The impact of pension on the health behavior of elderly rural residents: evidence from China

We used data from the China Rural Revitalization Survey (CRRS). CRRS is a nationwide large-scale rural tracking survey conducted by Rural Development Institute, Chinese Academy of Social Sciences. The research team conducts a tracking survey every 2 years. The first survey of farmers and villages was conducted in 2020, with face to face questionnaire survey conducted in 50 county (city) units and 156 township (town) units in 10 provinces (districts) across the country. In terms of sample selection, the research group comprehensively considered the level of economic development, regional location, and agricultural development situation and randomly selected sample provinces from the eastern, central, western, and northeastern regions. An equidistant random sampling method was used to select sample counties based on the per capita GDP at the county level in the province, and considered covering the entire province (region) as much as possible in space. Using the same sampling method, sample townships and villages were randomly selected based on their economic development level. Finally, randomly selected sample households were based on the roster provided by the village committee. The sample provinces specifically include Guangdong Province, Zhejiang Province, Shandong Province, Anhui Province, Henan Province, Heilongjiang Province, Guizhou Province, Sichuan Province, Shaanxi Province, and Ningxia Hui Autonomous Region.

A total of 15,554 samples were obtained from the face to face questionnaire survey. The content covers basic personal information, family situations, health status, and other information about farmers. This article selects the latest cross-sectional data released in 2020 for analysis (regression discontinuity design is locally random and does not require tracking data). Referring to existing research on the new rural social security pension [38], we selected samples aged 45 to 75 years old, excluded samples with missing covariates, and finally analyzes 5803 samples.

The outcome variable of regression discontinuity is health behavior. Four questions were selected as proxy variables for health behavior, namely “Do you consciously control salt intake?”, “Do you consciously control sugar intake?”, “Do you consciously learn health or wellness knowledge?”, and “Do you consciously control edible oil intake?”. The answer “Yes” was assigned a value of 1, and “No” was assigned a value of 0. The disposal variable is whether to receive the new agricultural insurance pension. “Yes” is assigned a value of 1, and “No” is assigned a value of 0. In addition, as mentioned earlier, due to certain delays in pension payments, relevant studies have selected ages after 60 years old, such as 60.5 years old. However, due to the lack of specific research months in the CRRS, precise age cannot be obtained. Therefore, this article chooses 61 years old as the cutoff point (Cutoff).

In addition, in order to improve the accuracy of regression discontinuity estimation, this article also sets control variable (covariate) based on questionnaire questions, including gender (questionnaire options include male and female), education level (questionnaire options include not attending school, primary school, junior high school, high school, technical secondary school, vocational high school, college, and undergraduate), marriage (questionnaire options include married, unmarried, divorced, and widowed), BMI (questionnaire requires respondents to fill in their weight and height), disabled (questionnaire provides options for physical disability, brain damage/intellectual disability, visual impairment, hearing impairment, language impairment, and no disability), chronic diseases (questionnaire provides options for hypertension, dyslipidemia, blood glucose abnormalities, heart disease, language disorders, cancer or malignant tumors, liver or kidney or stomach diseases, neurological or psychiatric disorders, memory disorders, other chronic diseases, no chronic disease), and household income (questionnaire asks respondents in detail about the income of 8 major categories and and 23 subcategories of households, and summarizes them to obtain the household income. The 8 major categories of income include net income from planting, breeding, forestry, fishery, non-agricultural, wage/work income, property income, and transfer income) [39‐43].

Table 1 summarizes the sample characteristics. The elderly individuals were divided into two groups based age, namely 45–61 years old (excluding 61 years old) and 61–75 years old (including 61 years old). There are 3816 samples in the 45–61 age group, with males accounting for 50.68%. In terms of education level, 48.45% had an educational level of junior high school, 40.75% had an educational level of primary school and below, 9.85% had an educational level of high school/technical secondary school/vocal high school, and only 0.94% had an educational level of college/undergraduate. 95.78% of people are married. The average BMI index is 26.02, which is higher than the normal range (18.5–24), indicating that the body is overweight. 8.78 and 38.92% of the elderly individuals have disabled and chronic issues, respectively. The household income of the elderly was 76,409.81 Chinese Yuan (CNY), which was 10.74 after being logged. For the 61–75 age group, there were a total of 1987 samples, with males accounting for 53.45%. The education level of this group is lower. 68.50% had an educational level of primary school and below, 23.86% had an educational level of junior high school, 7.50% had an educational level of high school/technical secondary school/vocal high school, and 0.15% had an educational level of college/undergraduate. 88.83% of people are married. The average BMI index is 23.50, indicating a normal body type. The household income of this elderly group was lower than the first one (63,195.57 CNY for average, and 10.46 after being logged).

