Did More Generous Health Insurance Improve Health Outcomes of The Elderly? Evidence From China

3 Background: Catastrophic Medical Insurance (CMI) has been piloted in China Since 2012 and 4 gradually implemented in various regions. Most studies focus on its impact on medical economic 5 risks, and few studies discussed the impact of CMI on health of the elderly. 6 Methods: This study used Chinese Longitudinal Healthy Longevity Survey (CLHLS) data to 7 explore the impact of CMI on health of the elderly. Difference-in-differences (DID) and Propensity 8 score matching-DID were employed to study the health impact of CMI. Heckman selection model 9 was used to study the potential mechanisms. 10 Results: We found that the implementation of CMI improved the mental health of the elderly, and 11 the effect was limited. Moreover, the positive effect of CMI on the health of the elderly was mainly 12 in the high-income group, and CMI had no significant effect on any health indicators of the low- 13 income group population. The potential channel of CMI health improvement was its ability to 14 increase hospitalization rate in the elderly. 15 Conclusions: Therefore, the health promotion and equity of CMI deserve more attention, and the 16 compensation level of CMI needs to be improved under the premise of sustainable and effective 17 supervision of the fund.


3
Background 1 The challenge of aging has prompted countries around the world to pay attention to aging 2 health. According to China's seventh population census in 2020, people aged 60 and above 3 accounted for 18.7% of the national population, an increase of 5.44% compared with the sixth 4 recall and language. Each question has three options, namely correct, wrong and not able to answer. 16 We recoded the answer "not able to answer" as wrong based on existing research [38,39]. Finally, 17 a continuous variable MMSE with a value range of 0-30 points was obtained by summing up all the 18 related variables. Individuals with higher MMSE scores had better cognitive function. 19 In addition, in terms of mental health, this study referred to the existing references and selected 20 index of positive and negative well-being to measure the mental health status of the elderly [40]. 21 There were three related questions about positive mental health, including how do you rate your life 22 9 at present, with 1=very bad, 2=bad, 3=so so, 4=good, 5=very good; Do you always look on the 1 bright side of things, with 1=always, 2=often, 3=sometimes, 4=seldom, 5=never; Are you as happy 2 as when you were younger, with 1= always, 2=often, 3=sometimes, 4=seldom, 5=never. The 3 answers to these three questions were added together to obtain the index of positive well-being 4 ranging from 3 to 15, which is the positive health variable used in our study. Individuals with higher 5 index had better mental health. Negative mental health involves three related questions, including 6 Do you often feel fearful or anxious? Do you often feel lonely and isolated? Do you feel the older 7 you get, the more useless you are, and have trouble doing anything? Each of the above three 8 questions has five options: 1 = always, 2=often, 3=sometimes, 4=seldom, 5 = never. The negative 9 score ranging from 3-15 were obtained by summing up the three variables. Individuals with lower 10 index of negative well-being had better mental health. 11

CMI 12
The key explanatory variable in this study was CMI, which is whether an area has implemented 13 CMI in the past year. To construct this variable, we collected the policy documents issued by 14 provinces and cities to implement CMI, and matched them with CLHLS data sets at the individual 15 level to explore the impact of CMI on health among elderly. The policy documents of different 16 regions from the official websites of each regional government or the medical insurance 17 administration. Our judgment is based on the year in which each province started to fully implement 18 CMI. The time distribution of each province implementing CMI was shown in figure1. Finally, we 19 selected five provinces as the treatment group, namely Liaoning, Jilin, Fujian, Hubei and Chongqing, 20 and took the other provinces as the control group. In this study, we controlled for variables that might confuse the relationship between health 5 outcomes and CMI, with variable selection based primarily on existing studies [14,22] and data 6 availability. Covariates included the following categories: sociodemographic characteristics, 7 socioeconomic status, and health-related behaviors. Sociodemographic characteristics included age, 8 gender, marital status and number of children, and whether the elderly live alone. Marital status is 9 divided into two categories, with 0 indicating the respondent was divorced, widowed or never 10 married, 1 indicating the elderly was married. Whether the elderly live alone or not is a dummy 11 variable, 0 represents the respondent live with his/her family. 1 indicated the elderly lived alone. 12 Socioeconomic status variables included years of education, type of job held before age 60, 13 household income per capita(logarithmic), and place of residence, where the job type before the age 14 11 of 60 is a dummy variable (1 = had a white-collar job, 0=others). There are two categories of 1 residence, where 0 means an individual lives in a rural area and 1 means an individual lives in an 2 urban area. Regarding the health-related behaviors, we controlled 3 variables: Smoking (1=smoke 3 at the present) and Drinking (1= drink alcohol at the present). Considering that chronic diseases is 4 degenerative diseases, the number of chronic diseases suffered by the elderly is also included in our 5 study as a control variable. 6

