Production of Physician Services under Fee-For-Service and Blended Fee-For-Service: Evidence from Ontario’s Natural Experiment N. H. Somé,1,2 R. A. Devlin,4 N. Mehta,3 G. Zaric,1,5 A. Thind,1,2 S. Shariff,2 A. Garg,1,2 L. Li,2 B. Belhadji,6 S. Sarma1,2 1

Department of Epidemiology & Biostatistics, University of Western Ontario, London, ON Institute for Clinical Evaluative Sciences, Toronto, ON 3 Department of Economics, University of Western Ontario, London, ON 4 Department of Economics, University of Ottawa, Ottawa, ON 5 Richard Ivey School of Business, University of Western Ontario, London, ON 6 Strategic Policy Branch, Health Canada, Ottawa, ON 2

April 30th, 2018

Abstract We examine family physicians’ response to financial incentives in the production of medical services in Ontario, Canada. We rely on health administrative data covering 2003-2008 fiscal years, a period of reform during which family physicians could choose between the traditional fee-for-service (FFS) and blended FFS known as the Family Health Group (FHG) model. Under FHG, FFS physicians are incentivized to provide comprehensive care and afterhours services. We develop a theoretical model to understand the behaviour of physicians who switched from FFS to FHG. Empirically, a two-stage estimation strategy teases out the impact of switching from FFS to FHG on physicians’ service production. In the first stage, we account for the differences between switchers and non-switchers using a propensity score matching model. In the second stage, we use panel-data regression models to account for both observed and unobserved heterogeneity affecting service production. Our results show that switching from the FFS to the blended FFS increases service production in the range of 2% to 5% per annum. These results are robust to a variety of model specifications and alternative matching methods. Our results underscore the importance of financial incentive mechanisms for influencing the production of physician services.

1

1 Introduction Like many developed countries, access to family physicians is a concern to a large number of residents of Ontario, Canada. In early 2000s, approximately 57% of Ontarians could not see their primary care provider the same day or next day when they are sick and 52% find it difficult to access care in the evenings or on weekends (Ministry of Health and Long-Term Care (MOHLTC), 2015). In response to this lack of adequate access, the government of Ontario has introduced primary care reform, including family physician’s compensation systems to include new incentives and requirements for after-hours care in order to influence physician’s service production. Prior to 2000, the vast majority of family physicians in Ontario have been paid according to a fee-for-service (FFS) payment system, under which a physician receives a fixed fee for each service provided in accordance with a provincially defined schedule of benefits. In July 2003, the government of Ontario introduced a blended FFS model known as the Family Health Group (FHG), whereby physicians continue to receive 100 percent of their FFS payments plus extra incentives to provide comprehensive care (i.e. health assessment, preventive care services such as immunizations and cancer screenings, diagnosis and treatment) and are required to provide services during after-hours (i.e., opening clinic outside of the regular office hours, such as evenings, weekends and holidays). Understanding how physicians respond to such incentives has important policy implications for the supply of healthcare services. Many analyses have focused on comparing some measure of physicians’ output across different payment systems (Campbell et al., 2007; Devlin and Sarma, 2008; Dumont et al., 2008; Glazier et al., 2009; Kantarevic et al., 2011; Kralj and Kantarevic, 2013; Li et al., 2014; Sarma et al., 2010; Sutton et al., 2010), but the literature to date has been unable to determine the extent to which financial incentives affect the mix of “incentivized” (i.e., comprehensive care and after-hours services) and “non-incentivized” medical services within the FFS environment.1 We undertake a comprehensive study of the production of “incentivized” and “non-incentivized” services following the introduction of the FHG in Ontario. We use Ontario health administrative data containing information on the number and the type of services individual physicians provide during regular- and after-hours periods, the fees In this paper, “incentivized” refers to those services in which physicians practicing in FHGs receive extra incentives, and “non-incentivized” services refers to standard FFS payments to physicians practicing under FFS and FHGs. 1

2

paid for those services, as well as physicians’ and their patients’ characteristics to evaluate the impact of switching physicians from FFS to FHG on total services, comprehensive care services, after-hours services, and non-incentivized services. To date, relatively little is known about the relationship between physicians’ compensation scheme and access to both after-hours and comprehensive care. Recent studies on Australian physicians’ shed some light on characterizing physicians who provide after-hours services (Broadway et al., 2016; Pham and McRae, 2015). Pham and McRae (2015) found that physicians who were employees rather than partners in a practice, female, older or lived in urban areas were less likely to provide services during afterhours. Physicians in solo practice or who were partners in the practice were more likely to provide after-hours services. Australian family physicians provided more after-hours services if their hourly earnings in that setting increases, according to Broadway et al.( 2016); but physicians with a high earning in weekday-daytime practice were unlikely to provide after-hours services. We develop a theoretical model of physician behaviour in a multitasking environment. In this model, physicians choose the total hours of work and then allocate those working hours to different services (comprehensive care, afters-hours care and non-incentivized services). The model generates gross revenue functions (representing total production of services) that depend on total hours, number of enrolled or rostered patients, and a wage index capturing the marginal return to an hour worked. The gross revenue functions predict the ability of physicians to produce the mix of services under FFS and FHG payments, allowing us to derive FHG participation decision rule. A rational physician chooses the payment system which maximizes his/her gross revenue for a given number of hours devoted to work. A comparison of physician’s gross revenues in FFS and FHG reveals that physicians are more likely to switch to FHG if the proportion of patients they are able to enroll is higher than the opportunity cost of an hour worked under FFS payment system. Our empirical strategy relies on a two-stage estimation approach. In the first stage, we account for observed differences between switchers and non-switchers using propensity score matching (PSM) to render control and treatment groups comparable. It is well known that PSM might produce biased estimates if the propensity score model is mis-specified. To ensure robustness of our results, we rely on recently developed doubly robust matching methods: covariate balancing propensity score (CBPS) method (Imai and Ratkovic, 2014) and entropy balancing (EB) method (Hainmueller, 2012). These methods are robust to the specification of the propensity score model (Fan et al., 2016; Zhao and Percival, 2016). In the second stage, we 3

estimate the impact of switching from FFS to FHG using panel-data regression models. We allow for unobserved physician-specific effects, year fixed effects, and a physician-specific trend using a high-dimensional fixed-effects model (Balazsi et al., 2018). To ensure robustness, we apply our estimation strategy using inverse probability weights resulting from the CBPS and EB matching methods. Physicians respond to financial incentives as predicted by our theoretical model. On average, physicians who switched to FHG produce more incentivized services than those who remained in FFS. Comprehensive care services increased by 3.5% and after-hours services increased by 4.8%. Although the FHG payment system does not provide physicians with additional incentives to provide non-incentivized services, these services also increased by 2.6%, suggesting a positive spill-over effect onto non-incentivized services. FHG physicians, on average, produced an additional $4,702 (all figures are in 2003 Canadian dollar unless otherwise stated) worth of comprehensive care services, $4,927 of after-hours services and $1,465 of additional nonincentivized services.

2 Institutional Context In Canada, health care falls under provincial jurisdiction, meaning that provinces or territories have the primary responsibility for organizing and delivering health care services to residents. The federal government ensures that each provincial/territorial health care system meets national standards enshrined in the Canada Health Act adhering to the principles of universality, comprehensiveness, accessibility, portability and public administration (Marchildon, 2013). Like other Canadian provinces, medically necessary physician and hospital services are publicly funded through taxes and privately delivered in Ontario. The fees for medical services in Ontario are set by negotiations between the Ontario Medical Association and the government of Ontario and administrated by the Ontario Health Insurance Plan (OHIP) (Marchildon and Hutchison, 2016). The fees paid are service-specific and are listed in a periodically updated Schedule of Benefits and Fees. OHIP compensates physicians for a wide range of health care services and patients do not pay out-of-pocket for the physicians’ services. In 2000, approximately 95% of family physicians in Ontario were paid predominantly by FFS (Sweetman and Buckley, 2014). As of April 1, 2016, 62% had opted for one of the reformed models (MOHLTC, 2016) with Family Health Groups and Family Health Organizations (FHO) 4

accounting for 87% of the 8,800 family physicians in the reformed models and 92% of the 10.6 million enrolled patients (MOHLTC, 2016). FHG was the first to be introduced in July 2003 and was optional. This paper focuses on studying the first stage in Ontario’s primary care reform in which the transition from FFS to FHG occurred (Kralj and Kantarevic, 2013). Under FFS, physicians receive a fixed fee for each service provided; under FHG they receive FFS payment for all services provided plus additional financial incentives for providing comprehensive care and after-hours services to enrolled patients. The comprehensive care premium (CCP) is 10% of the FFS value. The after-hours premium was initially equal to 10% of the FFS value, increasing to 15% in April 2005, then to 20% in April 2006 (MOHLTC Bulletin 11020). As per the FHG agreement, physicians are required to provide after-hours services up to at least one three hour-block per physician in the group unless exempted by the Ministry of Health and Long-Term Care. In addition, FHG physicians receive a comprehensive care capitation fee for each enrolled patient representing a small proportion of physician gross revenue (Kantarevic et al., 2011; Sweetman and Buckley, 2014). Beginning 2006, FHG physicians may also claim bonuses for certain preventive care when targeted levels of preventive care to specific patient populations are achieved as well as claim financial incentives for the management of diabetes and congestive heart failure as per the recommended guidelines.

