• No results found

One-Sided Commitment in Dynamic Insurance Contracts: Evidence from Private Health Insurance in Germany

N/A
N/A
Protected

Academic year: 2021

Share "One-Sided Commitment in Dynamic Insurance Contracts: Evidence from Private Health Insurance in Germany"

Copied!
33
0
0

Loading.... (view fulltext now)

Full text

(1)

One-Sided Commitment in Dynamic Insurance Contracts:

Evidence from Private Health Insurance in Germany

October 8, 2010

Annette Hofmann

Mark Browne

University of Hamburg University of Wisconsin-Madison Institute for Risk and Insurance Gerald D. Stephens CPCU Chair Faculty of Economics and Social Sciences in Risk Management and Insurance

Business Administration Department Wisconsin School of Business Von-Melle-Park 5 975 University Avenue 20146 Hamburg (Germany) Madison, WI 53706-3123 (USA) hofmann@econ.uni-hamburg.de mbrowne@bus.wisc.edu

Abstract This paper contributes to the literature on one-sided commitment by studying long-term private health insurance in Germany. Policyholders do not commit to renewing their private health insurance contracts, but insurers commit to oering renewal at a premium rate that does not reect revealed future information about the insured risk. Our empirical results are consistent with theoretical predictions: front-loading in premiums generates a lock-in of consumers, and more front-loadlock-ing is generally associated with lower lapsation. Due to a lack of consumer commitment, dynamic information revelation about health status implies that high-risk policyholders are more likely to retain their private health insurance contracts than are low-risk types.

JEL Classication: G22·I11·I18

(2)

1 Introduction

In many European countries, health insurance is oered by government-sponsored social insur-ance funds. However, these funds are subject to mounting problems of nancing the welfare state, and do not oer much, if any, consumer choice. Leading policy-makers look for alter-native strategies for creating incentives to increase eciency. In some countries, such as the United States, health insurance is mostly oered by private companies and premiums are risk-based. As rst discussed by Arrow (1963), in this environment, individuals face premium or reclassication risk.1 Premium risk is the risk of an increase in the health insurance premium when the policyholder's health deteriorates. Given that in the United States, a large part of the population is insured via private health insurance, premium risk is a major problem since long-term health insurance is not available at a guaranteed price.2 This situation shows that freely competitive health insurance markets do not seem to work eciently either. In these markets, policy-makers are also often concerned about selection, equity, and access issues. Thus, creating an environment that can overcome ineciencies in health insurance markets appears to be a challenging task.

In a private health insurance market, it is far from easy to design long-term contracts that protect consumers from premium risk in the long run. This reclassication risk mainly emerges because contracts tend to be incomplete, i.e., they do not specify a price for every possible state of the world. To date, the insurance economics literature oers two solutions to this problem. On the one hand, Pauly, Kunreuther, and Hirth (1995) propose solving the premium risk problem via guaranteed renewable insurance contracts. Since policyholders initially prepay guaranteed renewable premiums to cover losses of everyone in the pool who (will) become high risk, the leaving of a low-risk policyholder has no impact of the insurer's prots while the leaving of a high-risk policyholder is protable to the insurer. The German private health insurance system in its actual form follows this idea. Its working principle illustrates that the technical design of such guaranteed renewable contracts is relatively sophisticated and may require signicant regulation. On the other hand, Cochrane (1995) argues that the premium risk problem could also be solved via separate premium insurance that pays an indemnity in the event an individual becomes a high risk. Yet, given separate premium insurance, resulting health insurance contracts will still tend to be incomplete.

The present paper addresses the issue of long-term insurance contracts in a world where risk types evolve over time. This issue is especially important in the health and life insurance industries. The rst best insurance policy in this world would fully insure against both the short-term risk of an accident or illness and long-term reclassication risk (premium risk). As shown by Cochrane (1995), insurer commitment to future policy terms is not necessarily sucient to ensure that reclassication risk will be fully insured. If consumers cannot commit, then those who receive better than average signals about their risk types, i.e. those consumers who turn out healthier, will have incentives to switch to a competitor who oers better rates.

1 In health economics, reclassication risk is often referred to as premium risk. See Arrow (1963), Pauly, Kunreuther, and Hirth (1995) and Cochrance (1995).

(3)

Due to this selection bias, insurers cannot protably oer consumers full insurance against reclassication risk.

Is it reasonable to regulate competition in health insurance markets in such a way that in-surers must commit to an insurance contract and reclassication risk is insured? This paper investigates whether such one-sided commitment contracts solve well-known issues in health insurance. We analyze private health insurance in Germany using a unique data set. Our study contributes to the literature on dynamic contract theory in three ways. First, it con-rms theoretical ndings on dynamic contracts and one-sided commitment. Second, it is, to our knowledge, the rst to study the German private health insurance market using individual-level data. Third, private health insurance coverage in Germany is signicantly dierent from that in other countries which makes Germany interesting for a study of one-sided commitment. The German market environment is unique in several aspects.

One of these is that Germany has a social health insurance (SHI) system as well as a private health insurance system (PHI) coexisting side by side. In the statutory SHI system, nonprot sickness funds collect premiums from their policyholders and pay health care providers accord-ing to negotiated agreements. Those not insured through these funds, mostly civil servants and the self-employed, have private insurance.3 SHI premiums are not risk-based but depend on an individual's annual labor income. Children and spouses without labor income are in-sured without surcharge. Premiums are shared between the inin-sured and his or her employer. Sickness funds are required to set a uniform premium for all their policyholders. The most important features of the German SHI system are community rating and open enrollment, i.e., sickness funds are required to accept any individual who applies without making a risk assessment. This is to ensure coverage for all risk types.4

Germany's PHI system is characterized by long-term contracts. Private health insurers com-mit to oering renewal at a premium rate that does not reect revealed future information about a policyholder's risk type. As a consequence, policyholders face no reclassication risk. Participation in the German PHI system is restricted to an upper income group as well as to the self-employed and civil servants.5 If annual income is below the social security ceiling,

3 Less than 0.2 percent of the German population has no health insurance of any kind. They are generally the rich who do not need it and the very poor, who receive health care through social assistance. 4 Government regulation is an alternative way to insure premium risk. When community rating is required,

health insurers must impose a somewhat uniform price for all individuals who enroll in their health insurance plans. Community rating is then necessarily associated with (a) open enrollment and (b) compulsory insurance. Open enrollment is necessary to avoid cherry picking by insurers. Compulsory insurance is necessary to ensure cross-subsidization between high and low risks. This is because if insurance were not compulsory, low risks would prefer to purchase risk-based insurance (or remain uninsured) to avoid cross-subsidizing high risks. Government regulation in this form - a combination of community rating, open enrollment and compulsory insurance - is implemented in Belgium, the Netherlands, and Switzerland, and is part of Enthoven's (1988) proposal to reform the U.S. health care system. See Kifmann (2002), p. 15. Since our focus is on private health insurance in Germany, we do not further discuss Germany's SHI system. For a more detailed discussion, the reader is referred to Breyer (2004).

5 Self-employed and civil servants are not compulsorily insured in the SHI system. For the latter, there is an entirely tax-nanced plan for civil servants (called Beihilfe). Depending on marital status and the number of children, this plan covers 50-70 % of health care expenditures with the remainder being covered

(4)

individuals must remain within the community-rated social health insurance system. In con-trast to the SHI system where premiums depend on income, PHI premiums are risk-based, individually calculated, and unrelated to income. PHI premiums are mainly based on age, sex and health status of a policyholder when he or she enters the contract.

Long-term PHI contracts are calculated using a funding principle. Policyholders accumulate funding capital to compensate for higher expected health expenditures in the future. Therefore, as health expenditures increase with age, premiums necessarily exceed expected cost in early years and fall below expected cost in older years. German law requires, however, that private health insurers calculate premiums in such a way that they are constant over the insured's life-time. The precautionary savings element used to equalize premiums over time is the aging reserve.

