the range of common values (on support), with a large overlap in propensity scores between the treatment and control groups, as depicted in the figure.
Standard empirical results
The propensity score matches the hypothesis of common support
Two conditions need to be met before using the propensity score matching method, one is that the common support assumption is met and the other is that the parallel assumption is met (34). In this study, the kernel density function plot was used to test whether the common support hypothesis was satisfied. As shown in Fig. 3, the kernel density distribution of the propensity scores of the outworking experimental group and the outworking control group before matching differed significantly, and the kernel density distribution curves of the outworking group and the outworking control group after matching fit better, indicating that the overall matching effect was good and the common support hypothesis was satisfied.
Processed the self-selection problem
Propensity score matching was applied to estimate the net effect of outworking on the self-rated health of rural residents, which can mitigate to some extent the bias caused by self-selection. As can be seen from Table 2, the standard deviations of the samples after the nearest neighbor matching all dropped below 10%. The standard deviations of the variables dropped significantly. In other words, the differences in the characteristics of the samples were controlled to some extent. The results of propensity score matching have good explanatory power. In addition, propensity score matching was found to have an impact on solving residents' self-selection problems through the balance test.
Table 2
Sample balance test of propensity score matching.
Variable | Sample | Mean | Standard deviation% | Deviation reduction % | t-test |
Processing group | Control group | T-value | P-value |
Gender | Before | 0.7312 | 0.4757 | 54.1 | 99.1 | 33.55 | < 0.001 |
| After | 0.7311 | 0.7288 | 0.5 | 0.27 | 0.789 |
Age | Before | 1.6969 | 2.1533 | -58.1 | 98.5 | -35.54 | < 0.001 |
| After | 1.6970 | 1.7039 | -0.9 | -0.47 | 0.638 |
Marital status | Before | 1.6867 | 1.7930 | -21.7 | 93.4 | -14.41 | < 0.001 |
| After | 1.6864 | 1.6935 | -1.4 | -0.69 | 0.492 |
Education level | Before | 2.8473 | 2.5117 | 33.2 | 89.5 | 20.48 | < 0.001 |
| After | 2.8472 | 2.8119 | 3.5 | 1.86 | 0.062 |
Annual per capita household income | Before | 2.7663 | 3.0547 | -20.3 | 96.0 | -13.15 | < 0.001 |
| After | 2.7666 | 2.7551 | 0.8 | 0.41 | 0.682 |
Family size | Before | 2.0343 | 1.9963 | 5.2 | 70.6 | 3.32 | < 0.001 |
| After | 2.0341 | 2.0453 | -1.5 | -0.79 | 0.432 |
Type of drinking water | Before | 1.8745 | 1.9316 | -14.7 | 93.5 | -10.03 | < 0.001 |
| After | 1.8747 | 1.8710 | 1.0 | 0.46 | 0.642 |
Housing type | Before | 2.3625 | 2.3732 | -1.3 | 22.9 | -0.84 | 0.401 |
| After | 2.3624 | 2.3542 | 1.0 | 0.51 | 0.610 |
Toilet type | Before | 1.1943 | 1.2204 | -4.9 | -8.8 | -3.09 | 0.002 |
| After | 1.1944 | 1.1659 | 5.3 | 2.87 | 0.004 |
Chronic disease | Before | 0.0792 | 0.2094 | -37.7 | 95.6 | -21.75 | < 0.001 |
| After | 0.0792 | 0.0735 | 1.6 | 1.08 | 0.280 |
Two-week visit rate | Before | 0.0282 | 0.0511 | -11.7 | 81.1 | -6.97 | < 0.001 |
| After | 0.0282 | 0.0239 | 2.2 | 1.37 | 0.172 |
Before matching, PS R2, LR chi2, and the mean and median of the standard deviation were 0.111, 2880.57, 23.5, and 19.7, respectively; After matching, the corresponding values are 0.003, 35.50, 2.6, and 2.6, so the matching effect is good.
