Income and the Demand for Complementary Health Insurance in France

Transcription

1 Document de travai Working paper Income and the Demand for Compementary Heath Insurance in France Miche Grignon (McMaster University, Irdes) Bidénam Kambia-Chopin (Irdes) DT n 24 Apri 2009 Institut de recherche et documentation en économie de a santé IRDES - Association Loi de rue Vauvenargues Paris - Té. : Fax :

2 Institut d e r e c h e r c h e et d o c u m e n t a t io n en é c o n o m ie d e a s a n t é 10, rue Vauvenargues Paris Té: Fax: E-mai: diffusion@irdes.fr Director of pubication: Chanta Cases Scientific adviser: Thierry Debrand Copy editing: Franck-Séverin Cérembaut Layout compositor: Khadidja Ben Larbi Distribution: Suzanne Chriqui, Sandrine Bequignon Judgement and opinions expressed in this pubication are those of the authors aone and do not carry endorsement of Irdes.

3 - 1 - Miche Grignon, Ph.D 1, 2, 3 and Bidénam Kambia-Chopin, Ph.D 4 Abstract This paper examines the demand for compementary heath insurance (CHI) in the non-group market in France and the reasons why the near poor seem price insensitive. First we deveop a theoretica mode based on a simpe trade-off between two goods: CHI and a composite good refecting a other consumptions. Then we estimate a mode of CHI consumption and empiricay test the impact of potentia determinants of demand for coverage: risk aversion, asymmetrica information, non-expected utiity, the demand for quaity and heath, and suppy-side factors such as price discrimination. We interpret our empirica findings in terms of crossed price and income easticity of the demand for CHI. Last, we use these estimates of easticity to simuate the effect of various eves of price subsidies on the demand for CHI among those with incomes around the poverty eve in France. We find that the main motivation for purchasing CHI in France is protection against the financia risk associated with copayments in the pubic heath insurance scheme. We aso observe a strong income effect suggesting that affordabiity might be an important determinant. Our simuations indicate that no poicy of price subsidy can significanty increase the take-up of CHI among the near poor; any increase in the eve of subsidy generates a windfa benefit for richer househods. Keywords: Demand for heath insurance, Uninsured, Premium subsidies JEL cassification: D12, D81, I11, I18 1 McMaster University, Department of Economics and Department of Heath, Aging, and Society, Hamiton, Ontario, Canada. 2 Center for Heath Economics and Poicy Anaysis, McMaster University, Hamiton, Ontario, Canada. 3 Associate researcher, IRDES, Paris, France. 4 Economist, IRDES, Paris, France.

4 - 2 - Section 1 Introduction This study examines the demand for vountary heath insurance in France and, more specificay, the reationship between income and the quantity of private coverage individuas purchase on the nongroup market. Vountary heath insurance (VHI) pays an important roe aongside the pubic scheme in European countries (Mossiaos and Thomson, 2004, from now on MT2004). In some countries VHI serves some popuations, whether they are aowed to opt out of the pubic scheme (Germany) or excuded from it (Netherands); MT2004 suggest the name substitutive heath insurance in this case. In other countries, VHI covers anciary services not incuded in the pubic pan or, more importanty, extra-costs associated with better perceived quaity of care such as choice of physician and shorter waiting times for non-emergency hospita care (Austraia, Spain, and United Kingdom); this is referred to as suppementary heath insurance. Last, as is the case in France (and, to some extent for Medigap poicies in the US), VHI covers co-payments (incuding over-biing) of the pubic scheme and is caed compementary heath insurance, or CHI (MT2004). The rationae for mixed pubic-private systems is theoreticay straightforward: equity and adverse seection provide compeing arguments for universa coverage by a pubic singe insurer and singe payer (or even, in the US case, for tax exemptions for empoyer-sponsored pans); however, the disconnection in pubic systems between what one pays and the coverage one benefits from exacerbates the tendency toward mora hazard extent in any insurance scheme. The poitica economy iterature coins that tendency over-insurance and recommends that the pubic pan be imited to a basic eve of coverage of what is deemed necessary 5 (Besey and Gouveia,1994). Individuas then have the option to vountariy buy some individua heath insurance if they are wiing to get more than what this basic pan offers: the important thing here is that individuas have to pay the price of that insurance topping up the basic pan and, as a resut, woud purchase it ony if the vaue they get from it exceeds the price they are charged for it. In that normative sense there coud not be any overinsurance in VHI. In the case of suppementary VHI the definition of what is necessary and shoud be incuded in the basic pan is cear and refers to cinica notions (the second principe in the constitution of the British NHS states that access to services is based on cinica need, not abiity to pay, Department of Heath 5 In the German and Dutch cases the notion is that empoyees need compusory pubic coverage whereas the sefempoyed, managers, and professionas are free to buy coverage (or not). In that case, mora hazard is controed by the contribution rate: in short, ow and midde income cannot free ride on the rich as in a system financed through genera taxes.

