Adverse selection into health insurance in Bosnia and Herzegovina

(1)

Adverse Selection into Health Insurance in

Bosnia and Herzegovina

Author: Jacob-Jan Koopmans

Student number: 10452427

Master of Science in Economic, Specialization: Development Economics.

Supervisor: prof. dr. E.J.S. Plug

(2)

1

Abstract

This thesis evaluates whether the Bosnian and Herzegovinian health insurance system was characterized by adverse selection between 2001 and 2004. Most empirical studies on adverse selection in the context of health insurance focus on voluntary health insurance. However, the distinguishing feature of the Bosnian and Herzegovinian health insurance system is that health insurance is compulsory but not enforced. To evaluate whether there is adverse selection within the Bosnian health insurance system a fixed effects model based on the living in Bosnia and Herzegovina panel survey is developed in which the relationship between subjective or self-rated health-status and health insurance-status is estimated. Based on this method, there is no evidence of adverse selection into health insurance in Bosnia and Herzegovina.

(4)

3

I Introduction

Adverse selection is a widely debated topic in the economic scientific literature, especially in the context of health insurance. The theoretical idea of adverse selection in the context of insurance markets is that people that have a high risk are more likely to buy insurance than those that face lower risks. In the context of health insurance this implies that unhealthy people are more often insured than healthy people. The problem with this is that only unhealthy people will take on insurance which will increase the costs of providing insurance for the insurer. This will lead to a price increase which leads to a decrease in the value of the insurance for insurance policy holders and could make the insurance policy unattractive for the healthiest people in the insurance pool, resulting in a drop out of the healthiest people from the insurance pool. The result is that the insurer has to increase the price of their policy to compensate for the higher average risk per policy holder; this in turn leads to another drop-out of the healthiest policy holders. This vicious cycle can continue until the insurance policy is no longer

sustainable and ceases to exist. This mechanism is often revered to as the adverse selection death spiral (Cutler & Zeckhauser, Adverse Selection in Health Insurance, 1998). The underlying principles which might lead to an adverse selection death spiral may also lead to less severe outcomes. Due to the drop out of healthy individuals, the average risk per policy holder for the insurer becomes higher, to

compensate for this, insurance premiums need to be increased. This makes insurance more expensive for all policy holders. Moreover, assuming individuals are risk averse, the relatively healthy individuals are worse of as well since they would rather be insured than not if an insurance policy with an

actuarially fair insurance premium would have been offered to them. Therefore, adverse selection causes a decrease in surplus in society.

To counter adverse selection, and its potential negative consequences, a standard approach is to mandate compulsory health insurance. When a compulsory health insurance system is enforced and everyone has insurance, adverse selection is logically impossible. However, when a compulsory health insurance scheme is not fully enforced and people can, albeit illegally, remain uninsured or drop out of insurance, there is scope for adverse selection to occur. In Bosnia and Herzegovina such an, unenforced, compulsory health insurance scheme is available. This results in a large share of the population not having health insurance. Between 2001 and 2004, based on the living in Bosnia and Herzegovina survey (State Agency for Statistics (BHAS), Republika Srpska Institute of Statistics (RSIS), Federation of BiH Institute of Statistics (FIS), 2001-2004), almost 20 percent of the citizens remained uninsured. Therefore, even within the compulsory health insurance scheme in Bosnia and Herzegovina there can potentially be adverse selection. Therefore the research question of this thesis is:

Has adverse selection occurred within the health insurance system in Bosnia and Herzegovina between the years 2001 and 2004?

To evaluate whether there is adverse selection into health insurance in Bosnia and Herzegovina linear probability and fixed effects models to estimate the relationship between health and insurance status have been developed. These models are based on the data provided in the living in Bosnia and Herzegovina survey (State Agency for Statistics (BHAS), Republika Srpska Institute of Statistics (RSIS), Federation of BiH Institute of Statistics (FIS), 2001-2004) However, before discussing these models in

(5)

4 detail, it is important to get a better grasp of the concept and the implications of adverse selection. Therefore the next chapter focuses on the existing scientific literature on adverse selection both from an empirical and theoretical perspective. To get a better grasp of the institutional framework of the

Bosnian health insurance system, in chapter 3 the institutional framework of the Bosnian health care system will be reviewed. Before a model that estimates the relationship between health status and insurance status for Bosnian citizens can be constructed, it is necessary to focus on the possible proxies for health. Therefore in chapter 4 the relationship between self-rated health and ‘true’ health will be evaluated. To get a better grasp of the structure of the data, used in this thesis, chapter 5 discusses summary statistics and possible issues with the data. After this in chapter 6 different models for the identification of potential adverse selection will be constructed. The first models to be developed are linear probability models; these models provide a first insight into the effects of self-rated health on health insurance. These results are likely to be biased due to omitted variable bias. Therefore, to get more robust results, a fixed effects model will be estimated. The model section will conclude with heterogeneous effects to evaluate whether there are diverging relationships between health and insurance status among different sub groups in society. Finally in chapter 7 the results from the above mentioned models are presented.

The surprising result is that although theoretical models suggest that adverse selection can happen, adverse selection does not seem to be a significant issue in the Bosnian health insurance system.

(6)

5

II Adverse Selection

Adverse selection is a term most economists have some kind of understanding of. Often this term is being used in the context of insurance behavior. One of the first to mention adverse selection in the context of health insurance is Kenneth J. Arrow. He states that in a competitive insurance market with at least two groups of policy holders: “insurance plans could arise which charged lower premiums to preferred risks and draw them off, leaving the plan which does not discriminate among risks with only an adverse selection of them.” (Arrow, 1963) This implies that the non-discriminatory plan ends up insuring the people with the highest risk profile. However, adverse selection is applicable to more diverse topics. Essentially adverse selection is a problem of information asymmetry between the insurer and the potential insurance policy holder. Therefore to understand the impact of adverse selection, an understanding of information asymmetry is critical. An essential paper in this context is “The Market for “lemons”” by Akerlof (1970). Akerlof describes the consequences of information asymmetry using the second hand car market as an example. The idea is that buyers do not know the quality of second hand cars. Furthermore, sellers have an incentive to present cars as being of high quality to drive up the price they can ask for the car they are selling. When a high quality second hand car is on offer buyers will therefore not fully believe the seller if the seller presents his or her car as being a good car. These factors make buyers unwilling to pay a reasonable premium for a truly high quality car. Therefore an owner of a high quality car will receive a price which is too low to for the car. They will therefore not be selling their car. The result is that only the low quality cars are being offered on the market. And

therefore, “the bad “cars” drive out the good” (Akerlof, 1970, p. 489). Akerlof also connected this to the health care sector. For people 65 years and older it was at that time in the USA difficult to obtain medical insurance. His explanation is that insurers have less information on the health of the potential policy holders than the people which search for insurance. This leads the insurer to expect that the unhealthy people will insure themselves, whereas, healthy pensioners will be less likely to insure themselves. Therefore insurers have an incentive to set insurance policy prices high enough to

compensate for medical expenses expected to be made by the unhealthy. These high insurance policy prices will effectively drive the healthy out of the market as with these high insurance premiums the net present value of an insurance policy turns negative for healthy pensioners. Adverse selection is thus essentially a problem of asymmetric information between insurers and potential insurance policy holders. To evaluate the theoretical possibility of adverse selection more thoroughly, the next section focuses on some more extensive theoretical models of adverse selection.