3 The Theoretical Model We develop a model of physician behaviour under FFS and FHG in the context of a multitasking practice environment. We focus on modelling the supply side of health care services governing physician’s decision over the allocation of the total working hours across different services, and how it affects the volume of individual service production given the payment system. Our goal is to derive a participation decision rule indicating whether or not a physician chooses FHG compensation system. Physician services are categorized as: comprehensive care, 𝑄1, afterhours care, 𝑄2 and non-incentivized services, 𝑄3 (Section 4.2 provides details the aggregated services). 3.1 FFS physician problem A FFS physician receives a fixed fee for each service regardless of the volume of the services provided. We assume that the production of service 𝑗 provided by family physician 𝑖 is a

5

function of the hours devoted to produce 𝑗, ℎ𝑖𝑗 , the physician’s personal characteristics (age, gender, year of graduation, whether or not they graduated from an international medical school, and practice location,), and his/her patients’ characteristics (average comorbidity score, average age, proportion of male patients, proportion of patients living in rural areas) denoted by the vector 𝑋𝑖 , and a production shock 𝜀𝑖𝑗 . The production shock captures random elements that affect the time spent per service including the complexity of a particular service that are specific to physician. The quantity of service 𝑗 provided by the physician 𝑖 is specified as 𝛿 𝑄𝑖𝑗 = 𝒷(𝑋𝑖 )ℎ𝑖𝑗 𝜀𝑖𝑗 , 𝜀𝑖𝑗 > 0, 𝒷(𝑋𝑖 ) > 0,

𝑗 = 1,2,3

(1)

where 𝛿 determines the marginal return to time spent by the physician to produce a service. We assume that 𝛿 is between zero and one. This form of production function exhibits decreasing returns to hours spent in providing services which guarantees a finite interior solution for hours worked. The physician’s utility function is specified as a constant elasticity of substitution (CES) utility function defined over consumption, 𝐶 and leisure, 𝐿. This functional form is general enough to permit unrestricted responses to incentives, yet parsimonious in parameters, allowing for simple and direct interpretation of the results. Physician 𝑖 preferences are given by 𝜌

𝜌

1

𝑈(𝐶𝑖 , 𝐿𝑖 ) = (𝐶𝑖 + 𝐿𝑖 )𝜌 ,

𝜌 < 1,

(2)

Where 𝜌 determines the elasticity of substitution between consumption and leisure, 𝐿𝑖 = 𝑇 − ℎ𝑖 , ℎ𝑖 is total working hours and T is total amount of time available. The time constraint is ℎ𝑖 = ℎ𝑖1 + ℎ𝑖2 + ℎ𝑖3 , since ℎ𝑖 is the time allocated by the physician to produce the services. Under the FFS payment system, physician 𝑖’s budget constraint is given by 𝐶𝑖 = 𝔭1 𝑄𝑖1 + 𝔭2 𝑄𝑖2 + 𝔭3 𝑄𝑖3 + 𝑦,

(3)

where 𝔭𝑗 is the price of service 𝑗 = 1,2,3 and 𝑦 represents non-labour income. The prices of services are assumed to be exogenous. This is consistent with the publicly funded healthcare system in Ontario. Although physician fees are determined by the negotiation between the

6

MOHLTC and the Ontario Medical Association representing physicians, the fees are exogenous to an individual physician. We impose the following timing in order to solve the model and derive the optimal decision rules: i.

For each service j and each physician i, nature chooses 𝜀𝑖𝑗 ;

ii.

The physician observes 𝜀𝑖𝑗 , knows 𝒷(𝑋𝑖 ) and the price of services before he/she chooses ℎ𝑖𝑗 conditional on the total working hours, ℎ𝑖 ;

iii.

The physician chooses ℎ𝑖 and receives his/her payment from the MOHLTC.

We implicitly assume that the physician has complete information before making his/her choices. We solve the physician’s utility maximization problem in two steps. First, for a given total hours ∗ (ℎ𝑖 ). Then we worked, we solve the optimal time spent on each service ℎ𝑖𝑗 , denoted as ℎ𝑖𝑗

substitute the optimal value in the utility function to obtain an indirect utility function that depends on total hours worked, which we maximize for optimal hours. We rewrite the utility function by taking into account the budget constraint, the time constraint and the technology of production as: 1 𝜌

𝜌

𝛿 𝛿 𝛿 𝑈(ℎ𝑖1 , ℎ𝑖2 , ℎ𝑖3 ) = ((𝒷(𝑋𝑖 )(𝔭1 ℎ𝑖1 𝜀𝑖1 + 𝔭2 ℎ𝑖2 𝜀𝑖2 + 𝔭3 ℎ𝑖3 𝜀𝑖3 ) + 𝑦) + (𝑇 − ℎ𝑖1 − ℎ𝑖2 − ℎ𝑖3 )𝜌 ) .

The first-order condition for ℎ𝑖𝑗 , 𝑗 = 1,2,3 is 𝛿−1 𝛿 𝛿 𝛿 𝛿𝔭𝑗 ℎ𝑖𝑗 𝒷(𝑋𝑖 )𝜀𝑖𝑗 (𝒷(𝑋𝑖 )(𝔭1 ℎ𝑖1 𝜀𝑖1 + 𝔭2 ℎ𝑖2 𝜀𝑖2 + 𝔭3 ℎ𝑖3 𝜀𝑖3 ) + 𝑦)

𝜌−1

− (𝑇 − ℎ𝑖1 − ℎ𝑖2 −

ℎ𝑖3 )𝜌−1 = 0. We assume common shocks across services i.e. 𝜀𝑖𝑗 = 𝜀𝑖 , 𝑗 = 1,2,3. That means, the productivity of each physician is affected in the same way by the new technologies or new procedures and recommended treatment guidelines. Imposing common shocks assumption into the first-order conditions and solving for time devoted to each service gives

7

(4)

𝑝1 ℎ 𝑝1 + 𝑝2 + 𝑝3 𝑖 𝑝2 ∗ (ℎ𝑖 ) = ℎ𝑖2 ℎ 𝑝1 + 𝑝2 + 𝑝3 𝑖 𝑝3 ∗ (ℎ𝑖 ) = ℎ𝑖3 ℎ { 𝑝1 + 𝑝2 + 𝑝3 𝑖 ∗ (ℎ𝑖 ) = ℎ𝑖1

(5)

1

where 𝑝𝑗 = (𝔭𝑗 )1−𝛿 . Substituting back these optimal solutions into the utility function gives 1 𝜌

𝜌

𝑉(ℎ𝑖 ) = ((𝒷(𝑋𝑖 )𝑤𝐹𝐹𝑆 ℎ𝑖𝛿 𝜀𝑖 + 𝑦) + (𝑇 − ℎ𝑖 )𝜌 ) ,

(6)

Where 𝑤𝐹𝐹𝑆 = (𝑝1 + 𝑝2 + 𝑝3 )1−𝛿 determines the marginal return to an hour/day worked when that hour/day is optimally allocated across services. The physician optimal choice of the hours/days worked, ℎ𝑖∗ is derived from the indirect utility function 𝒷(𝑋𝑖 )𝛿𝑤𝐹𝐹𝑆 ℎ𝑖∗𝛿−1 𝜀𝑖 (𝒷(𝑋𝑖 )𝑤𝐹𝐹𝑆 ℎ𝑖∗𝛿 𝜀𝑖 + 𝑦)

𝜌−1

− (𝑇 − ℎ𝑖∗ )𝜌−1 = 0.

From this equation we cannot derive a closed-form of ℎ𝑖∗ . The optimal quantity of services produced is given by ∗ 𝑄𝑖𝑗 = 𝒷(𝑋𝑖 ) (

𝛿 𝑝𝑗 ) ℎ𝑖∗𝛿 𝜀𝑖 , 𝑝1 + 𝑝2 + 𝑝3

𝑗 = 1,2,3.

3.2 FHG physician problem A FHG physician receives the FFS payment and extra incentives for selected services provided to enrolled patients. Our model allows the FHG physician to choose the number of patients to enroll. Let 𝑛 denote the number of patients enrolled by the physician. For each enrolled patient, the comprehensive care service is remunerated at (1 + 𝜏1 )𝔭1 , where 0 < 𝜏1 < 1 is the comprehensive care premium; and after-hours care service is remunerated at (1 + 𝜏2 )𝔭2 , where 0 < 𝜏2 < 1 is the after-hours premium. For each non-enrolled patients, 1 − 𝑛, all the services are remunerated at FFS 𝔭3 . Note that the number of patients is normalized to one. Since FHG physician 𝑖 can enroll patients, we allow the technology of production of medical services to depend on 𝑛𝑖 (Woodward and Warren-Boulton, 1984)