There is migration within and between these two health insurance systems. However, switching possibilities from the private to the social system and vice versa are somewhat restricted. Individuals aged 55 or older are not allowed to switch to the SHI system in any case. Despite regular increases in the contribution ceiling, ever since 1975, the number of people switching to substitutive private health insurance has been higher every year than the number of people lost to the statutory social system. Since 1997, however, the number of people switching from social to private health insurance has seen a substantial increase. The German Association of Private Health Insurers argues that this is partly due to cutbacks in the statutory SHI system. There are no data available on the extent of migration within the PHI system. However, ineciencies from the lack of bilateral commitment may involve migration-inducing distortions. In view of many countries' current health insurance reform debate, it is of interest to study and evaluate these potential distortions.6

The remainder of the paper proceeds as follows. The next section reviews related literature. The basic structure of premium calculation in German private health insurance is explained in Section 3. We derive our main hypotheses according to the theory of one-sided commitment in Section 4. Section 5 contains the empirical analysis. Section 6 discusses policy implications of our study. Section 7 concludes.

2 Related Literature

Our paper is related to two strands of literature. The rst strand is the vast literature on dynamic contracts, learning and commitment. This theoretical literature has received rela-tively little empirical attention. Hendel and Lizzeri (2003), following Harris and Holmstrom

by an additionally purchased PHI contract. As civil servants would lose these entitlements while staying in the public system, there is a strong incentive to (partially) join the PHI system.

6 In the United States, for example, the House of Representatives recently approved legislation to man-date that all individuals be covered by health insurance, premium subsidies for low-income individuals, and creation of a government-run health insurer to compete with private health plans. Following this legislation, most individuals would be required to have health insurance that meets minimum standards specied by the government. See Harrington (2010).

(5)

(1982), test a theory of dynamic contracting in the U.S. life insurance market. They provide strong evidence of the existence and signicance of learning over time. Life insurance contracts involve front-loading (prepayment of premiums) resulting from lack of bilateral commitment to contracts. This front-loading creates a partial lock-in for consumers and more front-loaded contracts generally involve lower lapsation. Finkelstein, McGarry and Su (2005) support Hendel and Lizzeri's ndings using data on the U.S. long-term care insurance market where mortality risks are learned over time. However, there are several important dierences between life insurance contracts and health insurance contracts. A life insurance contract mainly in-sures an income stream; a health insurance contract inin-sures health care treatment. Learning about health is an important phenomenon. Life insurance contracts are comparatively simple and explicit when compared to health insurance contracts, and therefore asymmetric infor-mation is not an important issue.7 We look at more sophisticated private health insurance contracts evolving over time. In this context, asymmetric information may be more important because distortions due to dynamic information revelation on health status can be large. In contrast to Hendel and Lizzeri, who found that asymmetric information is not important in the case of life insurance, informational asymmetries may be an issue when contracts are more complex.

The second strand of literature this paper contributes to is the theory of adverse selection in insurance markets.8 Following this theory, high-risk agents can be expected to purchase more (comprehensive) insurance coverage. This prediction, the coverage-risk correlation, can also be expected to manifest itself in a greater tendency of high-risk agents to purchase insurance.9 A high-risk individual is one who generates higher expected insurance payouts due to a larger number of expected claims, a higher expected payout in the event of a claim, or both. Em-pirical evidence on adverse selection in insurance markets varies across markets and pools of policies. A signicant amount of empirical work nds evidence of adverse selection in insurance markets.10 Early research by Phelps (1976) reports no evidence of a signicant relationship between predicted illness of individuals and their choice of insurance coverage. Although Car-don and Hendel (2001) suggest that informational asymmetry in the U.S. health insurance

7 See Hendel and Lizzeri (2003), p. 299.

8 The theory of adverse selection in insurance markets was introduced by Rothschild and Stiglitz (1976) and has been extended in many ways. For a survey and discussion of theoretical adverse selection models, see Dionne, Doherty and Fombaron (2001).

9 The existence of a coverage-risk correlation itself is viewed as necessary for adverse selection to be present (and its absence as sucient for rejecting adverse selection). See, for instance, Chiappori and Salanié (2000). However, it may not always be sucient to conrm the existence of adverse selection given that this correlation is also suggested by moral hazard theory. See, for instance, the discussion in Cohen and Siegelman (2010), pp. 71-74. To test for adverse selection, individuals facing the same set of choices should be examined. To test for moral hazard, similar individuals facing dierent coinsurance rates should be examined, where price sensitivity can be measured by using the coinsurance variability across individuals. See, e.g., Cardon and Hendel (2001). A detailed discussion can be found in Cohen and Siegelman (2010). 10 For a detailed review, see Cutler and Zeckhauser (2000). They review a substantial amount of empirical adverse selection studies in health insurance. Virtually all of these studies support the hypothesis of informational asymmetry in favor of policyholders. While our focus is on health, adverse selection is shown to be present in several other insurance markets. See, for instance, Makki and Somwaru (2001) or Cohen (2005).

(6)

market may be unimportant, Browne (1992) provides statistical evidence that adverse selec-tion is present in the market for individual health insurance in the U.S.. His analysis conrms that cross-subsidization of high risks by low risks occurs in this market. Browne and Doerp-inghaus (1993) also focus on the U.S. market for individual health insurance. They nd that the characteristics of the insurance policies purchased by high and low risks and the premiums paid by high and low risks are similar, but that high risks derive more indemnity benets from their insurance contract than do low risks. This nding suggests that adverse selection in the market for individual health insurance results in a pooling of risk types. Browne and Doerpinghaus (1994) report similar results in a study of the Medicare supplemental insurance market in the United States. Cutler and Reber (1998) study dierent health insurance plans oered by Harvard University, which switched from subsidizing the most generous plans to oering a xed-dollar subsidy, thus increasing the annual cost of the most generous plan. The coverage-risk correlation was strongly conrmed by their analysis: the most generous plan was abandoned by the best risks.

Theoretical analysis by de Meza and Webb (2001), in which they assume that individuals have private information about both their risk type and risk aversion, concludes that selection based on risk aversion is advantageous if high risk aversion types both purchase more insurance and are of a lower risk type. The failure to condition on risk preferences may then mask adverse selection predicted by one-dimensional models. Fang, Keane, and Silverman (2008) report evidence of advantageous selection in the U.S. Medigap insurance market.

Finkelstein and McGarry (2006) nd no statistically signicant evidence of a positive cor-relation between insurance coverage and (ex post) realizations of loss in the long-term care insurance market in the United States; however, Browne (2006) shows that high-risk types are more likely to retain their insurance coverage, an indication of the presence of adverse selection in this market.

Although there is a signicant body of empirical work addressing adverse selection in health insurance markets, the main focus of this literature has been on testing predictions of static insurance models.11 An exception is Dionne and Doherty (1994), who study a two-period competitive insurance market with one-sided commitment and renegotiation. They show that an optimal renegotiation-proof contract may entail semi-pooling in the rst, and separation of risk types in the second, period. More importantly, they show that optimal contracts will exhibit highballing features, i.e., the insurer will typically make positive prots in the rst period, compensated by below-cost second period contracts. In this paper, we use unique private health insurance data and contribute to the smaller literature on dynamic contract theory, learning, and one-sided commitment. We empirically investigate contract dynamics in a world where consumers cannot commit to renewing a contract but insurers commit to oering renewal at a premium rate that does not reect revealed future information about the insured risk.12 We combine the two strands of literature, that on one-sided commitment and learning with that on adverse selection.