5 ). It is certainy not uncontroversia to determine what constitutes a cinicay necessary time to access a hip repacement on a case-by-case basis but at east the principe is cear. In CHI systems, such as the French, a heath care good or service is cassified as more or ess necessary depending on the eve of the co-payment in the pubic pan 6. In such a system, the deimitation of what is necessary and what beongs to individua responsibiity is a combination of characteristics of the good or service on one hand, and of the abiity to pay of the insured on the other hand. Because a co-payment affects utiization through a price effect, and since price effects work differenty at various income eves, in the French system the same service is deemed ess necessary at the bottom of the income distribution than at the top 7. Because CHI is amost never priced based on abiity to pay (this happens in some pubic arge empoyers but it remains exceptiona) but rather on a mix of fat rate and risk adjustment (oder insured tend to pay substantiay more), the combination of co-payments in the basic scheme and vountary CHI to reimburse those raises two equity issues: the cost of CHI reative to income is greater for those at the bottom of the income distribution if they decide to purchase it; if not the abiity to access heath care wi be diminished, generating an income gradient of utiization of medicay necessary goods and services. This transates in internationa comparisons of equity of heath care utiization: for instance the horizonta inequity indices of ambuatory care utiization (probabiity of any visit to a GP in the past 12 months, tota number of visits to a GP, probabiity of any visit to a speciaist and tota number of visits to a speciaist) show eves of pro-rich inequity between three and ten times higher in France than in the United Kingdom (van Doorsaer and Masseria, 2004). A ogica response to such an unintended consequence of using co-payments as a way to imit what is necessary and fas in the pubic responsibiity is to subsidize the purchase of CHI: if individuas are reimbursed a fraction of the cost of their CHI contract, fraction that diminishes when their income increases the government can make sure that no househod has to spend more than a given share of its tota income on CHI and that a househods who are wiing to buy have some CHI and can access necessary care. A first step in that direction was made in France in 2000 with the impementation of Couverture Maadie Universee Compémentaire (CMU-C), a means-tested entirey free CHI coverage (see a detaied description in, e.g. Grignon et a. 2008). A second step was the creation of Aide à acquisition d une compémentaire santé (ACS) in 2005, providing partia reimbursement of the cost of a non-group CHI contract for individuas iving in househods with income between 100% and 115% of the cut-off eve for CMU-C (see beow for a description of the scheme). 6 Since co-payments can be covered by a CHI poicy they cannot contro mora hazard but can spit tota cost of a service between a basic eve of coverage and what is the individua s responsibiity. 7 Another consequence of co-payments and CHI, and one that is more often mentioned in pubic debates is regressive financing: the same co-payment represents a arger share of a sma budget and poorer househods are hit harder Indeed any increase in the rate of co-payment hits harder at the bottom of the income distribution and we do not dispute it. However, in a static situation where co-payments are used to determine what is not medicay necessary one shoud not be more concerned by the share of co-payments in tota income than by the income reated differentia cost of buying a uxury car. As a resut, we posit here that the ony issue in a socia insurance system with co-payments is that of incomereated inequity in the utiization of heath care services.

6 - 4 - The impementation of CMU-C has demonstrated ceary that co-payment and CHI generate inequities in heath care utiization: the utiization of heath care by the 10% poorest in France went up to match the eve of utiization of those with a private CHI when CMU-C was impemented (Grignon et a. 2008). However, the second step seems ess of a success: despite what coud be seen as a generous eve of subsidy the take-up remains ow (at between 10% and 20% of the target popuation, see Franc and Perronnin 2007). This ack of success raises the foowing question: by how much shoud the purchase of a private CHI be subsidized by the pubic purse and where shoud the income cut-off be? In this research, we measure the efficiency of the subsidy: how high shoud the subsidy go to entice one individua to purchase a CHI contract, and what is the cost in windfa for those who receive the subsidy but were ready to purchase a CHI even without it? This is different from the efficiency of subsidizing SHI in the UK (Emmerson et a. 2001) or Spain (Lopez-Nicoas and Vera-Hernandez, 2008), measured as the gain in NHS resources when one individua uses their private SHI poicy to access private heath care. It is much coser to the `bang for the buck` approach of subsidies for private insurance in the US (Marquis and Long, 1995; Gied, 2001; Auerbach and Ohri, 2006; Gruber, 2007). We aso investigate the wefare consequences of the ack of CHI: after a, not everybody shoud be covered by CHI and ess than universa coverage for co-payment does not aways entai a wefare oss (Gruber, 2008, Monheit and Primoff Vistnes, 2006). To address that issue we estimate a function of demand for CHI and we disentange suppy-side effects from behaviora and affordabiity ones. Our findings contribute mosty to the debate on co-payments and CHI in heath care systems with a strong pubic singe payer, such as those found in Europe. However, it might aso be of interest to the debate in the US on poicies such as subsidy to purchase private pans on the non-group market or tax credit to increase offerings of empoyer-sponsored pans aimed at increasing heath insurance coverage (Gied 2001; Swartz, 2001; Zeenak, 2001). To estimate the demand function for CHI in France we use a inked survey-caims dataset that provides individua-eve information on CHI (premium paid and whether it is group or non-group), usua sociodemographics, attitudes toward risk and heath, and administrative caims data on heath care spending and co-payments eft by the basic pan. We mode the demand for compementary heath insurance as a simpified trade-off between two goods: CHI and a composite good refecting a other consumptions. The mode contains two important features: first, there is a minimum eve of the composite good beow which ife is not sustainabe, so that even an infinite eve of CHI cannot compensate for a consumption eve of the composite good beow that minimum; second, the minimum eve of CHI (the eve for which they want to receive an infinite eve of the composite good to be compensated and keep the same eve of utiity) is negative for