8

𝛾

𝑄𝑖𝑗 = {

𝛿 𝒷(𝑋𝑖 )ℎ𝑖𝑗 𝑛𝑖 𝜀𝑖𝑗 , 𝛿 (1 𝒷(𝑋𝑖 )ℎ𝑖𝑗

𝑗 = 1,2,

)𝛾

− 𝑛𝑖 𝜀𝑖𝑗 ,

𝑗=3

where 𝛿 represents the marginal return to time spent by the physician to produce a service and 𝛾 represents the marginal return for a physician to enroll a patient. We assume that 𝛿 and γ are between zero and one to ensure finite interior solutions for both time spent for each service and the number of enrolled/non-enrolled patients. It also ensures that optimal behaviour will not lead the physician to specialize in the production of only one type of service nor will he/she select only one type of patients. Table 1 provides a description of the comparison of FFS and FHG system that apply to family physicians in Ontario. The FHG physician budget constraint is 𝐶𝑖 = (1 + 𝜏1 )𝔭1 𝑄𝑖1 + (1 + 𝜏2 )𝔭2 𝑄𝑖2 + 𝔭3 𝑄𝑖3 + 𝑦. Maximizing the physician’s utility function keeping total hours worked ℎ𝑖 and the number of enrolled patients 𝑛𝑖 fixed, under common shocks assumption gives the optimal time devoted to each services as a function on ℎ𝑖 and 𝑛, denoted ℎ̂𝑖𝑗 (ℎ𝑖 , 𝑛𝑖 ), 𝑗 = 1,2,3 ℎ̂𝑖1 (ℎ𝑖 , 𝑛𝑖 ) =

ℎ̂𝑖2 (ℎ𝑖 , 𝑛𝑖 ) =

ℎ̂𝑖3 (ℎ𝑖 , 𝑛𝑖 ) = {

𝛾 1−𝛿 𝑛𝑖 𝑃1

𝛾 𝑛𝑖1−𝛿 𝑃1

+

+

𝛾 1−𝛿 𝑛𝑖 𝑃1 𝛾 𝑛𝑖1−𝛿 𝑃2 + (1 − 𝛾 𝑛𝑖1−𝛿 𝑃2 𝛾 𝑛𝑖1−𝛿 𝑃2 + (1 − 𝛾 1−𝛿

(1 − 𝑛) 𝛾 𝑛𝑖1−𝛿 𝑃1

+

𝛾 𝑛𝑖1−𝛿 𝑃2

1

𝛾 1−𝛿

𝑛𝑖 )

𝛾 1−𝛿

𝑛𝑖 )

ℎ𝑖 , 𝑃3

ℎ𝑖 , 𝑃3

𝑃3 𝛾 1−𝛿

+ (1 − 𝑛𝑖 )

ℎ𝑖 , 𝑃3

1

where 𝑃𝑗 = ((1 + 𝜏𝑗 )𝔭𝑗 )1−𝛿 , 𝑗 = 1,2; 𝑃3 = (𝔭3 )1−𝛿 . Substituting back the optimal time devoted to each service into the utility function gives an indirect utility function which depends on (ℎ𝑖 , 𝑛𝑖 ). Maximizing this function with respect to (ℎ𝑖 , 𝑛𝑖 ) gives an explicit form of 𝑛̂𝑖 1−𝛿

𝑛̂𝑖 =

𝑃𝛾+𝛿−1 1−𝛿

, with 𝛾 + 𝛿 ≠ 1

1 + 𝑃𝛾+𝛿−1

9

Where 𝑃 = 𝑃

𝑃3

1 +𝑃2

represents the price of non-incentivized services relative to incentivized-

services. Again there is no closed-form solution for optimal ℎ̂𝑖 and the implicit form suggests that ℎ̂𝑖 is correlated with epsilon. ℎ̂𝑖 solves 1−𝛿

1−𝛾−𝛿

0 = 𝛿 (1 + 𝑃 𝛾+𝛿−1 )

1−𝛿

𝜌−1

1−𝛾−𝛿

𝒷(𝑋𝑖 )𝑃31−𝛿 ℎ𝑖𝛿−1 𝜀𝑖 (𝒷(𝑋𝑖 )𝑃31−𝛿 ℎ𝑖𝛿 𝜀𝑖 (1 + 𝑃 𝛾+𝛿−1 )

+ 𝑦𝑖 )

− (𝑇 − ℎ𝑖 )𝜌−1 .

The optimal number of each type of service produced is given by 𝛿 𝛾 𝒷(𝑋𝑖 )ℎ̂𝑖𝑗 𝑛̂𝑖 𝜀𝑖𝑗 , 𝑗 = 1,2, 𝑄̂𝑖𝑗 = { 𝛿 𝛾 𝒷(𝑋𝑖 )ℎ̂𝑖𝑗 (1 − 𝑛̂𝑖 ) 𝜀𝑖𝑗 , 𝑗 = 3.

3.1 FHG Participation Decision A rational physician will choose the payment system which maximizes utility given the constraint on total hours worked (budget). We illustrate the physician’s optimal choice in Figure 1 illustrating the tradeoff between the number of enrolled patients and consumption/income conditional on ℎ𝑖∗ (the optimal hours worked under FFS). Derivations of the budget equations and the participation decision are presented in the Appendix A1. The budget line FFS is horizontal under FFS as there is no incentive for physicians to enroll patients. The budget curve FHG illustrates the tradeoff between the number of enrolled patients and income under FHG, holding the total working hours fixed at ℎ𝑖∗ . It passes through the origin because we ignore the small fixed payment under FHG. Note that n* represents the number of enrolled patients that renders physicians indifferent between FFS and FHG conditional on the total hours worked. When n< n* the physician will always choose FFS compensation system -- the efficient budget constraint is the bold straight line. This is consistent with the fact that a FHG physician with zero patient enrollment is compensated at least as FFS physician. When n>n* the FFS physician will switch to FHG to move to the upward sloping bold line. Figure 1 illustrates the potential selection bias associated with switching to FHG. Physicians who have ability2 to enroll more patients (such as physician E1) will tend to choose

2

This might include being able to make himself/herself available to patients during after-hours and having good reputation.

10

FHG moving from E1 to E2, while those who have difficulties3 to enroll patients (such as physician E0) will remain in FFS. Thus, a direct comparison of physicians’ behaviour across FFS and FHG payment systems will potentially confound the effects of FHG.

4 Data and Variables 4.1. Source of data The data for this study come from family physicians practicing in Ontario between April 1st 2003 (year of FHG introduction) and March 31st, 2008. We use several health administrative databases held at the Institute for Clinical Evaluative Sciences (ICES) in Ontario, Canada. ICES Physician Database (IPDB) gives information on physician’s demographic and practice characteristics: age, sex, full time equivalent (FTE) based on physician’s billings, year of graduation, whether or not they graduated from an international medical school (IMG), and practice location. The Corporate Provider Database (CPDB) provides information on physician’s compensation model and effective date of eligibility for billing under the Ontario Health Insurance Plan (OHIP). OHIP claims database contains our outcomes: the quantity index of comprehensive care services, the quantity index of after-hours services, and the quantity index of non-incentivized services representing all non-incentivized services. Construction of these indices is described in the next section. The OHIP database contains the fees paid to physicians for services provided, and the number of services performed by each physician, as well as their fee codes. These fee codes include the comprehensive care services that are eligible for the CCP premium under FHG listed in Table A2.1 in the appendix A2. The after-hours services are identified with the after-hours premium code Q012A along with the list of fee codes that are eligible for after-hours premium under FHG (see Table A2.2 in the appendix A2 for the fee codes). The rest of the OHIP fee codes submitted by family physicians are grouped as non-incentivized services. We obtain patient characteristics for each individual physicians’ patients annually using Registered Persons Database (RPDB) which is Ontario’s health registry database and the Client Agency Program Enrollment (CAPE) database. The CAPE allows us to match enrolled patients with physicians and also provides information on physician’s model type. If a physician was 3

These difficulties might include lack of staff, resources, administrative expertise and IT support (The College of Family Physicians of Canada, 2012).

11

affiliated with more than one model type in a given year, then we selected the most recent model that she/he joined that year. For patients in FFS, we virtually enrolled them to physicians based on the highest billing based on 18 core primary care fee codes submitted in a given fiscal year (Kiran et al., 2012). We obtained patient’s postal code, age and sex from RDPB. 4.2. Variables As discussed earlier, introduction of the FHG compensation system in 2003 led to three distinct categories of services: (i) comprehensive care services (j=1), (ii) after-hours services (j=2) and (iii) non-incentivized services (j=3). Each type of service is clearly identified by the OHIP fee codes listed in Tables A2.1-2. To render our estimation tractable, we aggregated the services to construct a quantity index of services for each of the three groups of services. The formula of the index numbers for services and prices is described in the Appendix A3. The aggregated service j, 𝑄𝑗 , is given by the volume of services provided within group j, weighted at base year level prices. Note that price variation is excluded from the index by considering the price paid for each service at the base year which is 2003. This avoids incorporating into the quantity measure the effect of the variation in prices due to switching to FHG. The price of the aggregated service j was also aggregated into indices. The price index for services provided within group j denoted 𝔭𝑗 , was calculated as a Laspeyres price index. This price index represents the fees paid for aggregated service j under FFS. We used the number of each service provided among group j=1,2,3 in the base year as weights. In this way, the quantity variations due to FHG switching are excluded from the price index. The total quantity of services provided was derived as a sum of the volume of three types of services: comprehensive care services, 𝑄1, after-hours services, 𝑄2 , and non-incentivized services, 𝑄3 , weighted by the corresponding price indices. The quantity variable captures the ability of the physician to produce medical services at different relative prices. To control for the possible influence of patient complexity on physicians’ service production, we used patients’ characteristics as covariates: average age, proportion of male patients in physician’s practice, proportion of patients living in rural areas, and patient’s average comorbidity score based on the Johns Hopkins Aggregated Diagnosis Groups (ADGs). We derived ADG for each patient based on their diagnosis codes from all health administrative databases available at the ICES using The Johns Hopkins ACG® System Ver 10 case-mix adjustment system 12

