Journal of Health Economics 40 (2015) 109–121

Contents lists available at ScienceDirect

Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase

Who joins the network? Physicians’ resistance to take budgetary co-responsibility Maurus Rischatsch Department of Economics, University of Zurich, Switzerland

a r t i c l e

i n f o

Article history: Received 15 September 2013 Received in revised form 1 December 2014 Accepted 6 December 2014 Available online 15 December 2014 JEL classification: C5 D8 I1

a b s t r a c t Managed Care (MC) is expected to provide health care at a lower cost than conventional provision. Therefore, Switzerland intends to promote MC by forcing health insurers to write MC contracts and introducing budgetary co-responsibility for ambulatory care physicians. A discrete choice experiment conducted in 2011 including 872 physicians reveals a strong preference heterogeneity with respect to network participation and alternative remuneration schemes. The number of physicians working in networks is unlikely to rise on a voluntary basis, while general practitioners are more likely to join networks than specialists with surgical activities. For physicians considering joining networks, cost savings are predicted to be higher than the estimated willingness-to-accept payments. © 2014 Elsevier B.V. All rights reserved.

Keywords: Physician reimbursement Managed Care Physician networks Latent-class logit Discrete choice experiment

1. Introduction Over the last decade health care expenditure in most OECD countries has grown at a faster rate than the gross domestic product (GDP). Even if the national health care systems are organized differently, most countries face similar challenges. Firstly, there is a rising demand for medical services. On the one hand, an aging society and other demographical changes lead to more people in need of medical treatments and the number of patients with chronic diseases or co-morbidities is on the rise. On the other hand, additional demand originates from the coverage of uninsured citizens. A prominent example for the latter is the United States under the Affordable Care Act. Secondly, the number of innovative, but expensive therapies has increased. As a consequence, health care expenditure measured as a share of GDP increased to 17% in the United States in 2012, the highest share worldwide according to statistics from the OECD (2013). By comparison, Western European countries typically spent between 10 and 12% on health care. Only the Netherlands, France and Germany paid a higher share than Switzerland, where health insurance premiums have

E-mail address: [email protected] http://dx.doi.org/10.1016/j.jhealeco.2014.12.002 0167-6296/© 2014 Elsevier B.V. All rights reserved.

increased by 3.6% per year over the last decade according to statistics from the Federal Office of Public Health. In the aftermath of the financial crisis and in the wake of the economic downturns, financing health care became an issue for many governments. An important contribution to lower health care costs is expected through a paradigm shift in physician reimbursement from fee for service to capitation. Under conventional fee for service, medical services are remunerated according to an administrated fee schedule, while under capitation a prospective payment is paid to the health care provider, e.g. a payment per enrollee per month. Transferring more cost responsibility to health care providers is seen as one way to reduce health care expenditure due to the avoidance of unnecessary medical treatments. For this reason, the so-called Alternative Quality Contract (AQC) was designed in the United States and is currently used by the Blue Cross Blue Shield of Massachusetts. Already more than two decades ago, the United Kingdom introduced its GP fundholding allowing larger general practices to work under a global budget for hospital referrals with the possibility to retain a surplus as a bonus.1

1 These two forms of risk sharing are discussed in more detail in the literature review.

110

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

Switzerland intends to introduce similar cost incentives for health care providers through the encouragement of Managed Care (MC). MC is understood as a model to introduce incentives for health care providers and patients with the goal to increase treatment quality and to reduce health care expenditure. While MC is the dominant form of health insurance in the United States, it is less established in Europe. Even if Switzerland was the first European country to allow MC contracts in its social health insurance in the 1990s, as discussed in Beck et al. (2009), the share of capitation policies has remained low (5.5% in 2011). Although half the population signed a insurance policy accepting some kind of restriction in return for a lower premium, the promotion of MC remains on the political agenda because several studies have shown that MC leads to lower health care expenditure (e.g. Berchtold and Hess, 2006; Beck et al., 2009; Reich et al., 2012). 1.1. Institutional background The Swiss health care system is financed by a dual system composed of a compulsory basic health insurance and an additional voluntary insurance (e.g. free hospital and physician choice, private rooms in hospitals). The mandatory coverage includes most health care services and is written by about 70 competing private health insurers. The health insurers are not allowed to make profits for the compulsory coverage but for the voluntary coverage. The premium for the basic insurance depends on gender and area of residence, but not on health risk and income. Premium subsidies are granted to low-income citizens and are funded through general taxes. To set incentives for individuals not to consume unnecessary medical services, two cost-sharing incentives are installed. First, individuals have to choose one of six deductible levels that affect the final premium level. Second, a co-payment of 10% – limited to CHF 700 (USD 840 in 2011) per year – is imposed for annual costs exceeding the deductible. Insurers must accept all applicants for the mandatory coverage but are allowed to use medical underwriting to reject applicants for supplementary coverage. To mitigate risk selection for healthy risks in the basic insurance, a national risk adjustment scheme is in place, compensating insurers with a riskier portfolio than the Swiss population. Recently, a shift from conventional insurance plans to MC policies took place. The main reason for a shift in the demand toward MC policies are the lower premiums granted in return for the acceptance of certain restrictions, e.g. accepting a gatekeeper and giving up free physician choice. As a consequence, health insurers have to contract with additional physicians, physician networks, and Health Maintenance Organizations (HMOs) to enroll their portfolio. Additional information about the Swiss health insurance system can be found in Trottmann et al. (2012). In Switzerland, ambulatory care is predominantly provided by independent private practice physicians. These are mostly paid through fee for service. Only a small number of ambulatory care physicians works in MC-type arrangements, where alternative remuneration models like capitation are used, but according to Berchtold and Peytremann-Bridevaux (2010) every second general practitioner and more than 400 specialists have cooperated with an established medical group in 2010. Contracting with these organizations is getting more important today with the increasing demand for MC plans, mainly because these are cheaper than conventional plans. A common contract form is what Reich et al. (2012) call a contract model with capitation. In this set-up, a health insurer contracts with either a group of (independent) physicians organized as a network or an HMO. These organizations agree to provide health care for the enrollees and to accept a global budget, which is calculated as a per patient per year payment. As Reich et al. (2012) emphasize, the global budget is in practice a virtual cost

target and not an actual payment to the organization. The medical services are reimbursed through conventional fee for service, but the global budget (called spending target in the remainder of this article) is used to introduce a bonus/malus system as discussed in Section 3. 1.2. Health care reform In 2012, Switzerland held a referendum with the intention to increase the quality and the efficiency of the health care system mainly through a better cooperation and coordination among health care providers. The government aimed at encouraging the nationwide development of MC networks. Among other changes, the legislative proposal encompassed that health insurers have to sign contracts with physician networks to govern their cooperation, data exchange, quality assurance, and the remuneration. In addition, the legal text contained that the health care providers organized in physician networks are financially responsible for the medical provision of the network-insured individuals. In other words, the implementation of a budgetary co-responsibility for ambulatory care physicians was intended. The referendum was rejected by a strong majority of voters (76%). The main reason for rejecting the referendum was not the implementation of budgetary co-responsibility. The physician community successfully campaigned against the referendum with the argument that the reform would abolish free physician choice because the reform also intended to impose a higher co-payment for patients that were not treated in a physician network (NZZ, 2012). Even if the health care reform was rejected, budgetary coresponsibility remains of central interest for both, private health insurers as well as the government. Therefore, a better understanding of physicians’ preferences for alternative reimbursement systems and their willingness to accept budgetary co-responsibility is of great interest. Hence, the objective of this article is to measure physicians’ preferences to work in networks and accept budgetary co-responsibility. Willingness-to-accept (WTA) values are derived using a discrete choice experiment including 872 ambulatory care physicians surveyed in 2011. The article is organized as follows. Section 2 gives a short literature review of physician preference studies and the design of payment incentives for physicians. Section 3 discusses how budgetary co-responsibility can be designed and describes how physicians’ risk aversion can be expressed in WTA values. Section 4 presents the modeling approach, derives expected preference tendencies for or against reimbursement attributes used in the experiment, and explains the design of the discrete choice experiment. In addition, physician’s utility derived from alternative reimbursement designs and the applied econometric model used to elicit the preference weights is discussed. Section 5 describes the survey data. The estimation results are interpreted in Section 6. The concluding Section 7 compares the estimated willingnessto-accept values with potential cost saving through global budgets and outlines the implications. 2. Literature review A broad international literature on incentive pay exists, which mainly revolves around the impacts of differently designed incentive schemes to improve the performance of the employees (see Ichniowski and Shaw (2003) and Gibbons (1996) for an overview). With respect to physician payment, the focus is on installing incentives regarding the effort of ambulatory care or hospital physicians, mainly by replacing conventional fee for service by global budgets. As discussed by Robinson (2001), physician remuneration