11 See the seminal theoretical work by Akerlof (1970), Rothschild and Stiglitz (1976) and Wilson (1977). 12 To our knowledge, there is no empirical study on adverse selection in the private health insurance market

(7)

3 Private Health Insurance in Germany

About 10 per-cent of all Germans have private health insurance. Only an upper-income group as well as the self-employed and civil servants are eligible for the private system. Employees can purchase PHI only if their annual labor income exceeds a certain threshold.13 If annual income is below this social security ceiling, they must remain within the compulsory community-rated public health insurance system. There are exceptions for learners and students who can more easily choose between both systems due to the absence of income restrictions.14 In case a policyholder retires, the income ceiling does no longer apply and she stays in the private system even though her income may have fallen below the ceiling.15 Only very few people in Germany have no health insurance at all (less than 0.2 per cent) and it is highly unlikely that any of these uninsured individuals dispose of earnings above the income ceiling.16

There are ve possible reasons a policyholder may opt out of a PHI contract. Individuals opt out over time (a) because their incomes have fallen below the contribution ceiling, (b) because they nd out they are low risk and switch to another more favorable health insurance contract involving less extensive coverage and lower prices (c) because their work status changes from self-employed to wage earner,17, (d) because they expatriate themselves, or (e) because they die. Of course, we cannot dierentiate between the causes for lapsation, but we may draw conclusions about relationships. In this view, our focus will mainly be on the (b), the second reason.

Private health insurance premiums are risk-based. The price depends on health status, sex, and age at contract entry as well as on the extent of PHI coverage. In particular, when a policyholder enters into a new PHI contract, the ph insurer conducts a risk assessment that may lead him to imposing a risk loading on the net premium due to an individual's poor health status at contract entry. However, there is no reassessment of risk type over time and the loading stays constant except when the policyholder can prove that the precondition

health insurance system. They look at why company-based sickness funds were able to attract many new customers during 1995-2000and study potential determinants of switching behavior. They nd no

evidence for selection by sickness funds in German SHI.

13 In 2010, for instance, annual income before taxes must exceed 49,950 Euros.

14 Most children are insured via their parents' PHI or SHI contract. However, they might change this status and switch systems or contracts when they enter the job market. For instance, it is possible to pause a PHI contract, when a 16-year-old decides to enter an apprenticeship or other work during which he will be insured in the SHI system. Therefore, there is a choice even for young people in choosing between systems or dierent PHI contracts.

15 Note that once a policyholder has entered a PHI contract, he or she cannot simply drop out and reenter the SHI system, regardless of how attractive this seems. The policyholder must stay in the PHI system as long as annual income is above the ceiling. However, it is possible to switch within the system and to choose any other PHI contract oered by either the same or another private health insurer.

16 See Thomson, Busse, and Mossialos (2002), p. 426.

17 In this case, individuals can either cancel or pause their PHI contract. Pausing may be attractive if the work status may change again in the future. Then, the individual can reenter the PHI contract at the same conditions. If, however, the person feels that he or she is a low risk and that the SHI system can do equally well, the person may cancel the PHI contract. We take this eect into account in our empirical analysis below.

(8)

that led the insurer to impose a risk loading is no longer present or has become unimportant. As a consequence, PHI premiums are not adjusted according to health status over time and individuals face no reclassication risk. They only face this risk if they decide to switch insurers. Then, again, the new insurer conducts a risk assessment and may impose some risk loading in the premium due to poor health at contract entry.18 When a policy lapses (i.e. when the insured dies or switches insurers) the excess in aging reserves forfeited at lapsation is captured by the remaining collective. As a result, there is implicit partial cross-subsidization between policies over time and some benet of survivorship. However, since there is always an individual risk assessment when entering a PHI contract with a new insurer, switching seems more attractive for lower-risk types.

PHI premiums are calculated using a funding principle. Policyholders accumulate funding capital to compensate for higher expected health expenditures in the future. Therefore, as health expenditures increase with age, premiums necessarily exceed expected cost in early years and fall below expected cost in later years. At contract entry, the PHI premium is calculated in such a way that it is, ceteris paribus, constant over the insured's life-time.19 The precautionary savings element used to equalize premiums over time is the aging reserve. From an actuarial viewpoint, a constant net premium is calculated such that accumulated aging reserves in early contract years are sucient to compensate high health care costs in later years.20 Basically, if policyholders wish to switch insurers, they cannot take any fraction of the aging reserve with them.21

Although expected health expenditures per risk change over time, premium adjustment is independent of individual health status due to a collective actuarial calculation method. As a result, reclassication risk is insured. The funding principle used to calculate PHI premiums is implemented via the so-called equivalence principle. This principle states that a lifetime constant insurance premium needs to be calculated such that the present value of expected premium income (P) is equivalent to the present value (P V) of expected calculated health

expenditures (CHE). It is shown in the equivalence equation:

E[P V(P)] =E[P V(CHE)], (1)

where P V(CHE) and E[P V(P)] are determined using actuarial methods, mortality tables,

and an actuarial interest rate.22 Using some exemplied progression data, Figure 1 illustrates

18 Impairments are usually assessed by a health questionnaire and/or doctor's report.

19 Premiums would be constant if, for instance, the insurer's collective and treatment cost do not vary. See Milbrodt (2005).

20 Insurers calculate with a maximum life expectancy of 102. Since life expectancy depends on age and risk type of a policyholder, premiums vary for dierent entry ages.

21 It should be noted that this fact has partly changed since 2009. Since January 2009 private health insurers oer an additional base rate (Basistarif). Under certain circumstances, switching into such a contract is possible without losing accumulated aging reserves. Since our empirical study is based on data that were collected before 2009, this new law has no impact on our study and we thus neglect it.

22 Following legislation, insurers commonly use a 3.5 per cent actuarial interest rate. The law claims that the interest rate used to calculate insurance premiums and aging reserves must not be higher than 3.5 per cent.

(9)

4

We proceed as follows. The next section introduces the calculation principle in German private health insurance. The idea is the same for the differentiated conventional packages and the new basic benefit package. We then discuss …

2. Premium Risk

Aufbereitung der Literatur, insbes. Kifmann-Diss (der Anfang) und die dort zitierte Literatur: Lock-in, Informationsprobleme…

3. Insuring Premium Risk in German Private Health Insurance a) Basis for Actuarial Calculation

The funding principle in calculating private health insurance premiums is implemented via using the so called equivalence principle. This principle states that a lifetime constant insurance premium has to be calculated such that the present value of expected premium income (EP) is equivalent to the present value (PV) of expected calculated health expenditures (ECHE). It is shown in the equivalence equation

PV [EP] =PV[ECHE]. 0,00 2000,00 4000,00 6000,00 8000,00 10000,00 12000,00 14000,00 16000,00 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 Age A m ount in E u ro Niedrige Kopfschadenreihe Hohe Kopfschadenreihe Prämie bei Eintrittsalter 30 und niedriger Kopfschadenreihe Prämie bei Eintrittsalter 30 und hoher Kopfschadenreihe Prämie bei Eintrittsalter 50 und niedriger Kopfschadenreihe Prämie bei Eintrittsalter 50 und hoher Kopfschadenreihe

Low health expenditures per risk High health expenditures per risk Premium for age at entry 30 and low health expenditures per risk Premium for age at entry 30 and high health expenditures per risk Premium for age at entry 50 and low health expenditures per risk Premium for age at entry 50 and high health expenditures per risk

Figure 1: Health expenditures per risk and net premiums. Source: Nell and Rosenbrock (2009).

how premiums are determined.23 As can be seen from the gure, higher health expenditures per risk imply higher premiums over the insured's life-time. Premiums are based on the age of the policyholder when he or she enters into a contract with a private health insurer. Of course, for a higher age at entry, premiums are increased since there is less time remaining for pre-nancing health expenditures at a higher age.