II.i Theoretical Evidence

An early theoretical study on adverse selection has been provided by Rothschild and Stiglitz (1976). They postulate a model in which individuals either have a high or a low income risk, furthermore, Individuals have an option to insure against income fluctuations and it is assumed that people are equally risk-averse. Insurers operate in a competitive market, are assumed to be risk neutral and have enough capacity to sell all insurance policies that bring a non-negative profit. A further assumption is that individuals know their own risk profile and are thus able to evaluate how likely they are to end up with a low income. However, insurers do not have any information on the risk status of individuals and can thus not distinguish between high and low risk individuals.

(7)

6 When there is just one risk category, insurers and consumers have the same information set. Thus there is no asymmetric information and there is therefore no adverse selection. This situation is depicted in the following graph.

Figure 1. Insurance market with 2 groups of the same risk category. Reprinted from “Equilibrium in Competitive Insurance Markets: An Essay on the Economics of Imperfect Information”, M.Rothschild and J.Stiglitz, 1976, The

Quarterly Journal of Economics, 90(4), p.633

In this graph, W1 and W2 represent income in state 1 and 2 respectivelly. The line represents the indifference curve between income in state 1 and 2 for individuals. The line E-F is the break even line for insurers. This break even condition arrises from the assumption of a competitive insurance market. All insurance packages on this break-even line are potential insurance packages to be sold to consumers. From this graph we can see quite straightforwardly that individuals, since they are risk averse, maximize their utility by choosing the intersection of the 45˚ line and their indifference curve. This is point α*. At this point the insurers brake even and total utility is maximized. It is clear that there is no adverse selection when there is only one group of potential policy holders.

When there are multiple groups the situation changes. In the graph below there are 2 groups which differ in their risk parameters. One group has a low risk and the other group has a high income risk.

(8)

7

Figure 2. Insurance market with 2 groups with different risk parameters. Reprinted from “Equilibrium in Competitive Insurance Markets: An Essay on the Economics of Imperfect Information”, M.Rothschild and J.Stiglitz,

1976, The Quarterly Journal of Economics, 90(4), p.635

The difference between figure 1 and 2 is that now represents the indifference curve of the high risk group and represents the indifference curve of the low risk group. Theoretically there are 2 different types of equilibriums when there are 2 groups of individuals with different risk parameters. These can be classified as a pooling equilibrium and a separating equilibrium. A pooling equilibrium implies that there is only one equilibrium for both groups where both groups are offered the same insurance policy. On the other hand, a separating equilibrium is characterized by a separate equilibrium per risk group where each group is offered a separate insurance scheme.

In practice only a separating equilibrium can is feasible. In figure 2 above, suppose there is equilibrium at α. The problem with this equilibrium is that although it is on the indifference curve it can not be an equilibrium. To see why, lets look at point in this graph. is prefered to by the low risk group. However, is prefered by the high risk group over . This implies that if is offered, the high risk group will not follow. More generally any insurance policy that offers a combination of inome in period 1 and 2 that lies above the indiverence curve of one group which lies under the indifference curve of the other is a reason for a separating equilibrium to arrise. It is thus better for the insurer to offer separate

insurance policies to the different risk groups. In the following graph the possible separating equilibria are to be found.

(9)

8

Figure 3. Insurance market with 2 groups and a separating equilibrium. Reprinted from “Equilibrium in Competitive Insurance Markets: An Essay on the Economics of Imperfect Information”, M.Rothschild and J.Stiglitz, 1976, The

Quarterly Journal of Economics, 90(4), p.636

In figure 3 we see a situation with multiple groups, just like in figure 2. The difference with the last example is that separating equilibria are to be found. Therefore the insurer markets two different insurance policies. One for the high risk group which has break even line E-H and one for the low risk group which has break even line E-L. When assesing the best options for each group separatelly, high risk consumers are best of when they take policy where their indifference curve is tangent to the break even line E-H. Low risk consumers on the other hand will choose insurance option . The problem is that is also a more profitable option for the high risk group. Therefore, when option is on offer all consumers will choose insurance package . The problem is that insurers will now make a loss and therefore can not offer option . Therefore, the insurer needs to separate the high risk group from the low risk group. The only way how this can be done when there is no information about risk status is by providing an insurance package to the low risk group which is non beneficial to the high risk group compared to the original choice by the high risk group. Therefore the insurer will offer option to the low risk group. So the two insurance policies on offer are now and . From this example it is easy to see that the high risk group is neither worse nor better of due to information assymetry, however, the low risk group would have been better of in a market with symetric information, in which they would have been able to take on a more preferential insurance policy.

However, sometimes an equilibrium can be found to the left of in for example . Whether this is profitable for the insurer depends on how the high and the low risk group relate to each other with respect to size and risk. If the high risk group is relativelly small and does not pose extreme risks, the premiums for the low risk group might be able to compensate for the loss an insurer makes on the high riks policies. An equillibrium can thus be located between the optimal choice for the low risk group, the individual choice of the high risk group and the worst outcome for the low risk group. So the equilibrium will in graph 3 be located above and and below . The larger and riskier the high risk group is

(10)

9 compared to the low risk group the further we move to the separating equilibrium characterised by option / .