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

through fee for service rewards inappropriate services, fraudulent up-coding of visits and procedures, and reciprocal referrals among specialists. On the other hand, capitation may lead to the refusal of expensive medical services and stimulate activities to avoid patients with chronic diseases. This makes clear that the introduction of budgetary co-responsibility for ambulatory care physicians to tackle rising health care expenditure entails certain drawbacks that might affect the treatment quality. For this reason, the Blue Cross Blue Shield (BCBS) of Massachusetts introduced a modified form of the conventional global budget models in 2009. The so-called Alternative Quality Contract (AQC) has three major components. First, a global budget with an annual spending growth limit for the next five years replaces the annually bargained budget for the subsequent year. Second, incentive payments to improve treatment quality are introduced, which was absent in the former global budget models. This offers medical groups the opportunity to make significant profits if they meet the clinical performance target set by the BCBS. Third, technical support for participating providers is granted. For a more detailed discussion on AQC see Chernew et al. (2011) who find that all medical groups were rewarded with a significant quality bonus in the first year of the AQC year. In the United Kingdom, the National Health Service (NHS) introduced the fundholding scheme in 1991 allowing larger general practices to work voluntarily under a global budget for hospital referrals. These practices received a bonus if costs were below the budget. As a consequence, they reduced unnecessary use of secondary care, increased their effort to purchase secondary care at lower rates, and decreased waiting times for hospital treatment. Nevertheless, the fundholding was abolished in 1999 because it was perceived as unfair and led to high transaction costs. Croxson et al. (2001) find empirical evidence that fundholders increased their use of secondary care in the year before entering the scheme to inflate their budgets. Dusheiko et al. (2006) find that the abolition of fundholding increased ex-fundholders’ admission rates for chargeable elective admissions up to 5%. Gaynor and Gertler (1995) investigate the determinants of the internal organization of medical group practices. Focusing on the influence of the physicians’ risk aversion on the choice of compensation method and physician panel size, they show that moral hazard is present and that moving from fee for service to capitation results in reduced physician effort, especially in large practices. They conclude that physicians’ risk aversion influences the choice between risk diversification through larger networks and potential lower productivity of the practice. Gaynor et al. (2004) study how primary care physicians in MC networks respond to financial incentives to contain medical expenditure in the United States, where MC organizations rely on a system of financial and non-financial incentives that encourage physicians to control costs. They find that group-based incentives are more effective in reducing medical expenditure for small physician panels because many group-based incentives rely to some extent on the power of peer pressure to abolish moral hazard. Budgetary co-responsibility to be effective therefore requires an optimal choice of physician panel size. The behavior of Swiss physician was analyzed in several studies including Rischatsch and Zweifel (2012) who investigate private-practice physicians’ willingness to apply some typical MC treatment concepts. They find that cost savings from the application of these concepts have to be substantial to be able to pay the compensations asked by the physicians for voluntary application. Rischatsch et al. (2013) measure the impact of financial incentives on the prescribing behavior of Swiss physicians. The study uses claims data to explain the disparity in the rate of generic substitution between dispensing and non-dispensing physicians. They show that the market share for generic drugs increases if physicians profit from selling the drugs they prescribe. In Rischatsch

111

(2014), dispensing physicians’ drug margin optimization is studied and empirical evidence is found that financial incentives from sub-optimal pharmaceutical price regulation in Switzerland lead to a cost-inefficient prescription of drugs if physicians are selling drugs on their own account. In general, these studies show that physicians are prone to financial incentives and that setting the right incentives can motivate them to work more cost-conscious. While most studies focus on measuring the success in reducing health care expenditure, the objective of this article is to analyze the required compensation payment to physicians for accepting a budgetary co-responsibility. The next section explains how incentives can be designed in physician payment mechanisms. 3. Physician payment mechanisms This section starts with the description of a general reimbursement mechanism that permits the incorporation of financial incentives to reduce health care expenditure. Depending on the choice of its parameters, the payment system reduces to conventional fee for service, the dominant form of physician reimbursement in Switzerland. Following Ellis and McGuire (1986), physician utility is assumed to depend on the network’s profit () as well as patient health benefit (B), i.e. (, B). For simplicity, patient benefit is ignored in the theoretical part, focusing only on network profit with ∂/∂ > 0. However, the effect of patient health benefit considerations on network participation is included in the empirical part. Assuming a network panel of N physicians, the network profit can be written as  = (1 − s)R + [sR + b(T − C)] 1(C ≤ T ) − [m(C − T )] 1(C > T ) + pE − C

(1)

where R indicates the network revenues, s the share of retained revenues by the health insurer (discussed below), T the spending target or global budget, and C the costs of the medical services provided. The expression 1( · ) is an indicator function, being one if the expression in parenthesis is true and zero otherwise. The share of cost savings (T − C) that the insurer pays to the network is given by b, while the share of costs (C − T) to be borne by the network is given by m. An additional source of profit is a payment per enrollee (p), which is multiplied by the total number of enrollees (E).2 For simplicity, only one medical service is considered, so that the network’s total costs are given by C = i min {ci (q), L}, where i indexes patients and ci (q) denotes patient-specific cost within a year depending on the quantity of care (q) provided. It is assumed that physicians can only influence ci by choosing q, that the annual number of patients is given exogenously, and that they do not reject patients. To cap the risk caused by patients with catastrophic costs (e.g. HIV patients), a so-called stop-loss limit (L) per patient per year is introduced, meaning that if ci > L only L contributes to C, not the actual amount ci . The health insurer pays the remainder. Network revenues are given by R = (1 + )C with  being the loading on costs to reward the network for the medical services provided. The spending target (T) is an budget for the subsequent year. It depends on the number of enrollees in the network and the predicted annual cost per enrollee (ˆc ). The target level is determined

2 In this study, we investigate physicians maximizing the network’s profit. The individual profit for each member is then given by the network-specific internal income allocation. Commonly, network members receive their own fee for service claims and participate in a gain/loss. How specific gains or losses are distributed among the members is not further discussed in this article.

112

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

by T = w(ˆc E), where w is a weighting factor taking into account network-specific characteristics (e.g., specialization and skill mix), forecasts of changes in demand, and expected price changes. The factor  represents the factor for the intended cost reduction compared to the non MC-insured collective.3 Gaynor et al. (2004) investigate an HMO payment mechanism where a certain share of the revenues is retained by the insurer until the end of the year (payment at risk) and only paid to the HMO if costs are below the spending target. Otherwise, the insurer keeps the retained fees. This type of incentive is represented in Eq. (1) by the parameter s. In Switzerland, a bonus/malus system is more common in practice that can be defined in absolute or relative terms. Because different types of physicians with different cost structures are investigated, relative bonus/malus payments are more appropriate than absolute payments. Thus, the parameter b denotes the share of the achieved cost savings that goes to the network as a gain-sharing bonus and m defines the share of the cost above a spending target that the network has to bear as a loss-sharing malus. Hence, in addition to the income contribution through , a network is able to increase profits or decrease losses by providing the optimal quantity of care and optimize costs. 3.1. State-dependent network profits One can define the state-dependent network profit to be 1 if costs are below the spending target (state 1) and 2 if costs exceed the spending target (state 2) with C1 < T < C2 . In this case, the expected profit is E[] =  [C1 + b(T − C1 )] + (1 − ) [C2 − s(1 + )C2 −m(C2 − T )] ,

(2)

where  denotes the probability that state 1 occurs. If no budget responsibility is introduced with s, b, m, and p being zero, the networks are reimbursed by conventional fee for service and Eq. (2) reduces to E[F ] = [C1 ] + (1 − )[C2 ].

(3)

One can single out the uncertainty caused by budgetary coresponsibility, which is reflected by E[B ] =  [b(T − C1 )] + (1 − ) [−s(1 + )C2 − m(C2 − T )] ,

Fee for service constitutes the reference reimbursement system in the discrete choice experiment because it is the dominant reimbursement mechanism for independent private practice physicians in Switzerland and does not include budgetary co-responsibility (see Section 4.2). The profit function given by Eq. (1) reduces to  = C.5 Thus, physicians may be incentivized to provide a higher quantity of care than optimal leading to an increase in ci , which constitutes the classical supplier-induced demand problem as discussed in McGuire (2000). 3.3. Alternative reimbursement designs Budgetary co-responsibility can be modeled in different ways. A first example is described and analyzed by Gaynor et al. (2004) and can be modeled using Eq. (1) with positive values for s, b, and , so that the network profit reads  = (1 − s)R + [sR + b(T − C)] 1(C ≤ T ) − C,

(6)

with R = (1 + )F and  = 0.25. The spending target is derived as discussed previously. The stop-loss provision is limited at USD 15,000 per patient per year and revenues are calculated based on Medicare’s fee schedule denoted by F. Bonuses are based on the performance of panels of doctors which vary between three and 30 physicians. Realized cost savings are divided in equal shares between the insurer and the network (b = 0.5). In the case that the actual costs exceed the spending target, the network’s loss sharing amounts to the retained revenues (s = 0.2). A second example replaces the retained revenues by losssharing malus. The gain-sharing bonus is sometimes designed to be more generous than the loss-sharing malus because insurers are better able to manage risks than are networks. Services are paid through fee for service (income contribution) where revenues are R = (1 + )C. In addition, a payment per enrollee is paid. With this specification, Eq. (1) reduces to  = C + b(T − C)1(C ≤ T ) − m(C − T )1(C > T ) + pE.

