Policy Research Working Paper 10073 The Financial Risk Reduction Provided by Ghana’s National Health Insurance Scheme Dhushyanth Raju Stephen D. Younger Social Protection and Jobs Global Practice June 2022 Policy Research Working Paper 10073 Abstract This paper estimates the monetary value of financial risk Scheme has no value to members. Indeed, the findings show reduction associated with membership in Ghana’s National that the insured pay significantly less for healthcare than Health Insurance Scheme, based on recent national house- the uninsured on average. But that average reduction does hold survey data. The paper compares the risk premiums for not translate into a reduced spread of consumption net of distributions of out-of-pocket healthcare expenditures with out-of-pocket healthcare expenditures. Thus, the benefit of and without insurance and find that the difference is small. the National Health Insurance Scheme is entirely a transfer This does not mean that the National Health Insurance benefit, not a reduction in financial risk. This paper is a product of the Social Protection and Jobs Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at draju2@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team The Financial Risk Reduction Provided by Ghana’s National Health Insurance Scheme Dhushyanth Raju Stephen D. Younger Key words: out-of-pocket expenditures, public healthcare, health insurance, financial risk protection, Ghana, Africa JEL codes: I13; I15; H51 Raju, World Bank, draju2@worldbank.org; Younger, Commitment to Equity Institute, sdy1@cornell.edu. The study was prepared under the World Bank’s advisory services and analytics task Ghana Human Development Public Expenditure Review (P175117). We thank Jeremy Barofsky, Jishnu Das, Jeffrey Hammer, Magnus Lindelow, Patrick Mullen, and Paul Andres Corral Rodas for comments. We also thank Enoch Oti Agyekum and Alex Nartey for comments and assistance in obtaining relevant data and documentation. 1 Introduction Most governments spend a significant share of their budgets to provide healthcare services to their populations. Studies that attempt to value these services in monetary terms for benefit-cost or benefit incidence analysis typically assess their transfer value. That is, they calculate the difference between the value of the service received and the amount beneficiaries must pay for it (which may be nothing). Barofsky and Younger (2022) discuss a variety of methods to do this, and many examples exist for Ghana. 1 But publicly financed health insurance provides an additional benefit in the form of financial insurance. Even if the premium that insurance beneficiaries must pay is actuarially fair so that, on average, the premium equals the expenditure the insurer makes on the beneficiaries’ behalf, the insurance still provides a benefit to the insured because it lowers the financial risk associated with adverse health outcomes. The insured will have lower variation in income net of healthcare expenditures less any premiums paid. It is less common to estimate the monetary value of this reduction in financial risk, a gap this study aims to fill for Ghana. 2 The out-of-pocket (OOP) expenditures necessitated by an adverse health shock pose a financial risk because they add variance to income net of those expenditures. As with any risk, it is possible to calculate a risk premium, that is, the quantity of money a risk-averse person is willing to pay to completely insure against it. Publicly provided health insurance reduces financial risk and so provides a benefit to those who are insured. This benefit is over and above the value of the healthcare services they receive less what they pay in premiums. We call the latter difference a transfer value of publicly provided health insurance. To simplify, the transfer is a difference in the mean; financial risk protection is a difference in variance and skewness. In contrast to most studies which calculate the transfer values, here we want to estimate the increase in utility of lower dispersion of income net of healthcare expenditures, which we call the financial insurance value of public health insurance. We should note that the existing literature often confounds these two effects. A recent extensive review of the literature finds that, at best, “financial protection” is used to refer to the combined transfer and financial risk reduction benefits, though it often refers only to the transfer 1 For recent benefit-cost analyses, see Asuming, Kanmiki, and Wong (2020) and Nketiah-Amponsah et al. (2020). For benefit incidence analyses, see Coulombe and Wodon (2012) and Younger, Osei-Assibey, and Oppong (2017). 2 Barofsky and Younger (2022) do make this calculation for Ghana. We have updated and modified their approach considerably. Appendix A discusses the differences. 2 (Das and Do 2022). This despite the fact that the early literature in the field of financial risk reduction is clear about the distinction. 3 The specific insurance we evaluate is Ghana’s National Health Insurance Scheme (NHIS). Since 2005, NHIS has been open to all Ghanaians, and the government’s stated goal is to enroll everyone, though actual membership falls far short of that. In 2017, the year of our data, 42 percent of the population was enrolled in NHIS. To enroll, individuals pay a small registration or an annual renewal fee, and an annual premium, though some may qualify for exemptions from one or both fees. These fees fund only a small part of the NHIS budget; most is funded through an earmarked value-added tax of 2.5 percent. Thus, the NHIS includes a larger transfer to its members on average. While this is sometimes called “financial protection,” it is not the focus of our study, which is financial risk reduction as defined above. Our approach is to calculate the risk premium associated with a distribution of net income with NHIS membership and without it. The difference in these two is the financial insurance value of NHIS membership, in monetary terms. To do this, we must estimate the distribution of net income with and without NHIS membership. We then calculate and compare the risk premiums for each distribution to obtain an estimate of the financial insurance value of NHIS. 4 Of course, we do not observe people’s expenditures with and without NHIS membership, so we must compare those with membership to those without, who serve respectively as “treatment” and “control” groups. Ideally, these would be randomized to avoid selection bias, but that is not possible with our data, 5 nor is it usually possible with a large insurance program such as NHIS. Instead, we create a control group with coarsened exact matching (Blackwell 2009). 6 3 Englehart and Gruber (2011) (in p. 97) state “Of course, the introduction of Part D will shift the mean level of out- of-pocket spending as well as its risks. The shift in the mean is simply a transfer from the government to the insured, and so should not enter these risk calculations.” Finkelstein and McKnight (2008) (in p. 1663), whose approach we follow, note that “Since our interest is in the social welfare benefits from Medicare rather than the private benefits from transfer payments, we assume for the purposes of our calculations that the beneficiaries “pay for” the actuarial expected cost of Medicare, so that the welfare benefits come only from reductions in the variance, and not the mean, of medical expenditures.” 