Pauly (1974) also used a similar model and came to the conclusion that the competitive equilibrium in insurance markets can be non-optimal. He postulates that a possible solution for overinsurance due to asymmetric information could be state intervention. However, it is crucial to realize that a state intervention will only work if it makes information available. The principle idea is that insurers should get relevant information about the risk parameters of individual potential insurance policy holders so they are able to make a better risk assessment about potential policy holders. A possible state

intervention is to mandate compulsory health insurance with a variable voluntary extended insurance. Suppose again there are two groups of consumers, a risky and a less risky. In this system the system of compulsory health insurance with a possible extension would be a pareto improvement over the competive equilibrium, provided the insurance would match the preferences of the low risk group. In this situation, as low risk consumers already have their optimal insurance quantity, they would not gain from taking additional insurance. Therefore, all people that take on aditional insurance are logically high risk consumers. High risk consumers, therefore, gain without predating on low risk consumers and low risk consumers also are more optimally insured. Therefore a system with compulsory insurance with supplementary insurance is a step towards a more optimal equilibrium compared to the competitive market equilibrium.

II.ii Empirical Evidence

The last papers have described the problem of asymmetric information and adverse selection

theoretically. However, it is of interest for us to know how relevant the problem of adverse selection is in practice. Therefore I will focus on some empirical papers that estimate whether there is adverse selection. A paper which is closelly related to the situation described above by Rothschild and Stiglitz is that of Cutler and Zeckhauser (1998). In this paper an in depth review of the insurance policies offered by Harvard University and the Group Insurance Commission (GIC) of Massachusetts is presented. In both situations multiple insurance policies are provided. Individuals choose an insurance plan based on cost-benefit calculation. People that expect to need medical care frequently would therefore opt for a more extensive insurance package. However general preferences also influence which health package will be chosen. For example, if people need to switch to a new general practitioner when they switch

insurances they might be less willing to opt for a different insurance. On the other hand, when preferences are very important it is therefore not likely that adverse selection will occur. When

preferences become less important cost benefit analyses is more likely to determine which package will be chosen and adverse selection becomes more likely. To determine whether there is adverse selection in the insurance package offered by the Harvard University. Cutler and Zeckhauser assessed the health-status of employees who dis-enrolled from the extensive package. They came to the conclusion that these employees where healthier and younger than employees which stayed in the insurance. In 1996 employees who dis-enrolled were 20% healthier than the average employee in the year before they left. Due to this adverse selection the high quality policy became untenable and was cancelled in 1997. With respect to the GIC a similar story can be told. When comparing the standard health care plan with a generous health care plan; in the standard plan about 25% of the population is over 45 year old;

(11)

10 however, in the more extensive plan 50% of the population is 45 years or older. Moreover they found that there are relatively more pregnancies and hearth attacks in the generous package than in the standard package. Furthermore, they observed that people that move from the generous health package to the basic health package spend on average less on medical expenses than people that remain in the extensive health package, whilst, people that move from the normal plan to the extensive plan do not spend more on medical expenses than the average policy holder in the extensive plan. Therefore they concluded that adverse selection is prevalent in the GIC and is caused by the drop out of relatively healthy individuals from the extensive package. Moreover, from both examples we must conclude that adverse selection is not just a theoretical argument but in some circumstances proves to be problematic in practice.

Cutler and Zeckhauser focused mainly on adverse selection. However, adverse selection is difficult to distinguish from moral hazard. Moral hazard, in the context of health insurance, implies that people change their behavior once they are insured and life a more risky life than those that are uninsured. The effect of moral hazard and adverse selection is the same. Both result in a higher risk profile for the average policy holder. This makes it difficult to separate adverse selection and moral hazard empirically. Abbering et al (Abbring, Chiappori, Heckman, & Pinquet, 2003) evaluate the possibilities to separate moral hazard and adverse selection in the context of automobile insurance that feature a bonus-malus scheme. They claim that dynamic aspects of risk and insurance choice can help to distinguish between adverse selection and moral hazard. The reason for this is that individuals do not know their accident probability when they take their first insurance; it takes time for consumers to learn about their own accident probability. Adverse selection can thus only happen over time when individuals have learned about their own accident probability. Adverse selection can therefore be observed through the changes in insurance a consumer decides upon: There is adverse selection when people that have risky behavior over time move towards more extensive insurance coverage. On the other hand, moral hazard can be identified by looking at the rate of accidents of consumers over time. When there is a bonus-malus scheme, consumers will be more risk averse when the premium is high than when the premium is low. So over time as the premium decreases a person becomes less careful but he or she will be more careful just after an accident, when the premium just increased. When this pattern can be identified there is moral hazard. There are, however, more ways to identify adverse selection separately from moral hazard.

Cardon and Hendel (2001) estimate a two stage empirical model of health insurance purchases. They use data from the National Medical Expenditure Survey (NMES). This survey includes data on insurance contracts offered to each individual, ex post claims by the insured and transfers. With the help of this database they estimate whether insurance take up can be explained by observables. They reason that if insurance choice cannot be explained by observable variables such as age and income there must be un-observable variables that explain insurance choice. A key assumption for the previous to hold is that the type of employment is unrelated to health-status. Theoretically it could be possible that people with bad health enroll in jobs which offer good health insurance because of their health-status. They find that after controlling for income and job characteristics, as proxies for skills, self-selection into jobs that offer insurance is very limited. Moreover, Cardon and Hendel find that observables explain most of the

(12)

11 insurance take up and therefore conclude that asymmetric information has no or possibly a very small effect on insurance take up. However, it is most certainly not a key factor in the insurance decision. Therefore adverse selection does not appear to be an issue.

In Siegelman (2004) some interesting reasons for the possible non-existence of adverse selection are presented. He states that neither theoretical nor empirical studies provide convincing support for the practical existence of adverse selection in insurance markets. One of the reasons why adverse selection might not be a practical concern is that there is simply no information asymmetry in insurance markets. In the classical adverse selection papers it is assumed that insurants have more information about themselves than insurers and can therefore make a better assessment of their own risk probability. Even though it is indeed true that mostly insurants have more information to base their risk assessment on than insurers they cannot always use this information efficiently. One of the reasons for this is that risk assessment is inherently difficult and insurers might be able to evaluate the given information more properly with the help of actuarial techniques. Therefore, it might be that even when there is asymmetric information there is no asymmetry between insurers and insurants in risk assessment. A different reason is that individuals simply do not have the right information to assess their own risk probability and neither do the insurers. This happens when risk is very random. When this happens there is naturally no information asymmetry.