7 - 5 - some individuas (reca they aready benefit from basic coverage and ack of CHI ony means copayments). Whereas most studies of demand for VHI mode a binary variabe indicating whether the individua is covered or not (e.g. Marquis and Long, 1995; Costa and Garcia, 2003; King and Mossiaos, 2005; Auerbach and Ohri, 2006; Sabia and Venteou, 2007) we mode the premium paid (incuding 0s for the non-covered). This does not aow us to estimate a price easticity of the demand for CHI in France but we want to mode the effect of income, tastes, and suppy-side features on the demand for CHI. A natura way of estimating the observed demand resuting from such an underying utiity maximizing behavior is the Tobit estimator: we use a the information avaiabe on the non- group market, incuding individuas without any CHI for which we treat the zero quantity as censored negative quantities. Even though our econometric strategy does not aow us to estimate the price-easticity of the demand for CHI, we use our mode and the parameters estimated for the demand function to simuate the response of the demand for CHI to changes in the subsidy. Our findings are as foows: The main motivation for purchasing CHI in France is protection against a financia risk (risk aversion). We aso find a very sma price effect on the decision to buy a poicy among the near-poor, thus confirming findings on the decision to buy VHI based on other heath care systems with a strong pubic payer (King and Mossiaos, 2001, in the UK; Buter, 2001, in Austraia; and Costa and Garcia, 2003, for Spain) but contrary to what is observed in the US (Auerbach and Ohri, 2006). We find a very strong income effect on the quantity of CHI demanded, confirming previous findings on the decision to buy non-group CHI in the US or in France (Sabia and Venteou, 2007): in our mode, individuas with an equivaized househod income beow 700 (approximatey USD900) per month are unikey to buy a CHI poicy even if the price was heaviy subsidized. Beyond that income eve, most consumers woud buy even with a sma subsidy. These findings suggest that subsidizing the purchase of CHI is unikey to be an efficient poicy to increase coverage: targeted individuas wi not buy anyway and those who aready buy CHI without subsidy wi benefit from a windfa profit, making the cost per unit increase in coverage very high. Section 2 Heath Insurance in France The market for vountary, private heath insurance in France is a market for compementary heath insurance (CHI): a ega residents of France are covered by a socia scheme (Sécurité sociae) financed out of ear-marked income tax (the socia contribution ). The basic pan provides medica, drug, and hospita insurance with amost no deductibe (some were introduced on ambuatory care in 2007) and covers denta and eye care (but not prostheses or prescription gasses). User-fees are associated with coinsurance and provider over-biing, and a stop-oss cause so that a expenses are

8 - 6 - covered above a specified eve. As a resut there is an inverted tunne in the socia scheme where ow eves of spending are fuy covered as are high eves, eaving a medium range of spending with userfees, representing in some cases as much as 90% of tota cost (for a detaied presentation of the French system, see Couffinha and Franc, 2008). Compementary pans work within the frame of rues and reguations devised by the socia scheme to seect the services and goods they cover, and even benefit from the prices negotiated at the nationa eve for a coinsurances: e.g. there is a nationa retai price for prescription drugs, the socia scheme reimbursing a given rate (which can be as ow as 20%) and the compementary scheme compementing the reimbursement to 100% without excuding any drug that is on the socia scheme s formuary, and incuding a margina number of drugs not reimbursed by the socia scheme. There is no rea competition between pans in terms of coinsurance since a CHI schemes provide fu coverage. CHI poicies differ in the amount they cover for over-biing, denta prostheses and prescription gasses (Bocognano et a, 1998; Couffinha et Perronnin 2004). Over biing is rare for GP services (ony 12% of GP over bi, Fennina and Geffroy, 2007), can be frequent (38% of speciaists) but aways imited in vaue ( 27 on average, EcoSante, IRDES, 2007) for some ambuatory care speciaties (ENT, eye speciaists, dermatoogists), very frequent and of a different magnitude for surgeons in private cinics (the socia scheme reimburses a fee but private surgeons charge 75 on average above it). It is aso important to note that CHI pans cover over-biing without any attempt at seecting procedures or providers and negotiating prices with providers. Overa, approximatey 78% of tota heath care expenditures are covered by the socia scheme, with 13% covered by CHI and 9% out-of-pocket (Fennina and Geffroy; 2007, Couffinha and Franc, 2008). Individuas without CHI sti have access to medica care and the socia scheme covers catastrophic expenditures. An individua s average expenditures not covered by socia insurance system are 421 per year (estimation by the authors based on a representative sampe of administrative data), substantia if not catastrophic. These expenditures, however, are highy concentrated among a sma number of individuas. In 1999, 84% of the popuation had CHI. Since the introduction of CMU-C in 2000, 9% are without any coverage and 2.5% of the popuation is covered by ow-quaity CHI pans (not covering any over-biing or denta prosthesis or prescription gasses, Franc and Perronnin, 2006). ACS was introduced in 2005 and it works as a voucher: any eigibe individua uses the voucher to get a rebate on the purchase of a non-group CHI contract 8 and the suppier of the contract gets reimbursed by the government. The voucher amounts to 75 per individua beow age 25, 150 per individua ages 25 to 59, and 250 per individua ages 60 and oder, to individuas iving in househods above the income cut-off for CMU-C 8 Amost a individua market contracts are eigibe: minor restrictions appy to make sure the contract foows the genera rues impemented by the pubic fund, namey a GP gatekeeper.