Statistical Methods 7
Difference-in-Differences model 8 As mentioned before, the implementation time of CMI in different provinces (municipalities 9 directly under the Central Government and autonomous regions) is inconsistent, so we adopted 10 difference-in-differences model to study the impact of CMI on health outcomes of the elderly, and 11 take it as the benchmark. The specific model is set as follows. 12 is the provincial fixed effect used to control for 21 confounding factors at the provincial level; is the error term of the model. 22 12 In addition, in order to answer our second question, we added interaction items of 、 1 and residence, income variable in the model (1) to verify the CMI effect differences to the 2 health of urban and rural residents and different income group population (model (2)). 3 Where is the × in model (1). Urban is a dummy variable, where 1 7 means that individuals live in cities and towns, and 0 means that individuals live in rural areas. 8 Therefore, coefficient 1 reflected the difference of health impact of CMI on urban and rural 9 residents. In order to further study the influence of CMI on the health of individuals in different 10 income group population, we divided the sample into three categories according to the income 11 quantile of the sample, and generated three corresponding binary variables. Considering that there 12 would be complete collinearity if all of them were added into the model, we took the low-income 13 group as the reference group and added the interaction terms of and Middle_income , 14 and High_income , respectively to observe the impact of CMI on the health of the elderly 15 with different income levels. 16

PSM-DID 17
To make individuals more comparable between treatment and control groups, we used 18 propensity score matching-difference-in-differences (PSM-DID) method based on the 19 benchmark [14] and take it as the robustness test. To be specific, we first screened the sample and 20 set it as balanced panel data, that is, every individual in the sample was interviewed in both two 21 waves. Then, the sample of the former wave (2011/12) and the covariate mentioned above were 22 13 used to establish logit model to obtain propensity score ( ( = 1| )), and the matching of 1 individuals between the treatment and control group in the common support was carried out by the 2 kernel matching strategy. Then the matched data and DID model were employed to estimate the 3 impact of CMI on the health of the elderly. Finally, the average treatment effects on the treated was 4 as follows. Characteristics of the study population 4 Table 1 compared the characteristics of individuals in CMI implementation and non-5 implementation groups. There were 551 and 4176 elderly people in the treatment and control group, 6 respectively. From the perspective of health outcomes, most of the elderly in the whole sample 7 reported general health status (39%), followed by those who believed good (35%). In terms of 8 activities of daily living (ADL) and Instrumental Activity of Daily Living (IADL), both types of 9 limitations were worse in the control group. In addition, there was a significant difference in IADL 10 between the two groups (P<0.01). MMSE score, which reflected individual cognitive function, 11 showed that the cognitive function of the elderly in the treatment group was slightly worse than that 12 in the control group (25.66 vs. 26.09), and the difference was significant at the level of 10% (P<0.1). 13 From the results of the indices of positive and negative well-being, there was significant difference 14 in index of positive well-being between the two groups (P<0.05), but there was no significant 15 difference in index of negative well-being between them (P>0.1). The average age of the elderly in 16 the whole sample was 82.13±9.61 years old, and years of education of the elderly in the treatment 17 group was significantly higher than that in the control group (2.41 vs. 2.08). Also, individuals in the 18 treatment group had more chronic diseases than those in control group (1.83 vs. 1.58). In addition, 19 there were significant differences between the treatment group and the control group in medical 20 insurance type, residence and smoking (P<0.05). 21 a For continuous variables, T test was used to see whether the differences of variables were significant between the treatment and control group; for categorical 1 variables, Chi-square test was used to see whether the differences of variables were significant between the two groups.  3 In order to further examine the impact of CMI on different subgroups, we added the interaction 4 item of and residence , and interaction items of and the binary variables 5 _ , ℎ_ . It can be seen that the impact of CMI on self-rated health, 6 IADL and MMSE Score were still insignificant (P>0.1). But CMI could significantly reduce the 7 number of ADL limitations of the elderly living in rural areas and middle-and high-income 8 groups(P<0.1). In addition, CMI could significantly increase the index of positive well-being and 9 reduce the index of negative well-being of high-income group population (P<0.1). 10 1