Siegelman’s arguments for the limited support of adverse selection are based on the ability to assess risk for consumers and insurance providers. His arguments are thus connected to information asymmetry. There are, however, other possible explanations for the weak empirical support of adverse selection. A primary assumption made in the classical model as described by Rothschild and Stiglitz (1976) is that all people have the same risk preferences. Meza and Webb (2001) depart from this assumption and

assume that people have different risk preferences. The model they develop assumes that there are two risk categories one with high risk and one with low risk. The key assumption is that people that are very risk averse both engage in precautionary activities more often and are inclined to insure themselves. Moreover, people that take precautionary actions against potential damage are likely to have a smaller risk to incure damage. On the other hand, people that are less risk averse are less likely to take

precautionary actions and therefore are expected to have a higher risk probability. Grönqvist (2004), empirically supports this idea in a dental insurance natural experiment. In this experiment people are devided into risk categories based on their oral health. These people are unexpectedly offered dental insurance. Whithin the risk categories there is still scope for differences in risk parameters, therefore Grönqvist uses historical oral health records to evaluate whether people that have bad historical health records compared to other people in the same risk category are more or less inclined to purchase dental insurance. He finds that in low risk categories there is advantageous selection wherease in high risk categories there is adverse selection. This is in line with Meza and Webb (2001) which argue that peope in low risk categories are often of the cautious type and therefore more inclined to take on insurance, wherease, people in the high risk categories are less cautious so they only take insurance when the net return of the insurance is higher then the net cost. Since there are multiple groups that either feature advantageous or adverse selection these two effects can cancel each other out in the agregate.

(13)

12 Godfried, Oosterbeek and van Tulder (2001), however, study a similar case but find adverse selection. They focused on the introduction of supplementary dental insurance in the Netherlands. The largest difference between the type of dental insurance of Grönqvist and Godfried et al, is the difference in premium structure. In Grönqvist premia are diversified. People were beeing offered dental insurance based on their oral health risk. However, in Godfried et al premiums were uniform and all people had the opportunitiy to take on insurance on the same terms. Therefore Grönqvist had to focus on adverse selection within risk categories wherease Godfried et al could focus on the whole sample. The situation in Godfried et all is therefore more closelly related to the classical situation as depicted in (Rothschild & Stiglitz, 1976), wherease, Grönqvist focusses on a more practical situation where insurers have some ,albeit, limited information on the risk status of customers. Godfried et al thus prove that adverse selection is a real threat in a situation where insurers have no information about the risk parameters of customers or are not able to use this information efficiently. Wherease, Grönqvist shows that even when insurers have information about the riskiness of customers there might be adverse selection, this adverse selection is however, at least partly countered by advantageous selection in the low risk categories.

II.iii Conclusion

With respect to adverse selection in health insurance markets most theoretical models conclude that adverse selection is likely to occur when there is information asymmetry between the insurer and potential policy holders and there are differences in the risk parameters of insurance policy holders. However, the empirical evidence on this issue is less conclusive. Based on his own empirical research Siegelman (2004) even goes so far as to state that information asymmetry does not characterize insurance markets in the real world, and therefore adverse selection does not occur either. This claim seems to be too bold because other research indicates that markets are often characterized by adverse selection caused by information asymmetry. It is possible to separate these empirical papers into two groups. The first group studies the possibility of adverse selection when there is strong information asymmetry, whereas the second group focuses on the possibility of adverse selection in a system with only limited information asymmetry.

Within the first group of papers, which study situations that feature strong information assymetry, empirical evidence is inconclusive on the existence of adverse selection. Both Cutler and Zeckhauser (1998) and Godfried, Oosterbeek and van Tulder (2001) study markets which are characterised by information assymetry and find adverse selection in their studies. On the other hand, the research of Cardon and Hendel (2001), who focus on a system in which information assymetry is very likely to characterize the market, does not show adverse selection to occur. Research on markets with limited information asymmetry suggests that if insurers have only limited information about their potential customers this may already lower the possibility of adverse selection. This is evidenced by Grönqvist (2004), who studies a situation where there is only limited information assymetry and does not find adverse selection in his full sample.

The suggestion of Meza and Webb (2001) that people have different risk profiles which simultaneously leads to adverse selection among people with high risk profiles and leads to advantageous selection in low risk profiles, may explain the unexpected results of Cardon and Hendel (2001). Advantageous

(14)

13 selection and adverse selection can in theory cancel each other out. This hypothesis is supported by Gronqvist, who finds adverse selection within high risk categories and advantageous selection in low risk categories. If the resuls of Meza and webb and Gronqvist are thus generalizable to other insurance markets, this may explain why, often, there is no sign of adverse selection despite the presence of information assymetry. In conclusion: whether there is adverse selection seems to depend on the degree of information assymetry and whether individuals differ in their risk preferences.

When the two assumptions of full information assymetry and equal risk aversion of consumers from Rothschild and Stiglitz (1976) are relaxed adverse selection is thus less likely. However, within the Bosnian health insurance market, the assumption on information asymmetry is likely to hold true. Therefore, we suspect that adverse selection is likely to occur. The reason for this is that there is no option to exclude people from health insurance, therefore any information the insurer has on the customer cannot be used efficiently. This leads to a situation similar to that of actual information asymmetry: the potential policy holder knows his risk status and can act accordingly, whereas the insurer cannot. However, if Meza and Webb’s (2001) proposition that customers vary in their risk preferences is true, there is reason to suppose that no adverse selection will be found within the total sample, even if it is prevalent in some sub samples. To test this would be out of the scope of this thesis. The empirical evidence on adverse selection is mixed. Cohen and Siegelman (2009) acknowledge this and suggest that insurance schemes should be assessed individually and conclusions with respect to adverse selection should not be based on generalized models. The purpose of this thesis is therefore to answer whether there is adverse selection into health insurance in Bosnia and Herzegovina, it is not to find a definitive answer to the question whether there is adverse selection in insurance markets in general. To be able to answer whether adverse selection occurs in Bosnia and Herzegovina it is

important to understand the role of health and health insurance institutions. Therefore the next chapter focuses on the institutional framework of the Bosnian and Herzegovinian health insurance market.

(15)

14

III The Bosnian health system

To be able to build and interpret a reliable model that estimates the relationship between health and insurance status. It is important to understand the structure of the Bosnian health and insurance system. Therefore in this section, first; the general healthcare system in Bosnia and Herzegovina will be reviewed and second; I will focus on the Bosnian and Herzegovinian health insurance market.