The parts of the premium not used for indemnity payments due to lower health expenditures per risk in younger years are accumulated by the insurer in the form of actuarial aging reserves. Due to interest eects, aging reserves tend to increase until policyholders reach a high age (see Figure 2) even though the health expenditures per risk curve crosses the premium curve signicantly before this time (see Figure 1). The accumulation of aging reserves ensures that, ceteris paribus, at each point in time the equivalence equation holds, i.e., the present value of expected premium income and existing aging reserves (AR) is equivalent to the present value

of expected calculated health expenditures

E[P V(P)] =E[P V(CHE)]−AR, (2)

which constitutes the general equivalence equation after entry. This generalized equivalence equation is the basis of German PHI premium calculation. It corresponds to equation (1) for AR = 0. Individual reclassication risk is insured because equation (2) does not include

any individual health information but follows directly from the actuarial calculation principle. Thus, when the policyholder's state of health deteriorates, the premium stays constant. The basic mechanism is as follows. After a few periods (theoretically, after each period),CHEand

23 The gure stems from Nell and Rosenbrock (2009). Premiums calculation in Figure 1 is notional and

based on values similar to Milbrodt (2005). The calculations include exit and mortality risk, as well as an actuarial interest rate.

(10)

 

Figure 2: Aging reserves for dierent ages of entry and health expenditures per risk. Source: Nell and Rosenbrock (2009).

mortality rates used for calculating present values generally dier and need to be recalculated. Then, if necessary, aging reserves are adjusted.24 Together with new mortality tables, the new premium P is calculated and a new equivalence equation as shown in equation (2) is

obtained. Individual reclassication risk is insured becauseCHE and mortality tables apply

to all policyholders in the insurer's collective, and adjustment of aging reserves is regulated, so that the premium does not vary according to individual health status.

In contrast to individual reclassication risk, we use the term collective reclassication risk to mean the risk that collective health expenditures exceed collective costs. This risk is borne by policyholders via premium loadings since it follows the development of health expenditures in the insurer's collective. Premiums need to be adjusted if overall premium income does not suce to cover overall health expenditures. While individual reclassication risk can be insured via partial risk pooling, i.e., using a collective actuarial calculation method, collective reclassication risk cannot be insured.

In a long-term perspective, dynamic information revelation as to health status informs pol-icyholders about their individual risk. Since the equivalence principle is based on collective actuarial calculation, and aging reserves do not include any information on individual health status, the system appears prone to risk selection. This is because aging reserves are objec-tively too high for low risks and too low for high risks. It seems likely that lower-risk types, once they discover that they are low risks, will be inclined to cancel their policy and look for cheaper and less comprehensive coverage elsewhere. Therefore, reclassication risk constitutes some implicit switching cost and partial risk pooling seems likely to entail risk selection in the German PHI market.

24 Methods for adjusting aging reserves are regulated. Financial resources for this adjustment stem from separate sources or insurers' nancial surplus.

(11)

4 The Model

German PHI premiums involve front-loading, i.e., prepayment of premiums. As a consequence, policyholders transfer income from early contract years to later years, during which they generally have worse health. Contracts are unilateral in the sense that policyholders can cancel their PHI policies, whereas insurers commit to the terms of the contract as long as the contract is in force. These contract characteristics are consistent with a dynamic model of one-sided commitment. To make theoretical predictions, it seems fruitful to adapt such a model framework, which we accomplish by relying heavily on Hendel and Lizzeri (2003). The key features of the model are (1) symmetric learning, i.e., information about risk type is revealed over time; (2) one-sided commitment of insurance companies; (3) buyer heterogeneity as to front-loading, i.e., consumers vary in income (growth); and (4) guarantee of full insurance against reclassication risk, i.e., premiums are independent of health status.

Consider a competitive two-period health insurance market with a high number of buyers and sellers. Buyers wish to insure potentially worse health states in the future. A health status can be described as a certain probability of suering medical loss and a certain severity of the loss in the event the individual experiences medical loss. In the rst period, individuals are identical with regard to loss probability and severity, i.e., they face identical medical expenditures m

in case of illness. Buyers have identical probability p of suering an illness.25 In period 2,

multiple health statuses are possible so that policyholders dier in both their probabilitypi of experiencing some medical loss and the severity or size of the loss. If an individual is in health statusiin period 2, he or she faces potential medical lossmi2, which may represent physician

visits, time and money invested in obtaining treatment, etc. We order health status so that

p1 < p2 < ... < pN and m12 < m22 < .... < mN2 . We also assume p ≤ p1 and m ≤ m12, i.e.

health worsens over time in both frequency and severity. The probability of being in health state iis given by πi. Individuals' utility function is given by u(c) when they consume c≥0 in each period. The utility function is strictly concave, indicating risk-aversion, and twice continuously dierentiable.

Period 1 involves three stages: rst, insurers oer contracts, second, buyers choose a contract, third, uncertainty about medical loss is revealed, and consumption takes place. Period 2 involves four stages: rst, uncertainty about a policyholder's health status is revealed. The realized health status of a policyholder can be observed by the insurer. The following stages are equivalent to stages 1 to 3 in the rst period.

A rst-period premium consists of a rst-period premiumP1 and coverage amountC1, and a

vector of premiums and coverage amounts(P1

2, C21)...(P2N, C2N) indexed by the second-period

health states. Thus a rst-period contract is a long-term contract to which the insurer unilat-erally commits. In contrast, the second-period contract is a short-term contract. Note that a short-term contract may depend on information revealed at the beginning of the second period. It consists of a premium and coverage amount (P2i, C2i) indexed by second-period

health status. There is one-sided commitment: insurers can commit to future premiums while

(12)

consumers freely choose between staying with their period 1 contract and switching to a com-petitor oering a short-term spot contract in period 2.

Consumer heterogeneity is given by dierences in the income process. We assume that con-sumers dier in income growth. While some minimum income y˜ is needed in order to enter

into a PHI contract, consumers with minimum income y˜ may dier in income growth. To

capture this, we assume that a consumer receives an income of y−g ≥y˜ in the rst period

andy+g in the second period, and the variation in consumers' income growth is represented

by g >0. In our model, y˜ represents the minimum income needed in order to enter the PHI

market.26

In competitive equilibrium, allocations maximize individuals' expected utility subject to a zero expected prot constraint and a set of no-lapsation constraints that capture individuals' inability to commit to the PHI contract. Solving this constrained maximization problem gives the set of prices and coverage amounts that must be available to individuals in a competitive PHI market.27

In equilibrium, premiums and coverage amounts that are fully contingent on health states,

(P1, C1), and (P21, C21)...(P2N, C2N)must maximize individuals' expected utility:

EU = (1−p)u(y−g−P1) +pu(y−g−P1−m+C1) (3) + N X i=1 πi (1−pi)u(y+g−P2i) +piu(y+g−P2i−mi2+C2i)

subject to a zero prot constraint (i.e., contracts break even on average)

P1−pC1+ N X i=1 πi P2i−piC2i = 0 (4)

and no-lapsation constraints imposed by lack of consumer commitment: for all possible future health states,i= 1, ...N, and for allP˜2i,C˜2i such thatP˜2i−piC˜2i >0, we have

(1−pi)u(y+g−P2i) +piu(y+g−P2i−mi2+C2i) (5)

≥(1−pi)u(y+g−P˜2i) +piu(y+g−P˜2i−m˜i2+ ˜C2i)

The no-lapsation constraints imply that an equilibrium contract is such that there is no other contract that is protable and oers potential buyers higher expected utility in any state of period 2. We refer to the actuarially fair premium that guarantees full insurance in health state i and guarantees zero expected prots asP2i(F I). The following predictions about the

market equilibrium set of contracts, which can be found in the Appendix, are an adaptation of Hendel and Lizzeri (2003) to our environment.

26 For simplicity, we assume there are no capital markets and that the borrowing rate is higher than the lending rate. Therefore, consumers with highergare more tightly constrained since their income in period

1 is lower.