(The Johns Hopkins University, 2011), a well-known measure of patients’ comorbidity status in the health services literature (Glazier et al., 2008). As the ADGs comprise 32 diagnosis groups, each patient has 32 indicator variables representing the presence or absence of each diagnosis group. We summed up ADGs for each patient, yielding an ADG score per patient up to 32. The average ADG score was defined as the average of ADG scores of physician’s patients. To calculate the proportion of patients living in rural areas per physician, we used patient postal codes from the RPDB to obtain the rurality index. Individual patients with a rurality index of 40 or higher are considered to reside in rural areas (Kralj, 2000; Matheson et al., 2012). When considering whether to switch from FFS to FHG, a FFS physician can estimate what would be her/his gross revenue if she/he opted for FHG based on the actual services provided to patients in the fiscal year 2002/03. We calculated this expected gain in income variable based on the 2002 algorithm used by the MOHLTC to advise FFS physicians interested to switch to the FHG model. Note that the expected gain in income is potentially an important variable to consider in the physicians’ switching decision. 4.3. Pre-reform descriptive statistics Since this study focuses on comparing physicians practicing in FFS and FHG models, we started with all physicians who were paid under these two models during 2003-2008, covering one year before and five years after the introduction of the FHG model.4 Then, we excluded physicians who were not present both in 2003 and thereafter or switched back from FHG to FFS. This ensures that all physicians had the option to choose the FHG model and only a sample of physicians chose to practice in FHG. To analyze the effect of the FHG reform, we divided the sample into switcher and non-switcher groups. The switcher group is defined as those physicians who switched from FFS to FHG in any year between 2004 and 2008 and remained in FHG. The non-switcher is defined as those who remained in FFS throughout our study period. After excluding missing data on variables, we have 4,590 switchers (FHG physicians) and 2,498 non-switchers (FFS physicians) for our empirical analyses.

4

We chose 2003 (fiscal year 2002/03) as a starting period because no physician signed up the FHG model during this year. Observations beyond 2008 were not analyzed as new payment models were introduced by the MOHLTC in 2007.

13

Descriptive statistics for the pre-reform period (2003) for switcher and non-switcher groups are presented in Table 2. The top part of the table provides physicians’ average outcomes: total services, comprehensive care services, after-hours services, and non-incentivized services in thousands of Canadian dollars. The rest of the table presents the physicians’ and patients’ characteristics. In Table 2, we see that FHG and FFS physicians’ production of services are statistically different in the pre-reform period. Compared to FFS physicians, FHG physicians produced 46% more comprehensive care services, 45% more after-hours services, and 6% more non-incentivized services before the introduction of FHG. This highlights the presence of potential selection problem influencing the decision to choose the FHG payment scheme. Specifically, physicians who switched to FHG payment model, on average, were more productive in terms of both incentivized and non-incentivized medical services compared to the non-switchers. The descriptive results show that the non-switchers were relatively older (average age: 52.04 vs. 48.1 years). This is perhaps not surprising as the habits and the sense of practice independence are likely to be stronger among older physicians. While 35% of FHG physicians were female, only 30% of those who remained in FFS were female. The total days worked and the proportion of physicians who worked at least half a day for the switcher group of physicians (250.1 days and 96%, respectively) are larger than for the non-switcher group of physicians (209.8 days and 78%, respectively). The difference in terms of the expected gain in income is striking: $192,300 in the switcher group and $130,600 in the non-switcher group. Both groups are also different in terms of physicians’ patient characteristics. Specifically, the switcher group, on average, had relatively younger patients and more comorbidity than the non-switcher group. Also, the switcher group had 13% of their patients living in rural areas and 45% of them are males, while the corresponding percentages for the non-switcher group were 16% and 48%.

5. Matching Strategies The main challenge with studying the impact of remuneration on production of medical services is that switchers may be systematically different from non-switchers (see Table 2), leading to a potential non-random selection bias. To account for such a bias, we rely propensity score matching (PSM) methods (Rosenbaum and Rubin, 1983) to ensure that the two groups are comparable. We first estimate the propensity score (i.e. the probability that an individual physician 14

will switch from FFS to FHG) using the Dehejia and Wahba (2002) algorithm along with the hit or miss method (Heckman et al., 1997). The hit or miss method allows us to select the additional variables that might affect the participation decision, interaction terms, as well as higher-order terms of covariates. We retain the set of pre-reform variables which provides an estimated propensity score that ensures covariate balancing between switcher and non-switcher physicians. The PSM method, however, may give rise to a biased estimated average treatment effect if the propensity score model is mis-specified (Kang and Schafer, 2007; Smith and Todd, 2005). Also, the choice of the balancing test could affect the set of covariates used to estimate the propensity scores (Lee, 2013). Controlling for these potential sources of bias implies using methods that ensure balance of covariates even when the propensity score model is not correctly specified. Thus, we use the CBPS and EB matching methods as an alternative to the PSM. CBPS and EB directly incorporate covariate balance in the estimation procedure, so these methods automatically ensure that covariates are balanced. CBPS uses a GMM framework to combine score condition (equivalent to fit a logistic model) and covariate balancing moment conditions (ensuring the covariate balancing). EB method provides directly the optimal weights under some prespecified covariate balancing constraints on first and second moments. EB may be seen as a generalization of the PSM (Hainmueller, 2012). 5.1 Matching Results Propensity scores are estimated using both logit and CBPS methods. As summarized in Table A2.3, our specification ensures covariate balancing between the two groups. The logistic regression results show that expected gain in income, total days worked, gender (female), whether or not the physician is an international medical graduate (IMG) are statistically correlated with the probability of switching. Average ADG score and the proportion of male patients are negatively correlated with the probability of switching. Two interaction terms are good predictors of the probability of switching: Total services in 2003*(FTE>=0.5), Age*female. Nine of the 14 geographic indicators for regional health areas, known as Local Health Integration Networks (LHINs), are positively correlated with the probability of switching. We included all the variables listed in Table A2.3 in the model even though they were not statistically significant because they either increased the prediction rate (hit-or miss) or ensured overall covariate balancing. We use the same covariates to estimate the propensity score model with CBPS. The results (in the second 15

column of Table A2.3) are qualitatively similar to the results of the logistic regression. However, CBPS reduces standard errors of covariates. Figures 2 to 4 summarize the distribution of the estimated propensity score for switchers and non-switchers before and after Kernel matching, CBPS and EB. Overall, the before matching distribution of propensity scores for switchers and non-switchers were very different -- the Kolmogorov-Smirnov (KS) test for equality of propensity score distributions reject the equality of the distributions of propensity score. The figures reveal that the estimated common support or overlapping region of the two distributions is large enough to perform matching (Heckman et al., 1999). Five observations in the switcher group were dropped from the analysis due to the lack of common support. The after-matching distribution of propensity scores for switchers and nonswitchers are very similar -- the solid and the dashed lines largely coincide (Figure 2-4), suggesting a drastic reduction of covariate imbalance after-matching. The after-matching KS test p-values are 0.105, 0.058 and 0.078 for kernel matching, CBPS, and EB methods, respectively, suggesting that the equality of the distributions for switchers and non-switchers cannot be rejected at the 5% significance level. Table 3 shows the after-matching summary statistics for both group. The results clearly suggest notable differences in the covariates (and outcomes) between switchers and non-switchers in the original sample (Table 2) have disappeared in the matched sample. Also, the results of the regression test (Table A2.4) show that overall the covariates are unaffected by the participation to FHG model conditional on estimated propensity score and higher-order terms (to account for nonlinearity that might exist between one covariate and participation). The p-value of regression test (after matching) for each covariate is greater than 5%, except for two or three variables out of 31 variables. We cannot reject that the participation to FHG model does not provide any information on covariates. The three matching methods employed seem to achieve balancing between the switcher and non-switcher groups. This result is not surprising for CBPS and EB methods – they are built to achieve covariate balancing. We match and reweight 4,585 switchers to 2,498 non-switchers. Our final panel dataset used to estimate the average treatment effect on treated of FHG reform contains 36,639 physician-year observations on 7,083 physicians practicing in Ontario between the years 2003-2008.