4 This is a standard approach in the literature on health insurance. For example, see Feldstein (1973), Feldstein and Gruber (1995), and Finkelstein and McKnight (2008). For an application to Ghana, see Barofsky and Younger (2022). 5 Powell-Jackson et al. (2014) do randomize health insurance coverage in one district of Ghana. 6 Finkelstein and McKnight (2008) use a difference-in-differences strategy to estimate the insurance value of Medicare in the United States, comparing those just under the qualifying age of 65 to those just over, before and after the introduction of Medicare in the 1960s. This approach is not available to us either, as everyone qualifies for NHIS and our data were collected well after NHIS was established. 3 The control group is not completely without publicly provided health insurance. While those not in NHIS must pay fees at public healthcare facilities, those fees only have to cover supplies, maintenance, medicines, and diagnostic costs. Capital expenditures and, importantly, staff salaries are covered by the general government budget regardless of a user’s NHIS affiliation. While not called health insurance, this support to public healthcare facilities does function like insurance: General revenue is collected from all taxpayers to provide benefits to those who draw unfortunate health outcomes. Our comparison estimates the financial insurance value of NHIS relative to this “reduced insurance” state, not to no insurance at all. The data we need from treatment and control groups are more detailed than those used in a typical evaluation in that we need to compare full distributions of OOP healthcare expenditures, not just their means. We create these distributions from the observed distribution of OOP healthcare expenditures for treatment and control groups, assuming that the cross-person distribution reflects each individual’s ex ante distribution of possible expenditures. Given these OOP healthcare-expenditure distributions for those with and without NHIS membership, we then calculate the risk premium associated with each distribution for each person in the sample using a stylized constant relative risk aversion (CRRA) utility function. The difference in those risk premiums is our estimate of the financial insurance value of NHIS membership. The data come from the Ghana Living Standards Survey (GLSS) 2016/17, fielded between October 2016 and September 2017. GLSS 2016/17 is the latest available round of the national household sample survey administered by Ghana’s statistical agency (GSS 2019). The survey collected information on household expenditures, including expenditures for healthcare for each person in the household. Our data, then, are individual level for each person in the survey, with 58,844 people considered in all. The remainder of the paper is organized as follows. Section 2 provides background information on government healthcare financing and spending in Ghana and for NHIS. Sections 3 and 4 describe how we calculate the distribution of OOP healthcare expenditures for those with and without NHIS membership. Section 5 calculates and discusses the risk premium associated with each distribution and their difference, which is the financial insurance value of NHIS membership. Section 6 concludes. 4 2 Background Government expenditure on healthcare in Ghana ranged between 54 to 108 purchasing power parity (PPP) dollars per capita over the past decade, with a decline through most of the decade (figure 1a). Ghana’s levels of government expenditure on healthcare were higher than the average for lower-middle-income countries or for Sub-Saharan Africa at the beginning of the decade, but similar to that for Sub-Saharan Africa and well below that for lower-middle-income countries at the end. Government expenditure on healthcare accounted for a larger share of overall government expenditure than the average level in other lower-middle-income countries at the beginning of the decade, but was at a similar level by the end of the decade (figure 1b). Figure 1. Government Healthcare Expenditures in Ghana a. Govt. healthcare expenditures per capita b. Govt. healthcare expenditures as a percentage of overall govt. expenditures Source: Statistics obtained from the World Bank’s World Development Indicators (WDI) databank. Note: SSA = Sub-Saharan Africa. LMICs = lower-middle-income countries. “SSA, average” excludes high-income countries in the region. Panel a: PPP dollars = current international dollars in purchasing power parity (PPP) terms. Panel b. The World Bank’s WDI databank does not have data for this indicator for Sub-Saharan Africa. The flow of funds in the public healthcare sector is complicated, as captured in figure 2, reproduced from Wang, Otoo, and Dsane-Selby (2017). Because this study focuses on the financial insurance value of NHIS membership, it is natural to focus on NHIS’s financing and expenditures. But it is important to keep in mind that the largest share of healthcare expenditures 5 is for salaries which, in public healthcare facilities, are paid directly from the central government budget (the line between “Ministry of Health” and “Health facilities” in figure 2). 7 Based on government administrative data, over the decade of the 2010s, the central government budget funded 57 percent of healthcare expenditures at public facilities, with 92 percent of those expenditures going towards salaries. “Internally generated funds,” fees paid by patients and NHIS, accounted for 22 percent of expenditures, 84 percent of which covered goods and services and 12 percent of which paid for salaries. The remaining financing came from international donors, 62 percent of which also funded expenditures on goods and services. Figure 2. Fund Flow in Ghana’s Public Healthcare System Source: Figure 2.6 (pg. 13) from Wang, Otoo, and Dsane-Selby (2017). Note: VAT = value-added tax. SSNIT = Social Security and National Insurance Trust. These patterns—with the central government covering salaries and other sources covering goods and services—were broadly true before the advent of NHIS during the “cash and carry” period and remain true today. Thus, OOP healthcare expenditures and NHIS expenditures were and are only meant to cover the costs of supplies, diagnostic and laboratory services, and, mostly, medicine in public healthcare facilities. 7 Capital expenditures, a much smaller share of the total expenditures, are also funded directly from the central government budget or from international donor support. 