Comprehensive evaluation of influences of outworking on self-rated health
Columns (1)-(4) of Table 3 investigate the impact of outworking on the self-rated health of rural residents. Columns (1) and (2) are the estimated results of oprobit regression, and columns (3) and (4) are the estimated results of PSM-DID. Column (1) and column (3) show that after controlling the fixed time effect and individual effect, the implementation of outworking significantly positively influences self-rated health. After controlling for personal characteristics, family characteristics, and medical characteristics, the regression results in columns (2) and (4) show that the self-rated health level of residents has decreased by 21.6 and 27.5%, respectively, significant at the statistical level of 1%.
The estimation coefficient of control variables is also worthy of note. Take column (2), for an example. The estimation coefficient of gender, education level, annual family income, and toilet type are significantly positive at the 1% level. The estimation coefficient of age, housing type, chronic disease, and two-week visit rate are significantly negative at the 1% level.
Table 3
The effect of outworking on the health status of rural residents
Variable | (1) | (2) | (3) | (4) |
Outworking | 0.685*** | 0.216*** | 0.375*** | 0.275*** |
| (0.046) | (0.050) | (0.052) | (0.055) |
Gender | | 0.088*** | 0.048 | 0.048 |
| | (0.026) | | (0.038) |
Age | | -0.551*** | | -0.582*** |
| | (0.022) | | (0.033) |
Marital status | | 0.027 | | 0.131* |
| | (0.037) | | (0.053) |
Education level | | 0.263*** | | 0.270*** |
| | (0.015) | | (0.023) |
Annual per capita household income | | 0.029*** | | 0.035* |
| | (0.009) | | (0.013) |
Family size | | 0.034* | | 0.034 |
| | (0.019) | | (0.028) |
Type of drinking water | | 0.096* | | 0.257*** |
| | (0.037) | | (0.063) |
Housing type | | -0.051*** | | -0.018 |
| | (0.016) | | (0.024) |
Toilet type | | 0.095*** | | 0.006 |
| | (0.024) | | (0.044) |
Chronic disease | | -1.419*** | | -1.554*** |
| | (0.035) | | (0.066) |
Two-week visit rate | | -1.209*** | | -1.159*** |
| | (0.061) | | (0.142) |
Individual constant | Control | Control | Control | Control |
Time constant | Control | Control | Control | Control |
N | 26738 | 26738 | 26738 | 26738 |
R2 | 0.0187 | 0.1521 | 0.0286 | 0.1290 |
Prob > F | < 0.001 | < 0.001 | < 0.001 | < 0.001 |
***, **, and * indicates significance on the statistical levels of 1, 5, and 10%, respectively. The numbers in the brackets are the standard errors of coefficient robustness.
Robustness test
To ensure the robustness of the above regression results, the PSM-DID method is further used to analyze the impact of outworking on the health status of rural residents. The combination of different propensity score matching methods can therefore better identify the net effect of outworking on the health of rural residents. Columns (1)-(5) of Table 4 investigate the robustness of five propensity score matching methods to study the impact of outworking on the health status of rural residents. The following five matching methods were estimated for radius matching, kernel matching, nearest-neighbor matching, mahalanobis distance matching, and local linear matching. Theoretically, regardless of the matching method used, the final estimation results are not very different. The estimated coefficient signs (β > 0) and significance levels (P < 0.05) are consistent across the different matching methods, so the positive effect of outworking on the health status of rural residents estimated in this paper is robust.
Table 4
Variable | Radius matching | Kernel matching | Nearest neighbor matching | Mahalanobis distance matching | Local linear matching |
Outworking | 0.21710*** | 0.21554*** | 0.21554*** | 0.21554*** | 0.21710*** |
| (0.04982) | (0.04981) | (0.04981) | (0.04981) | (0.04982) |
Control variables | Control | Control | Control | Control | Control |
Individual constant | Control | Control | Control | Control | Control |
Time constant | Control | Control | Control | Control | Control |
N | 26738 | 26738 | 26738 | 26738 | 26738 |
R2 | 0.1507 | 0.1521 | 0.1521 | 0.1521 | 0.1507 |
Prob > F | < 0.001 | < 0.001 | < 0.001 | < 0.001 | < 0.001 |
***, **, and * indicates significance on the statistical levels of 1, 5, and 10%, respectively. The numbers in the brackets are the standard errors of coefficient robustness.
