Impact of NRPS on the utilization of healthcare services
Table 2 shows the estimation results for outpatient and inpatient medical service utilization and discretionary over-the-counter drug utilization. The first stage results are about the relationship between age and access to pensions. The standardized age at the discontinuity point is 60 years old. From Fig. 1, the probability of receiving an NRPS pension jumped significantly around age 60, which is consistent with the design of the NRPS system. In addition, the first stage results in Table 2 show the effect of age on pension receipt, and the results are all significant, which provides ample evidence that the cutoff at age of 60 years old was appropriate as an "instrumental variable" for whether to receive a pension.
Table 2
|
Panel A: Inpatient
|
Panel B: Outpatient
|
Panel C: Discretionary drug purchase
|
|
First stage
|
Inpatient
|
First stage
|
Outpatient
|
First stage
|
Discretionary drug purchase
|
Pension
|
0.420***
|
0.099
|
0.358***
|
0.110
|
0.303***
|
0.329*
|
|
0.043
|
0.080
|
0.048
|
0.109
|
0.054
|
0.188
|
Bandwidth
|
|
3.379
|
|
2.739
|
|
2.299
|
Sample size
|
|
740;1187
|
|
645;891
|
|
556;763
|
Note: Robust criteria errors for clustering to the individual level are reported in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. The default minimum mean square error bandwidth is used in the table.
Following Calonico’s method[38], we calculated the optimal bandwidth under the minimum mean square error and estimated the regression results for the fuzzy discontinuity. From Table 2 Panel A, when the dependent variable is Inpatient care, the optimal bandwidth is 3.379 years. The treatment effect of discontinuity regression is not significant with the growth of 9.9 percentage point. From Table 2 Panel B, the optimal bandwidth is 2.739 years when the dependent variable is outpatient visits and the treatment effect of discontinuity regression is not significant with the growth of 11 percentage point. In Table 2 Panel C, when the dependent variable is discretionary drug purchases, the optimal bandwidth is 2.299 years and the discontinuity regression treatment effect is significantly positive. These regression results show that NRPS pension had no significant effect on inpatient and outpatient healthcare services utilization, but had a significant effect on discretionary drug purchasing behavior. Figure 2 visualizes a significant upward jump in the probability of discretionary drug purchasing behavior of individuals after receiving an NPRS pension, where the pension income shock increased the probability of discretionary drug purchases by 33 percentage points among rural residents. When rural residents receive the pension, their income increase one fixed amount compared with no-pension status. The results show that pension receipt significantly increases the probability of spontaneous drug purchase by rural residents. The effect of pensions on outpatient and inpatient services was also positive, but not significant.
Subsample analysis
We conducted a subsample analysis based on health status and income. Health status includes respondents' self-rated health status, whether they are depressed and whether they have chronic diseases. Based on their median household wage income plus net agricultural income, respondents were divided into high-income and low-income groups. In the subsample analysis, we found there were no significant changes in outpatient and inpatient service utilization at age 60. But we confirmed a significant increase in the probability of discretionary drug purchases from the NRPS income shock in healthier populations. Table 3 shows that the impact of the pension income shock on over-the-counter drug purchases was more pronounced among those who were free of depression and chronic illness and who self-rated their health status as moderate or healthy. This suggests that healthier individuals were more likely to increase their investment in their health through over-the-counter discretionary drug purchases after receiving NRPS compared to unhealthy individuals.
Table 3 also shows that self-rated moderate or healthy individuals had a significant 27 percentage point increase in the probability of discretionary drug purchases in response to the NPRS pension shock. The income shock effect on the group that answered poor health was not significant. Similar results were found for the subsample on depression and chronic diseases, where the effect of NPRS on the utilization of health services, including over-the-counter drugs, were not significant, but on the receipt of NPRS, the probability of discretionary drug purchasing behavior increased among rural residents who did not suffer from depression (28%) and did not have chronic diseases (25%). Of course, individuals in poor health, with depression and self-rated unhealthy might have increased prescribed drug purchases as part of curative inpatient and outpatient services. In contrast, discretionary over-the-counter drug purchases were mainly for symptom relief and disease prevention [41]. The receipt of NRPS relaxed the budget constraint on individual consumption behavior that provided funds for discretionary over-the-counter drug purchases, especially among those who were without chronic diseases and knowledgeable about their health.