9 - 7 - and beow 115% of the cut-off (120% since January 2007). The expected target was 2 miion peope (approximatey 3.5% of the popuation), but ony 240,000 had taken it up in November 2006 (and 330,000 overa after the increase in the cut-off income to 120%, according to the Fonds CMU). The average subsidy amounts to amost 40% of the average premium paid by ACS beneficiaries (Franc and Perronnin, 2007), and represents 25% of the average premium on the non-group market (ACS vouchers are used to purchase ower quaity contracts). Sti, even with this generous subsidy, the subsidy never reay took off. A number of factors potentiay expain this ow up-take. Aggregate data shows a oading fee in 2003 of approximatey 20% on average for CHI in France, which is somewhat higher than what is observed in other setting (e.g. Gruber (2008) reports 12% for the US) and might indicate one or both of a ow eve of competition or too many sma firms in the business. Out of what is paid out to individuas by suppementary insurers we estimate that approximatey 80% goes to reimburse users fees of the socia scheme with the remaining paying for services not covered by socia insurance such as in vitro fertiization or aternative medicines 9. Despite the stop-oss on catastrophic spending user s fee is concentrated on a subset of individuas. In our dataset, which is a representative sampe of individuas and administrative caims for reimbursement to the socia scheme, we are abe to describe the distribution of the costs eft to patients by the socia scheme, as we as the distribution of costs for a variety of services (hospita, GPs, speciaists, drugs, denta care, prescription gasses, transports). Among both those with and without CHI in 2004, the 20% top spenders represent 60% of tota user charges. The average yeary user charge in the top 20% is 1,327, versus 182 among the remaining 80%. Among those without CHI the 20% top spenders account for 80% of tota user fees, and ower averages ( 1,235 and 109 respectivey). User charge is more concentrated on hospita, denta and gasses: over these three types of service, the 10% top spenders account for 72% of charges, with an average of The estimation is as foows: the average user fee is 421 in 2004 and 60% of user fees are paid for by CHI. Mutipying 60% of 421 by 60 miion residents of France yied a tota paid on reimbursing user fees of biion for the year Over the tota outay from CHI in the Heath accounts for the same year ( biion ) this yieds a ratio of 80%

10 - 8 - Tabe 1: Concentration (p% top spenders accounting for X% tota spending), average spending among p% top spenders and (1-p)% others, for each type of service. Service Probabiity (top spenders) Share of tota Average spending --top Average spending others (in ) spenders (in ) Hospita 5 77% Denta prostheses 5 79% Prescription gasses 5 67% Drugs 20 58% Over-biing -- Speciaists 20 69% Over-biing GPs 20 60% Other 20 71% From these observations it seems cear that reducing the financia risk stemming from the user charges eft by the socia basic scheme coud be an important motivation to purchasing CHI. Individuas can access CHI through an empoyer-sponsored contract or on the non-group market. Sefreports (ESPS 2004, un-weighed, avaiabe on IRDES website 10 ) indicate that 39% of contracts are through an empoyer and 2% through a poo for sef-empoyed. Another 39% are obtained on the nongroup market, and 15% are mixed: these are contracts subscribed by retirees as maintaining the coverage they had through their previous occupation (insurers cannot deny coverage and cannot increase premiums by more than 50%). The non-group market is more important in France than in the US because individuas over the age of 65 are sti wiing to purchase private insurance. In France contributions paid for directy by empoyers to a CHI contract are not taxed (even though they coud be considered in-kind wages) but there is no tax credit for individuas purchasing CHI on the non-group market or on the empoyee s share of the contribution in the group market. Who are the non-covered for CHI in France? Based on our survey for 2004 the mean equivaized househod income of the non-covered is 844 per month, compared to 1,382 among those who buy CHI. Among those with an income per unit beow 1,000, the proportion of non-covered is 24%, versus 4% ony among those with an income with more than 1,900. However, 25% of the non-covered have equivaized incomes of more than 1,000, impying that income is not the ony cause of non-purchase of CHI (some individuas do not buy even though it is affordabe). Living in Paris is a main factor of noncoverage: 19% of Parisians do not purchase CHI, versus 7% of individuas in rura areas. Age is not a major factor of non-coverage, with 15% of those younger than 30 being non-covered, versus 11% among the 65 and over. 10

11 - 9 - Section 3 A mode of the Demand for Compementary Heath Insurance We deveop a mode of demand for compementary heath insurance that expicity incorporates a threshod of affordabiity. Assume that individuas maximize utiity over two goods: CHI, which they consume a quantity x, and a composite consumption, which they consume a quantity c, under a binding budget constraint based on current income y (no saving or borrowing): Max U(c; x) s.t. π.x + с = y (1) Where: с = the numéraire, π = the reative price of CHI In such a mode, heath care insurance and other consumptions can aways be traded off at the margin. To mode zero-expenditure on the heath insurance good, however, we must assume two things: first, individuas need a minimum eve of the composite good to survive; if c were to fa beow that eve, no amount of x coud offset the disutiity generated. Second, a positive utiity can be obtained even when x is negative, subject to c being arge enough to compensate. We generate rationa zero-expenditures on x based on the foowing utiity function: α α [ max(0;( c G )] ( x + 0 ) 1 U ( c; x) = x (2) where: G = the minimum consumption of the composite good needed to survive; G is a concept inked to affordabiity and sometimes referred to as eft to survive (Murray et a. 2000; Bundorf and Pauy, 2006). It reates to a genera perception of a hierarchy of needs (where CHI woud come ast) as has been suggested by Masow (1970). It says in substance that famiies become risk-averse when other needs are satisfied, and it is supported by some empirica evidence: Starr-McCuer, (1996) finds that uninsured househods save ess on average than insured ones, other things being equa, and even controing as far as possibe for seectivity (behaviora seection), which suggests that affordabiity expains more than aversion to risk of coverage and savings behaviors. 0 x = a eve of insurance coverage such that the margina rate of substitution between the consumption good and insurance is infinite. -x0 is a eve of insurance coverage beow the current eve offered by the pubic mandatory scheme (Sécurité sociae) that woud have to be reached to decrease utiity to 0 (or that woud require an infinite eve of c to be compensated for in utiity terms). This does not mean the mandatory scheme covers too much in any sense but simpy that it is above and beyond the sheer minimum individuas can cope with. Introducing that threshod beow the pubic scheme is the main innovation of our mode and the main rationae for being non-insured even though risk aversion is