According to the New Rural Pension Scheme and Urban and Rural Resident Pension Scheme, the pension can be obtained on a monthly basis for the elderly with registered residence who have reached the age of 60 and have not accessed the Urban Workers’ Social Insurance. Therefore, we regarded only those who have reached the age of 60 to receive pension as a “quasi natural” experiment, and uses a regression discontinuity design (RDD) to evaluate the impact of the New Rural Pension Scheme on the health behavior of elderly rural residents. Previous studies have also pointed out that in rural China, there are no other policies that use the age of 60 as a boundary [44]. Therefore, the effect of the cutoff point of 60 will only be the income impact brought by the New Rural Pension Scheme, and will not be confused by other policy effects.

The regression discontinuity method can not only more accurately estimate the results caused by policy shocks, but also control the estimation bias caused by endogeneity issues. It is a causal identification method that is closer to random trials than the instrumental variable method and the double difference method [45, 46]. Regression discontinuity is divided into Sharp RD (SRD) and Fuzzy RD (FRD). If there is a deterministic change from 0 to 1 in the disposal state around the discontinuity of the driving variable, it is an accurate regression discontinuity; if the disposal variable is only a probability jump around the discontinuity, then it is a fuzzy regression discontinuity. Due to policy differences, insured individuals over 60 years old may not necessarily receive the pension, and a small number of insured individuals under 60 years old have also received the pension [47]. Therefore, this article has adopted a fuzzy discontinuity approach. The disposal variable is defined as whether an individual receives the pension, which is equal to 1 (received) or 0 (not received). The driving variable age is set to z_i, and the outcome variable health behavior is set to HB_i (see eq. 1). The disposal effect estimated by the fuzzy discontinuity is the ratio of health behavior to the jump in age of the pension. Due to the emphasis on parameter estimation near the critical point in regression discontinuity, A non parametric method is adopted to estimate the numerator and denominator of eq. 1 in specific operations. We use Stata16.0 software in data processing and empirical analysis.

Before entering the regression analysis, this article visually displayed the relationship between driving variables and disposal variables, and the relationship between driving variables and outcome variables in the form of a graph. Displaying the relationship between driving variables and disposal variables, and the relationship between driving variables and outcome variables, has become a standard practice in regression discontinuity (RD) analysis, which helps to intuitively understand the meaning of RD [45]. Figure 1 shows the relationship between the age of the driving variable and whether to obtain the pension (the standardized age point 0 in the figure is 61 years old). It can be intuitively seen that the probability of receiving a pension has a significant jump around the standardized age of 0, which confirms that Chinese insured individuals over the age of 60 can have access to pension.

Figure 2 shows the estimated results of the principal regression, (a), (b), (c), and (d) represent the relationship between age as the driving variable and conscious control of sugar, salt, and edible oil intake, as ell as learning health or wellness knowledge. It can be intuitively seen that sugar, salt, and edible oil intake, as well as learning health or wellness knowledge, have a significant upward jump on the right side of the discontinuity, Based on Fig. 1, it can be preliminarily explained that the positive income impact brought by pension can improve the health behavior of rural residents to a certain extent.

We calculated the optimal bandwidth under the minimum mean square error using the method commonly used in existing research and estimates the regression results of fuzzy discontinuity. According to Panel A in Table 2, when the dependent variable is sugar intake control, the optimal bandwidth is 3.59 years, which includes 589 samples on the left side of the discontinuity and 642 samples on the right side of the discontinuity. The regression discontinuity treatment had a significantly positive effect (P value of 0.064, not significantly different from the 5% significance level). Panel B in Table 2 indicates that when the dependent variable is controlled salt intake, the optimal bandwidth is 3.62 years, with 589 samples entering the bandwidth on the left side of the discontinuity and 642 samples on the right side of the discontinuity. The regression discontinuity treatment effect is significantly positive at the 10% level. Panel C and Panel D in Table 2 indicate that when the dependent variable is controlling edible oil intake, the optimal bandwidth is 3.70 years, which includes 589 and 642 samples on the left and right sides of the discontinuity, respectively. When the dependent variable is learning health or wellness knowledge, the optimal bandwidth is 4.00, and the left and right samples on the discontinuity are 864 and 795, respectively. Although the effect of regression discontinuity processing is not significant, the coefficient is still positive. It can be seen that the income impact brought by pension can effectively improve the health behavior of rural residents. In addition, the first stage results in Table 2 report the impact of the driving variable age on pension receiving, with significant results at the 1% level, fully indicating that the age discontinuity can be used as a “instrumental variable” for determining whether to receive pension.