(7)

This second example constitutes the most common form of budgetary co-responsibility in Switzerland today. Therefore, the discrete choice experiment is modeled using Eq. (7).

(4)

with E[] = E[F ] + E[B ]. Assuming risk averse physicians, a von Neumann–Morgenstern (VNM) utility function (·) permits a comparison between certain and uncertain profits and the measurement of the compensation required to bear the higher risk through budgetary co-responsibility. The VNM utility function can be written as

(E[] − P) =  (1 ) + (1 − )(2 ),

3.2. Conventional fee for service

(5)

where P is the risk premium or willingness to pay to avoid the uncertainty (compare Zweifel and Eisen, 2012, Chapter 2). The risk premium is modeled in the reimbursement mechanism by the inclusion of a payment per enrollee (pE). This should compensate the physicians for the uncertainty introduced through the budgetary co-responsibility. A more detailed discussion of the compensation that serves as the price attribute in the discrete choice experiment is given in Section 4.4

3 For simplicity, only one group of enrollees is modeled. In reality, many different groups are applied depending e.g. on age, gender, and health care expenditure in the previous year. 4 Because the bonus/malus system allows networks to make profits and losses, prospect theory might be an alternative approach to model physicians’ preferences.

4. Modeling approach 4.1. Derivation of expected preference tendencies Referring to Eq. (7), a health insurer can transfer more or less budgetary responsibility to the network by choosing different values for b, m, and T. In the following, all parameters pertaining to the reimbursement design will be called reimbursement parameters. This section derives the expected signs for the preference weights under the assumption that ∂/∂ > 0. Because payments at risk and changes in mark-ups are not of interest in the Swiss context, these attributes are neglected. It is assumed that physicians are not completely certain which state, state 1 or state 2, they will end up in, hence 0 <  < 1. The first derivative of the expected profit with respect to the gainsharing bonus parameter is given by ∂E[]/∂b = (T − C1 ) > 0. Thus, physicians are predicted to prefer higher bonuses so that ˇ1 > 0 is expected.6 Increasing the loss-sharing malus parameter m is

5

In this case the reimbursement parameters s, b, and p are set to zero. We assume that C1 < T < C2 so that an indifference arising from C1 = T = C2 is excluded. 6

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

113

Table 1 Summary of attributes, expected signs, attribute levels, and units Attributes

Abbr.

E[sign]

Description

Attribute levels

Bonus Malus Stop-loss limit Panel size Target reduction Payment

BON MAL LIM PHY TAR PAY

ˇ1 ˇ2 ˇ3 ˇ4 ˇ5 ˇ6

Share of cost savings paid to network Share of excess cost borne by network Stop-loss limit per patient and year Number of physicians with same target Reduction compared to non-MC cost Risk payment per insured per month

0, 10, 20, 30, 40, 50 in % 0, 10, 20, 30, 40, 50 in % 10, 20, 30 in CHF 1000 10, 50, 100 physicians 0, 5, 10, 15, 20, 25 in % 0, 0.5, 1, 1.5, 2, 2.5 in CHF

> 0 < 0 < 0 ≶ 0 0

undesired by physicians because ∂E[]/∂m = − (1 − )(C2 − T) < 0. A higher malus parameter means that physicians have to bear a higher share of the costs above the target if they end up in state 2, and therefore ˇ2 < 0 is predicted. The stop-loss limit enters Eq. (7) through C = i min {ci , L} and therefore ∂C/∂L > 0 if ci ≥ L. Therefore, physicians are predicted to dislike an increase in the stop-loss limit, hence ˇ3 < 0. According to Gaynor et al. (2004), the number of physicians in a panel working under the same target is an important decision variable for physicians joining a network. Implementing network-level incentives introduces the problem of moral hazard and the intensity of incentives might depreciate as the panel size increases. Hence, the number of physicians working under the same target is included as a reimbursement parameter. However, Eq. (7) does not allow deriving predictions for the network size (ˇ4 ) but considering risk diversification and moral hazard one may expect a preference for medium size panels. A higher spending target is preferred by the physicians because ∂E[]/∂T = b + (1 − )m > 0 and therefore ˇ5 > 0 as long as at least b or m is positive. The spending target is modeled in the choice experiment as target reductions and therefore, the coefficient ˇ5 is expected to be negative. Finally, increasing the payment per insured per month p is predicted to have a positive effect on physician utility (ˇ6 ), because ∂E[]/∂p = E > 0. The predicted preference tendencies or expected signs of preference weights are summarized in Table 1. 4.2. Design of the choice experiment Physicians’ preferences for or against budgetary coresponsibility are measured using a discrete choice experiment (DCE). In the following, the choice of attributes and their levels is explained. 4.2.1. Attributes and their levels The gain-sharing bonus representing the share of potential cost savings (target minus actual cost) and the loss-sharing malus measuring the share of excess cost (actual cost minus target) to be borne by the network is modeled using the attributes BON and MAL with six levels ranging from zero to 50%. Higher values than 50% are unlikely to be implemented in Switzerland and therefore not considered. In the Swiss context, the stop-loss limit (L) for general practitioners is commonly about CHF 10,000 per patient per year (1 CHF ≈ 1.20 USD in 2011). For specialized physicians costs are more likely to reach this limit than for general practitioners. Therefore, the stop-loss limit (LIM) is modeled to have three levels: CHF 10,000, 20,000 and 30,000 per patient per year. For an international comparison, the HMO analyzed by Gaynor et al. (2004) limits the stop-loss provision at USD 15,000 (≈ CHF 13,630) for primary care physicians. The attribute PHY represents the size of the physician panel. On the one hand, its inclusion in the experiment is important because the intensity of incentives might decrease with panel size and imposing network-level incentives may introduce moral hazard. On the other hand, larger networks are better able to manage financial risks. This was already discussed in the literature review. Referring again to Gaynor et al. (2004), the physician panels there

range from three to 30 physicians. The experiment uses three levels to model network size: 10, 50 and 100 physicians. In Switzerland, only a few of the largest physician networks have a panel size of approximately 50 physicians. To model realistic scenarios in the Swiss context and to measure the preferences for larger networks, an upper level of one hundred physicians was included in the experiment. The levels where discussions with physicians and health insurance representatives to ensure realistic scenarios. The spending target (TAR) is included as percentage reductions of the reference budget determined upon the cost of comparable non-MC insured citizens. The reduction in the spending target ranges from zero to 25% using intervals of 5%. The highest level is believed to represent an upper bound for cost savings through MC based on the findings by Reich et al. (2012). Finally, monetary compensation for the additional financial risk is included as risk payments (PAY) per insured per month ranging from zero to CHF 2.50 per insured per month (see Table 1). 4.2.2. Experimental design The total of six attributes and their levels combine to form 11,664 (=64 × 32 ) possible combinations of alternative reimbursement designs. SAS’s statistical software package JMP was used to generate a D-efficient design. The coefficients were assumed to be zero and the design was specified to have 80 alternatives. The resulting alternatives were randomly split into four blocks, each with 20 alternatives. No fixed comparator alternative was used and the 20 alternatives randomly paired to constitute ten binary choices between the hypothetical network alternatives A and B (compare choice y1 in the sequential choice set-up discussed next). To reduce the cognitive burden for respondents, only up to three attributes were allowed to differ at the same time.7 4.2.3. Sequential choice set-up In each of the 10 choice scenarios of each block, respondents had to make two sequential decisions. First, two of the generated alternatives were presented. The respondent was asked to make a first binary choice (y1 ) between Network A and Network B differing only in the reimbursement attributes. This choice constitutes a forced choice because no opt-out alternative is given. Second, the respondents were asked to make a conditional choice (y2 |y1 ) between the previously chosen network (Network A or B) and the opt-out alternative representing to work independently. The second choice of each scenario is an unforced choice. The sequential choice set-up with a forced followed by an unforced choice was designed to mitigate strategic behavior by choosing always the opt-out alternative. Under y1 , respondents have always to trade off liked and disliked attributes so that no alternative is available for strict protesting. Without this sequential set-up, the discrete choice experiment might have resulted in

7 A first design was generate using NGENE. Unfortunately, the software was not able to handle the mentioned restriction, which is the reason why JMP was used instead.