12 greater at ow eves of income. Aso: 0-expenditure coud be generated with a positive or nu 0 x and suppy-side constraints such that a minima quantity of x1 > x0 has to be sod to each customer (Bradey, 2008) 11. Graphicay, such a utiity function is a standard Cobb-Dougas where c and x are substitute in the midde range and compements at ow vaues, ow being positive for c and negative for x. Figure 1 shows a first iso-utiity curve intersecting the horizonta axis (consumption of CHI is zero) before the optima soution in CHI (when the budget constraint is tangents to the curve) and a iso-utiity curve to the north-east (hence for a higher budget eve) intersecting the horizonta axis right when the budget ine tangents the curve. When income is arger (iso-utiity to the north-east) the optima bunde incudes a positive amount of CHI. Figure 1: Iso-utiity curves in the two-goods space with a positive minimum eve of composite good and a negative minimum eve of CHI. x Locus of optima bundes CHI-C 0 x 0 G-bar c 11 Bradey aso suggests another source of non-affordabiity namey individua variation in the price of insurance: if oading fees increase with some non-heath reated characteristics, then individuas with the same preferences and budget wi make different purchasing decisions. This woud of course be a rather crucia determinant of non-insurance if the poor were systematicay over-charged by a insurers (e.g. based on the fase assumption that the poor are ess carefu or more prone to mora hazard) and there woud not be much cause for subsidizing the price of compementary heath insurance in that situation. Such an income-based price discrimination does not seem to be observed on the market for CHI in France (see footnote 8 for empirica findings on this issue in the French case).

13 Figure 1 reads as foows: The iso-utiity curve crosses the axis when the sope is greater than the price ine at ow income eves but after at higher income eves (so that there are ony corner soutions beow a given income eve and inner ones above it). The determination of the eve of utiity (and, therefore, of the budget eve) at which the individua chooses to buy at east some CHI coverage is easiy derived from the utiity function: First, we derive the expression of c as a function of x (iso-utiity curves in the (c,x) space): 1 0 α 1 [ v( x + ) ] α α 0 α 1 U ( c; x) = v, c > G ( c G ) = v( x + x ) c = G + x (3) From (3), the sope of the iso-utiity curve is: dc dx U = v α 0 α α = v ( x + x ) α (4) The optimum vaue of x, x*, is given by: dc( x*) dx U = v = 1 1 α 1 α 0 πα α α 0 π v ( x * x ) π x* v x (5) α + = = 1 α Hence, there is a vaue v such that if v < v then x*( v ) < 0: there is no purchase of CHI for these accessibe eves of utiity. Hence, if income is not arge enough to grant v individuas wi rationay decide they need a negative eve of CHI; they woud even be happy to partiay opt out of the basic pan if that coud aow them to cut their contribution to the pubic pan. From equation (5), we find the minimum eve of utiity at which individuas start buying CHI as: as: α 0 πα v = x α (6). 1 The derivation above shows that rationa individuas can make the decision not to purchase CHI when their income is beow a eve that woud yied a utiity eve beow that minimum v. Of course, a that is observed is that beow that eve individuas do not buy CHI, and above it they buy some. What ooks ike a dichotomous discrete decision (and is very often modeed as such, as we wi describe in our method section) is here modeed as a continuous decision where the underying decision is observed with censoring.

14 In our mode, a individuas share the same rationa behaviour, most importanty the same underying reationship between income and quantity of CHI purchased. Our empirica estimation is aimed at estimating that underying reationship which refects affordabiity of CHI for a given eve of income. Individuas differ around that underying effect of income in their preferences for insurance, and this is refected in the parameter x 0 that is aowed to vary across individuas in our modes (as taste shifters). We detai beow in the next section the preference shifters that we use to characterize tastes and individua variation around affordabiity. We wi use the parameters (sope of the income effect and taste shifters) estimated in our econometric estimation to simuate the effect of a price subsidy on the quantity of coverage individuas, at a given eve of income, woud purchase on average. We use the mode to derive a response function of coverage to price and, therefore, to the eve of the subsidy and we are abe to get this without directy estimating the price-easticity of the demand for CHI. The derivation of the effect of the subsidy works as foows: The reationship between x* and y is straightforward, and we wi use it to evauate the parameters of interest {π, α, x0}: α α Substituting c = y πx in U ( c; x) = ( c G ) ( x + x 0 ) 1 yieds: U(x) = α α ( y πx G ) ( x + x 0 ) 1 Maximizing over x we get: πα( y πx * G ) 0 x * + x y πx * G α 1 1 α = πα 0 ( x * + x ) 1 α + (1 α)( y πx * G ) α 0 ( x * + x ) α = 0 Therefore, the demand curve is simpy: x 1 α 1 α = y G + αx π π 0 (7). We use (7) to simuate the impact of a price subsidy. The price subsidy is simuated as a reduction in π, the price of coverage (oading fee). A we need to do is to use (7) to cacuate the percentage of individuas at a given eve y with an x* at east equa to an arbitrary eve (what the government wants individuas to buy) for a eves of price beow the market price.