27 As a reference point, we assume that the constrained maximization problem is constructed via fully

contingent contracts, i.e., the contracts are fully contingent on future health states. As a result, there is no lapsation. However, the argument can be extended to non-contingent contracts.

(13)

• Under competition, PHI insurance is oered at a fair rate and all consumers obtain full insurance in all possible states of the world.

• In the later period, premiums involve a cap so that the actual premium is below the fair premium in this period. Consumers transfer income from the rst period, when they enjoy comparatively good health, to future states involving worse health. The initial overpayment in the premium creates a lock-in or commitment to the PHI contract. • Contracts are front-loaded as long as income growth is below a certain threshold. Less

front-loaded contracts appeal to buyers with lower rst-period income.

These predictions are drawn from equilibrium allocations via fully contingent contracts that involve no lapsation. However, there are comparable non-contingent contracts that allows us to derive comparative statics predictions.

• Non-contingent contracts with higher rst-period premiums (i.e., higher front-loading) are chosen by consumers with lower income growth and have a lower rate of lapsation. The PHI market we study is dierent from the life insurance market studied by Hendel and Lizzeri (2003). However, the basic mechanism behind the market forces is very similar. Ger-man PHI contracts insure individual reclassication risk. A policyholder faces reclassication risk only when he or she considers entering a PHI contract with a dierent insurer. In summary, we predict the following two key hypotheses for the German PHI market.28

1. Front-loading creates a lock-in of consumers: Contracts with higher front-loading have lower rates of lapsation due to a more severe lock-in.

2. Since high-risk types have a lower incentive to lapse for any given contract, ceteris paribus, the risk pool worsens over time.

5 Testing the Implications of the Model for the German PHI

Market

5.1 Data

We study a large-sample data set from a German private health insurer. We observe enrollment in a comprehensive private health insurance contract over a period of ve years,2001through 2005. To analyze potential ineciencies resulting from the lack of bilateral commitment in

this market, it is of interest to discover which individuals drop their PHI contract over time. Therefore, we create a subsample consisting of those5,681individuals who were enrolled with

28 Note that we observe a given number of policyholders over time. In practice, new policyholders joining the collective tend to dilute eects.

(14)

2001 2002 2003 2004 2005 Total enrollment 5,681 4,871 4,256 3,859 3,627 Enrollment male 4,143 3,558 3,110 2,830 2,651 Enrollment female 1,538 1,313 1,146 1,029 976 Average premium 1,935.14 2,511.01 2,565.22 2,600.55 2,613.86 Average premium male 1,817.13 2,377.74 2,433.83 2,473.67 2,492.16 Average premium female 2,253.01 2,872.15 2,921.80 2,949.50 2,944.42 Average loss 1,582.23 1,701.09 1,740.17 2,192.57 2,191.12 Average loss male 1,453.63 1,434.96 1,538.95 2,171.81 2,170.74 Average loss female 1,724.10 1,968.27 2,134.63 2,238.52 2,236.50

Figure 3: Data overview 20012005.

the insurer in2001, and then study the characteristics of those individuals over the following

ve sample periods. The subsample used for our statistical analysis contains 28,405 (i.e., 5,681·5) consumer-year observations.

The sample includes information on gender, age, tenure, pausation, individual health expendi-tures, and premiums paid, but information on other characteristics of interest, such as marital status, race, and income, is not collected by the insurer. A short overview of the sample is shown in Figure 3.

Women are higher risk type than men on average. In terms of loss frequency, over the ve periods observed, 88.62 per-cent of women made health insurance claims, whereas only 80.71 per-cent of men did so. Women are also higher risk type in terms of loss severity. The average annual treatment cost for a woman in our sample is 2,017.37 Euros; for men, the average health care cost is 1,646.58 Euros. As a consequence of their being a higher risk type, women pay, on average, about 20 per-cent more for their health insurance premium than do men. Figure 4 displays the distribution of birth year for the policyholders in our sample. Most policyholders were born between1965and1970, meaning that most were in their early thirties

in the study period 20012005. The average age of a policyholder in our sample is 32.6.

Since the regulatory framework of German PHI requires employed individuals to stay within the social health insurance system as long as their earnings do not exceed the PHI income threshold, the majority of PHI participants are around 30 years old.29 However, given that

newborn children tend to enter via their parents' PHI contract, this average can be lower when children are taken into account. We include children in our study and thus the average age of entry is23.84. Figure 3 shows how individuals drop out over time. We are interested in the

characteristics of the policyholders who cancel their policy. We group individuals according to their age. Age category 1 includes all individuals up to age 20. Age category 2 includes

all individuals over 20 up to 30. Age category 3 includes individuals over 30 up to 40. Age

(15)

Histogram of Birthyear Birthyear Frequency 1920 1940 1960 1980 2000 0 200 600 1000

Figure 4: Year of birth of policyholders.

category 4 includes individuals over 40 up to 50. Finally, age category 5 includes all individuals older than 50.30

We conduct our analysis in two steps. The rst step involves using an estimation model to obtain a proxy for risk type in our sample. We refer to this model as the Prediction Model. We obtain this proxy before we test our theoretical predictions, which is the second part of our analysis. We refer to the second model as the Test Model. In the rst-step modeling approach, we use a two-part model to predict expected medical expenditure, which is our proxy for a policyholder's risk type. In the second step, we use logistic regression to test our hypotheses. Each statistical method is explained in more detail below.

5.2 Predicting Risk Type via Expected Medical Expenditure

The list of variables we use for our analysis is given in Figure 5. Claim is a dichotomous variable equal to one if there has been a claim in period t; zero otherwise. We dene a claim as being observed when medical expenditures are nonzero. Lnloss is a numerical variable indicating the size of medical expenditures on the log scale for a given period. It is dened only in the event of a claim, i.e., when medical expenditures are nonzero. Female is a dichotomous variable equal to one if the policyholder if female; zero otherwise.31 Age indicates the age in years of a policyholder in a given period. Similarly, Agesquare is the square of Age. CHE is a numerical variable indicating the insurer's calculated expected health expenditures for a given individual in a given year. In other words, CHE is what the insurer expects to spend in t on a given policyholder taking into account the policyholder's age and gender. Contractyear is a numerical variable indicating the policy age. For instance, at contract entry, Contractyear is equal to one, meaning that the contract is in its rst duration period. Year is a count variable

30 Note that this nal age category makes sense when taking into account that, at some age over50, switching is no longer possible for policyholders in old age.

(16)

indicating the period, i.e.,2001through2005. Riskload is a numerical variable representing the

premium loading for a high-risk policy, that is, Riskload is positive when the insurer conducts a risk assessment at contract entry and nds that the policyholder is actually high risk. The size of Riskload depends on the specic precondition or illness the policyholder suers. Exposure is a numerical variable indicating the number of months a PHI contract was actually in force for a given period. This variable takes into account that (1) a contract may be paused for some time and thus there is no insurance coverage in case of a loss, and (2) an individual may have joined the pool later in the year. Enrol is a dichotomous variable equal to one if a policyholder is enrolled with the insurer in a given period; zero otherwise. Note that a policyholder may be enrolled in a PHI contract in a given period, but the contract might be paused. However, this applies to very few policyholders and does not aect our results. AC1 is a dummy variable equal to one if the age of a policyholder falls within age category 1 for a given period; zero otherwise. AC2 through AC5 are similarly dened. AC1_Fem through AC5_Male as well as CHE_Fem represent interactions terms. Medexp is dened below.

To test our hypotheses and, more specically, whether the market may suer from potential selection issues, it is necessary to classify all policyholders in the sample by their risk type. Although individual risk type is unobservable, claims, along with their frequency and severity, can be observed. We thus use (predicted) medical expenditures as a proxy for policyholder risk type. Having identied risk types, we next look at which policyholders drop their coverage using our risk type proxy.