114

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

Table 2 Example of the sequential choice scenario Network A

Network B

50%

30%

20%

20%

CHF 10,000

CHF 10,000

10 physicians

50 physicians

10%

10%

CHF 1.50

CHF 1.00

Choice 1 (y1 ): I prefer to join...

Network A 

Network B 

Choice 2 (y2 ): I prefer to... remain independent 

join preferred network above 

Share of cost savings paid to network as a bonus Share of excess cost borne by network as a malus Stop-loss provision limit per patient and year Number of physicians with same target Target reduction compared to non-MC cost Risk payment per insured per month

whether network participation differs. To control for geographical differences, practice location is considered discriminating between urban, suburban (SUBURB), and rural (RURAL) areas. In the past, numerous studies (and referendums) have shown that preferences differ between the language areas. Hence, an indicator variable for German-speaking physicians (GERM) is used, where French- and Italian-speaking physicians constitute the reference group. Finally, the stated self-assessment of the probability of producing a cost below the comparable practices (P(C < T), compare  in Eq. (2)) and a covariate indicating if respondents think that the cost discussion in Switzerland leads to a decrease in the treatment quality (HEALTH) is included. The latter variable is coded as minus one if the respondent believes the statement is wrong, zero if indifferent, and one if it is believed that the statement is true. The class-membership equation used to explain segmentation of physicians into classes is ˛ z = ˛1 SP1 + ˛2 SP2 + ˛3 PSY + ˛4 MALE + ˛5 AGE + ˛6 MARR

only collecting information about network opposition measured by the model constant regardless of the reimbursement attributes. The second choice is included to measure overall resistance. Table 2 shows an example of a choice scenario with its two sequential choices y1 and y2 .

+ ˛7 SUBURB + ˛8 RURAL + ˛9 GERM + ˛10 P(C < T ) + ˛11 HEALTH + ˛12 CONST and explained in the next section. 4.4. Applied finite mixture model

4.3. Specification of physician utility Preferences are measured based on the theory of stochastic utility maximization (Luce, 1959; Manski, 1977; McFadden, 1981). The utility of alternative payment mechanisms is defined as a function of alternative-specific attribute levels. While a physician’s utility from network participation is specified as being linear in the payment attributes, the utility from independent private practice (opt-out option) is given by an alternative-specific constant (compare Marshall et al. (2009)). More precisely, a constant indicating the network option chosen under the first choice is included in the second choice (NETW) to measure physician resistance against working in networks. An additional constant was included for the first binary choice, controlling for unmeasured preference in favor of network B versus A (CONSTB) even though the pertaining coefficient is expected to be insignificant due to the experimental design. The basic model specification of the deterministic part of the random utility is

The preference parameters of interest are estimated using a latent-class logit model. In addition, a random-coefficient logit model is estimated as a reference model, but not discussed in detail here. The latent-class logit model assumes that the observed choices come from two or more unobserved classes of physicians and where the proportions of class size are unknown a priori as discussed in Wedel and DeSarbo (2002). In contrast to the randomcoefficient model, the latent-class model incorporates preference heterogeneity among physicians by estimating class-specific coefficients. The latent-class logit model can be extended to include class-membership (also called concomitant) variables to explain class membership by physician-specific characteristics as done by Kamakura et al. (1994), Leisch (2004), and Gruen and Leisch (2008). The choice probabilities are given by Pin =

 T K   k=1

ˇ x = ˇ1 BON + ˇ2 MAL + ˇ3 LIM + ˇ4 PHY + ˇ5 TAR + ˇ6 PAY + ˇ7 CONSTB + ˇ8 NETW.

(9)

(8)

Different functional forms are tested and discussed in Section 6.1. Depending on the econometric model, Eq. (8) is modified. The stop-loss limits and the sizes of the physician panels are included as categorical attributes. To be able to interpret the network constant independently from the preferences for the reference levels of these attributes, they are coded using effects coding. For a discussion on effects coding see Louviere et al. (2000) and Bech and Gyrd-Hansen (2005). Preference heterogeneity is modeled using a finite number of physician classes with class-specific preference parameters. To explain class membership of physicians, several socio-economic variables are used. Physician’s specialization is the first segmentation variable and used to distinguish between general practitioners who constitute the reference category and are compared to specialists without (SP1) and with (SP2) surgical activities, as well as psychiatrists (PSY). In addition, physician gender (MALE), age (AGE), and marital status (MARR) are used to investigate

t=1

exp(ˇk xint )

J

j=1

exp(ˇk xjnt )



·

exp(˛k zn )

K

k=1

exp(˛k zn )

,

(10)

where the first term is the logit probability of physician n choosing alternative i at time t given to belong to class k (Pin|k ). The second term of Eq. (10) constitutes the class-membership function (Pnk ), which explains the composition of classes using physician-specific covariates and is given by Eq. (9). The model was estimated using the flexmix package in R written by Gruen and Leisch (2008) and parameter estimation is performed using the EM algorithm. 5. Data The discrete choice experiment was included in online survey and distributed using an e-mail newsletter. The anonymous survey was addressed to approximately 11,000 Swiss physicians working in ambulatory care. A pretest was conducted to check the respondents’ understanding of the description of the attributes and their levels and the questionnaire in general. Some comments disclosed the sensitivity of the subject. Only some minor changes were necessary (e.g. wording) before the main survey was finally fielded in August 2011. The survey was fielded on a Wednesday evening because most physicians have their administration day on Thursday and the timing was expected to increase the response rate. The

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

survey was not personalized because of the sensitivity of the topic and no reminders were sent out. The survey mode was chosen with the knowledge of some limitations of web surveys, but avoiding to mail a paper survey to thousands of physicians decreased response time and cost and a higher response rate was expected. The rate of return of 14% was slightly higher than expected based on previous physician surveys. The sample size of physicians to be addressed was in fact calculated using an expected response rate of 10%. This rate was anticipated based on the previous experience with physician surveys in Switzerland made by the author as well as the Swiss Medical Association. All forced and unforced choices combined, a total of 15,748 choices from 872 physicians are observed and used in the estimation of the latent-class model, some participating only in the questions prior to the experiment. The share of choices in favor of Network B versus A (forced choices) is 49.6%. This reflects the fact that no fixed comparator alternative is used and alternatives are randomly paired. The share of unforced choices in favor of the previously chosen network alternative versus remaining independent is 19.2%. This already reveals physicians’ strong preference to remain independent. The share of physicians that made at least once a choice in favor of a network and once in favor of the independent practice option is 37.4% compared to 3.5% that always preferred the network alternative and 59.1% that always preferred the independent practice option. This shows a strong resistance against physician networks in Switzerland. The distribution of respondents between blocks – not classes – in the final sample is more or less uniform with 212 physicians (3891 choices), 235 (4246), 225 (3959), and 200 (3652), respectively. General practitioners (including gynecologists and pediatricians) account for 57% of the respondents, specialists without surgical activities for 13%, specialists with surgical activities for 13%, and 17% are psychiatrists. The majority of physicians work in independent single practice (49%). Official statistics by Kraft (2010) report a share of 63% working in single practice, indicating an undersampling of single practice physicians. About one-third work in shared practice on their own account (30%), while only 5% work in shared practice on common account. Network participation in Switzerland is low and predominately for networks where physicians continue to work on their own account (14%). Commonaccount networks are the exception (2%). The average team size in shared practice is 2.8 physicians. Males are oversampled, with a share of 81% compared to 68% from the official source cited above. The average age in the sample is 54 years and equals the average age reported by the Swiss Medical Association for all members working in practices in 2010 listed in Kraft (2010). The respondents have 26 years of job experience on average. The majority (81%) is married. Respondents are located in urban (47%), suburban (26%), and rural (27%) areas and in the German-speaking (79%), French-speaking (19%), and Italian-speaking (2%) part of Switzerland. The parameter  in Eq. (2) denotes the probability of producing an annual cost below the spending target. In the survey, respondents were asked to assess their own probability of producing a total cost (including pharmaceuticals and referrals to specialists and hospitals) lower than comparable practices. Response options ranged from zero to 100% in steps of 10%. On average, physicians stated their probabilities to be at 52% (the median category is 50%). Of those surveyed, 10% stated that they were certain to be over the target, while 8% were sure to be below. These probabilities were used to investigate whether physicians who assess their probabilities to be high are more likely to join networks with budgetary co-responsibility. Physicians were asked if they believe that decreasing health care expenditure is only possible with decreasing treatment quality. As to this, 79% think that decreasing expenditure is impossible