6 NHIS membership is voluntary. Although the government encourages all to join, less than half of the population is affiliated in any given year (Nsiah-Boateng and Aikens 2018). To obtain membership, individuals must pay a registration or annual renewal fee which, over our analysis period, varied between zero and 8 cedis (zero to 4.6 PPP dollars), 8 and an annual premium of 15 cedis (8.6 PPP dollars) in rural areas and 22 cedis (12.6 PPP dollars) in urban areas (Nsiah-Boateng and Aikens 2018). The premium is waived for several selected groups, including those under 18 years old and those aged 70 years old or older, as well as the “indigent,” 9 Social Security and National Insurance Trust (SSNIT) contributors and pensioners, 10 pregnant women, and women with infants under 3 months (and the infants). The indigent, pregnant women, and women with young infants also do not have to pay the registration or renewal fee. The fees charged, even with increases in recent years, are far short of what would be actuarially fair. Instead, over our analysis period, NHIS received the bulk of its funding— roughly three-fourths—from a special 2.5 percent addition to the VAT called the National Health Insurance Levy (NHIL). A further one-fifth of its funding came from SSNIT, which pays NHIS fees for its members (Wang, Otoo, and Dsane-Selby 2017). NHIS covers most illnesses found in Ghana. In principle, members can receive healthcare free of charge from all public providers and some participating private providers. 11 In practice, NHIS has a long history of being slow to reimburse healthcare providers (Wang, Otoo, and Dsane-Selby 2017), leading some providers to refuse to treat NHIS members and others to charge them despite that they are insured—something to keep in mind as we examine the difference between reported OOP healthcare expenditures for NHIS members and nonmembers. 8 In 2017, the purchasing power parity conversion factor for cedis to international dollars was 1.75 cedis per dollar. We use this conversion factor wherever we report values in PPP dollars. The conversion factor was obtained from the World Bank’s World Development Indicators databank. 9 NHIS defines an “indigent” as an individual who (a) “does not have any visible source of income,” (b) “does not have a fixed place of residence,” (c) “does not live with a person who is employed and has [a] fixed place of residence,” or (d) “does not have a consistent source of support from another person” (https://nhis.gov.gh/Faqs/the- benefits-of-the-national-health-insurance-scheme-2?msclkid=60d5acdcab8811ec95eecede25b33dfb; accessed on March 24, 2022). 10 SSNIT is a statutory organization that administers a national social security scheme. The scheme offers old-age, disability, and survivor benefits to its members. 11 Thirty-six percent of all participating healthcare providers are private. Because staff salaries at private healthcare providers are not paid by the government, NHIS reimburses these providers more generously than it does public healthcare facilities, but not by enough to cover the entirety of provider salaries. 7 Anecdotal reports suggest that this problem is more severe at private providers because they must worry about covering staff costs as well as medicines, supplies, and diagnostic tests. Ghanaians who are not covered by NHIS must pay for healthcare on a fee-for-service basis. However, as we noted above, at public healthcare facilities, salaries and capital expenditures are covered by the central government budget, so the fees cover only nonsalary, noncapital costs. 3 Calculating Out-of-Pocket Healthcare Expenditures GLSS 2016/17 asks about healthcare expenditures for each individual if in the past two weeks she or he consulted a healthcare provider, purchased medicines, or was admitted to a hospital. For outpatient services, survey respondents can report what they paid for registration at the healthcare facility, consultation, diagnosis (laboratory fees, etc.), and medicines, or they can report the total amount they paid. 12 In our calculations, we use the larger of the sum of the individual items or the reported total. For inpatient services, survey respondents only report what they paid for “staying in a hospital or health facility.” 13 Survey respondents can also report “medicines and medical supplies” purchased outside the context of a healthcare visit. 14 Finally, they can report “total medical expenses.” 15 To calculate total OOP healthcare expenditures, we add the total of outpatient expenses, inpatient expenses, and medicine expenses and take the larger of this sum or the “total medical expenses” reported. In a few cases, respondents report exactly the same amount for medicines acquired in the context of a consultation and medicines acquired without a consultation. 16 For these, we set the value for OOP healthcare expenditures outside the context of a consultation to zero to avoid double counting these expenditures. All the raw variables in GLSS 2016/17 have some very large positive outliers. For each of the raw variables and again for our calculated total OOP healthcare expenditures, we replace any value that is larger than the respondent’s household consumption (not per capita) or five 12 These are questions 9 through 12 and question 13 in Section 3A of the GLSS 2016/17 questionnaire. We have excluded question 14 (“any other services”) and question 15 (“travel”) from our total out-of-pocket healthcare expenditures as these are not healthcare expenditures per se. 13 Question 20 in Section 3A of the GLSS 2016/17 questionnaire. 14 Question 22 in Section 3A of the GLSS 2016/17 questionnaire. 15 Question 23 in Section 3A of the GLSS 2016/17 questionnaire. 16 Questions 12 and 22 in Section 3A of the GLSS 2016/17 questionnaire, respectively. 8 standard deviations larger than the mean share of that type of OOP healthcare expenditures in total household consumption with the mean share of those expenditures in household consumption times the respondent’s household’s consumption. That is, for respondent i in household h and OOP type j, we replace OOPihj with the sample mean value of OOPihj / Ch ( ) times Ch, where Ch is household consumption. In all, we change 112 healthcare expenditure records (0.2 percent of the sample) with these procedures. An important problem with this estimate of OOP healthcare expenditures is that it applies only to expenditures incurred in the past two weeks. But healthcare expenditures in Ghana are not regular or frequent, so if we want to know how much is spent in a year, simply multiplying the reported amounts by 26 will be inaccurate. Survey respondents who happened to have incurred healthcare expenditures in the past two weeks are not the only people who will incur healthcare expenditures through the year, nor will their annual healthcare expenditures be anywhere near 26 times what they spent in the past two weeks. To address this, we divide the sample into deciles based on consumption per adult equivalent and into NHIS members and nonmembers; 20 subgroups in all. 17 For each respondent, we then draw and sum 26 random values from the distribution of the share of OOP healthcare expenditures in total household consumption for her/his subgroup whether or not she or he reports healthcare expenditures in the past two weeks. We then multiply the sum of the 26 random draws by total household consumption to obtain our estimate of the individual’s annual OOP healthcare expenditure. This assumes that every member of a subgroup has an equal probability of incurring a healthcare expenditure, though the probability can vary across subgroups. In doing so, we ignore the possibility of serial correlation of OOP healthcare expenditures, that is, the possibility that those who report OOP healthcare expenditures in the past two weeks are more likely to have OOP healthcare expenditures in the rest of the year than those who do not. However, simply multiplying observed OOP healthcare expenditures in the past two weeks by 26 assumes perfect serial correlation throughout the year, something that seems much farther from the truth than our assumption of zero correlation. This is particularly so 17 We also tried dividing the sample by groups that might be expected to have greater demand for healthcare: (income quartile)-by-(infants/children/adults/the elderly (60+)/women who have been pregnant in the past year)-by- (urban/rural); 40 subgroups in all. This produces results of a similar magnitude to our main results but with the risk premium being higher for NHIS members than nonmembers. 