Finally, we divided the sample into low- and high-income groups using the mean value of income. The results in Table 3 show that receipt of NRPS mainly influenced the over-the-counter drug purchases of the low-income group, with the probability of discretionary drug purchases by the low-income group increasing by about 50 percentage points. This is consistent with our theoretical framework where income reflects the ability to access related resources and is an important factor influencing the utilization of discretionary drug services. The budget constraint is weaker for the high-income group, while the demand for discretionary drugs is suppressed for the low-income group before receipt of the NRPS. The pension income brought by the NPRS relaxed the budget constraint for the low-income group.
Table 3
|
|
Drug purchase
|
Bandwidth
(Sample size)
|
Inpatient
|
Bandwidth
(Sample size)
|
Outpatient
|
Bandwidth
(Sample size)
|
Depression
|
Not depression
|
0.276*
|
3.227
|
0.060
|
3.367
|
0.243
|
2.210
|
(0.166)
|
(422; 649)
|
(0.100)
|
(432; 681)
|
(0.221)
|
(317; 434)
|
Depression
|
0.182
|
3.254
|
0.145
|
3.789
|
0.034
|
3.248
|
(0.156)
|
(301; 490)
|
(0.115)
|
(341; 565)
|
(0.132)
|
(296; 470)
|
Self-rated Health
|
Moderate or Healthy
|
0.270*
|
3.029
|
0.024
|
3.675
|
0.149
|
2.694
|
(0.152)
|
(472; 686)
|
(0.070)
|
(556; 841)
|
(0.126)
|
(443; 575)
|
Bad
|
0.222
|
2.804
|
0.281
|
3.101
|
0.053
|
3.071
|
(0.244)
|
(207; 327)
|
(0.202)
|
(214; 379)
|
(0.193)
|
(213; 363)
|
Chronic diseases
|
No chronic diseases
|
0.251*
|
3.481
|
0.002
|
4.279
|
-0.029
|
2.295
|
(0.143)
|
(442; 683)
|
(0.066)
|
(558; 824)
|
(0.171)
|
(333; 418)
|
Otherwise
|
0.162
|
3.097
|
0.180
|
3.593
|
0.156
|
3.960
|
(0.172)
|
(285; 480)
|
(0.132)
|
(324; 557)
|
(0.123)
|
(359; 611)
|
Income
(net of pension)
|
Low
|
0.488**
|
2.632
|
0.270*
|
3.029
|
0.270*
|
3.029
|
(0.229)
|
(667; 928)
|
(0.152)
|
(472; 686)
|
(0.152)
|
(472; 686)
|
High
|
-0.013
|
4.537
|
0.077
|
4.062
|
-0.026
|
4.629
|
(0.108)
|
(514; 757)
|
(0.089)
|
(424; 696)
|
(0.076)
|
(528; 764)
|
Robust criteria errors for clustering to the individual level are reported in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01.
Robustness tests
To test the robustness that over-the-counter drug utilization was driven by the pension, we conducted continuity, data heaping, placebo and regressions using different bandwidths and cutoff tests.
1. Continuity test
The prerequisite of applying discontinuity regression is that all predetermined variables are free from discontinuity at the cutoff. All covariates were tested by discontinuity regression, and the regression results are shown in Table 4 and Fig. 3. We first test the continuity of each predetermined variable at the cutoff, and the regression settings are the same as in the previous regressions, except that the dependent variables are replaced with covariates. In Table 3, the estimation results for variables (1)-(9) show that there is no discontinuity at the cutoff for all the antecedent variables.
Table 4
|
|
Pension
|
Bandwidth
|
Sample size
|
(1)
|
Health insurance
|
-0.03014
|
2.312
|
556; 763
|
(0.04525)
|
|
|
(2)
|
Depression
|
0.00078
|
3.649
|
781; 1261
|
(0.09942)
|
|
|
(3)
|
Education
|
-0.61758
|
2.251
|
556; 763
|
(.68006)
|
|
|
(4)
|
Marital Status
|
-0.36904
|
3.248
|
718; 1119
|
(0.28038)
|
|
|
(5)
|
Number of chronic diseases
|
0.24649
|
2.405
|
581; 788
|
(0.34909)
|
|
|
(6)
|
Medication needs
|
-0.15099
|
3.811
|
819; 1321
|
(0.10045)
|
|
|
(7)
|
Sex
|
-0.10119
|
4.861
|
1188; 1667
|
(0.0665)
|
|
|
(8)
|
Self-rated health
|
-0.01528
|
3.998
|
862; 1379
|
(0.08584)
|
|
|
(9)
|
Income
(net of pension)
|
0.42971
|
2.804
|
655; 941
|
(0.31961)
|
|
|
2. Data heaping test
The fulfillment of the continuity assumption requires that there is no data heaping in the sample population at the cutoff [42]. To test for data heaping, a histogram of the data distribution is shown in Fig. 4, and it is clear from Fig. 4 that there is no data heaping.