5.2.1 Statistical Method

Health care data are unique in two main respects. First, they are often truncated or left-censored in the sense that they tend to feature a large proportion of zeros that must be accounted for in the econometric modeling. Zero values may represent an individual's lack of health care utilization or that there has been no health care expenditure in a given period. Second, health care data are also often skewed to the right and thus do not follow a normal distribution.

To account for the peculiarities of health care data, we adopt an approach often used in actuarial econometrics. Following the traditional actuarial literature based on the individual risk model, the response, i.e., the insurance claim, can be decomposed into two components: a frequency (number) component and a severity (amount) component.32 More formally, letri be a binary variable indicating whether theith individual had an insurance claim and letyi describe the amount of the claim given there was a claim. Then, the claim can be modeled as

(claim recorded)i =ri×yi (6) constituting the two-part or frequency-severity model.33

32 See Bowers et al. (1997), ch. 2. 33 See Frees (2010), p. 424.

(17)

Type Min. Max. Mean S.E. Prediction

Variables Claim (Claim = 1) Binary 0 1 0.5503 0.497

Lnloss Numerical 0.924 12.35 6.822 1.252

Female (Female = 1) Dummy 0 1 0.271 0.444

Age Numerical 0 86 32.583 15.755

Agesquare (Age*Age) Numerical 0 7396 1309.852 985.264

CHE Numerical 0 7764.9 1168.607 1013.96

CHE_Fem (CHE*Female) Numerical 0 7764.9 413.51 974.229

Contractyear Numerical 1 66 9.746 10.037

Riskload Numerical 0 3212.76 62.632 25.56

Exposure Numerical 0 12 10.02 3.553

Year Numerical 1 5 3 1.414

Response

Variable Enrol (Enrol = 1) Binary 0 1 0.785 0.411

Male (Male = 1) Dummy 0 1 0.729 0.444

AC1 (age category 1 = 1) (omitted) 0 1 0.232 0.422

AC2 (age category 2 = 1) Dummy 0 1 0.106 0.308

AC3 (age category 3 = 1) Dummy 0 1 0.359 0.48

AC4 (age category 4 = 1) Dummy 0 1 0.196 0.397

AC5 (age category 5 = 1) Dummy 0 1 0.106 0.308

AC1_Fem (AC1*Female) Interaction 0 1 0.108 0.31

AC2_Fem (AC2*Female) Interaction 0 1 0.018 0.131

AC3_Fem (AC3*Female) Interaction 0 1 0.071 0.257

AC4_Fem (AC4*Female) Interaction 0 1 0.043 0.203

AC5_Fem (AC5*Female) Interaction 0 1 0.031 0.174

AC1_Male (AC1*Male) Interaction 0 1 0.124 0.329

AC2_Male (AC2*Male) Interaction 0 1 0.089 0.284

AC3_Male (AC3*Male) Interaction 0 1 0.289 0.453

AC4_Male (AC4*Male) Interaction 0 1 0.153 0.36

AC5_Male (AC5*Male) Interaction 0 1 0.075 0.263

Contractyear Numerical 1 66 9.746 10.037

Medexp Numerical 0.219 8.77 4.14 1.757

Variables in the Prediction Model

Control Variables

Variables in the Test Model

Control Variables

Figure 5: Response and explanatory variables in the empirical models.

We use two equations to model these random variables: The rst equation is a logit equation that estimates the probability that an insured person will need any medical treatment during a given period. The second equation is a linear regression for the logarithm of annual medi-cal expenditures during a given period. The log transformation nearly eliminates the typimedi-cal undesirable skewness in the distribution of medical expenditures, making the model more ro-bust.34 Our aim is to predict expected logarithmic medical expenditures which is our proxy for policyholder risk type. Note that we could also predict medical expenditures (without the log scale) but this would require a retransformation to the normal scale. A shortcoming of this procedure is that the error terms in the log expenditures equation are often not normally distributed but still skewed so that normal retransformation estimates are biased. Therefore, the smearing estimate, developed by Duan (1983), is often used to estimate the

retransfor-34 In particular, the logarithmic transformation yields nearly symmetric and roughly normal error distribu-tions, for which the least squares estimate is ecient. See Duan et al. (1983).

(18)

mation factor. The smearing estimate is given by the sample average of the exponentiated least squares residuals. However, estimating expected expenditures in this way complicates our analysis without providing further insights. Since there is a positive relationship between expected losses and expected logarithmic losses (and since we only need a risk classication proxy here) we use predicted expected losses on the log scale as a proxy for risk type.

We employ a two-part (frequency-severity) model where the rst part is a binary model and the second part is a continuous model on positive expenditures only. This is because the mechanism that determines zero or nonzero expenditures might not be the same as the mechanism that determines the amount of positive expenditures. The model exhibits two separate equations: one for the probability of any insured medical expenditure, and one for the level of (log) positive insured medical expenses. Again, the basic idea is to decompose one observed random variable, here insured health care cost per policyholder, into two observed random variables, a frequency part variable to estimate the probability of an individual exhibiting positive medical expenditures, and a severity part variable to estimate loss severity given there has been a loss. Specically, our rst equation is a logit equation for the dichotomous event of zero versus posi-tive annual medical expenditure. Following Manning et al. (1987), we express the expectation of the response as a function of explanatory variables to be the probability of a claim, i.e.,

P rob(Claimi = 1) =

exp(x0iβ)

1 +exp(x0iβ) (7)

where

x0iβ=β0+β1x1,i+...+βkxk,i (8)

and where Claimi is a binary variable equal to one if there is a claim for a given insured individual i; zero otherwise. The unknown parameters βj are estimated by using maximum likelihood techniques.

Similarly, our second equation is a linear regression on the log scale for positive medical expenditure given that the policyholder receives any medical services:

Lnlossi|Lossi>0 =x0iδ+i. (9) As explained above, we use logistic medical expenditure, given by Lnlossi, as a dependent variable in order to account for the typical undesirable skewness in the distribution of medical expenditures. Since the logarithm is an increasing function, we can just as well use expected medical expenditures on the log scale as a proxy for risk type as compared to using non-log values.

5.2.2 Estimation Results

Assuming thatβandδare not related, i.e. under independence, the two parts of our model are

estimated separately to result in a prediction of expected medical expenditure per policyholder in a given sample period. The estimates for loss frequency and severity, respectively, are shown in Figure 6.

(19)

 

Figure 6: Loss frequency (1) and severity (2) estimations.

The two-part model estimations conrm that women are altogether higher risk than men. The coecient on loss probability (Claim) is positive and signicant at the5 per-cent level.

The coecient on loss severity (Lnloss) is positive but not signicant. Higher age tends to have a positive (but decreasing) eect on loss severity and a negative (but increasing) eect on loss probability. This suggests that treatment cost increases with age but policyholders tend to visit physicians less often. Our estimations conrm the intuition that the insurer's expected calculated health expenditures are positively correlated with both loss probability and loss severity. A higher exposure during a given period, that is, a longer policy duration, is associated with higher loss probability and severity. Finally, a higher risk loading due to bad health status has a positive eect on both loss probability and loss severity. As also clear from the estimated parameters of the twopart model for loss severity prediction, theR2

does not explain very much of the variation in loss probability and severity. However, this is not surprising as insured losses are largely random events and thus a considerable amount of unexplained variation is expected.35

Using these estimates and combining them in a new single estimate, Medexp, which is simply the product of the frequency and the severity components, we obtain an estimate for expected

35 Note that we are not using the prediction equation to test any hypotheses. Therefore, collinearity is not a concern at this point of our analysis.

(20)

medical expenditures (on the log scale) for each policyholder. We will use this estimate as an independent variable in the second part of our analysis when we test our hypotheses regarding the German PHI market.