9 in Ghana where most healthcare is curative for non-chronic conditions, usually for fever, upper respiratory infections, or birthing assistance. Finally, in our calculations, 415 people, 0.8 percent of the sample, had estimated annual OOP healthcare expenditures greater than their household’s consumption per adult equivalent. For these, we reduced estimated OOP healthcare expenditures to household consumption per adult equivalent. 18 One further detail of our calculations is important to the results. In the raw data, NHIS members actually have slightly higher OOP healthcare expenditures in the past two weeks than do nonmembers because they are much more likely to seek healthcare. To account for this when we draw from the OOP healthcare expenditure distributions for a subgroup, we adjust the probability associated with each nonzero OOP healthcare expenditure by nonmembers upward by the ratio of the share of NHIS members who sought healthcare in the past two weeks over the share of nonmembers who sought care. 19 Table 1 reports the probability of healthcare seeking by decile of the welfare distribution and NHIS member status. Note that we apply this adjustment Table 1. Probability of Seeking Healthcare in the Past Two Weeks by NHIS Membership Status Welfare decile Nonmembers Members (1) (2) st 1 (poorest) 0.040 0.082 2nd 0.054 0.124 rd 3 0.065 0.103 4th 0.060 0.108 th 5 0.061 0.108 6th 0.052 0.124 th 7 0.067 0.115 8th 0.058 0.123 th 9 0.077 0.115 10th (richest) 0.084 0.127 Note: Welfare = household consumption per adult equivalent. only to the OOP healthcare expenditures associated with a health consultation or hospitalization, not the separately reported OOP expenditure for “medicines and medical supplies” purchased 18 Barofsky and Younger (2022) do simply multiply by 26, an important difference with our calculations here. This makes their distribution of OOP healthcare expenditures much more skewed, increasing the risk. 19 And, of course, we reduce the probability of zero OOP healthcare expenditures to keep the sum of the probabilities equal to one. 10 outside the context of a healthcare visit. Most of this expenditure is likely to be direct purchases from a pharmacy. Because NHIS does not cover the costs of medicines and supplies acquired outside the context of a consultation, we think it unlikely that NHIS covers the OOP expenditure reported as such. 20 4 Estimating the Distribution of Income Net of OOP Healthcare Expenditures There are several options for identifying NHIS members. 21 We use only those who were able to show a valid NHIS card to the GLSS 2016/17 interviewer. This produces an estimated 9.3 million beneficiaries, compared to 10.8 million registered beneficiaries in 2016 and 12.2 million in 2017 based on program administrative data obtained from the government. To generate treatment and control groups, we use coarsened exact matching (Blackwell 2009). We match NHIS members and nonmembers on region and area of residence (urban versus rural), the log of household consumption per adult equivalent, and several individual-level characteristics that increase the likelihood of NHIS membership because they reduce or eliminate the annual registration and membership fees: those aged 70 years old or older; a child under 18 years; a woman who is currently pregnant, has been pregnant in the past year, or currently has an infant under 3 months old; a contributor to SSNIT; or a recipient of a SSNIT pension. Table 2 reports the results from our matching exercise. L1 is a measure of distance between NHIS members and nonmembers, bounded by 0 (perfect match) and 1. The ex-ante data show that the two groups are well-matched except with respect to region of residence and the log of household consumption per adult equivalent. After matching, some imbalance with respect to consumption remains, but it is considerably less than in the ex-ante comparison. Overall, the criteria used for the match generate 141 strata, 95 of which are successfully matched. Only 313 individual records out of 58,844 respondents could not be matched and are excluded from the analysis. 20 Indeed, the distribution of the values reported for “medicines and medical supplies” purchased outside the context of a healthcare visit are almost identical for members and nonmembers. 21 Questions 1, 3, and 6 in Section 3B of the GLSS 2016/17 questionnaire. 11 Table 2. Ex-Ante and Ex-Post Match of NHIS Members and Nonmembers a. Ex-ante results Multivariate L1: 0.326 Univariate Imbalance L1 Mean Min. 25% 50% 75% Max. Urban 0.0472 0.0472 0 0 0 0 0 Infant <3 months) 0.0040 –0.0040 0 0 0 0 0 Child (<18 years) 0.0679 0.0679 0 0 0 0 0 Pregnant woman 0.0293 0.0293 0 0 0 0 0 Elderly (≥70 years) 0.0203 0.0203 0 0 0 0 0 SSNIT pensioner 0.0029 0.0029 0 0 0 0 0 SSNIT contributor 0.0187 0.0187 0 0 0 0 0 Region 0.1266 0.1266 0 1 1 1 0 Ln(hh cons. pae) 0.0616 0.0599 0 0.0428 0.0839 0.0940 –1.081 b. Ex-post results Multivariate L1: 0.286 Univariate Imbalance L1 Mean Min. 25% 50% 75% Max. Urban 0.0451 0.0451 0 0 0 0 0 Infant <3 months) 0.0042 –0.0042 0 0 0 0 0 Child (<18 years) 0.0871 0.0871 0 0 0 0 0 Pregnant woman 0.0300 0.0300 0 0 0 0 0 Elderly (≥70 years) 0.0176 0.0176 0 0 0 0 0 SSNIT pensioner 0.0028 0.0028 0 0 0 0 0 SSNIT contributor 0.0163 0.0163 0 0 0 0 0 Region 0.0247 0.0247 0 0 0 0 0 ln(hh cons. pae) 0.0447 0.0055 0 0.0200 0.0035 0.0139 0 Note: SSNIT = Social Security and National Insurance Trust. ln(hh con. pae) = log of household consumption per adult equivalent. Min. = minimum. Max. = maximum. Pregnant woman = a woman who is currently pregnant, has been pregnant in the past year, or currently has an infant under 3 months old. To achieve this match, some reweighting of the sample is required. In all calculations that follow, we weight each observation by its sampling weight times the weights generated by the matching procedure. To estimate a distribution of (possible) healthcare expenditures for each individual, we first divide the sample into NHIS members and nonmembers. For each group, we regress the log of OOP healthcare expenditures on the log of household consumption per adult equivalent, draw 100 random observations of the residuals, and add those 100 draws to the predicted value from 12 the regression to generate 100 possible healthcare expenditures. 