Heterogeneity
Gender
This part tests for any heterogeneity in the effects of outworking on the health of rural residents in terms of gender. The sample was then divided into females(14, 022) and males(12, 716).The results in Table 5 show that outworking significantly improves the health of females(0.245) and males(0.240). Moreover, family size(0.048) significantly affects females' health but has no significant impact on that of males, and the type of drinking water(0.110) significantly affects males' health but has no significant impact on that of females. Age, education, annual average household income, housing type, toilet type, chronic disease, and two-week visit rate all have significant impacts on the health of males(-0.527, 0.243, 0.029, -0.052,0.125,-1.486 and-1.166, respectively) and females(-0.578, 0.270, 0.029, -0.050,0.065,-1.349 and-1.228, respectively).
Age
Individuals experience different physical and psychological conditions across their life stages. Therefore, the heterogeneity in the influence of the effects of outworking on the health of rural residents in terms of age cannot be ignored. In this study individuals aged between 18 and 29 years were classified as the young group(8, 704), those aged 30 to 44 years were classified as the middle-aged group(7, 559), and those aged 45 to 60 years were classified as the older group(10, 475).
As shown in Table 5, outworking has a significant impact on the health of the young(0.386) and middle-aged groups(0.169) but have no significant impact on that of the older group. In addition, the type of drinking water(0.207) only has a significant impact on the health of the middle-aged group, Annual average household income(0.041) and toilet type(0.156) only have a significant impact on the health of the older group. Housing type significantly affects the health of the young(-0.066) and older groups(-0.051). Gender significantly affects the health of the middle-aged(0.111) and older groups(0.166).Education(0.342, 0.266, and 0.221, respectively), chronic disease(− 2.560, − 1.757, and − 1.190, respectively) and two-week visit rate(-1.477, -1.509, and − 0.994, respectively) have significant impacts on health in all age groups.
Income
The unbalanced development of China has resulted in a serious income gap, thereby giving rise to a possible heterogeneity in the impact of outworking on the health of rural residents in terms of family income. This section divides the sample into the low-(5, 343), low and middle-income (5,354), middle-(5, 350), middle and high-income (5,351), and high-income(5, 340) groups based on annual average household income.
Table 5 shows that outworking significantly influences the health of adults from low-income, middle–income and high–income households, but does not significantly affect that of adults from low and middle income middle-and high-income households. Age, education, chronic disease, and Two-week visit rate significantly affect the self-rated health of all groups, whereas the effects of the other variables vary in significance across all groups.
Table 5
Heterogeneity analysis of gender, age, and income.