The state Bosnia and Herzegovina comprises of two large entities and one independently administered region. The two entities are the Republic Srpska1 (from now on also referred to as the Republic) and the Federation of Bosnia and Herzegovina (from now on also referred to as the Federation). The

Independently administrated region is called Brčko and is much smaller than the entities and left out of the discussion in this section. However, the healthcare system of Brčko can be seen as a scaled down version of the healthcare system in the republic Srpska. Furthermore, the data from the district Brčko is included in the models in the following model section. Both entities are independently responsible for the health care system within their territory (Cain, Duran, Fortis, & Jakubowski, 2002, p. 17). In the republic Srpska health care planning is centrally led by the ministry of health and social welfare in Banja Luka (Cain, Duran, Fortis, & Jakubowski, 2002, p. 18), whereas, the Federation is divided into 10 cantons which are responsible for their own healthcare system. Central coordination of the cantonal health ministries is the responsibility of the central Ministry of Health of the Federation of Bosnia and

Herzegovina located in Sarajevo. In figure 4 below the organizational structure of the healthcare system in the Federation of Bosnia and Herzegovina has been depicted. Figure 5 presents the organizational structure of the healthcare system in the Republic Srpska.

1

The literal translation of the Republic Srpska would be the Republic Serbia. However, this translated name could cause confusion with the neighboring country the Republic of Serbia. In this document the Republic Srpska will therefore not be translated.

(16)

15

Figure 4: Organization chart of the healthcare system in the Federation of Bosnia and Herzegovina. The left hand column represents the health provision system at the cantonal level. The cantonal ministries of health are responsible for the provission of healthcare as well as for the cantonal health insurance fund and the institute of

public health. In the right hand column, on the other hand, we see the health care and insurance system at the federal level. The federal ministry of health is not directly responsible for health care provision but cooporates with

the cantons. However, the federation is responsible for the federal health insurance fund and public health institutions. However, the main observation should be that the cantons have great autonomy with respect to healthcare and insurance policy. Reprinted from Health care systems in transition: Bosnia and Herzegovina. J.

(17)

16

Figure 5: Organization chart of the healthcare system in the Republic Srpska. In the Republic Srpska health care provision and insurance is centralized. The Ministry of Health and Social Welfare is directly responsible for healthcare provision and insurance. Reprinted from Health care systems in transition: Bosnia and Herzegovina. J.

Cain,A. Duran,A. Fortis, E. Jakubowski,2002, European Observatory on Health Care Systems, 4(7),p.21.

III.i The Bosnian & Herzegovinian health insurance market

The Bosnian health insurance market is neither a fully enforced compulsory health insurance system nor a fully flexible voluntary insurance system. The result is that many people are not covered by health insurance despite that insurance is obligatory by law. From a legislative point of view Bosnia and Herzegovina has a compulsory health insurance system (United Nations High Commissioner for

Refugees, 2001, p. 9). However, in practice a significant share of the population is uninsured. The table below gives insight into what percentage of the sample was insured between 2001 and 2004.

Insurance 2001 2002 2003 2004

Yes 78.10 78.61 81.18 85.85

No 21.90 21.39 18.82 14.15

Table 1: Percentage of sample with health insurance.

The data in this table make it clear that a significant share of the population is uninsured. Therefore, we must conclude that the compulsory health insurance scheme does not result in a fully insured

(18)

17 population. Therefore, the question arises why some people do not have health insurance and whether this can be a sign of adverse selection?

For adverse selection to occur a requirement is that insurance is costly. If this is not the case everyone would get insurance. Therefore an important question we should ask ourselves is how people pay for their health insurance. The health insurance schemes are mainly funded by deductions from salaries, benefits and pensions. Working people that have health insurance thus pay their fees through their employer, whereas, pensioners pay insurance premiums through a deduction from their pension benefits.

Within the Federation every canton can set its own tariff for health insurance. However the maximum contribution rate from net salaries is 18%. This contribution rate consists of 13% paid by the employee and 5% by the employer (Cain, Duran, Fortis, & Jakubowski, 2002, p. 41).

The system is more complicated in the republic. The table below provides an overview of the contributions by inhabitants in the republic (Cain, Duran, Fortis, & Jakubowski, 2002, p. 42).

Sector of work Health insurance contribution

Employees 15% of net wage=7.5% by employer+7,5% by employee

Self-employed 15% of net wage Pensioners 4% of net pension

Unemployed Paid by unemployment fund

Farmers 16% of estimated property tax with a minimum of KM20 per month

Table 2: Health insurance contributions in the Republic Srpska.

From this table it is clear that not everyone in the republic is taxed at the same level.

Furthermore, health insurance premiums do not always buy the same quality of health treatment. People in The republic Srpska and The Federation of Bosnia and Herzegovina do not have the same coverage of health expenses. Moreover even within the federation there can be differences in health insurance coverage between the cantons (Zukic, 2010, p. 2). Therefore the value of health insurance can be different from person to person. Furthermore co-payments are regularly required to be able to access health care in the Republic and the Federation (Cain, Duran, Fortis, & Jakubowski, 2002, p. 44), this decreases the value of an insurance policy and if co-payments differ between different locations, insurance policies will differ in value between different locations.

Based on this short review of the health care institutions in both entities we can answer the question whether there is scope for adverse selection into health insurance in Bosnia and Herzegovina. There is health insurance available in all regions and there are no barriers to entrance on the basis of health status. Moreover, having insurance is obligatory by law. Despite this obligation by law, a significant share of the population does not have health insurance (see table 1). Because regulations regarding health insurance are not fully enforced, people can in practice still choose not to be insured. They might opt to do so because health insurance is costly to the consumer: up to 15% of net wages may be

(19)

18 regulations are not strictly enforced means that adverse selection may occur in the Bosnian and

Herzegovinian health insurance system despite the fact that health insurance is in theory compulsory. To evaluate whether adverse selection actually occurs in reality, the relationship between health status and insurance status will be estimated in chapter 6 and 7. This relationship may be different within each entity due to the differences between the healthcare sectors in the Republic and the Federation. Therefore, in chapter 6 and 7 a model will be developed which can estimate whether the effects of health status on insurance status differs between the entities or different subgroups of their respective populations. However, before looking more closely at the relationship between health and insurance status it is important to understand how health status is measured exactly and what the pros and cons are of different proxy’s for measuring the health status of an individual. This will be discussed in the next chapter.

(20)

19

IV Self-rated health and true health-status.

To empirically verify whether there is adverse selection into health insurance, the focus of this thesis is on the relationship between the health status of individuals and insurance take up. There are, however, multiple options to define and measure health. In the living in Bosnia and Herzegovina survey (LIBIH), self-assessed health status is available as a proxy for true health status; however, no direct clinical health measurements are available. Therefore, it is of interest to evaluate whether self-assessed health status is a feasible proxy for true health status.