22 This allows healthcare expenditures to increase, on average, with consumption, which is to be expected. The regression for nonmembers has an elasticity of OOP healthcare expenditures with respect to household consumption per adult equivalent of 0.53 and an R-squared statistic of 0.14. That for NHIS members has an elasticity of 0.66 and an R-squared statistic of 0.18. In both the NHIS member and nonmember samples, considerable variance remains in the error term, so there is significant variation in the 100 estimated healthcare expenditures. Figure 3 shows the inverse of the cumulative distribution function for estimated annual OOP healthcare expenditures, by NHIS membership status. The distribution for both NHIS members and nonmembers is highly skewed with a roughly constant increase in expenditures up to about the 90th percentile and a sharp increase above that. This is typical for healthcare expenditures by the uninsured. One might expect, however, for healthcare expenditures of the insured to be much less skewed since the insurance should limit the need for catastrophic healthcare expenditures. While that is true to some extent in Ghana, the effect is rather limited. OOP healthcare expenditures for NHIS members still skew strongly after about the 90th percentile. (See appendix B for a comparison to results for Medicare in the United States.) On the other hand, OOP healthcare expenditures by the insured are consistently lower than those for the uninsured throughout the distribution, with the difference increasing across the distribution. 22 We repeated the analysis with 1,000 draws, with similar results. 13 Figure 3. Distribution of Annual Out-of-Pocket Healthcare Expenditures by NHIS Membership Status, 2016/17 4000 3000 Cedis per Year 2000 1000 0 0 20 40 60 80 100 Percentile of Healthcare Expenditure Distribution Nonmember NHIS member Table 3 reports mean annual OOP healthcare expenditures, by NHIS membership status. Overall, such expenditures for nonmembers are 344 cedis (196 PPP dollars) with a standard deviation of 525 cedis (299 PPP dollars). For NHIS members, the mean is 289 cedis (165 PPP dollars) with a standard deviation of 523 cedis (298 PPP dollars). NHIS membership reduces mean OOP healthcare expenditures but not the standard deviation, suggesting that having NHIS coverage is not so much reducing catastrophic healthcare expenditures as it is everyday healthcare expenditures. 23 That is, NHIS is providing an effective transfer, but not financial insurance. 23 The literature calculates “catastrophic” OOP healthcare expenditures in a variety of ways. We present a set of standard results in appendix C where we find very little difference in catastrophic payments by NHIS members and nonmembers. 14 Table 3. Estimated Mean Annual Out-of-Pocket Healthcare Expenditures by NHIS Membership Status, Cedis Welfare decile Nonmembers Members Mean Std. dev. Mean Std. dev. (1) (2) (3) (4) Overall 344 525 289 523 1st (poorest) 162 169 91 108 2nd 209 199 150 153 3rd 250 273 177 189 4th 214 241 175 217 5th 216 232 226 281 6th 278 294 242 269 7th 283 320 308 335 8th 393 467 295 329 9th 396 495 406 488 10th (richest) 877 1,084 678 1,202 Note: Welfare = household consumption per adult equivalent. Std. dev. = standard deviation. Many other studies estimate the effect of NHIS membership on mean OOP healthcare expenditures, usually based on cross-sectional data such as our sample. Garcia-Mandicó et al. (2021) is unique among these studies as the authors use data from the 2005/06 round of the GLSS, whose fielding straddled the rollout of NHIS at that time. Thus, they use a difference-in- differences strategy comparing survey respondents in the same districts, some of whom gained access to NHIS during the survey-fielding period and some of whom did not. The survey data they use, GLSS 2005/06, are very similar to the GLSS 2016/17 data we use, though it is likely that all OOP expenditures for medicines are reported in one variable in GLSS 2005/06, while they are separated between OOP expenditures in the context of a consultation and other OOP expenditures on medicines in GLSS 2016/17. They find a mean decrease in OOP healthcare expenditures of 18 percent. Powell-Jackson et al. (2014) investigate a randomized allocation of free healthcare to children ages 6–59 months in one district of Ghana, Dangme West, and find that this comprehensive insurance reduced OOP healthcare expenditures by 27 percent on average. By comparison, we estimate a mean reduction of 16 percent for the overall Ghanaian population, 20 percent for children ages 6–59 months across Ghana, and 46 percent for children ages 6–59 15 months in rural Greater Accra region. 24 Thus, our matching method produces results similar to better-identified studies. Powell-Jackson et al. (2014) also report the distribution of OOP healthcare expenditures in their treatment and control groups (reproduced in figure 4) which are similar to the results in figure 3 in that the overall distribution shifts down with health insurance, but is not noticeably less skewed. Figure 4. Distribution of Out-of-Pocket Healthcare Expenditure in Treatment and Control Groups Source: Figure 2 (pg. 313) from Powell-Jackson et al. (2014). Table 3 also shows that the reductions in OOP healthcare expenditures are not uniform across the welfare distribution (as opposed to the healthcare-expenditure distribution presented in figure 3). While we do not see a clear pattern, the seventh decile has a small increase and that in the tenth decile (the richest decile) is much larger than the rest. Of greater concern from the perspective of insurance, the standard deviation is not uniformly lower for NHIS members across the welfare distribution, and is actually greater in the fifth, seventh, and tenth deciles. 24 Okoroh et al. (2018) provide an extensive literature review of the effect of NHIS membership on OOP healthcare expenditures. They find estimates, all based on cross-sectional data, ranging from 40 percent to 99 percent reductions in OOP healthcare expenditures for NHIS members. 16 5 Valuing the Risk Inherent in the Distribution of Net Income We assume that each person must satisfy a per period budget constraint of c= y − m where y is income, m is OOP healthcare spending, c is other expenditure, and utility is determined under a constant relative risk aversion utility function, as follows:  c (1−ε ) − 1    if ε ≠ 1 U (c) =  (1 − ε )  .  ln(c) if otherwise  We assume that each person’s income, y, is her or his household’s consumption per adult equivalent. Essentially, we assume that each person in a household gets an equal share of real income before healthcare expenditures, and that income is best measured by household consumption. In what follows, we place a lower bound on c, ensuring that OOP healthcare expenditures do not drive it below some percent of household consumption per adult equivalent. We assume that the burden of healthcare expenditures for an individual is not shared through the household; only her or his utility falls with the health shock. Call the distributions of healthcare expenditure calculated in section 4 Pk (m) , where k = [ 0,1] indicates those without and with NHIS membership, respectively. The difference between income minus P0 (m) or P1 ( m) determines the change in risk exposure from NHIS membership. Expected utility is given by EU ( = y, γ , Pk ( m ) ) ∫ u ( max [ y − m, γ y ])P ( m ) dm, M 0 k where M is the maximum healthcare expenditure and γ is an assumed minimum consumption value under which a person’s consumption does not fall regardless of the cost of healthcare. Previous studies, mostly in high-income countries, set this limit between 20 percent and 40 percent of household expenditure. 