16

6 Estimating Average Treatment Effects on Treated After completing the first step of accounting for pre-reform observed differences between switchers and non-switchers using PSM, CBPS and EB, we proceed to the second stage and use panel data regression models to account for both observed physician and practice characteristic and unobserved physician-specific heterogeneity affecting the productivity. Physician-specific linear trends help ascertain the robustness of our results. The general form of the regression model estimated is 𝑦𝑖𝑡 = 𝜃𝑡 + 𝑐𝑖 + 𝑔𝑖 𝑡 + 𝑋𝑖𝑡 𝛽 + 𝜑𝐹𝐻𝐺𝑖𝑡 + 𝑢𝑖𝑡

(7)

where 𝑦𝑖𝑡 represents the (log) of services produced by physician 𝑖 at year 𝑡, the parameter 𝜃𝑡 is year fixed-effects, the parameter 𝑐𝑖 is physician-specific unobserved heterogeneity, 𝑔𝑖 is a physician-specific linear trend and, 𝑋𝑖𝑡 is a set of physicians’ observables characteristics including age, age squared, number of working holidays and weekends, indicator variable that indicates if FTE>=0.5 day (denoted FTE>=0.5), gender (female), IMG, age-female interaction term, age and (FTE>=0.5) interaction term, 14 geographic indicators for regional health areas (LHINs); and their patient characteristics: average ADG score, average age of patients, proportion of patients in rural areas, and proportion of male patients; as well as the prices of the different services. The regressor of interest, 𝐹𝐻𝐺𝑖𝑡 , is one if the physician 𝑖 at year 𝑡 switched to FHG and zero if remained in FFS, 𝑢𝑖𝑡 is the error term. Applying ordinary least squares (OLS) may produce a biased estimate of 𝜑 because the fixed-effects 𝜃𝑡 , 𝑐𝑖 , and 𝑔𝑖 are potentially correlated with 𝐹𝐻𝐺𝑖𝑡 . That is, the decision to switch to FHG is correlated with the unobserved differences in productivity between physicians, implying 𝐸(𝑐𝑖 |𝐹𝐻𝐺𝑖𝑡 = 1) ≠ 0. We first use a difference-in-difference regression model to control for 𝑐𝑖 (ignoring 𝑔𝑖 ), assuming parallel trends (same 𝜃𝑡 for both switchers and non-switchers). Then, we relax the parallel trends assumption by allowing each physician to follow his/her own trend, through 𝑔𝑖 the physician-specific trend coefficient. We use high-dimensional fixed-effects

17

(HDFE) estimation technique5 to consistently estimate the average treatment effects of switching from FFS to FHG model.

7 Estimated Results We estimate model in (7) in three different manners: pooled OLS; physician fixed-effects (FE); and unobserved physician- and year fixed-effects, and physician-specific trend using HDFE. In each case we use unweighted and weighted regressions, where weights come from the PSM at the first stage. The results of the average impact of switching from FFS to FHG on switchers are reported in Table 4. The results from the weighted HDFE (FE) model indicate that on average physicians who switched to the FHG model increase significantly the volume of total services, comprehensive care services, after-hours services, and non-incentivized services by 2.2% (4.4%), 3.5% (8.8%), 4.8% (9.9%), and 2.6% (5.9%) respectively.6 The significant positive impact on non-incentivized services is somewhat surprising since the FHG physicians are not incentivized to produce more of those services. As a comparison, in the second panel we present the results for weighted pooled OLS. The impact of the reform is statistically significant, however the estimated effects are very large in magnitude ranging from 4.0% to 30.3%. This result suggests that ignoring the selection bias and unobservable heterogeneity may lead to biased estimates. For comparison, we also present in Table 4 the results of the unweighted regressions. While these results indicate that, on average, switching to FHG model has a significantly positive impact on switchers, the estimates of the impact on switchers’ service production from pooled OLS and FE are greater in magnitude than those from the weighted versions of these models.7 This result highlights the importance of accounting for selection bias as those who switched to FHG are systematically different from those remained in FFS. The estimated coefficient on 𝜑 suggests that physicians respond to incentives brought about by the FHG reform in a predictable manner. On average, physicians who switched to FHG 5

It is an iterative approach for the estimation of models with more than two fixed-effects (Guimarães and Portugal, 2010). For the implementation in STATA, we use the reghdfe program written in Stata by S. Correia, 2017 available at http://scorreia.com/software/reghdfe/ (accessed November 2017). 6 The reported percentages are calculated by taking the exponential value of the estimates of 𝜑 from Table 4 and subtracting one and multiplied by 100. 7 The results from the weighted and unweighted HDFE models are very similar, suggesting that accounting for unobservable heterogeneity is crucial in identifying the average treatment effect on the treated.

18

model produce more services than those who remained in FFS. Comprehensive care services increase by 3.5%, suggesting a potential improvement in the provision of comprehensive primary care. After-hours services increase by 4.8%, suggesting a potential improvement in patient access to care outside of the regular working hours. Non-incentivized services also increase by 2.6%, suggesting that access to care may have improved for non-enrolled patients and/or some type of spill-over effects resulting from the provision of more incentivized services to the enrolled patients. Based on the preferred HDFE specification, we calculate the additional services produced by switchers relative to non-switchers. The results suggest that a switcher facing a 10% increase in the fee of comprehensive care services produces, on average, an additional volume of comprehensive care worth of $4,702 per year relative to a non-switcher. A switcher facing a more than 10%8 increase in the fee of after-hours services produces, on average, $4,927 more volume of after-hours services per year than a non-switcher. Regarding non-incentivized services, the spillover effects creates an additional $1,465 per year of non-incentivized services for switchers relative to non-switchers. The additional services produced by the FHG physician represent 1.03 FFS equivalent services for incentivized services, a similar result is found for non-incentivized services.9 In total, switching to FHG adds substantial extra total services of $43,448 per year, representing 1.11 FFS equivalent services. 7.1 Robustness Checks The main results presented in the previous sub-section were estimated by a weighted regression model using the inverse probability weighting technique based on PSM matching. As an alternative to PSM at the first stage, we use CBPS and EB methods which are relatively robust to model specification (Fan et al., 2016; Zhao and Percival, 2016). The results based on weights from CBPS and EB matching methods are presented in Table 5. The results on FE and HDFE regression models are very similar to those reported in Table 4. Switching to FHG compensation system significantly increases total services, comprehensive care services, after-hours services and non-incentivized services.

8

Recall that the after-hours premium increased from 2003 to 2008 from 10% to 20% (MOHLTC Bulletin 11020). FFS equivalent services is the ratio between the average volume of services produced by FHG physicians and the average volume of services produced by FFS physicians. The volume of services are predicted based on the HDFE model for each type of service. 9

19

7.2 Heterogeneous Impact Table 6 provides results on the average impact of switching from FFS to FHG model by gender (first panel of the table), graduation cohort (second panel of the table) and practice location (third panel of the table). The results show that, on average, the reform impact is greater for female physicians relative to male physicians – suggesting that female physicians respond more to financial incentives relative to males. The experienced physicians’ cohort (who graduated before year 1970) is more sensitive to FHG model than the other graduation cohort group. The third panel shows some heterogeneous response of physicians to the FHG reform according to their practice location. What is more interesting is that across all subsamples considered, the impact of switching to FHG model is positive implying that financial incentives encourage more incentivized medical services.

8 Discussions and Conclusions After accounting for the selection bias, we estimated the impact of switching family physicians from the traditional FFS to FHG model on their productivity in terms of the volume of total medical services, comprehensive care medical services, after-hours medical services and nonincentivized medical services. Our results show that physicians who switched from the traditional FFS to the blended FFS increased their service production by 2% to 5%: comprehensive care medical services increased by 3.5%, after-hours services by 4.8%, and non-incentivized services by 2.6%. This means that incentivizing some services may have a positive spill-over effect on others. These results have important policy implications for the provision of healthcare services to the population. In an era of increased demand for medical services, governments can use financial incentive mechanisms to increase the supply of services to some extent. Given that altering the number of physicians is generally expensive and time consuming, policy makers can use financial incentives to increase the supply of services at the intensive margin. However, there is a limit in terms of the quantity of medical services that can be produced with the existing physicians. We find that the value of the extra production of total services by physicians who switched FHG is $43,448 per year, and their production of services is 1.11 times higher than their FFS counterparts. Our results suggest that a 10% increase in incentives leads to a 4% increase in comprehensive care

20

medical services; a more than 10% increase in incentives leads to about 5% increase in after-hours medical services. This rather inelastic response, however, imposes limits rather quickly on the use of these financial incentives as a mechanism to increase physician services. Although a formal cost-benefit analysis is beyond the scope of this paper, our results suggest that financial incentives may be more effective in the production of targeted medical services (e.g. service production during after-hours) or temporary increased demand for medical services rather than as a permanent solution to increased demand for medical services. These financial incentives may lead, however, to other savings in the overall health care system by, for instance, improving after-hours care and reducing, say, emergency department visits. Our results are robust to the alternative matching approaches employed but rely on the assumption that the matching equation in the first stage is correctly specified. This may not hold if any unobservable factors also influence selection into FHG. However, we believe that our first-stage regressions and alternative matching procedures are reasonable specifications in identifying the true effect of switching to FHG.

Acknowledgements Funding for this research by the Canadian Institutes of Health Research operating grant (MOP– 130354) is gratefully acknowledged. This study was undertaken at the ICES Western site. ICES is funded by an annual grant from the Ontario Ministry of Health and Long-Term Care (MOHLTC). Core funding for ICES Western is provided by the Academic Medical Organization of Southwestern Ontario (AMOSO), the Schulich School of Medicine and Dentistry (SSMD), Western University, and the Lawson Health Research Institute (LHRI). The opinions, results and conclusions are those of the authors and are independent from the funding sources. No endorsement by ICES, AMOSO, SSMD, LHRI, CIHR, or the MOHLTC is intended or should be inferred. We thank Jasmin Kantarevic for sharing his codes to estimate the expected gain in income variable.