Robustness test 2
Although the DID model controlled as many factors as possible, due to the limitations of two 3 periods of data, we could not verify whether the assumptions of DID method, namely common trend 4 was met. Therefore, we referred to the practices of existing studies. [14] The results of PSM-DID 5 model were used as the robustness test of our benchmark results. Firstly, the samples were screened 6 into panel data, and then the samples were matched according to the characteristics of the baseline 7 survey (2011/12 wave). Then the effect of CMI on health of the elderly was studied using the DID 8 method. Table A1 and Figure A1 list the balance test results and the kernel density curve of 9 propensity score before and after matching. It can be seen from Panel A in Table A1 that after  10 matching, all observable characteristics between the treatment and control group were well balanced. 11 Variables with significant differences between the two groups became statistically insignificant 12 (P>0.05) after matching. 13 the Pseudo R square of probability model, the joint significance of covariates, the mean and median 1 of the standardized deviation were all significantly lower after matching. As can be seen from the 2 probability distribution density function in Figure A1, the probability distribution density curves of 3 the two groups were close to coincidence after matching, indicating that the samples of the treatment 4 and control group were more balanced after matching. 5 Table 4 reported the PSM-DID estimation results. It can be seen that CMI could significantly 6 improve self-rated health and index of negative well-being among the elderly. Specifically, the self-7 rated health of the elderly in treatment group was improved by 0.117 units, and the index of negative 8 well-being decreased by 0.375 units in the treatment group, and the above effects were significant 9 at the level of 5% (P<0.05). Moreover, the estimation results showed that CMI could reduce the 10 number of ADL and IADL limitations, and increase MMSE score and index of positive well-being 11 by 0.253 and 0.094 points, respectively. However, these effects were not statistically significant 12 (P>0.1). 13

Potential Mechanisms analysis 14
In order to verify the potential mechanisms of CMI influencing the health of the elderly, we 15 used the Heckman Selection model to estimate the impact of CMI on the total healthcare utilization 16 and inpatient healthcare utilization of the elderly. The results are shown in Table 5. for the total 17 healthcare utilization, the inverse Mills ratio was significant at the level of 1%, indicating that there 18 was an obvious selection effect in patients' decision making. But there was no significant selection 19 effect on hospitalization. After controlling for provincial information and related covariates, we 20 found that CMI mainly increased the hospitalization rate of the elderly (P<0.01), but did not 21 significantly affect their total hospitalization expenditure, probability of visit, and total medical 22 expenditure (P>0.1). 1