In order to ensure the credibility of the above conclusions, this study has done a series of robustness tests, including continuity test, different bandwidth analysis and placebo test.

Continuity testing is an important prerequisite for obtaining consistent estimates in regression discontinuity. This study adopts two testing methods. Firstly, it tests whether each antecedent variable is continuous at the discontinuity. The regression setting is the same as the previous text, but the dependent variable is replaced with the antecedent variable. As shown in Table 3 (1)–(7), the estimated results are not significant, indicating that all the predetermined variables have no discontinuity at the discontinuity. The second method of testing the continuity hypothesis is to observe the distribution of driving variables to test whether age is manipulated. If there is data heaping at the discontinuity, it will affect the regression discontinuity results [48]. Figure 3 shows that there is no data accumulation at the age of 61 at the discontinuity, so it can be considered that there is no discontinuity at the discontinuity and no trace of manipulation. Therefore, based on the above two test results, it can be seen that the continuity assumption of regression discontinuity is satisfied.

The optimal bandwidth used in the benchmark regression results shown in Table 2 is symmetric about both sides of the discontinuity, and Table 4 uses different bandwidth on both sides for estimation as a robustness test. As shown in Table 4, when the dependent variable is controlling sugar intake, the left bandwidth of the discontinuity is 3.516, which accommodates 589 samples, and the right bandwidth of the discontinuity is 5.839, which accommodates 976 samples. The regression discontinuity results are significant at the 10% level. When the dependent variable is controlling salt intake, the left bandwidth of the discontinuity is 3.482, which accommodates 589 samples, and the right bandwidth is 6.175, which accommodates 1120 samples. Although the regression results are not significant, the p-value is 0.105, which is close to the 10% significance level. In addition, this article also customized 3.5 years and 4.5 years as bandwidth, and repeated regression discontinuity. When the dependent variable was controlled for sugar intake, both had a significance level of 10%; When the dependent variable is controlling salt intake, the bandwidth is positively significant at the 10% level for 3.5 years. Although the bandwidth is not significant at 4.5 years, the p-value is 0.106, which can be considered close to the 10% significance level and the coefficient is also positive. In summary, the regression results under different bandwidths can demonstrate that the pension income can improve the health behavior of rural residents.

Drawing inspiration from the previous using placebo testing, to demonstrate that health behavior is only affected by the income shock of pension funds, rather than other mechanisms, then making “pseudo discontinuity” at other ages away from 60 years old should not result in significant regression results. This article also selected 56, 57, 58, 62, 63, and 64 years old as “pseudo discontinuity” for placebo testing. As shown in Panel A and Panel B in Table 5, the regression results are not significant, indicating that the improvement of health behavior near the age of 60 is indeed the effect of the pension benefits.

We selected income and gender for further heterogeneity analysis.

The difference in economic level is an important reason for the difference in farmers’ health [49]. Table 6 objectively selected household income and subjectively selected household income satisfaction as measurement indicators for income heterogeneity analysis. In terms of household income, according to the three-way method, the sample is divided into low-income, middle-income, and high-income. The regression discontinuity effects of Panel A and Panel B both show that low-income families are more likely to improve their health behavior under the impact of receiving pension benefits. In terms of household income satisfaction, based on the question “Are you satisfied with your family’s current income level?”, the sample is divided into satisfied, commonly, and dissatisfied. The results of Panel A and Panel B show that rural residents who believe that their household income is average are more likely to improve their health behavior after receiving pension benefits. From the coefficient perspective, The coefficient of health behavior of rural residents who are dissatisfied with their family income under the impact of the new rural social security pension is relatively equal to or higher than the coefficient of satisfaction with their family income (Panel A) or higher (Panel B). The regression effects of Panel C and Panel D, although not statistically significant, still have a similar distribution as the coefficients of Panel A and Panel B. The above results indicate that the exogenous impact of the pension income is more likely to improve the health behavior of medium and low income rural residents, and therefore has positive policy implications.

As shown in Panel A, Panel B, and Panel C in Table 7, when the dependent variable is controlling for sugar, salt, and edible oil intake, the regression discontinuity effect of the male rural resident sample is significantly positive at the 5% level, while it is not significant for women. When the dependent variable is learning health knowledge or wellness knowledge (Panel D), although the regression discontinuity effect is not significant for both men and women, the coefficient of the male sample is still positive. The above results indicate that the new rural pension income has a more significant improvement effect on the health behavior of male rural residents.