21

References Balazsi, L., Matyas, L., Wansbeek, T., 2018. The estimation of multidimensional fixed effects panel data models. Econom. Rev. https://doi.org/10.1080/07474938.2015.1032164 Broadway, B., Kalb, G.R.J., Li, J., Scott, A., 2016. Do Financial Incentives Influence GPss Decisions to Do After-Hours Work? A Discrete Choice Labour Supply Model. SSRN Electron. J. https://doi.org/10.2139/ssrn.2746275 Campbell, S., Reeves, D., Kontopantelis, E., Middleton, E., Sibbald, B., Roland, M., 2007. Quality of Primary Care in England with the Introduction of Pay for Performance. N. Engl. J. Med. https://doi.org/10.1056/NEJMsr065990 Devlin, R.A., Sarma, S., 2008. Do physician remuneration schemes matter? The case of Canadian family physicians. J. Health Econ. https://doi.org/10.1016/j.jhealeco.2008.05.006 Dumont, E., Fortin, B., Jacquemet, N., Shearer, B., 2008. Physicians’ multitasking and incentives: Empirical evidence from a natural experiment. J. Health Econ. https://doi.org/10.1016/j.jhealeco.2008.07.010 Fan, J., Imai, K., Liu, H., Yang, N., Yang, X., 2016. Improving covariate balancing propensity score: A doubly robust and efficient approach. Glazier, R., Moineddin, R., Agha, M.M., Zagorski, B., Hall, R., Ontario, I. for C.E.S. in, 2008. The impact of not having a primary care physician among people with chronic conditions. Institute for Clinical Evaluative Sciences, Toronto, Ont. Glazier, R.H., Klein-Geltink, J., Kopp, A., Sibley, L.M., 2009. Capitation and enhanced fee-forservice models for primary care reform: a population-based evaluation. Can. Med. Assoc. J. 180, E72-1119. https://doi.org/10.1503/cmaj.081316 Guimarães, P., Portugal, P., 2010. A simple feasible procedure to fit models with highdimensional fixed effects. Stata J. https://doi.org/The Stata Journal Hainmueller, J., 2012. Entropy balancing for causal effects: A multivariate reweighting method to produce balanced samples in observational studies. Polit. Anal. https://doi.org/10.1093/pan/mpr025 Imai, K., Ratkovic, M., 2014. Covariate balancing propensity score. J. R. Stat. Soc. Ser. B Stat. Methodol. https://doi.org/10.1111/rssb.12027 Kantarevic, J., Kralj, B., Weinkauf, D., 2011. Enhanced fee-for-service model and physician productivity: evidence from Family Health Groups in Ontario. J. Health Econ. 30, 99–111. https://doi.org/10.1016/j.jhealeco.2010.10.005 Kiran, T., Victor, J.C., Kopp, A., Shah, B.R., Glazier, R.H., 2012. The relationship between financial incentives and quality of diabetes care in Ontario, Canada. Diabetes Care. https://doi.org/10.2337/dc11-1402 Kralj, B., 2000. Measuring rurality for purposes of health-care planning: an empirical measure for Ontario. Ont. Med. Rev. Kralj, B., Kantarevic, J., 2013. Quality and quantity in primary care mixed-payment models: Evidence from family health organizations in Ontario. Can. J. Econ. https://doi.org/10.1111/caje.12003 Li, J., Hurley, J., Decicca, P., Buckley, G., 2014. PHYSICIAN response to pay-for-performance: Evidence from a natural experiment. Heal. Econ. (United Kingdom). https://doi.org/10.1002/hec.2971 Marchildon, G.P., 2013. Canada: Health system review, Health Systems in Transition. Marchildon, G.P., Hutchison, B., 2016. Primary care in Ontario, Canada: New proposals after 15 years of reform. Health Policy. https://doi.org/10.1016/j.healthpol.2016.04.010 22

Matheson, F.I., Dunn, J.R., Smith, K.L.W., Moineddin, R., Glazier, R.H., 2012. Development of the Canadian Marginalization Index: a new tool for the study of inequality. Can. J. public Heal. 103, S12-6. Ministry of Health and Long-Term Care (MOHLTC), 2015. Patients first: A proposal to strengthen patient-centred health care in Ontario. MOHLTC, 2016. The 2016 Annual Report of the Office of the Auditor General of Ontario: Physicians billing. Pham, M., McRae, I., 2015. Who provides GP after-hours care? Health Policy (New. York). https://doi.org/10.1016/j.healthpol.2015.01.005 Sarma, S., Devlin, R.A., Hogg, W., 2010. Physician’s production of primary care in Ontario, Canada. Health Econ. https://doi.org/10.1002/hec.1447 Sutton, M., Elder, R., Guthrie, B., Watt, G., 2010. Record rewards: The effects of targeted quality incentives on the recording of risk factors by primary care providers. Health Econ. https://doi.org/10.1002/hec.1440 Sweetman, A., Buckley, G., 2014. Ontario’s Experiment with Primary Care Reform. Univ. Calgary, Sch. Public Policy Res. Pap. 7, 1–35. The College of Family Physicians of Canada, 2012. Best Advice: Patient Rostering in Family Practice. The Johns Hopkins University, 2011. The Johns Hopkins ACG® Case-Mix System Version 10.0 Release Notes. Woodward, R.S., Warren-Boulton, F., 1984. Considering the effects of financial incentives and professional ethics on ‘appropriate’ medical care. J. Health Econ. 3, 223–237. Zhao, Q., Percival, D., 2016. Entropy balancing is doubly robust.

23

Figure 1: Optimal Choice along the efficient Budget Constraint

Figure 2: Distribution of the propensity scores (kernel matching)

24

Figure 3: Distribution of the propensity scores (CBPS)

Figure 4: Distribution of the propensity scores (EB)

25

Table 1: Comparison of Ontario’s FFS and FHG payment models FFS

FHG

Group size

1

At least 3

Patient enrollment requirement

No

No, but encouraged through incentives

After-hours requirement Services

No

At least one three hour-block

Comprehensive care services

Remunerated at price 𝔭1

Remunerated at price (1 + 𝜏1 )𝔭1

After hours services

Remunerated at price 𝔭2

Remunerated at price (1 + 𝜏2 )𝔭2

non-incentivized services

Remunerated at price 𝔭3

Remunerated at price 𝔭3

Fixed payment

No fixed payment

(small) Comprehensive care capitation

Fixed bonuses payment

No bonuses

bonuses for preventive care (flu shots to seniors, pap smear, mammogram, childhood immunizations, and colorectal-cancer screening) were introduced in late 2006

Table 2: Descriptive statistics on Switchers and Non-switchers (fiscal year 2002/03) Switcher (FHG physicians)

Non-switcher (FFS physicians)

390.28 174.45 169.61 46.21

279.96* 119.63* 116.77* 43.56*

Physicians characteristics 48.10 0.35 0.96

52.04* 0.30* 0.78*

250.10 192.30 0.12

209.80* 130.60* 0.16*

Outcomes Total services CC services AH services Non-incentivized services Age Female FTE>=0.5 Total days worked Expected gain in income IMG

Patients characteristics 3.29* Average ADG score 3.39 40.41* Av. Age of patients 38.88 0.16* Prop. of patients in rural areas 0.13 0.48* Prop. of male patients 0.45 Number of Phys. 4,590 2,498 Legend: Significance level: *p<0.05. Note: CC=Comprehensive care, AH= After-hours, FFS=Fee For service, FHG= Family Health Group

26

Table 3: After matching: Descriptive statistics on sampled physicians by group as of 2003 (fiscal year 2002/03) Treatment (FHG physicians)

Total services CC services AH services Non-incentivized services Age Female FTE>=0.5 Total days worked Expected gain in income IMG LHIN02 = South West LHIN03 = Waterloo Wellington LHIN04 = Hamilton Niagara Haldimand Brant LHIN05 = Central West LHIN06 = Mississauga Halton LHIN07 = Toronto Central LHIN08 = Central LHIN09 = Central East LHIN10 = South East LHIN11 = Champlain LHIN12 = North Simcoe Muskoka LHIN13 = North East LHIN14 = North West Average ADG score Av. Age of patients Prop. of patients in rural areas Prop. of male patients Expected gain in income squared Age squared Total services in 2003 squared Total services in 2003*(FTE>=0.5) Age*(FTE>=0.5) Age squared*(FTE>=0.5) Age*female Number of Phys. Legend: Significance level: *p<0.05.

390.28 174.45 169.61 46.21

Control (FFS physicians) Kernel matching Outcomes 396.33 178.47* 173.96* 43.902*

Physicians characteristics 48.10 48.01 0.35 0.35 0.96 0.95 250.1 247.11* 192.30 196.69* 0.12 0.12 0.08 0.03

0.08 0.03

0.07 0.07 0.05 0.05 0.08 0.09 0.13 0.13 0.14 0.15 0.12 0.13 0.04 0.04 0.12 0.12 0.04 0.04 0.03 0.03 0.02 0.01 Patients characteristics 3.38 3.39 38.59 38.88 0.12 0.13 0.46 0.45 Higher-order and interactions terms 46475 48703* 2407 2400.40 180606 1.9e+05* 386.70 392.68 46.14 45.90 2309 2292.20 15.63 15.40 4,590 2,498

27

CBPS

Entropy Balancing

391.79 176.10 171.45 44.24

390.23 174.35 169.54 46.34

47.91 0.36 0.95 245.95 193.9 0.12

48.10 0.35 0.96 250.10 193.90 0.12

0.07 0.03

0.08 0.03

0.07 0.06 0.08 0.13 0.15 0.13 0.04 0.12 0.03 0.03 0.01

0.07 0.05 0.08 0.13 0.14 0.12 0.04 0.12 0.04 0.03 0.02

3.39 38.67 0.12 0.45

3.39 38.88 0.13 0.45

47584 2391.50 1.8e+05 387.83 45.84 2289.4 16.14 2,498

46469 2407 180580 386.60 46.13 2309 15.63 2,498

Note: FFS=Fee For service, FHG= Family Health Group

Table 4: Impact of switching from FFS to FHG on physicians’ service production Weighted regressions (PSM )