115

without degrading treatment quality, whereas only 11% think that reducing costs is possible without affecting quality. The remainder is indifferent. About 58% think that the cost discussion is a threat to the treatment quality, while 25% think that this statement is not true. Physicians see the highest potential for reducing health care expenditure via the prescription of generic drugs. Nevertheless, they do not like to be forced to prescribe the generic drug as found in Rischatsch and Zweifel (2012). A majority of 77% stated that they consider cost when choosing treatments for their patients while 9% do not consider cost when deciding about medical services to provide. In addition, respondents were asked if the they are currently paid to some extent through capitation, a bonus or a malus. About 7% of the sampled physicians are paid to some extent through capitation. Only a small share of Swiss ambulatory care physicians have experience with alternative payment mechanisms. Information campaigns addressed to physicians could lower the expected resistance against cost sharing. 6. Results The optimal number of latent classes is not known a priori. The literature suggests to estimate latent-class models with different numbers of classes and to keep the model with the lowest Bayesian information criterion. Based on the Bayesian information criterion shown in Table 3, the model with four classes fits the data better than the same model with three and five classes, respectively. Modeling different functional relationships for the bonus and the malus, the specification with continuous quadratic relations leads to the best goodness of fit compared to a linear continuous or categorical specification. Therefore, this model specification is retained and discussed in the following. The estimated coefficients for the model with three classes are shown in Table 5. The choices are allocated to the class with the highest posterior probability, while class membership is fixed for each physician. The estimation leads to segment sizes of 17%, 21%, 40%, and 21% of observed choices, respectively. To circumvent being trapped in a local maximum instead of reaching the global maximum of the likelihood, the estimation process was repeated five times using different starting values. 6.1. Estimated preference weights Table 4 shows the estimated class-specific coefficients together with their standard errors (se) and p-values (pv). The upper part presents the preference parameters for the reimbursement attributes (ˇ), while the lower part lists the parameters pertaining to the class-membership variables (˛). All statistically significant preference weights have the expected signs as predicted in Section 4.1. This section starts with a discussion about the major forces behind the physician segmentation. Subsequently, the discussion turns to the estimated preferences for the reimbursement attributes. 6.1.1. Physician segmentation Overall network aversion measured by the network constant is revealed to be a major driver for physician segmentation. Using the results of an initial estimation, the classes were ordered according to the network aversion and the same model was re-estimated. This facilitates the interpretation of the class-membership variables because the reference class is now the one that is most likely to join a network if reimbursement attributes are neglected. While the first class has no statistically significant preference for or against networks in general and the second class weakly dislikes

116

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

Table 3 Information criteria for model selection Number of latent classes (k)

Functional form of bonus/malus (k = 4)

k

Bayesian

Akaike

ICL

Relationship

Bayesian

Akaike

ICLa

3 4 5

14,410.06 14,399.97 14,417.19

13,950.19 13,756.16 13,589.43

14,501.77 14,682.50 14,780.04

Cat. bonus/malus Cont. bonus/malus (linear) Cont. bonus/malus (quadratic)

14,503.53 14,476.36 14,399.97

13,675.77 13,893.86 13,756.16

14,772.27 14,753.25 14,682.50

a

a

ICL, integrated complete likelihood.

network participation, the third and the fourth class strongly dislike networks. Table 5 shows that if only three latent classes are modeled, physicians from the third and fourth class are grouped together, keeping the size of the first two classes unchanged. Looking at the reimbursement attributes elicits that the fourth class is likely to be composed of protesters that are not considering the reimbursement attributes when choosing between network alternatives. Drawing conclusions for this class should be limited to the size of the class measuring the share of physicians protesting against network participation. Therefore, the parameters for this class are not discussed. The class composition discussion is always in comparison to the reference class representing physicians with the lowest network aversion measured by the network constant. A strong impact on physician segmentation is found to be the physician’s specialty. Predicting the class membership for each individual physician permits to investigate class composition. The model estimation results in classes of the sizes of 16.5% physicians (class 1), 20% (class 2), 42.5% (class 3), and 21% (class 4). Looking at the distribution of physicians among classes based on their specialty shows that 37.3% of general practitioners are selected into the third class, while

24.8%, 21.6% and 16.3% are assigned to the second, first, and fourth class, respectively. 39.5% of specialists without surgical activity are predicted to be in the third, 26.3% in the second, 22.8% in the fourth, and 11.4% in the first class, while specialists with surgical activity are split into the third (46.0%), fourth (37.2%), first (10.6%) and second (6.2%) class. Most psychiatrists are estimated to be in the third class (59.7%) followed by the fourth class (22.8%). The remaining two classes contain only 9.4% (class 2) and 8.1% (class 1) of all sampled psychiatrists, respectively. Comparing the composition of each class across specialties shows that the first class is composed of 41.8% general practitioners, 22.1% specialists without and 20.6% with surgical activity, and 15.6% psychiatrists. In the same order of speciality, predicted shares are 37.2%, 39.5%, 9.3%, and 14.1% for the second class, 20.4%, 21.6%, 25.2%, and 32.7% for the third class, and 16.5%, 23.0%, 37.5%, and 23.0% for the fourth class. The estimated class-membership coefficients shown in Table 4 support these findings. General practitioners have the highest likelihood to be in the reference class. Physicians without surgical activities are less likely to be in this class and have about the same probability to be in the class with a weak or a strong aversion. In contrast, physicians with surgical activities are more likely

Table 4 Estimation results of latent-class logit model with four classes Latent-class logit Class 1

Class 2

Class 3

Class 4

Attributes

ˇ

se

pv

ˇ

se

pv

ˇ

se

pv

ˇ

se

pv

BON BON.SQ MAL MAL.SQ LIM20.EC LIM30.EC PHY50.EC PHY100.EC TAR PAY CONSTB NETW

0.48 −0.06 −0.06 −0.04 0.10 −0.26 0.31 0.14 0.03 0.42 −0.20 −0.04

0.09 0.02 0.09 0.02 0.07 0.07 0.06 0.06 0.07 0.05 0.07 0.19

0.00 0.00 0.51 0.03 0.18 0.00 0.00 0.03 0.63 0.00 0.00 0.81

0.80 −0.11 −0.72 0.07 0.17 −0.41 0.28 0.01 −0.72 0.43 −0.12 −0.60

0.10 0.02 0.10 0.02 0.07 0.08 0.06 0.07 0.07 0.05 0.07 0.23

0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.92 0.00 0.00 0.07 0.01

0.92 −0.14 −0.50 0.03 0.25 −0.76 0.43 0.42 −0.42 0.17 −0.31 −5.20

0.11 0.02 0.10 0.02 0.08 0.09 0.07 0.08 0.06 0.06 0.07 0.32

0.00 0.00 0.00 0.17 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

0.21 −0.04 −0.37 0.01 0.27 −0.27 −0.01 −0.79 −0.06 0.01 0.27 −5.82

0.16 0.03 0.14 0.02 0.11 0.08 0.08 0.13 0.09 0.07 0.09 0.74

0.18 0.18 0.01 0.55 0.01 0.00 0.86 0.00 0.46 0.85 0.00 0.00

˛

se

pv

˛

se

pv

˛

se

pv

0.78 −0.68 0.15 −0.44 −0.24 −0.85 1.05 0.43 0.56 −0.01 0.49 2.91

0.42 0.58 0.51 0.39 0.09 0.38 0.34 0.33 0.37 0.04 0.15 1.01

0.06 0.24 0.76 0.26 0.01 0.02 0.00 0.19 0.13 0.79 0.00 0.00

0.71 0.85 1.59 −0.33 −0.08 −0.61 0.68 0.34 −0.20 −0.03 0.89 1.98

0.39 0.40 0.39 0.36 0.07 0.35 0.31 0.28 0.29 0.04 0.14 0.90

0.07 0.03 0.00 0.36 0.30 0.08 0.03 0.24 0.49 0.44 0.00 0.03

0.87 1.38 1.40 −0.62 0.10 −0.68 0.76 0.16 −0.53 −0.01 0.77 −0.07

0.43 0.42 0.43 0.39 0.09 0.38 0.34 0.33 0.31 0.04 0.16 1.05

0.04 0.00 0.00 0.11 0.25 0.07 0.02 0.64 0.09 0.90 0.00 0.94

Class membership

Reference class

SP1 SP2 SY MALE AGE MARR SUBURB RURAL GERM P(C < T) HEALTH CONST Pnk Class sizes: choices Class sizes: physicians

2697 (17.1%) 144 (16.5%)

3370 (21.4%) 174 (20.0%)

6369 (40.4%) 371 (42.5%)

3312 (21.0%) 183 (21.0%)

Note: ˇ, ˛ are the class-specific coefficients, se are the standard errors, and pv are the p-values. The suffixes SQ and EC indicate the square-term and the effects-coded variables. BON, BON.SQ, MAL, MAL.SQ, TAR, and P(C < T) are in 10%-points, PAY in CHF per insured per month, and AGE per 5 years.