There is no obvious way to measure health-status. Most often it is considered impractical or too expensive to examine all subjects by medical specialists. Therefore, an expert opinion on the health-status of the subjects is often missing. Furthermore, even if a survey includes a medical examination, this examination is also not likely to be a perfect proxy for true health status. Therefore, surveys generally include a measure of self-rated health-status. This measure tells us something about how people perceive their own health. In the case of the LIBIH, such a self-rated health variable is available as well; this self-rated health-status provides us with a health-status ranging from very poor to excellent for each respondent. The LIBIH thus provides a good overview of the perceived health-status of Bosnian-Herzegovinian citizens relative to their peers. However: is self-rated health status a feasible proxy for true health status?

One way to evaluate self-rated health status as a proxy for true health is by looking at the relationship between self-assessed health-status and the risk of needing medical assistance. If there is a strong relationship, self-assessed health-status should have a strong correlation with health-status as being examined by a physician. It is likely that a medical health evaluation gives a better prediction of future medical expenditures than self-assessment. There are many reasons why an individual could grade his own health differently than a physician. One important aspect which influences self-assessed health-status is mental health (Singh-Manoux, Martikainen, Jane, Zins, Marmot, & Goldberg, 2006). If a person is very happy, or for example in love he is likely to be more positive about his health-status compared to a person who just lost his spouse. Therefore, medical health assessments which emphasize medical conditions that can be treated and are not influenced by the mental well-being of a subject are likely to be a better predictor of health expenses than self-assessments. Moreover, self-assessed health might not give homogeneous health assessments across individuals (Myers, 1982). Individuals that are objectively equally healthy might grade their health in a different way (Bound, 1991). Furthermore, health status is connected with labor market outcomes. Individuals might say they have poor health to justify why they are out of work (Bound, 1991). For the above mentioned reasons medical health assessments might be more valuable than self-assessed health status. However in the LIBIH survey there is no data on medical health-assessments. Therefore the focus will be on self-assessed health-status. There are, however, other variables that might give a more objective measurement of health status available within the LIBIH survey data. Objective health measurements are objectively measurable quantitative variables that have a relationship with true health status. An example of this kind of

objectively measurable variables, available within the LIBIH survey dataset, is the number of doctor visits of a person. However, with objective health measures there is the danger of not including circumstances

(21)

20 which are relevant to the true health status of an individual other than the objective measure in

question (Bound, 1991). This also is the problem with the number of doctor visits. If only the number of doctor visits are used to judge health status, there can be large variations in actual health of people that have equal numbers of visits. For example, a person with a chronic illness might visit a doctor as often as a person with hypochondria. However, the true health status of the chronically ill person is probably worse than that of the person with hypochondria who is in truth a healthy person (aside from his or her mental condition). If doctor visits would be used as the sole proxy for true health status, these people would seem equally healthy, despite the difference in true health. Aside from these considerations, self-assessed health-status is the preferred proxy for true health over doctor visits because healthy people are not likely to visit doctors. If doctor visits are used in the models in chapter 6 and 7 this implies that healthy people are left out of the model. As discussed above, the idea of adverse selection is that healthy people do not take on insurance whereas the unhealthy are insured. Leaving out healthy people from the models is therefore going to bias the results severely. Therefore it is better to use self-assessed health as this reflects the true health for especially healthy to moderately healthy individuals more accurately than number of doctor visits does.

Another reason why self-rated health is preferred is because there is a moral hazard problem. People that are insured are likely to over-consume healthcare because the price they have to pay for healthcare is lower than the true price of healthcare which has to be paid by uninsured people. If this is true it becomes difficult or impossible to compare health status of insured and uninsured people, when health status is based on the number of doctor visits. Thus self-rated health status is preferred as a proxy for true health status over doctor visits.

One of the first to analyze the relationship between self-assessed health and medically assessed health status were Suchman et all (1958). They analyzed the validity of health questionnaires. One of their targets was to estimate the validity of self-reported health as a proxy for health-status as determined by a medical professional. Their target group was, however, not representative of the whole society as they used panel data representing retired workers of 65 years and older. In their set up they found significant correlation between self-assessed health and medically assessed health. However there are also large discrepancies. They report that one out of 5 people which rated their own health as unfavorable received a favorable health-status by the physician. On the other side of the spectrum they found that one out of three people reported their health as favorable whereas they received an unfavorable health rating by the physician. However, they do find a significant relationship between self-ratings and

medical ratings at the 0.01 significance level. Also with respect to various specific medical conditions self-ratings and medical ratings differ. Therefore they conclude that self-ratings represent “perceived” health rather than objective health.

They, however, also analyze the change in objective and subjective health over time. They hypothesize that people whose health has improved between 1952 and 1954, measured on the basis of a physicians health assessment, report to worry less about their health. On the other hand people who perceive a decrease in their health are likely to worry more about their health. This hypothesis is supported is supported by the results of their analysis. Also other factors representing self-health-ratings were in line with the change in “objective” health. So on the basis of Suchman et all it is justifiable to assume that

(22)

21 when a person reports a more unfavorable self-rated health-status than previously, his or her health did indeed deteriorate. In general the same conclusion has been reached by Friedsam and Martin (1963), although, their data represented a slightly younger population aged 50 years and older. One of the classic works in this field is by Mossey and Shapiro (1982). Their data source was the Manitoba longitudinal study on aging. This data source made it possible to estimate the validity of self-rated health-status with respect to mortality while controlling for the effect of the objective health-status. Mossey and Shapiro found that self-health-ratings were a significant predictor of mortality with people claiming to be unhealthy having a higher chance to pass away than healthy people. More recently Burström and Fredlund (2001) estimated the relationship between self-rated health and mortality among adults in Sweden. New in their approach was that they distinguished between socio-economic groups. They found that there was a significant relationship between self-rated health and mortality within all socio-economic groups. However, the effect was stronger for higher socio-economic groups. A more resent paper that studied the relationship between self-rated health and physical, social and mental functioning is (Navaddat, Kinmonth, Sanderson, Surtees, Bingham, & Khaq, 2011). They conclude that physical functioning is more strongly related to self-rated health than mental health and social functioning. The conclusion of the above mentioned research is that self-reported health-status is indeed a proper indicator of true health. Therefore it is reasonable to select self-reported-health-status as a proxy for true health-status from the LIBIH survey data. This variable and other relevant data from the LIBIH survey will be discussed in more detail in the next chapter.