5.3 Testing the Implications of the Model

5.3.1 Statistical Method

We consider the individual's decision to enroll in a PHI contract as a discrete choice. An individual may either stay with the original contract or enter a new PHI contract with some other (private or public) insurer.36 The net expected utility of being enrolled in the PHI contract as compared to switching to some other contract is assumed to be given by some linear index function

EUi=x0iφ+i, (10) wherexi denotes individual i's characteristics, φis the vector of parameters to be estimated and represents the impact of these characteristics on the decision to stay enrolled in the private health insurance contract, andi is a random error term. We do not observe the net expected utility of being enrolled in the private health insurance contract in a given period, but we do observe whether the net expected utility is positive, meaning that the individual decided to be enrolled in the contract in a given period. Thus, using logistic regression analysis, the probability that an individual is enrolled in a given periodt is modeled as:

P rob(Enroli,t = 1) =P rob(EUi,t >0) =

exp(x0iφ)

1 +exp(x0iφ), (11)

where Enroli,t is a dummy variable equal to one if individual i is insured in period t; zero otherwise. The logit modeling approach here consists in estimating the probability that a policyholder drops out of his or her PHI contract in periodt, depending on individual

char-acteristics. We use predicted medical expenditure (on the log scale),M edexp, as proxy for a

policyholder's risk type. 5.3.2 Empirical Results

Logistic regression results for the test equation are shown in Figure 7. The test equation is estimated four times for each data period in the sample based on 100 per-cent enrollment in the previous period, i.e., we evaluate lapsation for every period from 2002 through 2005.

Therefore, in our test equation below, the number of observations available for testing in period

36 Note that this is true for over 99.8 per cent of all Germans since the proportion of uninsured people in Germany is below 0.2 per cent. Therefore, we ignore the decision to drop health insurance coverage.

(21)

(2002) (2003) (2004) (2005)

VARIABLES Enrol02 Enrol03 Enrol04 Enrol05

AC2_Male 9.733**** 7.688**** 4.291**** 5.157**** (0.777) (0.755) (0.641) (0.832) AC3_Male 11.79**** 9.263**** 5.938**** 6.616**** (0.909) (0.880) (0.708) (0.965) AC4_Male 13.46**** 10.69**** 7.195**** 7.568**** (1.026) (0.985) (0.784) (1.065) AC5_Male 14.14**** 11.00**** 7.006**** 7.034**** (1.133) (1.087) (0.904) (1.191) AC2_Fem 8.903**** 5.691**** 3.127**** 2.893*** (0.809) (0.967) (0.842) (0.964) AC3_Fem 10.71**** 8.855**** 5.016**** 5.756**** (0.896) (0.882) (0.770) (0.986) AC4_Fem 12.82**** 10.21**** 6.216**** 7.628**** (1.065) (0.988) (0.861) (1.128) AC5_Fem 10.49**** 10.74**** 6.736**** 7.995**** (1.884) (1.279) (0.956) (1.262) Contractyear 0.0240** 0.0200** 0.0266** 0.00618 (0.0106) (0.00965) (0.0104) (0.0118) Medexp 5.442**** 4.306**** 3.096**** 3.717**** (0.382) (0.362) (0.302) (0.460) Constant -17.23**** -13.81**** -9.334**** -9.752**** (1.181) (1.126) (0.897) (1.235) Observations Pseudo R2 5681 0.81 4871 0.79 4256 0.77 3859 0.75 Standard errors in parentheses

**** p<0.001, *** p<0.01, ** p<0.05, * p<0.1

Figure 7: Enrollment choice estimation results (test equation).

t will correspond to enrollment in t-1 (see Figure 3). Age category 1, i.e., policyholders up to

20years old, is the reference age category and left out of regression. Interpretations will refer

to age category 1 as the reference age group. A Wald test for the coecients for Contractyear and Medexp suggests that including these variables in the regression creates a statistically signicant improvement in the t of the models. Variance ination factors are below 3 for all covariates.

We now address the hypotheses derived in Section 4. One-sided commitment in a market environment such as German PHI insurance suggests the following. (1) Front-loading creates a lock-in of consumers: contracts with higher front-loading suer lower rates of lapsation due to a more severe lock-in. (2) Low-risk policyholders have a higher incentive to lapse. Since low-risk types have higher lapsation, ceteris paribus, the risk pool worsens over time. Dynamic information revelation entails adverse selection.

In regard to the rst hypothesis, we expect that more front-loaded contracts will experience lower rates of lapsation. Since we do not have information about accumulated prepayments from policyholders, we need a proxy for front-loading and we use contract duration, i.e., our variable Contractyear, for this purpose. This is reasonable because accumulated prepayments (aging reserves) increase with contract duration, at least up to a certain point which is generally

(22)

after age of50.37 As illustrated in Figure 2, for a given age of a policyholder, an older contract

(younger entry age) has accumulated more aging reserves which constitutes a higher nancial loss in case of lapsation. Thus, for a given age (in our case we control for age categories), the coecient of Contractyear should be positive indicating lower lapsation (higher probability of retaining coverage). Our regression results conrm this predicted eect as can be seen in Figure 7. The coecient of Contractyear is positive and mostly signicant in all four test periods.

An alternative way of testing for the relationship between front-loading and lapsation is to construct an interaction term consisting of Contractyear and Age; we call this interaction term Contract_Age. Results for this alternative model are shown in Figure 8. As expected, the coecient of Contract_Age is positive and highly statistically signicant.

 

(2002) (2003) (2004) (2005)

VARIABLES Enrol02 Enrol03 Enrol04 Enrol05

Female -1.508**** -1.378**** -1.320**** -0.654** (0.170) (0.190) (0.227) (0.258) Contract_Age 0.00186**** 0.00190**** 0.00161**** 0.00103**** (0.000232) (0.000225) (0.000221) (0.000239) Medexp 1.655**** 1.624**** 1.633**** 1.664**** (0.0563) (0.0593) (0.0696) (0.0929) Constant -2.860**** -2.959**** -2.770**** -2.200**** (0.130) (0.154) (0.185) (0.214) Observations Pseudo R2 5681 0.68 4871 0.69 4256 0.69 3859 0.67 Standard errors in parentheses

**** p<0.001, *** p<0.01, ** p<0.05, * p<0.1

Figure 8: Alternative test equation model.

Another way to interpret front-loading would be to look at the entry age of policyholders. Figure 1 shows that for a given risk type, i.e., identical CHE, a higher entry age is associated with a higher premium, which is due to the fact that the insurer needs larger prepayments to account for health care cost in old age. A higher entry age is thus associated with a more severe lock-in. Therefore, we would expect lapsation to be lower for higher entry ages in our sample. As can be seen from the alternative regression shown in Figure 9, our data conrm this relationship, i.e. the coecient ofEntryageis positive and highly statistically signicant

indicating a higher tendency of a given risk type to retain coverage when he or she enters later. Finally, the lock-in eect also implies that individuals at a somewhat early contract stage are expected to have higher lapsation since they lose fewer prepaid premiums when they switch insurers at an early contract stage. Conversely, policyholders at an older contract stage will have an incentive to retain their contract. By ordering lapsation by the year a PHI contract was entered, Figure 10 graphically illustrates the lock-in eect: younger contracts, i.e., those

37 Figure 2 shows that at some point accumulated aging reserves are decreasing. But, due to interest eects, aging reserves tend to increase until high ages of policyholders (see Figure 2).

(23)

(2002) (2003) (2004) (2005)

VARIABLES Enrol02 Enrol03 Enrol04 Enrol05

Female -0.954**** -0.790**** -0.860**** -0.381 (0.173) (0.190) (0.220) (0.257) Entryage 0.105**** 0.0818**** 0.0716**** 0.0469**** (0.00798) (0.00753) (0.00810) (0.00813) Medexp 2.323**** 1.953**** 1.847**** 1.788**** (0.113) (0.0920) (0.0963) (0.114) Constant -6.276**** -5.012**** -4.266**** -3.084**** (0.354) (0.309) (0.318) (0.316) Observations Pseudo R2 5681 0.72 4871 0.71 4256 0.70 3859 0.68 Standard errors in parentheses

**** p<0.001, *** p<0.01, ** p<0.05, * p<0.1

Figure 9: Impact of Entry age of policyholders. contracts that have accumulated fewer aging reserves, have higher lapsation.