25 The risk premium represents the quantity of money a risk-averse household would be willing to pay to completely insure against a given financial risk distribution. The risk premium for a household is 25 See Finkelstein and McKnight (2008), Engelhardt and Gruber (2011), Shigeoka (2014), Barofsky (2015), and Limwattananon et al. (2015). 17 π k= Ek ( y − m ) − CEk M   −1  M   ∑ ( max [ y − m, γ y ]) Pk ( m )  − u  ∑ ( max [ y − m, γ y ]) Pk ( m )   , =  m 0=   m 0  where Ek ( y − m ) is the expected value of nonhealthcare expenditure for distribution k, and CEk is the household’s certainty equivalent for the same distribution of healthcare expenditure. The difference in risk premiums between those with and without NHIS coverage, (π 1 − π 0 ) , is the monetary value of financial risk protection provided by NHIS. To ensure that we are capturing only the difference in risk premiums—not any difference due to the fact that net consumption is higher for the insured because of the transfer implicit in NHIS membership—we adjust each observation by the difference in the overall mean of the nonmember and member samples, thus making the overall means equal in the adjusted data. Finally, because of the randomness of the healthcare expenditure draws, we bootstrap the calculation 200 times and average the results. Table 4 reports the mean values for (a) estimated expected values of consumption net of healthcare expenditures (which are independent of the risk aversion parameter), (b) estimated certainty equivalence values, and (c) estimated risk premiums by NHIS membership status. The striking result here is that the change in risk premium from being uninsured to being insured is tiny—less than one percent of mean household consumption per adult equivalent except at a very high coefficient of relative risk aversion. Thus, the financial insurance value of NHIS membership is very small. 26 This does not mean that NHIS membership has no value to its members, only that its value is not coming from reduced financial risk from potentially catastrophic healthcare expenditures made necessary by health shocks. Instead, the value is from lower mean OOP healthcare expenditures for the insured. 26 The results in table 4 assume a minimum consumption of 20 percent of household consumption per adult equivalent. Results assuming a minimum of 1 or 40 percent are smaller still. 18 Table 4. Mean Values for Expected Value of Net Consumption, Certainty Equivalents, and Risk Premiums by NHIS Membership Status, Cedis per Year ρ Certainty equivalent Risk premium Difference in risk Nonmember Member Nonmember Member premiums (1) (2) (3) (4) (5) (6) 0.5 4,171 4,174 10.8 7.8 3.0 1.0 4,158 4,165 23.4 16.8 6.6 1.5 4,143 4,154 38.3 27.3 11.0 2.0 4,125 4,142 56.5 40.0 16.6 3.0 4,074 4,107 107.1 74.9 32.2 4.0 3,998 4,054 183.3 127.4 55.9 5.0 3,901 3,987 281.1 194.3 86.8 Expected value of net consumption = 4,182* Note: Assumed minimum consumption is 20 percent of household consumption per adult equivalent. * The expected value of net consumption is the same for NHIS members and nonmembers by construction. 6 Conclusion Publicly funded health insurance provides both a transfer value—the difference between the value of health services received and what a beneficiary must pay for them—and a financial insurance value, that is, the reduced financial risk that comes from having to pay potentially large sums for healthcare. In this paper, we have estimated the latter in monetary terms which, we imagined, would be an input to benefit-cost and benefit incidence analyses. As it happens, the estimated benefits are so small that their use for either purpose does not seem worthwhile. This is not to say that NHIS membership is not valuable for beneficiaries. Several studies suggest that there is a significant transfer value associated with NHIS membership either as a whole or for specific treatments it covers (for example, see Okorah et al. 2018; Asuming et al. 2020; Nketiah-Amponsah et al. 2020; García-Mendicó et al. 2021), and we find comparable transfer benefits in our calculations. 27 But the financial insurance value is quite limited largely because the distribution of OOP healthcare expenditures for the insured differs very little from that for the uninsured, except that it has a lower mean. The variance and skewness—the aspects that insurance usually protects against—are similar. This is not an unusual result. In their review of the literature, Das and Do (2022) state: “Should [public health insurance] programs still be regarded as health insurance schemes? We believe that the answer is “probably not,”…” 27 Dadzie, Raju, and Younger (2022) analyze the distributional consequences of the transfer implicit in NHIS. 19 Why might NHIS provide little insurance value even as it provides a significant transfer? One explanation, based on anecdotal evidence, is that healthcare facilities are reluctant to provide high-cost medical care to NHIS members because NHIS is notoriously slow to reimburse them. In a wide-ranging review of the literature on NHIS, Christmals and Aidem (2020) found eight papers reporting slow reimbursement as a complaint of healthcare facilities. Agyepong et al. (2014) note that slow or nonexistent reimbursement for medicines is a particular problem sometimes involving catastrophic payments for NHIS members. Thus, when NHIS members have a large adverse health shock, they may end up paying for services even though they have insurance coverage. On the other hand, because NHIS is intended to pay for almost all healthcare without copay, those with less expensive healthcare needs may in fact receive free care. This would explain the fairly consistent difference between member and nonmember OOP healthcare expenditure distributions in figure 3, but also the long tail in the NHIS member distribution that insurance should, in theory, eliminate. A second explanation, again based on anecdotal evidence, is that nonmembers suffering a serious illness or injury may be able to join NHIS ex post (Gajate-Garrido 2013). This would reduce the top end of the nonmember distribution in OOP healthcare expenditures, biasing downward our estimate of the gap between NHIS member and nonmember expenditures at the highest percentiles of the OOP healthcare expenditure distribution. While this is a concern, our estimated distributions of OOP healthcare expenditures for NHIS members and nonmembers are similar to those in studies with randomized assignment to free health insurance (Powell-Jackson et al. 2014) or difference-in-differences estimates spanning the introduction of NHIS in 2005 (García-Mandicó et al. 2021). Those same studies give us some confidence that our matching estimation based on a single cross-section provides a reasonable control group for NHIS membership. A further limitation of our study is that we rely on a stylized utility function to estimate the risk premiums associated with the distribution of healthcare expenditures with and without NHIS membership. While this is standard in the literature, we are not aware of any empirical evidence to support this assumption. Further, our simple individual-level utility function ignores the possibly complex interdependence of utilities within households, including risk sharing among household members. We try to accommodate this by checking the sensitivity of our 20 results to the assumed lower limit of nonhealthcare consumption for an individual but, again, we are not aware of any empirical evidence that suggests this is adequate. Finally, we use reported healthcare expenditure information, but we know nothing about the severity of the illness or injury that drove people to seek healthcare in the first place, the impact that has on respondents’ welfare, or the effect that has on respondents’ decisions to seek care and willingness to pay for it. 21 References Agyepong, Irene Akua, Daniel Nana Yaw Abankwah, Angela Abroso, ChangBae Chun, Joseph Nii Otoe Dodoo, Shinye Lee, Sylvester A. Mensah, Mariam Musah, Adwoa Twum, Juwhan Oh, Jinha Park, DoogHoon Yang, Kijong Yoon, Nathaniel Otoo, and Francis Asenso-Boadi. 