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

117

Table 5 Estimation results of latent-class logit model with three classes Latent-class logit Class 1

Class 2

Class 3

Attributes

ˇ

se

pv

ˇ

se

pv

ˇ

se

pv

BON BON.SQ MAL MAL.SQ LIM20.EC LIM30.EC PHY50.EC PHY100.EC TAR PAY CONSTB NETW

0.49 −0.06 −0.07 −0.03 0.09 −0.26 0.31 0.13 0.02 0.42 −0.19 −0.01

0.10 0.02 0.09 0.02 0.07 0.07 0.06 0.06 0.07 0.05 0.07 0.19

0.00 0.00 0.44 0.04 0.18 0.00 0.00 0.04 0.80 0.00 0.00 0.95

0.82 −0.11 −0.71 0.07 0.16 −0.42 0.30 0.03 −0.71 0.43 −0.14 −0.73

0.10 0.02 0.10 0.02 0.07 0.08 0.06 0.06 0.06 0.05 0.07 0.23

0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.61 0.00 0.00 0.04 0.00

0.66 −0.10 −0.38 0.02 0.17 −0.47 0.21 −0.05 −0.27 0.14 −0.06 −5.06

0.07 0.01 0.07 0.01 0.05 0.04 0.04 0.04 0.04 0.04 0.04 0.28

0.00 0.00 0.00 0.20 0.00 0.00 0.00 0.21 0.00 0.00 0.12 0.00

˛

se

pv

˛

se

pv

0.77 −0.66 0.11 −0.45 −0.23 −0.85 1.03 0.45 0.46 −0.01 0.51 2.91

0.42 0.58 0.50 0.39 0.09 0.37 0.34 0.33 0.36 0.04 0.15 1.00

0.06 0.26 0.83 0.25 0.01 0.02 0.00 0.17 0.20 0.75 0.00 0.00

0.76 1.08 1.51 −0.42 −0.01 −0.61 0.71 0.27 −0.35 −0.02 0.85 1.81

0.36 0.37 0.36 0.33 0.07 0.32 0.28 0.26 0.26 0.03 0.12 0.83

0.04 0.00 0.00 0.21 0.92 0.06 0.01 0.30 0.17 0.54 0.00 0.03

Class membership

Reference class

SP1 SP2 PSY MALE AGE MARR SUBURB RURAL GERM P(C < T) HEALTH CONST Pnk Class sizes: choices

2728 (17.3%)

3496 (22.2%)

9524 (60.5%)

Note: ˇ, ˛ are the class-specific coefficients, se are the standard errors, and pv are the p-values. The prefixes SQ and EC indicate the square-term and the effects-coded variables. BON, BON.SQ, MAL, MAL.SQ, TAR, and P(C < T) are in 10%-points, PAY in CHF per insured per month, and AGE per 5 years.

to be in the third class compared to the first and second class and most likely in the fourth class. A psychiatrist is most likely segmented into the third class. While gender does not affect the segmentation, older physicians are less likely to be in the second compared to the first class, but equally likely to be in one of the other classes. Married physicians are most likely in the reference class. A possible explanation is that networks allow physicians to work part time and to avoid days on emergency call, which might be important for physicians with family. While physicians located in urban and rural areas disclose the same membership probability for all four classes, physicians located in suburbs are less likely to be in the reference class. German-speaking respondents are significantly less likely to be in the fourth class compared to their French- and Italian-speaking colleagues. Otherwise, language has no explanatory power with respect to the segmentation. While the self-assessed probability of producing cost below comparable colleagues has no statistically significant effect on segmentation, physicians’ believe that the cost discussion negatively affects the treatment quality is positively correlated with the likelihood of being in a class with a stronger aversion. In addition to the latent-class model, preference heterogeneity among physicians is also modeled using a random-coefficient model. For a model description see Train (2003) and for a comparison between the latent class and the continuous mixture of logit models see Hess et al. (2011). The estimated parameters of the mixing distributions are shown in Table 6. The estimates support the findings of the latent-class model that general practitioners and specialists without surgical activity have the lowest network aversion measured by the constant followed by those with surgical activity, and psychiatrists. While gender has no explanatory power with respect to class membership in the latent-class model, the random-coefficient model estimates that males dislike

networks more than females. Older and married physicians are more likely to join networks, while suburban and rural physicians dislike networks more than urban physicians. German-speaking physicians are not statistically different from their French- and Italian-speaking colleagues. A higher self-assessed probability of being able to produce cost below comparable practices decreases the resistance to join a network. Finally, physicians who believe that the cost discussion leads to a lower treatment quality also dislike working in a network. 6.1.2. Preferences for reimbursement designs Focussing on the first three classes, all main- and square-term effects for the bonus turn out to be statistically significant with a positive but decreasing impact on the choice probability. For the malus, all main-effects coefficients take on the expected sign. For the first class, the main effect is insignificant while the quadratic effect is significant. The opposite is true for the third class. To interpret the effects-coded attributes the estimates are transformed.8 Increasing the stop-loss limit from CHF 10,000 to 20,000 has no statistical significant effect on accepting an alternative reimbursement scheme for the first two classes. An explanation for this finding is that even if a higher stop-loss limit increases the network’s revenue at risk, it also permits to achieve higher cost savings compared to non-MC insured patients. The group of physicians with a strong dislike for network participation is less likely to join a network if the stop-loss limit is increased to CHF 20,000.

8 To interpret the effects-coded attributes (ˇ), the coefficients have to be transˇj + ˇi , where i is the attribute level of interest and j is an index formed as ˇ˜i = j

for all attribute levels included. See Louviere et al. (2000) for a more detailed discussion.

118

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

Table 6 Estimation results of random-coefficient logit model Random-coefficient logit Mixing distribution Attributes BON MAL LIM20.ECa LIM30.ECa PHY50.ECa PHY100.ECa TAR CONSTB PAY NETW ×SP1 ×SP2 ×PSY ×MALE ×AGE ×MARR ×SUBURB ×RURAL ×GERM ×P(C < T) ×HEALTH

Censored normal Censored normal Normal Normal Normal Normal Censored normal Normal fixed Normal Fixed Fixed Fixed Fixed Fixed Fixed Fixed Fixed Fixed Fixed Fixed

Mean

Std.Dev.

WTA



se

pv



se

pv

Mean

Median

0.15 0.36 0.24 −0.60 0.41 0.01 −0.38 −0.12 0.31 −4.23 −0.07 −2.12 −2.71 −0.29 0.06 1.00 −0.88 −0.38 0.68 0.07 −1.19

0.02 0.02 0.04 0.04 0.03 0.03 0.13 0.03 0.03 0.36 0.14 0.19 0.20 0.13 0.03 0.13 0.12 0.11 0.13 0.02 0.06

0.00 0.00 0.00 0.00 0.00 0.81 0.00 0.00 0.00 0.00 0.62 0.00 0.00 0.03 0.03 0.00 0.00 0.00 0.00 0.00 0.00

0.30 −0.05 0.03 0.51 0.20 0.66 1.64 0.04

0.03 0.02 0.05 0.04 0.04 0.04 0.16 0.05

0.00 0.04 0.50 0.00 0.00 0.00 0.00 0.49

3.71

0.13

0.00

−0.67 1.15 0.34 2.78 −2.38 −1.13 0.00 0.40 – 13.48 0.22 6.74 8.62 0.91 −0.20 −3.18 2.82 1.22 −2.17 −0.21 3.80

−0.49 1.15 0.34 2.78 −2.38 −1.13 1.53 0.40 – 13.48

BON, MAL, TAR and P(C < T) per 10%-points, PAY in CHF per insured per month, and AGE per 5 years. Displayed  coefficients for MAL and TAR are for the negative values. a WTA calculation based on the binary-coded coefficients.