19 0 1960 1970 1980 1990 2000 2010 St ar t of phi cont ra ct 1930 1940 1950 1960 1970 1980 1990 2000 2010 0 500 1000 1500 2000 2500 St ar t of phi cont ra ct

Contracts lapsed through 2005

Figure 10: Contracts lapsed through 2005 and start of PHI contract.

The second hypothesis is that low-risk policyholders have a higher incentive to lapse. Our results conrm this hypothesis. The coecient of M edexp is positive and highly signicant

over all tested periods. As a result, high-risk policyholders those with higher predicted medical expenditure M edexp are more likely to retain their PHI contract (which implies

that low-risk types have a higher tendency to lapse). As a consequence, the risk pool worsens over time, ceteris paribus.

Constantly increasing low-risk type lapsation implies worse pools over time and dynamic in-formation revelation eventually leads to adverse selection in our sample. In the theoretical section of this paper, we assumed perfect information. The insurer can observe individual health status via loss frequency and loss severity. Even if there was no important informa-tional asymmetry between the parties, partial risk pooling and a lack of consumer commitment

(24)

would distort ecient behavior. The positive and highly signicant coecient ofM edexp

im-plies that high risks are more likely to retain their PHI contract than are low-risk types. Such a worsening of the collective over time, ceteris paribus, is indicative of adverse selection. Finally, it is interesting that the coecient ofF emaleis negative (see Figures 8 and 9). There

are two possible explanations for this nding. Women might be higher risk since the cost of pregnancy is covered by the contract and included in the premium. This would be reected in higher premium dierences between men and women for age in the range of mid twenties to late thirties. Indeed, in our sample, age categories 2 and 3 exhibit a stronger average premium dierence between genders. Another probable explanation is that the coecient represents an income eect. Since women tend to have lower income than men on average, and, as shown by our two-part model estimations, are also a higher risk type, their premiums are higher. Note that in our sample, women do indeed pay higher premiums (see Figure 3). InterpretingF emale

as a proxy for income, we would generally expect females to be more likely to cancel their PHI contract because their incomes are more likely to have fallen below the PHI contribution ceiling. This may be why their probability to retain a PHI contract is less than that of males. Second, if a woman drops out of employment to stay at home, she generally will be insured via her husband's insurance contract. This might have an impact on females' tendency to retain, or not, PHI coverage, too.

One of the theoretical predictions we made earlier is that in a dynamic one-sided commitment market where contracts are not contingent on future health state, the equilibrium will be such that the contract with higher rst-period premium is chosen by consumers with lower income growth, and that this contract has lower lapsation. Given the structure of the German PHI market, premiums tend to increase with age of entry. Given that income growth is generally higher in age categories 1 and 2 compared to age categories 3 and 4, we would expect that age categories 3 and 4 (with lower income growth) will be required to pay higher prices (compared to age categories 1 and 2), and will exhibit a higher tendency to retain the high-priced contract (implying lower lapsation). This is exactly what we observe in the data: the coecients of age categories are positive and highly signicant, and they tend to increase from age category 2 up to age category 5 (see Figure 7).

6 Discussion

Private health insurance in Germany is regulated in such a way that insurers must oer long-term contracts at a guaranteed rate, which involves front-loading of premiums and insurance of premium risk. The insurer accumulates aging reserves to smooth premiums over time. If policyholders want to switch insurers, they cannot take any fraction of this accumulated capital stock with them. As a result, this form of public regulation creates a lock-in eect implying that switching policyholders will bear a nancial loss even at early contract stages. To increase competition, it seems intuitive that aging reserves should be transferable between insurers. But transferring a policyholder's accumulated aging reserves to a new insurer would

(25)

trigger a new risk selection issue: low-risk policyholders would switch in order to save pre-miums, and, for the remaining collective, premiums would have to be adjusted. This raises the question of whether potential transfer amounts should be made contingent on the health state of the switching policyholder. If they were, higher (lower) transfers should accompany high-risk (low-risk) types so as to subsidize (account for) the higher (lower) premiums to be paid to the new insurer. However, such a scheme appears fraught with diculty in practice in view of incentive structures in the PHI market.38 Thus, it is still an open question as to how an ecient transfer system could be designed.

How can competition be improved without triggering risk selection? Eekho, Jankowski, and Zimmermann (2006) argue that changing the equivalence equation to include a risk-adjustment of aging provisions may deter risk-selection.39 Indeed, it seems reasonable to adjust the equivalence principle to account for individual risk. Then, the present value of health care expenditures and the present value of premiums will not depend only on collective risks, but also on individual ones. The risk-adjusted aging reserves become:

ARr =E[P Vr(CHE)]−E[P Vr(P)] (12) when we follow a morbidity-oriented equivalence principle. The new aging reserve depends on the health status (risk class) of a policyholder. Consider now a general increase in health care expenditures. Under the status quo without risk-adjusted aging reserves, impairment of individual health has no impact on individual premiums: individual premium risk is insured. But collective reclassication risk, namely, the risk of a worsening of the collective, is not insured but, instead, nanced by policyholders via premium markups. In the case of such risk-adjusted aging reserves, it is not clear how increased health care costs could be broken down in such a way that there is an actuarially fair premium adjustment for a given risk class. Furthermore, the morbidity-oriented equivalence principle is not without problems. For instance, if a policyholder suddenly contracts a chronic disease, the insurer has to adjustARr as well as P Vr(CHE) accordingly. If the policyholder wants to switch contracts, the insurer would have to transferARr. But how isARrcalculated and nanced? The morbidity-oriented equivalence principle may create a new calculation model leading to problems unrelated to the portability of aging reserves.

Since the morbidity-oriented equivalence principle may be inappropriate, searching for an alternative mechanism appears to be a fruitful line of inquiry. Intuitively, low-risk policyholders should be given a lower transfer amount than high-risk policyholders when they switch insurers. However, the big question is: Where would these amounts come from and how could a transfer mechanism be designed such that the risk allocation can be maintained? Nell and Rosenbrock (2009) argue that the current system in Germany might be stabilized when it becomes possible to solve adverse selection issues arising from the lack of portability of aging reserves. Risk selection could be corrected by transferring not only the accumulated aging reserves of a PHI

38 For instance, note that risk types cannot easily be veried at court. As a result, a new insurer would have an incentive to declare a policyholder a high risk and ask for a large transfer, while the former insurer would benet by declaring the switching policyholder to be healthy.

References

Related documents

• All health insurance plans in the Marketplace present their price and benefit information in plain language.. 3 Things to Know about

Certain statements contained in this presentation may be statements of future expectations and other forward-looking statements that are based on the management’s current

The Australian PMI ® average wages index rose again in May (seasonally adjusted), indicating a faster pace of increase in wages across the manufacturing sector,

The learning mechanisms identified suggest that through knowledge creation within an organisation and knowledge transfer across the supply chain and broader network of

Many studies have been conducted to determine the critical success factors for project execution in infrastructure related construction projects.. However, there is

Here, we demonstrate a recipe to fabricate a liquid phantom for near infrared spectroscopy and diffuse correlation spectroscopy with well-defined optical (absorption and scattering)

The linguistic whitewashing in the NewSouth edition expresses notions of “otherness” and “difference” bell hooks has deconstructed in her essay, “Culture to

We show that estimates of potentially available cropland (PAC) are highly sensitive to assumptions about: (i) the minimum estimate of current popula- tion density below which land