(23)

22

V

Data

As mentioned in the introduction I make use of the “Living in Bosnia and Herzegovina” survey (Living in Bosnia i Herzegovina Survey, 2001-2004). The first wave of the survey was launched in 2001 and has been followed by 3 consecutive waves in 2002, 2003 and 2004. The survey is set up as a panel data survey and focuses both on households and individuals. For this thesis the variables of primary

importance are self-assessed health and insurance status. Self-assessed health is available from wave 2 onwards. It is an ordinal variable that runs from 1 to 5 with higher values representing better health. The labels corresponding to the values of self-rated health status are provided in the appendix in table A1. In total there are 18539 observations for self-rated health. Furthermore there, are 13 observations that did not provide an indication on their health status. These observations have been dropped from the

dataset; since this is a very small number this is not likely to be problematic. In the table below the frequency of health statuses has been depicted. The variable follows a relatively bell shaped pattern. This is a loose indication of a normal distribution of people over the various health statuses. However, there are relatively many people with excellent health. The average health status is therefore slightly higher than the median health status.

Figure 6: Frequency of health statuses.

Another variable for health status is discrete health status which is an adaptation of the previously discussed self-rated health status variable. To test whether the results of the following models are robust to changes in the variable for health, I construct a dummy variable for health which can be 0 or 1. When this variable is 0 it represents very poor or poor health, when the variable is 1 it represents fair to excellent health.

Insurance status is also an important variable since it will be the dependent variable in the following models. Insurance status is a dummy variable that can be either 0 or 1, where 0 implies that a subject does not have health insurance, whereas 1 implies that a subject is insured. In the following figure the

0 1000 2000 3000 4000 5000 6000 7000 8000

Very poor Poor Fair Good Excelent

F r e q u e n c y

(24)

23 development of the average insurance status in the years 2001 to 2004 has been depicted.

Figure 7: Percentage of the population with health insurance by year.

It should be noticed that there is a clear trend in insurance status. In the later years a larger share of the population has health insurance than in the earlier years. For this reason there is a need to include year dummies in the primary regression models in the later sections.

The table below presents the summary statistics for the core variables of the models to be developed. The variable names are presented in the first column. In the second column the shorter labels for the variables are presented. These will be used in the equations and tables in the following sections.

Variable

Label in

equations Observations Mean

Std.

Dev. Minimum Maximum

Insurance status Y 27869 0,80 0,40 0 1 Self-rated health status H 18539 3,32 1,15 1 5 Discrete health status Ĥ 18539 0,76 0,43 0 1 Age Age 29548 42,27 19,68 0,01 97,09 Sex Sex 31413 0,52 0,50 0 1 Income* I 34937 211,45 521,59 0 16800 Education** Edu 28660 2,57 1,61 0 6 Year Year 4 X X 1 4

Table 3: Summary statistics.

*Income is an aggregated variable. The variables that contribute to monthly income are presented in table A2 in the appendix. ** A higher education value for education level indicates a higher education. The education levels

corresponding to the values of the variable for education are presented in table A3 in the appendix. 75,00 77,00 79,00 81,00 83,00 85,00 87,00 89,00 2001 2002 2003 2004 Per ce n tage wi th h e al th in su ran ce Year

(25)

24 The structure of Insurance status, Health status and Discrete health status have already been explained. However, its means need some explanation. The mean of insurance status is 0.80 this implies that 80% of the population has health insurance. For health status the mean is 3.32, this implies that on average people report a fair to good health status. The same applies to discrete health status. However, the mean of 0.76 implies that 76% of the sample has a fair to excellent health. The other variables that will return in the models to test for adverse selection are Age, Sex, Income and Education. Age was a given variable in the dataset. However, for some people it was missing, therefore a new variable for age has been created which is the difference in years between the date of birth and the date of interview. If the date of birth or the date of the interview was missing, the original given value for age has been used. Sex is simply the variable that denotes whether a person is male or female. If Sex is 0 the person a female and if sex is 1 the person is a male. The mean of sex is 0.52 this thus implies that 52% of the sample consists of males. Income is the total monthly income of a person. In the appendix in table A2 the different sources of income that constitute the variable for total income have been mentioned. Education is a 7 point ordinal scale where a higher number represents higher education. In table A3 in the appendix, the different education levels have been depicted. The average education level is 2.57 this implies that on average people have an education level between secondary compulsory schooling and secondary technical schooling. Finally year is a dummy variable representing the years in the survey. Here year 1 represents 2001, whereas, year 4 represents 2004. The average and standard deviation for year have not been presented since these have no useful meaning.

V.i Attrition

Attrition is a common problem in panel datasets. The problem with attrition is that it might induce attrition bias. Also in the LIBIH panel survey attrition is an issue. Between 2001 and 2004 the number of participants in the survey that answered the question whether they have health insurance decreased from 9337 to 4859. However, to evaluate whether there is a potential for attrition bias we have to compare people that remained in the sample for all years with those that dropped out. One potential source of attrition bias is that unhealthy people might be more likely to pass away than healthy people. This would result in a non-representative panel with respect to health status. Another potential source for bias might be that people with and without insurance do not have the same willingness to

participate in the survey and do not have the same probability to drop out of the survey. For example, it might be that some share of the uninsured are not insured because of their unwillingness to conform themselves to the norms in society. These people might also be more likely to be unwilling to participate in all waves of the survey. Therefore, the primary objective is to evaluate whether people that remain insured are equally healthy compared to those that drop out of the survey and analyze whether, people that drop out of the survey are equally likely to be insured compared to those that completed the survey. To do this the key statistics of survey participants that completed all 4 years are compared with those that dropped out of the survey. This will be done on the basis of the answers they provided in the second year of the survey because one of the primary variables: subjective health is only available from year 2 onwards.

(26)

25 The table below indicates whether there is a statistical difference between the key variables for people that remain in the survey from year 2 to year 4 which are called stayers and people that leave the survey after 1 or 2 years which are called leavers.

Variable Mean(0) (Leavers) Mean(1) (Stayers) Mean(0)-Mean(1) Two sided p-value

Insurance status 0,797 0,785 0,011 0,446 Health status 3,064 3,296 -0,232 0,000 Discrete health status 0,659 0,756 -0,097 0,000 Age 48,872 43,073 5,800 0,000 Sex 0,444 0,524 -0,079 0,001 Income 221,838 208,442 13,396 0,366 Edu 2,854 2,542 0,312 0,000

Table 4: Statistical difference between the averages of key variables for leavers and stayers in year 2. Leavers are those that do not complete all years of the survey, stayers are those that completed all years of the survey between year 2 and 4. Two sided p-values for the difference in means between stayers and leavers have been depicted in the last column. A p-value below 0.05, which is presented in green, implies that there is a statistical

difference in the average of the variable for leavers and stayers.