An increase from CHF 10,000 to 30,000 has a negative impact for all classes, the strongest for those physicians that dislike networks most. Physicians prefer medium and large panel sizes but do not distinguish between a panel size of 50 and 100. A reduction of the spending target is perceived negatively by the second and the third class but has no effect on the choice probability of the reference class. The reference class seems to believe in their ability to achieve cost reductions, to meet the spending targets, and receive a gain-sharing bonus. Finally, the price attribute in the experiment is statistically significant for all classes with the exception of the protesting class four. Therefore, no willingness to accept is calculated for the fourth class in the next section. The first two classes place the same positive value on the payment per insured per month, while the same payment to the third class has a significantly lower impact on network participation. 6.2. Willingness-to-accept values The previously discussed preference weights can be used to predict how much it would cost an insurer (a network or an HMO) to win physicians over to work under certain reimbursement designs and take budgetary co-responsibility. To answer this question, class-specific willingness-to-accept (WTA) values are derived and discussed. The stop-loss limit and the size of the physician panel enter the utility function as categorical variables using effectscoded variables, while the reduction of the spending target is modeled as a linear continuous variable. In these cases, the WTA values are given by the ratio of the attribute-specific coefficient and the one pertaining to the payment attribute. The estimates of the effects-coded variables are first transformed to dummy-coded variables (see Bech and Gyrd-Hansen (2005) again). In contrast, the bonus and the malus are modeled as quadratic continuous variables. Therefore, the derivation of the WTA values is different. The willingness to accept is given by (ˇ1 xm + ˇ2 (xm )2 )/ˇ3 , where ˇ1 and ˇ2 are the estimated coefficients pertaining to the main and the square term of attribute m, respectively, and ˇ3 is the preference weight pertaining to the payment attribute. Figs. 1 and 2 illustrate the derived WTA values that are discussed next. The standard errors

used for the calculation of confidence intervals (see Table 7) are estimated using the Delta method. All WTA values are measured in CHF per insured per month with CHF 1 ≈ USD 1.20 in 2011. Again, the payment coefficient for the fourth class is statistically insignificant and therefore, compensation is not discussed for this group. 6.2.1. Bonus and malus While physicians from the first class do not ask for a statistically significant compensation to join a network without budgetary responsibility, the physicians from the other classes do. The total compensation required depends on the sum of an initial payment for giving up independent practice and working in a network and the compensations required for changes in the reimbursement attributes. The initial payments derived using the network constants amount to CHF 1.38 and CHF 31.11 per insured per month for the second and the third class, respectively, while physicians in the third class are likely physicians with surgical activities. The positively valued bonus reduces the total compensation while the introduction of a malus further increases the amount required to compensate physicians for a voluntary participation. Starting the discussion with the willingness to accept a malus, the values amount to CHF 0.23–2.91 for the first class, CHF 1.51–4.47 for the second class, and CHF 2.85–11.20 for the third class. The compensation asked by the first class is relatively moderate in comparison to the one asked by the third class (see Fig. 1 and Table 7). Depending on the bonus level, compensation for accepting an alternative remuneration reduces between CHF 1.02–2.40 (class 1), CHF 1.59–3.31 (class 2), and CHF 4.68–8.92 (class 3). Theoretically, the lower valuation of a bonus of 50% compared to 40% is counterintuitive and more a result of the functional form estimated. Comparing symmetric bonus-malus combinations reveals that the gain-sharing bonus overcompensates the loss-sharing malus for combinations up to 40% in the case of the first and up to 30% in case of the third class. The required compensation is positive for higher levels. The opposite is true for the second class where only a combination of 10% is valued positively. For a comparison, the WTA values from the random-coefficient model are shown in Table 6.

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

Bonus

15

119

Malus

11.2

Willingness to accept (in CHF)

10

9.57 7.64 5.4

5

3.6 2.85

−2.3 −2.88

−2.22 −3.24

−2.4 −3.31

1.22

0.64

0.23

0

2.91 1.98

1.51

0

4.47

4.19

2.71

−1.02 −1.59

−1.76 −2.67

−4.68

−5 −6.32

Class 1 2 3

−7.65 −8.47

−8.92

−10 50

40

30

20

10

0 Level (in %)

10

20

30

40

50

Fig. 1. Class-specific WTA values for different bonus and malus levels.

Willingness to accept (in CHF)

15

Limit of CHF 20k

Limit of CHF 30k

Panel of 50 phys.

Panel of 100 phys. Target reduc. of 10pp

10 7.63

5 2.53 1.58

1.04

1.66

1.51

0.18

0.17

0 −1.83

−1.32

−1.42

−0.08

−0.68

−5

Class

−7.52

−7.58

1

−10

2 3 LIM20

LIM30

PHY50

PHY100

TAR

Attributes Fig. 2. Class-specific WTA values for stop-loss limits, panel size, and spending target reduction.

Table 7 Class-specific willingness-to-accept values Class-specific WTA values Class 1

Class 2

Attributes

WTA

95%-CI

BON10 BON20 BON30 BON40 BON50 MAL10 MAL20 MAL30 MAL40 MAL50 LIM20 LIM30 PHY50 PHY100 TAR NETW

−1.02 −1.76 −2.22 −2.40 −2.30 0.23 0.64 1.22 1.98 2.91 0.17 1.04 −1.83 −1.42 −0.08 0.10

−1.45 −2.45 −3.03 −3.23 −3.14 −0.12 0.06 0.52 1.19 1.93 −0.37 0.46 −2.61 −2.14 −0.40 −0.78

−0.59 −1.06 −1.40 −1.57 −1.45 0.59 1.21 1.92 2.77 3.90 0.72 1.61 −1.05 −0.69 0.24 0.98

Class 3

WTA

95%-CI

−1.59 −2.67 −3.24 −3.31 −2.88 1.51 2.71 3.60 4.19 4.47 0.18 1.51 −1.32 −0.68 1.66 1.38

−2.07 −3.47 −4.19 −4.26 −3.75 1.03 1.91 2.61 3.10 3.31 −0.29 0.86 −1.97 −1.33 1.22 0.37

−1.10 −1.87 −2.30 −2.36 −2.00 1.99 3.52 4.60 5.28 5.62 0.65 2.15 −0.66 −0.04 2.10 2.39

WTA

95%-CI

−4.68 −7.65 −8.92 −8.47 −6.32 2.85 5.40 7.64 9.57 11.20 1.58 7.63 −7.58 −7.52 2.53 31.11

−7.67 −12.54 −14.62 −13.95 −10.66 0.74 1.56 2.39 3.18 3.84 −0.27 2.22 −13.13 −12.98 0.90 11.39

−1.69 −2.77 −3.22 −2.99 −1.98 4.96 9.25 12.89 15.97 18.56 3.44 13.04 −2.03 −2.06 4.16 50.83

Note: WTA values are shown in CHF per insured per month, where 1 CHF ≈ 1.20 at 2011 exchange rates. The 95%-confidence intervals (95%-CI) were estimated using the delta method. The effects coding was considered when calculating the WTA values. WTA for the target reduction is in 10%-points.

120

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121

6.2.2. Stop-loss limit, physician-panel size and target reduction Increasing the stop-loss limit from CHF 10,000 to 20,000 per patient per year has no statistically significant effect on network participation as indicated by the confidence intervals in Fig. 2. As discussed previously, this imposes a higher risk but it also permits networks to achieve higher cost savings that translate into higher gain-sharing bonuses. However, increasing the limit from CHF 10,000 to 30,000 requires a compensation payment to the tune of CHF 1.04 (class 1), CHF 1.51 (class 2), and CHF 7.63 (class 3). The third class that is more likely composed of physicians with surgical activity shows a willingness to accept that is five and seven times higher than the one of the other two classes. Increasing the panel size of physician working under the same target from 10 to 50 or 100 physicians, respectively, decreases the required compensation for all three classes. While physicians from the first and second class lower their demand by CHF 1.83 and CHF 1.32 for a panel of 50 physicians, those from the third class are willing to give up CHF 7.58. The compensation reduces by CHF 1.42 (class 1), CHF 0.68 (class 2), and CHF 7.52 (class 3) if the panel size increases from 10 to 100 physicians. Acceptance of a lower spending target requires a compensation payment for the second and the third class but not for the first. The compensation amounts to CHF 1.66 per 10%-points (pp) reduction for the second class and CHF 2.53 for the third class.

7. Discussion and conclusion In Switzerland, increasing health care expenditures have fueled political debates on possible regulatory changes to promote Managed Care (MC). Several studies have shown that gatekeeping patients to health care providers and remunerating physicians through capitation instead of conventional fee for service contributes to lower health care expenditure and a higher treatment quality. While MC constitutes the dominant form of health insurance in the United States, it was less dominant in Switzerland until recently. This has changed between 2008 and 2013 during which period the market share for gatekeeping policies with capitation has more than doubled. Even though MC is gaining popularity in Switzerland, two major changes in regulation are discussed to further promote its market share. First, some politicians want to force health insurers to write MC contracts. Second, the introduction of budgetary co-responsibility for ambulatory care physicians is believed to encourage physicians to better control costs. Co-responsibility embodies that financial responsibility is shared between the health insurer and the physicians arranged in physician networks. A grouping of physicians to networks is necessary for risk diversification because single events are less influential with an increasing number of patients enrolled in the network. If working in physician networks under a global budget is voluntarily, a pivotal question is if the required payments to compensate physicians for the additional financial risk are lower than the potential cost savings. In this article, the required willingness-to-accept values are derived and compared with potential cost savings from the health insurer’s perspective. The discrete choice experiment, completed by 872 physicians in 2011, reveals a strong preference heterogeneity among Swiss ambulatory care physicians. On the one hand, a first group of physicians – mainly general practitioners – is indifferent between working in a network or in independent practice as long as no budgetary responsibility is imposed, but preferences change depending on the design of the reimbursement mechanism. On the other hand, a major share of physicians – mainly specialists with surgical activities – are unwilling to work in a physician network regardless