Unweighted regressions Outcomes Log Total services Log CC services Log AH services Log non-incentivized services Observations

Pooled OLS 0.1255***

FE 0.0680***

HDFE 0.0158***

Pooled OLS 0.0389***

FE 0.0432***

HDFE 0.0216***

(0.0099)

(0.0068)

(0.0053)

(0.0116)

(0.0070)

(0.0057)

0.3132***

0.1259***

0.0267***

0.1432***

0.0841***

0.0341***

(0.0181)

(0.0098)

(0.0074)

(0.0195)

(0.0103)

(0.0080)

0.3282***

0.1380***

0.0362***

0.1640***

0.0940***

0.0471***

(0.0189)

(0.0108)

(0.0079)

(0.0228)

(0.0136)

(0.0093)

0.1558***

0.0916***

0.0215***

0.2644***

0.0570***

0.0254***

(0.0189)

(0.0106)

(0.0077)

(0.0266)

(0.0126)

(0.0082)

36,639

36,639

36,639

36,639

36,639

36,639

7,083 7,083 7,083 7,083 7,083 7,083 Physicians Legend: Significance level: * p<0.10; **p<0.05; *** p<0.01. Robust standard errors in parentheses. Note: HDFE: High-dimensional fixed-effects, FE: Fixed effects, FFS=Fee For service, FHG= Family Health Group. OLS regressions include prices of the different services, age, age squared, number of working holidays and weekends, dummy FTE>0.5, gender (female), IMG, age-female interaction term, age-(FTE>0.5) interaction term, average ADG score, average age of patients, % of patients in rural areas, % of male patients, 14 geographic indicators for regional health areas (LHINs) (LHIN number one: Erie St. Clair is used as reference), and an intercept. Fixed Effects regressions include the same explanatory variables as in OLS except IMG, female, age-female interaction term and age-(FTE>0.5) interaction term. High-dimensional Fixed effects regressions include the same explanatory variables as in Fixed Effects regressions except age, age squared, and the dummy FTE>0.5.

Table 5: Impact of switching from FFS to FHG on physicians’ service production Weighted regressions (CBPS) Outcomes Log Total services Log CC services Log AH services Log non-incentivized services Observations

Pooled OLS

FE

HDFE

Weighted regressions (EB) Pooled OLS FE HDFE

0.0420***

0.0397***

0.0223***

0.0490***

0.0433***

0.0215***

(0.0091)

(0.0075)

(0.0057)

(0.0118)

(0.0070)

(0.0056)

0.0803***

0.0654***

0.0298***

0.1591***

0.0868***

0.0335***

(0.0125)

(0.0103)

(0.0079)

(0.0196)

(0.0103)

(0.0079)

0.0902

0.0769***

0.0362***

0.1808**

0.0939***

0.0460***

(0.0135)

(0.0117)

(0.0085)

(0.0229)

(0.0137)

(0.0094)

0.1077***

0.0704***

0.0235***

0.2456***

0.0575***

0.0254***

(0.0172) 36,639

(0.0110) 36,639

(0.0083) 36,639

(0.0269) 36,639

(0.0130) 36,639

(0.0083) 36,639

7,083 7,083 7,083 7,083 7,083 7,083 Physicians Legend: Significance level: * p<0.10; **p<0.05; *** p<0.01. Robust standard errors in parentheses. Note: HDFE: High-dimensional fixed-effects, FE: Fixed effects, FFS=Fee For service, FHG= Family Health Group. CBPS=Covariate Balancing Propensity Score, EB=Entropy Balancing. OLS regressions include prices of the different

28

services, age, age squared, number of working holidays and weekends, dummy FTE>0.5, gender (female), IMG, agefemale interaction term, age-(FTE>0.5) interaction term, average ADG score, average age of patients, % of patients in rural areas, % of male patients, 14 geographic indicators for regional health areas (LHINs) (LHIN number one: Erie St. Clair is used as reference), and an intercept. Fixed Effects regressions include the same explanatory variables as in OLS except IMG, female, age-female interaction term and age-(FTE>0.5) interaction term. High-dimensional Fixed effects regressions include the same explanatory variables as in Fixed Effects regressions except age, age squared, and the dummy FTE>0.5. Table 6: Heterogeneity of responses Observations

[Physicians]

Outcome Variables

Log (total services)

Log (CC services)

Log (AH services)

Log (nonincentivized services)

Gender Males Females

24,320

0.0686***

0.1095***

0.1204***

0.0806***

[4,711]

(0.0114)

(0.0148)

(0.0189)

(0.0168)

12,319

0.0974***

0.1409***

0.1526***

0.1163***

[2,372]

(0.0144)

(0.0197)

(0.0211)

(0.0247)

Graduation cohort Grad. year < 1970 Grad. year 1970-1980 Grad. year 1980 -1990 Grad. year > 1990

6,502

0.1313***

0.1630***

0.1691***

0.1192***

[1,319]

(0.0314)

(0.0362)

(0.0379)

(0.0367)

9,995

0.0810***

0.1230***

0.1334***

0.0711**

[1,879]

(0.0172)

(0.0212)

(0.0229)

(0.0279)

11,923

0.0508***

0.0866***

0.0976***

0.0959***

[2,273]

(0.0144)

(0.0215)

(0.0321)

(0.0241)

8,219

0.0911***

0.1626***

0.1726***

0.0881***

[1,612]

(0.0167)

(0.0243)

(0.0256)

(0.0279)

Location South West Ontario Central Ontario South East Ontario Northern Ontario

8,251

0.0462***

0.1208***

0.1386***

0.0177

[1,709]

(0.0146)

(0.0221)

(0.0269)

(0.0252)

19,571

0.0561***

0.0699***

0.0776***

0.0924***

[3,713]

(0.0113)

(0.0147)

(0.0199)

(0.0192)

5,680

0.1217***

0.1910

0.1991***

0.0947***

[1,129]

(0.0269)

(0.0345)

(0.0363)

(0.0327)

3,137

0.1542***

0.1995***

0.2146***

0.1555***

[735]

(0.0306)

(0.0435)

(0.0464)

(0.0392)

Legend: Significance level: * p<0.10; **p<0.05; *** p<0.01. Robust standard errors in parentheses. Note: FFS=Fee For service, FHG= Family Health Group. We use weighted Fixed Effects regressions controlling for the prices of the different services, age, age squared, number of working holidays and weekends, average ADG score, average age of patients, % of patients in rural areas, % of male patients, and an intercept.

29

Appendix A1: Deriving participation decision From Section 3.2 we derive the budget constraint equations of physicians under the different models as the total quantity of services provided which reflects physicians’ gross earnings or consumption. Under FFS payment system, the total quantity of services provided by the physician i (denoted by 𝐶𝑖𝐹𝐹𝑆 ) is ∗ ∗ ∗ 𝐶𝑖𝐹𝐹𝑆 = 𝔭1 𝑄𝑖1 + 𝔭2 𝑄𝑖2 + 𝔭3 𝑄𝑖3 = 𝒷(𝑋𝑖 )𝑤𝐹𝐹𝑆 ℎ𝑖∗𝛿 𝜀𝑖

(8)

where 𝑤𝐹𝐹𝑆 = (𝑝1 + 𝑝2 + 𝑝3 )1−𝛿 represents the marginal return to an hour when that hour is 1

optimally allocated across services, and 𝑝𝑗 = (𝔭𝑗 )1−𝛿 . While, under FHG payment model, the predicted total quantity of services provided (denoted by 𝐶𝑖𝐹𝐻𝐺 ) by the physician 𝑖 is 𝛾

𝐶𝑖𝐹𝐻𝐺 = (1 + 𝜏1 )𝔭1 𝑄̂𝑖1 + (1 + 𝜏2 )𝔭2 𝑄̂𝑖2 + 𝔭3 𝑄̂𝑖3 = 𝒷(𝑋𝑖 )𝑤𝐹𝐻𝐺 ℎ̂𝑖𝛿 𝑛̂𝑖 𝜀𝑖 −𝛾

(9)

1−𝛿

where 𝑤𝐹𝐻𝐺 = (𝑃1 + 𝑃2 + 𝑃𝛾+𝛿−1 𝑃3 )

represents the marginal return to an hour/day worked 1

when that hour/day is optimally allocated across services, and 𝑃𝑗 = ((1 + 𝜏𝑗 )𝔭𝑗 )1−𝛿 , 𝑗 = 1,2; 𝑃3 = 1

(𝔭3 )1−𝛿 . Equations (8) and (9) represent the budget line for a FFS physician and the budget curve for a FHG physician, respectively, as drawn in Figure 1. Comparing the total production of services (gross earnings) under the two different payment models, at ℎ𝑖∗ (optimal hours worked under FFS), gives rise to a participation to FHG decision 1

𝑤𝐹𝐹𝑆 𝛾

which is: the physician 𝑖 switches to FHG, if and only if 𝑛̂𝑖 > (𝑤

𝐹𝐻𝐺

30

) = 𝑛∗ .