2016. “The ‘Universal’ in UHC and Ghana’s National Health Insurance Scheme: Policy and Implementation Challenges and Dilemmas of a Lower Middle Income Country.” BMC Health Services Research 16: 504 (2016). doi: 10.1186/s12913- 016-1758-y. Agyepong, Irene A., Geneieve C. Aryeetey, Justice Nonvignon, Francis Asenso-Boadi, Helen Dzikunu, Edward Antwi, Daniel Ankrah, Charles Adjei-Acquah, Reuben Esena, Moses Aikins, and Daniel K Arhinful. 2014. “Advancing the Application of Systems Thinking in Health: Provider Payment and Service Supply Behavior and Incentives in the Ghana National Health Insurance Scheme—A Systems Approach.” Health Research Policy and Systems 12:35. doi: 10.1186/1478-4505-12-35. Asuming, Patrick Opoku, Edmund Wedam Kanmiki, and Brad Wong. 2020. “Cost Benefit Analysis of Priority Health System Strengthening Interventions in Ghana.” Ghana Priorities, Copenhagen Consensus Center, Lowell, MA. Barofsky, Jeremy. 2015. “Estimating the Impact of Health Insurance in Developing Nations: Evidence from Mexico’s Seguro Popular.” Harvard School of Public Health, Cambridge, MA. Barofsky, Jeremy, and Stephen D. Younger. 2022. “The Effect of Government Health Expenditure on the Income Distribution: A Comparison of Valuation Methods in Ghana.” In Commitment to Equity Handbook: Estimating the Impact of Fiscal Policy on Inequality and Poverty, 2nd ed., Vol. 2, edited by Nora Lustig. Washington DC: Brookings Institution Press; New Orleans, LA: CEQ Institute, Tulane University. Blackwell, Matthew, Stefano M. Iacus, Gary King, and Giuseppe Porro. 2009. “cem: Coarsened Exact Matching in Stata.” Stata Journal 9 (4): 524–46. Christmals, Christmal Dela, and Kizito Aidem. 2020. “Implementation of the National Health Insurance Scheme (NHIS) in Ghana: Lessons for South Africa and Low- and Middle- Income Countries.” Risk Management and Healthcare Policy 13: 1879–904. doi: 10.2147/RMHP.S245615. Coulombe, Harold, and Quentin Wodon. 2012. “Benefit Incidence of Public Health Spending for Public and Faith-Inspired Health Facilities in Ghana.” MPRA Paper 45390, University Library of Munich, Germany. Das, Jishnu, and Quy-Toan Do. 2022. “Two Decades of Health Insurance in Low- and Middle- Income Countries.” Manuscript. 22 Dadzie, Christabel, Dhushyanth Raju, and Stephen D. Younger. 2022. “Social Protection Program Spending and Household Welfare in Ghana.” Unpublished Manuscript. Engelhardt, Gary V., and Jonathan Gruber. 2011. “Medicare Part D and the Financial Protection of the Elderly.” American Economic Journal: Economic Policy 3 (4): 77–102. Feldstein, Martin. 1973. “The Welfare Loss of Excessive Health Insurance.” Journal of Political Economy 81 (2): 251–80. Feldstein, Martin, and Jonathan Gruber. 1995. “A Major Risk Approach to Health Insurance Reform.” Tax Policy and the Economy 9. doi: 10.1086/tpe.9.20061828. Finkelstein, Amy, and Robin McKnight. 2008. “What Did Medicare Do? The Initial Impact of Medicare on Mortality and Out-of-Pocket Medical Spending.” Journal of Public Economics 92 (7): 1644–68. Gajate-Garrido, Giselle, and Rebecca Owusua. 2013. “The National Health Insurance Scheme in Ghana: Implementation Challenges and Proposed Solutions.” IFPRI Discussion Paper 1309, International Food Policy Research Institute, Washington, DC. Garcia-Mandicó, Sílvia, Arndt Reichert, and Christoph Strupat. 2021. “The Social Value of Health Insurance: Results from Ghana” Journal of Public Economics 194: 104314. doi: 10.1016/j.jpubeco.2020.104314. GSS (Ghana Statistical Service). 2019. Ghana Living Standards Survey (GLSS) 7: Main Report. Accra: GSS. Limwattananon, Supon, Sven Neelsen, Owen O’Donnell, Phusit Prakongsai, Viroj Tangcharoensathien, Eddy Van Doorslaer, and Vuthiphan Vongmongkol. 2015. “Universal Coverage with Supply-Side Reform: The Impact on Medical Expenditure Risk and Utilization in Thailand.” Journal of Public Economics 121: 79–94. Nketiah-Amponsah, Edward, Sheetal Silal, Timothy Awine, and Brad Wong. 2020. “Cost Benefit Analysis of Selected Malaria Interventions in Ghana.” Ghana Priorities, Copenhagen Consensus Center, Lowell, MA. Nsiah-Boateng, Eric, and Moses Aikins. 2018. “Trends and Characteristics of Enrolment in the National Health Insurance Scheme in Ghana: A Quantitative Analysis of Longitudinal Data.” Global Health Research and Policy 3:32. doi: 10.1186/s41256-018-0087-6. Owen O’Donnell, Eddy van Doorslaer, Adam Wagstaff, and Magnus Lindelow. 2008. Analyzing Health Equity Using Household Survey Data: A Guide to Techniques and Their Implementation. Washington, DC: World Bank. Okoroh, Juliet, Samuel Essoun, Anthony Seddoh, Hobart Harris, Joel S. Weissman, Lydia Dsane-Selby, and Robert Riviello. 2018. “Evaluating the Impact of the National Health 23 Insurance Scheme of Ghana on Out of Pocket Expenditures: A Systematic Review” BMC Health Services Research 8 (1): 426. doi: 10.1186/s12913-018-3249-9. Powell-Jackson, Timothy, Kara Hanson, Christopher J. M. Whitty, and Evelyn K. Ansah. 2014. “Who Benefits from Free Healthcare? Evidence from a Randomized Experiment in Ghana.” Journal of Development Economics 107: 305–19. Shigeoka, Hitoshi. 2014. “The Effect of Patient Cost Sharing on Utilization, Health, and Risk Protection.” American Economic Review 104 (7): 2152–84. Wang, Huihui, Nathaniel Otoo, and Lydia Dsane-Selby. 2017. Ghana National Health Insurance Scheme: Improving Financial Sustainability Based on Expenditure Review. Washington, DC: World Bank. Younger, Stephen D., Eric Osei-Assibey, and Felix Oppong. 2017. “Fiscal Incidence in Ghana.” Review of Development Economics 21 (4): e47–e66. 24 Appendix A Comparison with Barofsky and Younger (2022) The approach we take in this paper is based on Barofsky and Younger (2022). While conceptually similar, we have made substantial modifications to their approach, listed here: 1. Barofsky and Younger use household-level data on income and healthcare expenditures, counting a household as “covered” by NHIS if half its members or more have a valid NHIS card. We use individual-level data on healthcare expenditures and NHIS membership (also verified by having an NHIS card) and assume that an individual’s income is her or his household’s consumption per adult equivalent, the standard welfare measure for poverty and inequality analysis in Ghana. 2. Barofsky and Younger multiply reported two-week healthcare expenditures by 26 to obtain annual OOP healthcare expenditures. As described in the paper, we draw 26 observations from the welfare decile-specific distribution of reported two-week healthcare expenditures and add them to estimate annual OOP healthcare expenditures. Our distribution is thus much less skewed. 