of how the reimbursement scheme is designed. Between these extremes, two additional physician classes are identified: one class that asks for a relatively low compensation to join a network without taking budgetary responsibility and one class that has to be strongly compensated. Using the estimated willingness-to-accept values and assuming a reimbursement design with a symmetric bonus and malus, a stop-loss limited of CHF 10,000 per insured per year, a panel of 50 physicians, and a spending target that is 20% below the conventionally insured collective, compensation payments between CHF 3.30–4.97 (low-compensation class) and CHF 26.34–33.47 (high-compensation class) are required, depending on the magnitude of the bonus and the malus. Again, one class of physicians accepts such a design without asking for a compensation, while a group of protesters is reluctant to change its willingness to join a network in return for the risk payments modeled in the choice experiment. Comparing the estimated willingness-to-accept values with the potential cost savings permits to assess if the introduction of budgetary co-responsibility contributes positively or negatively to the health insurers’ profit as long as independent practice without financial responsibility remains an alternative. Reich et al. (2012) estimate cost containments for capitated health insurance policies to be 21%. According to official statistics provided by the BAG (2011) average annual claims cost per MC insured adult (above 26 years) amounted to CHF 1,922 in 2011, excluding co-payments and deductibles paid by the insured. Assuming that this MC collective has 20% lower costs than the conventionally insured collective due to efficiency gains – adjusted for the composition of the portfolios – about CHF 40 per insured per month can be saved by the insurer with a target reduction of 20%. Comparing these savings with the estimated willingness-to-accept values reveals that health insurers are likely to achieve net savings by contracting with physician networks. Assuming 600 enrollees per physician – a realistic value for Switzerland – a physician from the low-compensation class and very likely a general practitioner would receive CHF 3.42 per insured per month (or CHF 2052 per month) for the previously discussed reimbursement design in combination with a bonus and a malus of 20%. According to Reichert (2010), the average monthly income for a Swiss general practitioner is CHF 16,100. Therefore, the payment would lead to a substantial increase in income by 13%. This shows that it would be profitable for both, general practitioners and health insurers, to contract on such a reimbursement design as long as physicians prove to be able to reduce treatment cost to meet the spending target. Again, considering the estimated efficiency gains by Reich et al. (2012) this seems possible. Nevertheless, the large share of physicians with a relatively high compensation demand (42.5%) and the strong aversion group (21%) show that a major acceptance of budgetary coresponsibility is unlikely to be achieved on voluntary basis. Acknowledgements The author would like to express his thanks to Dr. med. I. Cassis (FMH) for making the data collection possible. Support by Esther Kraft is also gratefully acknowledged. Special thanks go to Prof. Dr. P. Zweifel and my former colleagues M. Trottmann, H. Telser, Y. Schneider, and Ph. K. Widmer for their very helpful comments. The findings, interpretations and conclusions expressed in this article are entirely those of the author. References BAG, 2011. Statistik der obligatorischen Krankenversicherung 2011. In: Statistik der Krankenversicherung., pp. 70.

M. Rischatsch / Journal of Health Economics 40 (2015) 109–121 Bech, M., Gyrd-Hansen, D., 2005. Effects coding in discrete choice experiments. Health Economics 14 (10), 1079–1083. Beck, K., Käser, U., Trottmann, M., Rotz, S.v., 2009. Effizienzsteigerung dank Managed Care? (Efficiency increase through Managed Care?). Data Master 5, 15–21. Berchtold, P., Hess, K., 2006. Evidenz für Managed Care. In: Technical Report. Swiss Health Observatory., pp. 16. Berchtold, P., Peytremann-Bridevaux, I., 2010. Integrated care organizations in Switzerland. International Journal of Integrated Care, 11. Chernew, M.E., Mechanic, R.E., Landon, B.E., Gelb Safran, D., 2011. Private-payer innovation in Massachusetts: the ‘Alternative Quality Contract’. Health Affairs 30 (1), 51–61. Croxson, B., Propper, C., Perkins, A., 2001. Do doctors respond to financial incentives? UK family doctors and the GP fundholder scheme. Journal of Public Economics 79, 375–398. Dusheiko, M., Gravelle, H., Jacobs, R., Smith, P., 2006. The effect of financial incentives on gatekeeping doctors: evidence from a natural experiment. Journal of Health Economics 25, 449–478. Ellis, R.P., McGuire, T.G., 1986. Provider behavior under prospective reimbursement. Journal of Health Economics 5, 129–151. Gaynor, M., Gertler, P., 1995. Moral hazard and risk spreading in partnerships. RAND Journal of Economics 26, 591–613. Gaynor, M., Rebitzer, J.B., Taylor, L.J., 2004. Physician incentives in health maintenance organizations. Journal of Political Economy 112 (4), 915–931. Gibbons, R., 1996. 1996 Incentives and careers in organizations. In: NBER Working Paper Series 5705. Gruen, B., Leisch, F., 2008. Flexmix version 2: Finite mixtures with concomitant variables and varying and constant parameters. Journal of Statistical Software 28 (4), 1–35. Hess, S., Ben-Akiva, M., Gopinath, D., Walker, J., 2011. Advantages of latent class over continuous mixture of logit models. In: Working paper. Ichniowski, C., Shaw, K., 2003. Beyond incentive pay: insiders’ estimates of the value of complementary human resource management practices. Journal of Economic Perspectives 17, 155–180. Kamakura, W., Wedel, M., Agrawal, J., 1994. Concomitant variable latent class models for conjoint analysis. International Journal of Research in Marketing 11 (5), 451–464. Kraft, E., 2010. 30’273 Ärztinnen und Ärzte für die Schweiz. Schweizer Ärztezeitung 92 (12), 440–444. Leisch, F., 2004. Flexmix: a general framework for finite mixture models. Journal of Statistical Software 11 (8), 1–18. Louviere, J.J., Hensher, D.A., Swait, J.D., 2000. Stated Choice Methods. Analysis and Applications. University Press, Cambridge. Luce, D., 1959. Individual Choice Behavior. Wiley and Sons, New York.

121

Manski, C.F., 1977. The structure of random utility models. Theory Decision 52, 229–254. Marshall, D.A., Johnson, F.R., Kulin, N.A., Özdemir, S., Walsh, J.M., Marshall, J.K., Bebber, S.V., Phillips, K.A., 2009. How do physician assessment of patient preferences for colorectal cancer screening tests differ from actual preferences? A comparison in Canada and the United States using a stated-choice survey. Health Economics 18 (12), 1420–1439. McFadden, D., 1981. Econometric models of probabilistic choice. In: Manski, C., McFadden, D. (Eds.), Structural Analysis of Discrete Data with Applications. The MIT Press, Cambridge, MA, pp. 198–272. McGuire, T.G., 2000. Physician agency. In: Culyer, A., Newhouse, J. (Eds.), In: Handbook of Health Economics, vol. 1. Amsterdam, North Holland, pp. 461–536 (chapter 9). NZZ, 2012. Wuchtiges Nein zur Managed-Care-Vorlage. In: Neue Zürcher Zeitung 17. Juni. OECD, 2013. Health at a Glance: Europe 2013. OECD Publishing. Reich, O., Rapold, R., Flatscher-Thoeni, M., 2012. An empirical investigation of the efficiency effects of integrated care models in Switzerland. International Journal of Integrated Care 12, 1–12. Reichert, M., 2010. Einkommensverhältnisse der freien Ärzteschaft der Schweiz in den Jahren 2006 [neu] und 2005 [Re-Evaluation]. Schweizer Ärztezeitung 91, 479–487. Rischatsch, M., 2014. Lead me not into temptation: drug price regulation and dispensing physicians in Switzerland. European Journal of Health Economics 15, 697–708. Rischatsch, M., Trottmann, M., Zweifel, P., 2013. Generic substitution, financial interests, and imperfect agency. International Journal of Health Care Finance and Economics 13 (2), 115–138. Rischatsch, M., Zweifel, P., 2012. What do physicians dislike about Managed Care? Evidence from a choice experiment. European Journal of Health Economics 14 (4), 601–613. Robinson, J.C., 2001. Theory and practice in the design of physician payment incentives. The Milbank Quarterly 79, 149–177. Train, K.E., 2003. Discrete Choice Methods with Simulation. University Press, Cambridge. Trottmann, M., Zweifel, P., Beck, K., 2012. Supply-side and demand-side cost sharing in deregulated social health insurance: which is more effective? Journal of Health Economics, 231–242. Wedel, M., DeSarbo, W.S., 2002. Mixture regression models. In: Hagenaars, J.A., McCutcheon, A.L. (Eds.), Applied Latent Class Analysis. Cambridge University Press, pp. 366–382. Zweifel, P., Eisen, R., 2012. Insurance Economics. Heidelberg, Springer.