Discussion 2
In order to reduce the risk of catastrophic health expenditure for urban and rural residents, 3 China has issued a document to establish and promote the implementation of the catastrophic 4 medical insurance since 2012. The gradual implementation of CMI in different regions provides a 5 good basis for studying the effect of the implementation of it. Existing studies discussed the effect 6 of CMI in lowering medical economic risks, ignoring its influence on health outcomes. The study 7 based on a natural experiment framework, using the CLHLS data and a number of health indicators, 8 evaluated the effect of CMI on health outcomes among the elderly and its heterogeneity. Also, its 9 potential mechanism was studied. The conclusion can provide useful references for improving the 10 medical insurance system design in China and other developing countries. 11 Our study found that the implementation of CMI can improve the mental health status of the 12 elderly to some extent, which is consistent with the conclusions of previous studies [31]. From the 13 perspective of system design, CMI mainly compensates for the compliance medical expenses that 14 still need to be paid by individuals after being reimbursed by the basic medical insurance. Studies 15 have pointed out that CMI can reimburse an additional 10% of the expenses on the basis of the basic 16 medical insurance [42]. This medical cost sharing mechanism of CMI can further improve the 17 accessibility of hospitalization services and hence increase the probability of hospitalization. One 18 study found that CMI increased the frequency and length of hospital stays, promoted the utilization 19 of medical services for inpatients [26]. In addition, CMI could promote the subjective healthcare 20 accessibility of the elderly, increases the probability of timely treatment when the elderly was ill[31], 21 the promotion in healthcare utilization accessibility could improve the health outcomes of the 22 20 elderly, especially in the aspect of mental health. 1 Our study also demonstrated that CMI had a limited effect on the health improvement of the 2 elderly. CMI had no significant impact on the physical health of the elderly, including ADL, IADL 3 and cognitive function. Existing studies on the effect of URBMI and NCMS found individuals 4 would be excluded from the scope of compensation due to high deductibles, Low reimbursement 5 rate and limited coverage [20]. Therefore basic medical insurance contribute only modestly to health 6 improvements [23]. CMI provides extra compensation for URBMI and NCMS patients with high 7 medical costs. Although the level of security is improved to a certain extent, the protection effect is 8 limited [43], which may also be the reason for the limited health improvement effect. Moreover, 9 since health is a stock, the impacts of the cost-sharing mechanism of CMI on health outcomes of 10 the elderly deserve further study [17]. 11 Finally, our study showed that there was heterogeneity in the health improvement effect of 12 CMI among different groups. The positive effect of CMI on health was mainly concentrated in the 13 elderly with higher income level, and there is no significant improvement effect on the middle-and 14 low-income group, especially for the low-income group population. The coverage of CMI and its 15 segmented compensation design suggested that people who pay more out-of-pocket medical 16 expenses get more subsidies from it. As studies have shown, out-of-pocket health expenditures 17 increase as affordability increases [44]. For individuals, in order to get additional compensation from 18 CMI, they have to reach the high deductible. For the low-income group population, to achieve the 19 threshold means that they need to undertake a certain proportion of out-of-pocket medical expenses. 20 And those constrained by their income could not enjoy CMI benefits because it was hard to pay the 21 threshold for them. A study about the NCMS also found that participants with low-and middle-22 income were more likely to avoid using medical services [20]. Therefore, higher income group will 1 benefit more from CMI [26]. In contrast, the benefits of the CMI are not fully available to low-2 income group population, so there is no significant improvement in health outcomes. The income 3 heterogeneity of the health effect of critical illness insurance is consistent with previous studies on 4 the general population [30]. 5 Our study had several policy implications. First of all, CMI should focus on its health 6 improvement effect in the future since the ultimate goal of health insurance is to improve the health 7 outcomes of participants. For patients with serious diseases, the government should focus on their 8 health demand and give personalized medical insurance reimbursement plan; Second, CMI should 9 take fairness into consideration during the process of practice, focusing on welfare improvement for 10 low-income group population. Although CMI has provided some priorities in view of the extreme 11 poverty population in the process of implementation, such as to reduce the deductible, cancel the 12 cap line, increase reimbursement ratio by 5-10% and so on. However, some people with lower 13 income level are still excluded from the scope of policy security due to budget constraints. In view 14 of this problem, it is suggested to design the compensation policy of CMI according to the gradient 15 of residents' disposable income, so as to make the policy security more accurate. Finally, although 16 most areas are introduced CMI mainly for hospitalized patients with serious illness, but for some 17 patients with specific diseases, the outpatient medical expenses are also burdensome. These people 18 remain at high risk of falling into catastrophic health spending. Therefore, it can be considered to 19 broaden CMI coverage, but need to pay attention to individual excessive demand and induced 20 demand from staff in health facilities. 21 Our study had some limitations. Firstly, considering the implementation of CMI to the 22 improvement of individual health needs some time, but data used in this study cannot capture the 1 health effect of CMI for a long time, so this is one of the limitations of this study, and it is also a 2 research direction in the future. Secondly, the influence of CMI on health of other group population 3 is also a research direction we will pay attention to in the future. 4 Conclusions 5 1  Figure A1. The distribution of density of propensity score before and after matching 2