Appendix A2: Tables Table A2.1: Comprehensive care fee codes OHIP fee Codes

Definition of the fee codes

A001 A003 A007 A008 A888 A901 A902 C010 C882 G365 G538

MINOR ASSESS.-F.P./G.P. GEN. ASSESS. -F.P./G.P. INTERMED.ASSESS/WELL BABY CARE-F.P./G.P./PAED. MINI ASSESSMENT-F.P./G.P. PARTIAL ASSESSMENT EM.DEPT EQUIVALENT GENERAL/FAMILY PRACTICE-HOUSECALL ASSESSMENT HOUSECALL ASSESS - PRONOUNCEMENT OF DEATH IN HOME SUPPORT CARE-F.P./G.P.-HOSP TERMINAL CARE IN HOSP.G.P/F.P D./T. PROC.-GYNAECOLOGY-PAPANICOLAOU SMEAR D&T IMMUNIZATION-WITH VISIT, EACH INJECT INJECTION OF UNSPECIFIED AGENT - SOLE REASON (FIRST INJECTION) INFLUENZA AGENT +VISIT INJECTION OF INFLUENZA AGENT - SOLE REASON INDIVIDUAL CARE PER 1/2 HR COUNSELLING-ONE OR MORE PEOPLE-PER 1/2HR. ANNUAL HEALTH EXAM-CHILD AFT. 2ND BIRTHDAY. HIV PRIM CARE INDIVID CARE 1/2 HR OR MAJOR PART PALLIAT CARE SUPPORT INDIVID CARE 1/2 HR OR MAJOR PART DIABETIC MANAGEMENT FEE

G539 G590 G591 K005 K013 K017 K022 K023 K030

Table A2.2: After Hours care fee codes OHIP fee Codes

Definition of the fee codes

A001 A003 A004 A007 A008 A888 K005 K013 K017 Q012A

MINOR ASSESS.-F.P./G.P. GEN. ASSESS. -F.P./G.P. GEN.RE-ASSESS-F.P./G.P. INTERMED.ASSESS/WELL BABY CARE-F.P./G.P./PAED. MINI ASSESSMENT-F.P./G.P. PARTIAL ASSESSMENT EM.DEPT EQUIVALENT INDIVIDUAL CARE PER 1/2 HR COUNSELLING-ONE OR MORE PEOPLE-PER 1/2HR. ANNUAL HEALTH EXAM-CHILD AFT. 2ND BIRTHDAY. AFTER HOURS PREMIUM

31

Table A2.3: Propensity score estimates

Expected gain in income Expected gain in income squared Total services in 2003 Total services in 2003 squared Total services in 2003*(FTE>=0.5) Age*(FTE>=0.5) Age squared*(FTE>=0.5) Age*female Age Age squared Female FTE>=0.5 IMG Total days worked Average ADG score Av. Age of patients Prop. of patients in rural areas Prop. of male patients LHIN02 = South West LHIN03 = Waterloo Wellington LHIN04 = Hamilton Niagara Haldimand Brant LHIN05 = Central West LHIN06 = Mississauga Halton LHIN07 = Toronto Central LHIN08 = Central LHIN09 = Central East LHIN10 = South East LHIN11 = Champlain LHIN12 = North Simcoe Muskoka LHIN13 = North East LHIN14 = North West intercept Observations

Logit model Coefficient Std. error 0.0081*** 0.0005 -9.82e-06*** 2.45e-06 -0.0004 0.0019

CBPS model Coefficient Std. error 0.0073*** 0.001 -9.029e-06*** 0.0000 0.0024 0.0015

-5.81e-06*** 8.78e-07

-6.640e-06***

0.0000

0.0060*** 0.0549 -0.0007 -0.0317*** 0.0416 -0.0005 1.9172*** -1.4781 -0.5107*** 0.0064*** -0.1023* 0.0009 -0.2016 -0.9526*** 0.4809*** -0.0222

0.0019 0.0553 0.0005 0.0068 0.0515 0.0005 0.3374 1.3994 0.0883 0.0007 0.0580 0.0050 0.1292 0.3369 0.1787 0.1982

0.0041*** 0.0688 -0.0008 -0.0324*** 0.0399 -0.0005 1.8809*** -1.7048 -0.4813*** 0.0052*** -0.0752 0.0025 -0.1579 -1.0563*** 0.4399** -0.0347

0.0016 0.054 0.0005 0.0065 0.0506 0.0005 0.3236 1.3686 0.0849 0.0005 0.0504 0.0044 0.1127 0.2992 0.1738 0.1889

0.0462

0.1709

0.0628

0.1675

0.8230*** 0.2939* 0.2849* 0.4828*** 0.3854** 0.5171** 0.7115*** 0.9089*** -0.0144 0.1792 -2.6942**

0.2062 0.8214*** 0.1961 0.1745 0.2825* 0.1703 0.1610 0.3095* 0.1572 0.1630 0.458*** 0.1589 0.1639 0.3793** 0.1594 0.2073 0.4856** 0.1999 0.2072 0.6352*** 0.1626 0.2216 0.845*** 0.2016 0.1955 -0.0968 0.178 0.2374 0.0967 0.2149 1.3684 -2.6721 1.3406 7,088 7,088 * p<0.10; **p<0.05; *** p<0.01

32

Table A2.4: Regression test before and after matching regression test PSM

CBPS Before After

Before

After

0.0933

0.0571

0.1649

0.0799

0.0998

0.1447

0.1687

0.1082

0.1203

0.1447

0.4007

Total services in 2003 Total services in 2003 squared Age

0.0262

0.0624

0.0258

0.0676

0.0399

0.0711

0.0629

0.0646

0.0311

0.1779

0.0629

0.1221

0.2274

0.1724

0.1659

0.3319

0.2274

0.2947

Age squared

0.2902

0.1551

0.1929

0.5920

0.2902

0.2845

0.0608

0.0729

0.0374

0.0869

0.0608

0.111

0.0774

0.2027

0.0407

0.4347

0.0774

0.4565

0.0649

0.1800

0.0343

0.2095

0.0649

0.4276

0.2548

0.4606

0.1173

0.1445

0.2548

0.1054

Total days worked

0.0006

0.0047

0.0002

0.2626

0.0045

0.0043

Average ADG score

0.2595

0.3997

0.2472

0.6066

0.2595

0.3733

Av. Age of patients Prop. of patients in rural areas

0.0128

0.0077

0.0248

0.0258

0.0128

0.0090

0.0287

0.0912

0.0021

0.1230

0.0287

0.0217

Prop. of male patients

0.0027

0.5885

0.0173

0.1089

0.0027

0.9843

Female

0.1519

0.456

0.0027

0.1046

0.1519

0.0733

FTE>=0.5

0.537

0.7767

0.3857

0.9941

0.537

0.8837

IMG LHIN02 = South West

0.6049

0.5954

0.6614

0.1470

0.6049

0.4881

0.6309

0.1246

0.3939

0.4694

0.4181

0.6280

0.0304

0.0530

0.3482

0.2845

0.3621

0.3872

0.6259

0.5961

0.5303

0.3207

0.6259

0.5995

0.3488

0.2118

0.2617

0.0504

0.3488

0.1526

0.6044

0.7468

0.6262

0.8566

0.6044

0.8532

0.3333

0.2687

0.2832

0.3036

0.3333

0.3166

Covariates Expected gain in income Expected gain in income squared

Total services in 2003*(FTE>=0.5) Age*(FTE>=0.5) Age squared*(FTE>=0.5) Age*female

LHIN03 = Waterloo Wellington LHIN04 = Hamilton Niagara Haldimand Brant LHIN05 = Central West LHIN06 = Mississauga Halton LHIN07 = Toronto Central LHIN08 = Central

Before

After

0.0799

Entropy

0.2143

0.2839

0.2751

0.5929

0.2143

0.0927

LHIN09 = Central East

0.2488

0.2804

0.2688

0.6095

0.2488

0.3766

LHIN10 = South East

0.4443

0.2581

0.5263

0.2734

0.4443

0.1554

LHIN11 = Champlain

0.6569

0.7655

0.5793

0.8409

0.6569

0.8339

0.0000

0.0510

0.0123

0.2456

0.0331

0.0331

0.3433

0.0135

0.0000

0.0165

0.0033

0.0074

0.0668 7,083

0.0521 7,083

0.0949 7,083

0.7465 7,083

0.0668 7,083

0.0701 7,083

LHIN12 = North Simcoe Muskoka LHIN13 = North East LHIN14 = North West Number of Phys.

33

Note: We consider a polynomial of degree 7 or smaller in the estimated propensity score to compute the F-tests for the covariates.

Appendix A3: Indices Quantities: The formula for calculating the quantity index of type j=1,2,3 service for a physician i at year t is given by 𝑁

𝑄𝑖𝑗𝑡 = ∑ 𝑞𝑖𝑘𝑡 𝑝𝑘,2003 𝑘=1

where 𝑝𝑘,2003 is the fee for service k in 2003, 𝑞𝑖𝑘𝑡 is the number of service k performed by physician i at year t=2003-2008 and N is the number of fee codes or services of type j. Prices: The Laspeyres price index formula for service j at year t is ∑𝑛

𝑝𝑘,𝑡 𝑞𝑘,2003

𝑘=1

𝑘,2003 𝑞𝑘,2003

𝔭𝑗𝑡 = ∑𝑛 𝑘=1𝑝

,

with 𝑝𝑘,𝑡 the fee for service k (of type j) in year t and 𝑞𝑘,2003 the number of service k (of type j) in 2003.

34