Variable | Male | Female | Young | Middle-aged | Older | Low–income | Low and middle income | Middle–income | Middle and high income | High–income |
| (1) | (2) | (3) | (4) | (5) | (6) | (7) | (8) | (9) | (10) |
Outworking | 0.245*** | 0.220** | 0.386*** | 0.169** | 0.152 | 0.323*** | 0.129 | 0.348*** | -0.060 | 0.338*** |
| (0.059) | (0.095) | (0.085) | (0.081) | (0.104) | (0.099) | (0.108) | (0.115) | (0.118) | (0.127) |
Gender | | | -0.082 | 0.111** | 0.166*** | 0.145** | 0.042 | 0.055 | 0.100* | 0.095 |
| | | (0.054) | (0.050) | (0.039) | (0.059) | (0.059) | (0.059) | (0.059) | (0.060) |
Age | -0.527*** | -0.578*** | | | | -0.588*** | -0.624*** | -0.483*** | -0.554*** | -0.507*** |
| (0.030) | (0.033) | | | | (0.049) | (0.049) | (0.050) | (0.052) | (0.051) |
Marital status | 0.049 | -0.015 | -0.042 | 0.027 | -0.035 | -0.031 | 0.098 | 0.105 | -0.062 | -0.010 |
| (0.048) | (0.057) | (0.059) | (0.075) | (0.083) | (0.077) | (0.082) | (0.082) | (0.087) | (0.084) |
Education level | 0.243*** | 0.270*** | 0.342*** | 0.266** | 0.221*** | 0.161*** | 0.290*** | 0.290*** | 0.301*** | 0.269*** |
| (0.021) | (0.023) | (0.037) | (0.027) | (0.022) | (0.034) | (0.034) | (0.034) | (0.034) | (0.034) |
Annual average household income | 0.029** | 0.029** | 0.020 | 0.027 | 0.041*** | | | | | |
| (0.013) | (0.014) | (0.019) | (0.018) | (0.014) | | | | | |
Family size | 0.013 | 0.048* | 0.015 | 0.034 | 0.033 | 0.097** | 0.155*** | 0.025 | 0.040 | -0.138*** |
| (0.026) | (0.027) | (0.040) | (0.036) | (0.026) | (0.042) | (0.041) | (0.041) | (0.041) | (0.042) |
Type of drinking water | 0.110** | 0.080 | -0.003 | 0.207** | 0.075 | 0.055 | 0.107 | 0.231*** | 0.006 | 0.091 |
| (0.052) | (0.053) | (0.077) | (0.066) | (0.056) | (0.080) | (0.081) | (0.082) | (0.083) | (0.088) |
Housing type | -0.052** | -0.050** | -0.066** | -0.033 | -0.051** | -0.071** | -0.080** | 0.000 | -0.035 | -0.062* |
| (0.022) | (0.023) | (0.033) | (0.029) | (0.023) | (0.035) | (0.035) | (0.034) | (0.036) | (0.036) |
Toilet type | 0.125*** | 0.065* | 0.016 | 0.035 | 0.156*** | -0.009 | -0.014 | 0.150*** | 0.147*** | 0.108** |
| (0.034) | (0.034) | (0.051) | (0.046) | (0.033) | (0.063) | (0.060) | (0.055) | (0.052) | (0.045) |
Chronic disease | -1.486*** | -1.349*** | -2.560*** | -1.757*** | -1.190*** | -1.548*** | -1.287*** | -1.518*** | -1.334*** | -1.462*** |
| (0.050) | (0.049) | (0.183) | (0.074) | (0.040) | (0.085) | (0.079) | (0.078) | (0.075) | (0.075) |
Two-week visit rate | -1.166*** | -1.228*** | -1.477*** | -1.509*** | -0.994*** | -1.720*** | -1.224*** | -1.101*** | -1.022*** | -1.232*** |
| (0.095) | (0.081) | (0.224) | (0.123) | (0.073) | (0.178) | (0.149) | (0.139) | (0.127) | (0.117) |
Individual constant | Control | Control | Control | Control | Control | Control | Control | Control | Control | Control |
Time constant | Control | Control | Control | Control | Control | Control | Control | Control | Control | Control |
N | 14022 | 12716 | 8704 | 7559 | 10475 | 5343 | 5354 | 5350 | 5351 | 5,340 |
R2 | 0.137 | 0.1636 | 0.0722 | 0.0854 | 0.0749 | 0.1505 | 0.1526 | 0.1511 | 0.1567 | 0.159 |
Prob > F | < 0.001 | < 0.001 | < 0.001 | < 0.001 | < 0.001 | < 0.001 | < 0.001 | < 0.001 | < 0.001 | < 0.001 |
***, **, and * indicates significance on the statistical levels of 1, 5, and 10%, respectively. The numbers in the brackets are the standard errors of coefficient robustness.