2. VENTA ON-LINE
2.4. ANÁLISIS DE VENTAJAS E INCONVENIENTES DE VENDER ON-LINE
3.6.1 I
NITIAL ESTIMATESTable 3.3 shows the estimates of the effects of the NRCMS on utilization of formal medical care, preventive care, folk doctors and inpatient care38 based on logit models since all these outcomes are dummy variables. For each dependent variable, estimation results are first shown based on models with only demographic characteristics, and then additional covariates that may be correlated with the insurance status are added in the models. The only significant effect occurs for preventive care, and the probability of using preventive care is 1.4% higher for the insured compared to the uninsured with only demographic characteristics, while the magnitude of the coefficient is largely reduced after controlling for other covariates but still remains to be significant at the 1% level.
35 Since the propensity score distribution is very similar across the treatment and comparison groups (Figure A 3.2), NN without replacement can achieve good matching quality. In addition, NN without replacement will make use of more uninsured people from the comparison group to construct the counterfactual outcome and thereby reduces the variance of the estimates.
36 We use NN matching instead of kernel matching or Mahalanobis distance matching because NN can achieve balance with the maximum amount of covariates. Since omitting important variables can seriously increase bias in estimates, the outcomes will be more likely to be independent of treatment under the NN matching.
37 The sensitivity of results to three other choices of caliper (0.005, 0.0025, 0.00125) is shown in Appendix (Table A 3.2). We also try NN without replacement in both ascending and descending order (Table A 3.3). The estimates remain similar across different choices of caliper and order in which observations get matched.
38 The sample size of inpatient care use is particularly small compared to other dependent variables since the survey question is only
Tables 3.4 and 3.5 take a further look at the utilization of different levels of health facilities based on both actual utilization and stated household preference. In Table 3.4, the probability of using village clinics is 12% to 17%
higher among the insured people compared to the uninsured ones. The same significant effect is also found in Table 3.5 using stated preference data. Insured households are around 13% to 20% more likely to use village clinics, and the effect is significant at 1% level. Meanwhile, the likelihood of using town, county and city hospitals is around 11%, 19% and 2.5% lower among the insured, respectively. The estimates in Table 3.4 are quite similar to the ones in Table 3.5 with the same signs and similar magnitudes of the coefficients. However, the results based on stated household preference (Table 3.5) are more precise given the larger sample size and lower variance.
Therefore, in the following sections only results based on stated household preference are presented.
[Tables 3.3, 3.4 and 3.5 about here]
The results from logit estimation provide suggestive evidence that the NRCMS is related to an increase in the use of preventive care and village clinics and a decrease in the use of county and city hospitals. However, the results here may not be reliable if the treated and comparison groups differ prior to treatment in ways that matter for the outcomes under study. Since the NRCMS is offered on a voluntary basis, it is very likely that the participants and non-participants may be different in terms of unobserved individual characteristics that influence both their decision to participate in the program and their levels of outcomes. For instance, the insured people may deliberately select themselves into the NRCMS because they expect to use many health services and benefit more from the insurance. Therefore, simply employing logit models conditional on observables may provide a misleading estimate of the effect of the program due to omitted variable bias.
3.6.2 I
MPACT ESTIMATESTo reduce the self-selection bias related with observational data, DID and NN matching are used to control for the observed and unobserved characteristics between the insured and uninsured subjects. Table 3.6 shows the treatment effects of the NRCMS based on six waves of data (1991, 1993, 1997, 2000, 2004 and 2006). We first show the results based on DID estimation only and then employ NN matching before DID to make the selected non-treated group as similar as possible to the treated group in terms of their observed characteristics. For both models, the NRCMS appears to have negative impact on the use of folk doctors, with the insured 2% to 3% less likely to use health services provided by folk doctors. Both the use of village clinics and county hospitals appear to
be around 10% higher for the insured compared to the uninsured, although the coefficient for county hospitals is only significant at 10% level under the DID combined with NN matching model. The insured households also appear to be around 16% to 20% less likely to use town hospitals.
[Table 3.6 about here]
To investigate whether the significant results are driven by the change of the sample or the addition of covariates, a series of logit regressions are estimated for all significant outcome variables (e.g. folk doctors, village clinics, town hospitals and county hospitals). The estimation starts with the model with only demographic characteristics and subsequently adds different groups of independent variables such as income, education status, occupation, chronic disease, health risk preference and community health facility characteristics (from Model 1 to Model 7).
The treatment effects of the NRCMS on the use of town hospitals and county hospitals are significant and robust across all models. However, there is only a statistically significant effect of the NRCMS on the use of folk doctors and village clinics in the last model. It is possible that the significance is driven by the change of the size or composition of the sample since Model 7 has the least number of observations compared with previous models.