The first thing to notice is that there are some variables that have different averages for leavers and stayers. Most noticeably there is a difference between health status of leavers and stayers. People that remain in the sample from year 2 onwards are, at the 5% significance level, statistically healthier than those that drop-out of the sample. Moreover, leavers are on average almost 6 years older than stayers. As mentioned before, this is something we expect as older people with worse health are more likely to pass away and are therefore unable to participate in the following years of the survey. A further observation is that females are overrepresented in the group of leavers. It is hard to find a specific reason for this. Moreover, people that drop-out of the survey are on average higher educated than those that remain in the survey. The above mentioned differences might make the results of this thesis difficult to generalize to the overall population as the survey might become non representative for the entire population over time. Moreover, due to the drop out of relatively unhealthy individuals, the average health status in the sample may increase over time. This increase in average health status makes it difficult to evaluate whether there is an actual increase in health status in the whole

population. Therefore, there is likely to be attrition bias. This limits the generalizability of the models that are developed in the next chapter; the results may not be applicable to the whole population in Bosnia and Herzegovina.

(27)

26

VI Models for the identification of adverse selection

To evaluate whether there is adverse selection into health insurance in Bosnia and Herzegovina the relationship between self-rated health and insurance status will be estimated. To do so, initially linear probability models will be developed. However, in a linear probability model, omitted variables are likely to introduce a bias to the results. To mitigate this omitted variable bias, the models will be adapted to include a fixed effect. The results of the fixed effect models might however, not be applicable to the whole sample due to the, before mentioned, differences between the healthcare system in the Republic and the Federation. Therefore, the fixed effect model will be specified for each entity separately. Furthermore, the model will also be adapted to allow a focus on the specific subsamples: Serbians living in the Federation, Bosnians living in the Republic, farmers and public sector workers. Before any models are discussed in detail, some basic statistics about the relationship between health and insurance status are reviewed.

Adverse selection essentially implies that the unhealthy become insured and the healthy remain uninsured. It is therefore interesting to know whether on average the insured are unhealthier than the uninsured. Moreover, it might be interesting to know whether people that get insurance have different health characteristics compared to the total sample. The same also applies to those that opt-out of insurance. These basic statistics are presented in the table below.

Average self-rated health status

Two sided p-value for

difference with average health status of total sample.

Insured people 3.289 0.0044

Uninsured people 3.269 0.0553

Newly insured people 3.425 0.0001

Newly uninsured people. 3.453 0.0000

Total sample 3.318 1.0000

Table 5: Average subjective health-status by insurance status. Subjective health status is given on a 5 point scale where a higher number implies better health. The third column presents the p value for the difference between the average subjective health of a specific subgroup and the average subjective health of the total sample. In the

table Insured people are people that were also insured in the previous year. Uninsured people are people that were uninsured in the last year and remained uninsured. Newly insured people are those that did not have insurance in the previous year but do in the current year. And newly uninsured people are those that dropped out of insurance between last year and the current year. If a p value is below 0.05, it is presented in green and implies

that there is a statistical difference between the given sub sample average and the total sample average. The averages are the averages over all 4 years. The total sample average thus represents the average health status for

the full sample over all 4 years.

The third column presents the results from a student’s t-test of statistical difference between the within group averages and the total sample average. When entrants into health insurance are unhealthier than people that are already insured, this is a very rough indication of adverse selection. On the other hand when they are healthier than the average insured person this suggests that there might be

advantageous selection. The opposite applies to those that drop-out of insurance. If drop-outs are healthier than the average insured person this is a loose suggestion of adverse selection whereas, if they

(28)

27 are unhealthier this suggests that advantageous selection is more likely. Both forces can either move in the same direction or oppose each other. The table above presents only basic averages and we do not control for relevant covariates. Nevertheless, the results from the table can provide some first insight in the relationship between self-rated health status and insurance status.

Table 5 presents average self-rated health statuses for different subsamples. The statistical difference, of the average self-rated health status from a given subsample, with the total sample average is given by a p value where a p value smaller than 0.05 indicates that a within group average is statistically different from the total sample average. These values have been marked green. There is statistical evidence at the 5% level, that there is a difference between the health-status of insured people and the total sample. Insured people are on average a little less healthy than the overall sample. On the other hand there is no evidence that uninsured people are less healthy than the full sample whereas, they seem to be less healthy than insured people. This might be caused by the smaller number or uninsured people which causes the difference between het average health of uninsured people and the full sample to be insignificant. Furthermore, people that are new entrants into health insurance after previously not having been insured are relatively healthier than the overall sample and those that have insurance. The same applies to those that drop out of insurance: drop outs are healthier than those that are insured and the average person in the total sample. To say a little more about whether this may suggest adverse selection or not, we need to know whether newly insured people and drop outs of insurance statistically have different health statuses. Since the average subjective health of entrants into insurance and drop-outs are very close we do not expect them to be statistically different from each other. This is indeed true. With a p value of 0.461 there is no reason to suspect that entrants and drop-outs differ in health status. Therefore, when there are as many people dropping out of insurance as people that take insurance there is no effect on the average health status of those that are insured. Whether there is adverse selection thus depends on the ratio of people that drop out of insurance and those that get insurance. From the data section we know that the number of insured people increases over the years, if this trend continues and the average health status of those that drop out of insurance and those that take insurance remains the same, the average health-status of insured people would increase. This would suggest advantageous selection. However, these results are inconclusive because of the lack of control variables. Therefore, to determine whether there is adverse selection, or given the previous results even advantageous selection, more rigorous statistical methods are needed.

VI.i Linear probability models

The first approach to evaluate the possibility of adverse selection into health insurance is by estimating a naive linear probability model with limited control variables. This approach gives a first insight into whether adverse selection is likely. The primary question is whether there is a relationship between the independent variable, self-rated health-status and the dependent variable insurance-status which is abbreviated as Y. However the results of OLS regressions are very sensitive to omitted variable bias. To limit this, control variables need to be selected. The control variables to be selected should be those that possibly influence both the insurance decision and health. The primary variables to control for are age, sex and year. These variables will be explained below.

Adverse selection into health insurance in Bosnia and Herzegovina