15 Our empirica mode provides an estimate for the sope of the income effect, therefore providing a vaue for (1-α)/π in equation (7). The empirica mode aso provides a vaue for the second member on the 0 right-hand side of equation (7), which is the taste shifter x pus a constant, and we use that vaue for each individua in our sampe to cacuate an individua vaue for x*. 0 In order to derive numerica vaues for x, and then, x* we need to assign vaues to two parameters: G and π. G is given by the CMU-C cut-off income, and we use a reasonabe assumption for the oading fee π. Section 4 Data and Method 4.1 Data Our dataset is a survey on heath, heath care, and heath insurance inked to administrative caims data on expenditures on heath care for each type of service (hospita stays, visits, denta care). The survey was conducted in 2004, and administrative caims data covers the period January to December of For each type of services administrative caims data indicates the tota amount spent during the year by the individua, as we as the share reimbursed by the socia scheme. Hence, we know the tota amount of user s charges paid by the individua or their compementary insurance. We drop a individuas with CHI obtained through their empoyer (even partiay) and restrict our popuation to those with a non-group contract (incuding retirees) and those with no CHI at a. The reason for doing it is that individuas who benefit from some contribution from an empoyer (or spouse s empoyer) cannot aways te the true vaue of the premium paid and woud not know the share of the contribution (if any) they paid out of their own pocket. We therefore keep the 9% of the popuation who are not covered as we as 33% of the popuation who purchased their poicy on the non-group market ony (33% is the product of 39% and 84%: 84% are covered privatey and 39% of those covered privatey purchased on the non-group market without any contribution from an empoyer). Overa, then our sub-sampe is comprised of 42% of the initia popuation; non-covered represent 21% of our study sampe (versus 9% of tota popuation, when those covered on the group market are incuded). An impication of restricting our sampe to those individuas is that we negect the seection process in the group market and assume a individuas offered some empoyer-sponsored contract (either directy or as spousa benefits) took it up. We therefore assume that those without CHI are individuas who are

16 on the market for non-group contracts and renounced these contracts, but never had access to group contracts that they decined. This is certainy a strong assumption if one beieves individuas may reject an empoyer pan they deem too generous and expensive, or seect jobs according to their offering good quaity empoyer-sponsored CHI. The ony data avaiabe in our sampe regards civi servants: in France, a civi servants are offered a group contract that they are free to reject and, as a resut, it is the ony group for whom we can be sure they were offered such a pan; our survey indicates that, out of the 396 respondents who report working as civi servants or for pubic entities, 24 ony are without CHI. Generaizing from civi servants we make the assumption that individuas wi tend to take up an empoyer s pan which is amost certainy cheaper than any non-group coverage 12. Not a respondents who reported some non-group coverage answered the questions on CHI, income, or heath. We treat missing vaues for independent variabes as foows. For categorica variabes (education or heath status) we incude a observations but define one category as missing. For continuous variabes we test two strategies: (1) we categorized the variabe, created a missing category just as for other categorica variabes, and incuded a observations; and (2) retained the continuous specification of the variabe but dropped a observations with a missing vaue. We excude 721 observations with missing information on the premium paid for CHI, which precude constructing the dependent variabe (19% of those with non-group CHI). As described beow we contro the impact of excusions based on a sampe seection mode. See tabe 2 for the impact of various excusions on our anaysis sampe. Tabe 2: Sampe size and excusions Sampes Tota sampe Subset with information on premium Individuas with no group CHI (incuding those with no CHI) Individuas with a non-group CHI poicy Subset with information on income and premium Subset with information on a independent variabes 5,106 4,385 3,644 3,618 3,762 3,041 2,658 2, The pan offered to civi servants is not very generous and it might therefore be somewhat presumptuous to generaize from this sma popuation. On the other hand the pan offered to civi servants is not the most cost effective of a group pans.