Therefore, Models 1-6 are re-run based on the sample of Model 7 for the use of folk doctors and village clinics (Appendix). The significance of folk doctors seems to be related to the change of the sample in the last model, while the significance of village clinics is merely due to the addition of facility characteristics 39. Since facility characteristics are important covariates that may confound the effect of the NRCMS on health care utilization and needs to be controlled for in the complete model, the significant treatment effect on the use of village clinics should also be regarded as robust.
[Tables 3.7, 3.8, 3.9 and 3.10 about here]
Similar results are also observed when we combine ITT analysis with DID models (Table 3.11). The NRCMS is found to decrease the likelihood of using folk doctors and town hospitals by 3% and 12-15%, respectively, and increase both the probabilities of using village clinics and county hospitals by more than 10%. All the estimates are significant at the 1% level. The robustness check still shows that the significance for village clinics, town
39 Table A 3.4 (Appendix) shows that all the treatment effects become significant when we use the sample based on Model 7, indicating that the significance of folk doctors is mainly driven by the change of the sample size or composition. Table A 3.5 shows that the significance of village
hospitals and county hospitals are consistent across various specifications with different groups of independent variables40.
[Table 3.11 about here]
Tables 3.12 and 3.13 present the DID estimates for different subgroups. In Table 3.12, the impact of the NRCMS is estimated for different income groups. The results suggest that the decrease in the use of folk doctors is mainly driven by low- and middle-income households, suggesting that the poor are less likely to use informal health services with insurance coverage. The rich have seen a significant increase in the use of village clinics and a decrease in the use of town hospitals. It suggests that the NRCMS seems to be more favourable to the rich in obtaining more convenient access to health services from village clinics. Table 3.13 presents the heterogeneous impact of the NRCMS across regions since the insurance program differs geographically41. Larger impacts of the NRCMS are observed for eastern and western provinces. The insured from the eastern provinces are significantly less likely to use folk doctors, village clinics and town hospitals while significantly more likely to use county hospitals and private clinics. On the other hand, people from western provinces have seen significant increase in the use of village clinics, county hospitals and city hospitals while significant decrease in the use of town hospitals and private clinics.
[Tables 3.12 and 3.13 about here]
In conclusion, although the NRCMS appears to have no effect on the use of formal medical care, preventive care and inpatient care, it seems to be associated with a change in type of health facility used when ill. The insured are less likely to use town hospitals, but more likely to use village clinics and county hospitals42. The shift from town hospitals to village clinics and county hospitals is inconsistent from the design of the insurance benefit package.
The NRCMS provides little coverage for outpatient care delivered by village clinics, and the rate of reimbursement is generally lower in county hospitals than in town hospitals. However, similar results have been found in
40 Tables A 3.6-A 3.9 (Appendix) show that the significance for the use of town and county hospitals is consistent across different models with various independent variables, while the significance for the use of folk doctors and village clinics only appears in the last model. Similar as before, Table A 3.10 (Appendix) suggests that the significance for the use of folk doctors is driven by the reduction of the sample size, and Table A 3.11 (Appendix) suggests that the significance for the use of village clinics may be due to the addition of facility characteristics.
Therefore, the treatment effects on the use of village clinics, town hospitals and county hospitals should be robust.
41 The eastern provinces are the richest compared with the other two areas while the western provinces are the poorest. Eastern provinces include Liaoning and Shandong provinces. Middle provinces include Heilongjiang, Henan, Hubei and Hunan provinces. Western provinces include Guangxi and Guizhou provinces.
42 Further robustness checks are shown in Tables A 3.3 and A 3.12 (Appendix). Table A 3.3 presents the results based on different calipers and matching orders for nearest neighbour matching. The significance on the use of village clinics disappears in all four models. In Table A 3.12, we adjust the standard errors by clustering at the community level. The standard errors become much larger, leading to smaller t-statistics and less significant results. The significant effects on the use of village clinics and county hospitals all disappear. However, there may be problems with too few clusters since the number of clusters here is only 28. According to , the number of clusters is considered to be too small if it ranges from less than 20 to less than 50 clusters for balanced clusters and even more for unbalanced clusters.
previous studies as well (Babiarz et al., 2010; Wagstaff et al., 2009a). The results suggest that the NRCMS not only leads to more use of primary care, but also helps patients to receive more specialized and better quality medical care provided by high-level health facilities. The decrease in the use of town hospitals given relatively more generous reimbursement rates may indicate that insured patients are more concerned about the accessibility and quality of health care compared to its price.