3. Placebo test
The placebo test shows that the health care burden is affected only by the pension income shock, and not by other mechanisms, when the regression results are insignificant at other ages away from 60. The regression results should be insignificant when we make a "pseudo cutoff" at other ages away from 60 years old. In Table 4, the regression results for "pseudo-cutoff" ages 57, 58, 59, 61, 62, and 63 were not significant, which suggests that the increase in the probability of discretionary drug purchase near the age of 60 was due to NRPS.
Table 4
|
(1)
|
(2)
|
(3)
|
(4)
|
(5)
|
(6)
|
Cutoff
|
57
|
58
|
59
|
61
|
62
|
63
|
|
Probability of discretionary drug purchase
|
Pension
|
-0.36698
|
0.69171
|
0.87341
|
0.32878
|
0.28203
|
-2.1045
|
(0.6513)
|
(0.60854)
|
(1.112)
|
(0.65295)
|
(0.28669)
|
(6.7784)
|
Bandwidth
|
3.233
|
2.614
|
2.096
|
3.038
|
3.119
|
2.930
|
Sample size
|
1047;768
|
586;700
|
416;666
|
821;1081
|
953 ;1090
|
994;1015
|
|
Inpatient
|
Pension
|
-1.0819
|
-0.01068
|
0.0613
|
0.24706
|
0.06453
|
-0.30709
|
|
(0.66936)
|
(0.23101)
|
(0.66866)
|
(0.29373)
|
(0.22644)
|
(0.57976)
|
Bandwidth
|
3.079
|
4.188
|
2.143
|
5.188
|
3.096
|
3.687
|
Sample size
|
985;710
|
1241;1226
|
416;666
|
1266;1777
|
953 ;1090
|
1217;1236
|
|
Outpatient
|
Pension
|
-1.7666
|
0.35675
|
0.27627
|
1.1342
|
-0.06784
|
-0.04335
|
|
(1.954)
|
(0.30082)
|
(0.64699)
|
(1.4924)
|
(0.22001)
|
(0.32914)
|
Bandwidth
|
2.551
|
3.191
|
1.544
|
3.542
|
3.013
|
4.035
|
Sample size
|
781;565
|
827;907
|
333;469
|
946;1223
|
934 ;1065
|
1302;1333
|
4. Regression using different bandwidths
Our optimal bandwidth for the baseline regression was symmetric about both sides of the cutoff. Different bandwidths on both sides of the cutoff were applied as a robustness check [37]. Table 5 shows that the regression results were still significant with a bandwidth of 2.861 years on the left side of the cutoff and a bandwidth of 2.380 years on the right side. We also customize 4 years as the bandwidth, and the results remain robust.
Table 5
Estimation results at different bandwidths
|
Panel A
|
Panel B
|
Panel C
|
Discretionary drug purchase
|
Inpatient
|
Outpatient
|
Pension
|
0.28673*
|
0.17292*
|
0.10006
|
0.07948
|
0.11012
|
0.09868
|
(0.14995)
|
(0.09106)
|
(0.08782)
|
(0.06601)
|
(0.10165)
|
(0.06991)
|
Bandwidth
|
2.861;2.380
|
4
|
2.688;3.904
|
4
|
2.867;2.930
|
4
|
Observations
|
663;788
|
862;1379
|
645;1349
|
862;1379
|
663;1019
|
862;1379
|
Robust criteria errors for clustering to the individual level are reported in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01.
5. Cutoff
For rural residents, there is no other policy that uses age 60 as a decomposition [43], so the effect of the age 60 cutoff can be attributed to the NRPS income shock.