3. Barofsky and Younger match NHIS member to nonmember households by household income quartiles. Having already calculated decile-specific distributions of OOP healthcare expenditures, we match individuals in the entire sample. 4. Barofsky and Younger use quantile regression of OOP healthcare expenditures on NHIS membership at each percentile of the healthcare expenditure distribution to estimate the difference in OOP healthcare expenditures (the “treatment effect”) at that percentile. This generates their distribution of possible healthcare expenditures. We estimate OLS regressions of the natural log of OOP healthcare expenditures on the natural log of household consumption per adult equivalent separately for the (matched) samples of NHIS members and nonmembers and then draw 100 random values from that distribution to estimate each person’s distribution of possible OOP healthcare expenditures. 5. Barofsky and Younger use data from GLSS 2012/13, while we use data from GLSS 2016/17. We have run our models using GLSS 2012/13 data as well and find the results are very similar to what we report in this study. 25 Despite substantial differences in the exact calculations, both papers find a small financial insurance value from NHIS membership, though Barofsky and Younger do find somewhat larger values than those reported here. For the first three quartiles of the household income distribution, they find financial insurance values ranging from –6 to 80 cedis (–3.4 to 45.7 PPP dollars) per year per household, depending on the assumed floor for consumption and the risk aversion parameter. The results for the top quartile are greater, ranging from –9 to 301 cedis (–5.1 to 172.0 PPP dollars) per year per household. Given that average household size in the GLSS 2016/17 data is 3.8, these results are quite close to ours (which are for individuals, not households). 26 Appendix B Distribution of Out-of-Pocket Medical Expenditures, Medicare-Eligible, United States Finkelstein and McKnight (2008) report the estimated OOP medical expenditures for those age 65 and older in the United States before and after the introduction of Medicare in 1963. Figure B1 presents their results, excerpted from their paper. “Hypothetical” shows a mechanical adjustment of medical expenditures based on what Medicare covered when it was introduced, while “quantile treatment adjusted” shows the results for a matching procedure similar to what we use in this paper. As in Ghana, medical expenditures are sharply skewed both with and without Medicare insurance, but, in the United States, the gap between the two is considerably larger at the top deciles. So, unlike Ghana’s NHIS, the United States’ Medicare did offer significant reductions in financial risk for its participants. Figure B1. Distribution of Medicare-Eligible Out-of-Pocket Medical Expenditures in the United States, 1963 Source: Figure 9 (pg. 1664) of Finkelstein and McKnight (2008). 27 Appendix C Catastrophic Out-of-Pocket Health Expenditures of NHIS Members and Nonmembers It is common in the literature to use reductions in catastrophic OOP healthcare expenditures as one measure of the financial protection that health insurance offers its members. While there are many measures of “catastrophic” payments in the literature, the two most common seem to be 1. Total household OOP healthcare expenditures divided by total household income or consumption; or 2. Total household OOP healthcare expenditures divided by total household income or consumption less “essential” expenditures, which are usually taken to be food expenditures. Different authors apply different cutoff values for these measures, with 10 percent being common for the first and 40 percent for the second. 28 We have reservations about these measures because large outliers in the OOP healthcare expenditure variable will have a much more influential effect on catastrophic payments than they do, say, on mean OOP healthcare expenditure. This is a particular concern in our data from GLSS 2016/17 where OOP healthcare expenditures are recorded by hand in two side-by-side columns (see below graphic), one for cedis and the other for pesewas (one one-hundredth of a cedi). It would be easy enough for a data entry operator to record an entry like this as 1700 cedis rather than 17.00 cedis: 10 How much did (NAME) pay for consultation? (IN NEW CURRENCY) GH¢ GHp 17 00 But even without data entry errors, survey respondents are known to “telescope” large payments that happened before the recall period into their response. Either of these errors will bias upward 28 See chapter 18 of O’Donnell et.al. (2008) for an extensive review and recommendations. 28 the estimate of catastrophic expenditures. In addition, there is an uncomfortable arbitrariness in selecting the cutoff value for what constitutes “catastrophic.” Despite these concerns and for the sake of comparison to other studies, we report calculations of two standard measures used to calculate catastrophic OOP healthcare expenditures: the ratio of OOP healthcare expenditures to total household expenditures and the ratio of OOP healthcare expenditures to total household nonfood expenditure. To avoid the decision of selecting a cutoff value for “catastrophic” OOP healthcare expenditure levels, we present the inverse cumulative density functions of these ratios. Each reader may choose her or his favored cutoff on these graphs. One difference between our ratios and the standard ones in the literature is that we use individual OOP healthcare expenditure over household expenditure. We want to compare NHIS members with nonmembers, and membership is for individuals, not households. Many households have both members and nonmembers, so it is impossible to classify households as NHIS (non)members. This difference should lower the cutoff one chooses for catastrophic OOP healthcare payments, but should not influence (much) the inverse cumulative density functions we present. Figure C1 shows the inverse cumulative density function of individual OOP healthcare expenditure divided by total household expenditure (panel a) or by total household nonfood expenditure (panel b). Both functions are so close that, no matter the cutoff one chooses to determine that OOP healthcare payments are catastrophic, there will be little difference in the measure for NHIS members and nonmembers. This reinforces our conclusion in the main text that NHIS provides little in the way of financial risk insurance even as it does provide a significant transfer to members. 29 Figure C1. Inverse Cumulative Density Function of the Ratio of Individual Out- of-Pocket Healthcare Expenditures to Household Expenditure a. With respect to total household expenditure 1 (OOP healthcare exp.)/HH expenditure .1 .2 .3 .4 .5 .6 .7 .8 .9 0 0 20 40 60 80 100 Percentile of (OOP healthcare exp.)/HH expenditure Nonmember NHIS member b. With respect to total household nonfood expenditure (OOP healthcare exp.)/HH nonfood expenditure 3 2.5 2 1.5 1 .5 0 0 20 40 60 80 100 Percentile of (OOP healthcare exp.)/HH nonfood expenditure Nonmember NHIS member Note: HH = household. 30