17 Estimation strategy Dependent variabe: The typica econometric mode of demand of non-group heath insurance (in the US: Marquis and Long, 1995; Auerbach and Ohri, 2006; in France: Saiba and Venteou, 2007) examines the dichotomous decision to purchase or not purchase insurance. In the French case, the proportion of individuas with CHI is higher and it is of greater interest to understand the quantity demanded rather than the probabiity of having any coverage. On the right hand side of the equation are found: income, taste shifters (education, heath status), and a price variabe. US studies construct the price variabe as a premium for a standard pan with $1,000 deductibe (Auerbach and Ohri, 2006). The premium is imputed on individuas based on their individua (age, gender, heath status) and oca (state eve) characteristics (medica price index and poicies affecting community rating). In the case of CHI in France, Saiba and Venteou (2007) identify a premium effect but it is not cear exacty how it is cacuated and they do not present any easticity resut. Our empirica approach is to mode the quantity of coverage demanded rather than the probabiity of being uninsured (in the U.S., Thomas (1995) uses the same strategy). This is simiar to estimating the atent variabe underying the binary choice of being insured or not. However, it puts emphasis on different dimensions of the demand function: in the binary choice modes, price is measured as the premium paid by an individua with a given eve of risk for a standardized contract (and eve of coverage). The price easticity refects mosty the underying risk of the individua and, ess importanty any oca (state) reguations affecting rating (e.g. community rating). In such specifications, price infuences demand in two opposite directions: as any price of a standard good, a higher premium yieds a ower quantity demanded; but, simutaneousy, since a higher risk-adjusted premium refects a higher need, a higher premium yieds a higher eve of demand. Econometricay, the price variabe is not exogenous in these modes. In this strategy the true price of insurance is the oading fee (premium divided by expected benefit) and demand is the quantity of coverage, defined as expected benefit given the parameters of the pan such as deductibes and co-insurance rates. However, we cannot repicate Thomas s strategy: we do not know the detaied parameters of the pan each individua buys and, as a resut, we cannot cacuate the expected benefit 13. We use the premium paid per person covered by the contract as our dependent 13 We are abe to construct some rough measure of the eve of coverage: a sub-sampe of respondents was interviewed four years before and asked to describe the guarantees on their poicy (at the time of this prior interview, which means there is no certainty it is the same poicy for which the premium is known). Based on these sef-reports we are abe to

18 variabe (for the measurement of the variabe see beow): as a resut we mode the consumption of insurance (unit price by voume) rather than the quantity. One consequence is that we do not have any rea price variabe: we do not observe the oading fee at the individua eve. We use proxies based on the risk adjusters (mosty age, gender and famiy size) of CHI operators and regiona dummies to contro for variations in medica and denta prices and for variations in the unit price of coverage charged to the individua 14. We wi estimate the price easticity of the demand for CHI based on our caibration of the utiity function: the econometric equation wi yied vaues for the iso-utiity curves and the budget ine which wi aow us to cacuate the sope of the demand ine (quantity-price response). Athough we do not observe the parameters of an individua s CHI pan and, as a resut, cannot estimate the expected benefit, our dataset does ink the premium paid for coverage with a set of variabes rarey observed in the same dataset: heath care expenditures, income, demographics (education, occupation, heath status). Cacuation of the premium was, in most cases, straightforward. The individua is covered by one contract ony and we know how many individuas are covered by the non-group contract (these are members of the same househod). We cacuate the vaue of CHI consumed by that individua as tota premium paid divided by the number of persons covered. Some cases are trickier though, when the same individua is covered by severa contracts. In such a case we cacuate, for each of these contracts, the vaue of insurance per person in the contract, and we sum these vaues to measure the tota vaue of consumption of CHI by that individua. We excude individuas with at east one empoyersponsored CHI therefore we cacuate the tota vaue of non-group CHI per individua. In these cases the vaue of the variabe Covpers in the mode is the average over a contracts (usuay two) of the number of individuas covered by each contract. The average tota premium on the non-group market is 527 per year. categorize the poicies into four categories (ow coverage, medium, high coverage with an emphasis on denta care, and high coverage with an emphasis on eye care). Our cacuations show a strong and positive correation between premiums and quaity and, more importanty, no price discrimination effect: for a given eve of quaity, age, and famiy size, the premium paid is independent of househod s income. We therefore reject the hypothesis that insurers charge higher premiums on poorer househods for the same eve of quaity. Data are not presented here but are avaiabe on request. 14 This is based on an assumption that the oading fee does not vary across regions or départements in France. A recent study shows variations in the average premium for a given poicy according to ocation in France (60 miions de consommateurs, 2008) but this does not prove that the oading fee per se varies: we interpret these variations as refecting variations in the price of medica services across regions rather than variations in oading fees. Such an interpretation is reasonabe because it is now more and more often the case that individuas purchase CHI onine, by mai, or through brokers and woud not, as a resut, be constrained to pay a higher price per unit of coverage than in a neighboring region.

19 Independent variabes: Income is our main variabe of interest in the econometric equation. It refects the effect of the budget ine on the iso-utiity curve in the two-goods space. Individuas who did not report on the vaue of their income are imputed the median of their income cass when this information is avaiabe in the survey. We define househod income as income per consumption units based on the CMU equivaence scae (which is simiar to the OECD scae) which assigns a weight of 1 for the first individua in the househod,.5 fors the second one,.3 fors the third and fourth ones, and.4 thereafter. When working with categorica income we define the foowing seven categories: income beow 700 per month (the cut-off for ACS); 700 to 999; 1,000 to 1,299; 1,300 to 1,599; 1,600 to 1,899; 1,900 to 2,199; and 2,200 and above. Since the premium paid by an individua is a function of the eve of coverage (how much of the overbiing is covered by the pan) and of the expected cost of the individua, we enter the main variabes used by insurers to charge the insured: age and famiy size. The shape of the iso-utiity curve depends on the utiity of being covered. We describe it as foows: Risk reduction: Individuas purchase insurance to reduce the financia risk of having to spend for treatment when sick (they protect their weath). Two main theoretica frameworks of risk reduction ead to different ways of measuring the gain of CHI in reducing risk. Expected utiity: Under the assumption that the amount spent on medica care in a given year is a random variabe, individuas with a concave utiity of weath are better off with fu coverage and are therefore wiing to pay a certain actuariay fair premium to reduce the uncertain oss generated by user charges. In this standard expected utiity framework, the poor are more wiing to purchase insurance (under the standard assumption that the utiity function is of the decreasing absoute risk aversion type) and we expect that introducing such a variabe wi increase the positive effect of income on the amount of CHI an individua purchases. We cacuate the vaue of risk reduction according to the expected utiity theory as foows: We introduce the risk reduction motive in our demand equation as the risk premium: P(Y) = Y U -1 [(1-p)U(Y)+pU(Y-D)] with Y the income (weath) of the individua, p the probabiity to be in the top spenders popuation, and D the average amount of spending within that popuation. Foowing our estimation in section 2 we use 0.2 for p and 1,235 for a vaue of D. We mode U as Y 1/2 in our baseine scenario.

20 Another mode of risk reduction is the prospect theory (Khaneman and Tversky, 1978): according to that theory the vaue of risk reduction is independent of income and increases at a decreasing rate with the vaue of damage (yeary user charge). We enter it in our demand equation as foows: we use the administrative data to cacuate expected vaues of spending over various sub-popuations defined by heath status that we impute to individuas according to their heath status. Attitude toward risk and uncertainty: Barsky et a (1997) and Monheit and Primoff Vistnes (2006) have demonstrated that attitudes toward risk and preferences regarding insurance are important determinants of the purchase of group insurance as peope seek jobs offering empoyer-sponsored insurance. We enter variabes describing genera atttudes toward risk in the mode; however none of these variabes reached significance in.our estimations and we utimatey dropped these from our preferred mode. Commitment to spend on heath care: individuas bind themseves into consuming heath care (e.g. denta prostheses) that wi be beneficia in the ong run but is not needed in the short run. They anticipate that they wi need to spend on denta care or prescription gasses and they use the CHI as an ear-marked saving device. The main reason for such a costy behavior (they have to pay the 20% oading fee on top of medica costs) is that they do not trust themseves in spending the money on these goods or services (e.g. if they had saved ahead of time). To account for such a motive, we enter the individua amount of user charges during the caendar year of the interview in the right hand side of our demand equation. Our mode does not encompass Nyman s (1999) suggestion that individuas purchase insurance in part to gain access to treatments that they woud never be abe to afford with their income or even their accumuated savings and credit. Such a motive is credibe in the American context but seems unikey in the case of CHI in France because individuas in need of a very expensive treatment get fu coverage through the socia scheme. Specification: In order to provide the intuition of the reationships we want to estimate, we start with a simpe OLS (mode 1 beow). However, because we want to mode a demand function with unobserved (censored) negative utiities, the Tobit is the best-suited estimator and wi be our preferred strategy. We use the sampe of individuas who are not covered on the group market, which incudes those with a CHI poicy and those who answered they had no coverage at a. Our dependent variabe takes the vaue of the premium for those with a CHI poicy on the non-group market and who agreed to provide a vaue for the

21 premium and a vaue of 0 for those who are not covered. The probem with that popuation is that a substantia proportion of the popuation was not abe (or not wiing) to report a vaue for the premium paid for their CHI poicy and, as a resut, cannot be used in the mode. A first consequence is that noncovered individuas (for whom the dependent variabe is aways known) represent a arger share of the sampe of individuas with non-group CHI and information on premium (incuding those with no CHI) than of the true popuation (approximatey 31% instead of 26%). Finay, the non-covered represent 27% of the sampe used to run the mode 2 beow (subset with information on a variabes). We aso run a Heckman sampe-seection mode: a first equation (Logit) estimates the probabiity that an individua wi not report the vaue of the premium for their CHI poicy on the foowing variabes: age, gender, education, sef-assessed heath, and site and type of conurbation (rura, urban ower than 200,000, urban arger than 200,000, Paris). We then use the inverse Mi s ratio (IMR) in an OLS estimate of the premium, estimated on the sub-popuation with no coverage or a coverage with a reported premium (mode 3 beow). We aso run a sensitivity check on the effect of non-responses on the income variabe: in the OLS version, we incude income as a categorica (rather than continuous, with seven categories) variabe and we run two estimations, one restricted to the popuation with a reported income, and one on the tota popuation with non reported income being one suppementary category. Resuts (not reported here but avaiabe on request) confirm that excuding observations with missing income does not change the coefficients and significance vaues. In the sampe-seection mode, we aso tried with a first step where the non response bias on premium is controed for a dummy variabe taking the vaue 1 for individuas who refused to provide information on their income. This improves substantiay the fit of the first step but eaves the coefficients on income unchanged in the second step. One thing we were not abe to assess empiricay is the assumption in the Tobit that the same determinants are at pay for the seection process (to buy or not a CHI) and conditiona consumption (once the decision has been made, how much of CHI to purchase). It is possibe that individuas anticipating higher premiums are deterred from seeking CHI in the first pace, and, as a resut, some characteristics, such as age woud have a negative impact on the probabiity and a positive one on conditiona consumption. This is not ikey for CHI in France, however, since we observe that individuas who pay higher premiums (e.g., the edery) are aso more ikey to be covered. Moreover, because the origin of censoring might come from a suppy side issue (the transaction costs of suppying a ow eve of CHI might be too high) as we as from the demand side issue tested so far, we re-estimate the demand function with a censoring threshod at 200 instead of 0 (it appears that very few contracts are worth ess than 200 per year and per person in our dataset). The findings are not