A critical issue within performance measurement is assigning responsibility (also called "attribution") for the services or set of services that are or are not provided to a beneficiary. How an episode is constructed, as illustrated by the building block approach, and then used (e.g., resource use profiling, pay for performance, larger payment and performance measurement reform), implies different issues related to attribution.
An episode-based approach that cuts across the continuum of care would require that accountability for the episode is assigned to an entity or group of entities. The accountable entities would then assume the responsibility for the quality and resource use for a range of services provided during the episode. The accountability could be reinforced in a range of ways, including measurement and reporting of performance and resource use for episodes of care; financial incentives for performance and resource use for episodes of care, or at the extreme, episode-based payment adjusted for performance. In reviewing the literature, we sought to understand how attribution has been addressed either in practice or in concept within the area of performance measurement, as a means to inform the implications of different types of episode constructions and applications.
Our review of the published literature on assignment of accountability for an episode of care revealed two distinct approaches that have been used or proposed for use to assign responsibility:
- Prospective assignment-where an entity assumes accountability for a patient before the episode begins (similar to what occurs under capitation),or
- Retrospective assignment--where an entity is assigned accountability following the episode based on analyses of care patterns.
The main distinction between these two approaches is that under prospective assignment, providers and patients are aware of the accountability before services are delivered, whereas under retrospective assignment accountability is assigned after care is delivered. Prospective assignment of accountability would likely be necessary for prospective payment approaches (the extreme end of the spectrum in the "building block" approach to reform). The retrospective or prospective method for assigning accountability could be used for any of the other approaches discussed.
The entities with accountability can be individual providers, integrated provider organizations, or "virtual groups" - that is, a group of independent providers that create a relationship for the purposes of coordination across the episode of care. Prospective designation allows for some choice by physicians and patients about which providers should be responsible for which patients' episodes of care (Davis, 2007; Pham et al., 2007), but it also creates the possibility of risk selection by incentivizing providers to assume accountability for healthier, more-profitable patients. Several methods for retrospective attribution have been proposed in the literature and are described below.
(1) Individual physician(s). The accountable physician(s) could be identified retrospectively through analysis of claims data, although current provider identifiers imperfectly identify individual physicians and their practice specialty.29 Commonly proposed criteria include a count of Evaluation and Management (E&M) visits or costs, physician specialty type, or some combination thereof (Cheryl Damberg, personal communication, 8/6/2008).
One study examined attribution for a year of care for Medicare beneficiaries to individual physicians (Pham et al., 2007). The major finding was that dispersion of care among multiple providers made retrospective attribution of accountability difficult. Although the study focused on retrospective assignment of accountability, the dispersion of care observed would also likely prevent physicians from prospectively claiming responsibility for patients who receive much of their care from other physicians. - Four assignment algorithms were tested, mirroring assignment methods currently used in pay-for-performance programs. The four algorithms tested were:
- Plurality provider algorithm - beneficiary assigned to physician who provided the greatest number of E&M visits in a given year;
- Plurality primary care physician algorithm - same as (1) only assignment was restricted to primary care physicians;
- Majority provider algorithm - beneficiary assigned to physician who provided >50 percent of E&M visits in a given year;
- Multiple provider algorithm - beneficiary assigned to all providers providing at least 25 percent of E&M visits.
Table 6 summarizes the results of application of these four assignment algorithms. The percentage of all beneficiaries with at least one E&M visit who were assigned to beneficiaries ranged from a low of 65 percent (majority provider algorithm) to a high of 97 percent (multiple provider algorithm). For all four algorithms, on average most Medicare patients a physician treated in a year were not assigned to that physician. That is, most beneficiaries in a physician's practice population received a minority of their E&M services from that physician. This was particularly true for specialists. Under the plurality provider algorithm, which assigns patients to either specialists or generalists, primary care physicians were assigned 39 percent of the beneficiaries for whom they provided services, while specialists were assigned only 6 percent of the beneficiaries they treated. (The study did not test algorithms based on costs rather than visit counts, which would be expected to assign more beneficiaries to specialists.) Care was highly dispersed: under the plurality provider algorithm, in one year the typical beneficiary saw two primary care physicians and five specialists, collectively from four different practices. Higher dispersion of care among physicians was found for patients with more chronic conditions. Many patients also changed physicians year-to-year (based on claims analysis) under all four algorithms.
Pham et al. conclude that the dispersion of care observed make it difficult to hold individual physicians accountable for a year of patient care. Episodes of care may be more highly concentrated among physicians, making attribution of accountability for an episode more feasible than for an entire year of care. However, the results also indicate that attribution is very sensitive to the algorithm used, and that each approach likely involves tradeoffs between a number of criteria that may be important.
A RAND study examined the effects of 13 different retrospective attribution rules, in an application where the Symmetry ETG tool was used to construct resource use measures using commercial data from four health plans in Massachusetts (Mehrotra et al., 2007). The 13 rules assignment rules differed on characteristics such as the basis of assignment (e.g. costs versus visits) and whether the episode was assigned to only one or multiple physicians. This study found both significant variation in the fraction of episodes that could be assigned to a physician and also the level of agreement in which physician was held responsible. For example, comparing the results of two different rules found that 50 percent of the episodes were assigned to different physicians. The results demonstrate that different assignment methods can lead to substantially different results on various criteria.
MedPAC conducted several similar analyses, testing assignment of accountability for episodes of care, measured using ETGs and MEGs, to individual physicians. They found that most episodes could be assigned to individual physicians using a threshold of 35 percent of E&M visits. They also explored attribution to multiple providers, but found that few episodes had more than one physician providing more than 35 percent of E&M visits (MedPAC, 2006; MedPAC, 2007d). Some specialties saw a broad range of types of episodes, while other specialties mainly saw a small number of episode types (MedPAC, 2007d).
(2) Individual physician – hospital care only. A variant on assignment of accountability to an individual physician is assignment of services provided during a hospital stay to the attending physician. This model was tested in the study of Physician DRGs mandated by Congress with the implementation of IPPS (Jencks and Dobson, 1985; Mitchell, 1985; Mitchell et al., 1987; Welch, 1989) . The analysis showed that spending on physician services for surgical cases was relatively homogeneous, but that spending for medical cases varied widely. Thus, assignment of responsibility for hospital-based physician services to individual attending physicians would be likely to cause substantial financial risk for the attending physician (Mitchell, 1985; Welch, 1989) . This finding was one of the major reasons that Physician DRGs were not considered further. Subsequent proposals and analyses focused on spreading the financial risk more broadly.
(3) Hospitals. Another attribution approach that has been proposed is to hold hospitals accountable for episodes of care that include a hospitalization in addition to physician services and/or services from other providers, such as skilled nursing facilities (Jencks and Dobson, 1985; Welch, 1989) . One issue that has been raised with this approach is that hospitals may not be able to influence physician and/or post-acute provider care provision (Welch, 1989). One solution that has been tested is gainsharing, which is a payment arrangement by which hospitals incentivize physicians (Wilensky et al., 2006). However, there are several legal restrictions against gainsharing (Wilensky et al., 2006). These regulations are motivated by a concern about the incentives created for skimping on care, selection of healthy patients, and kickbacks from hospitals to physicians for referrals. MedPAC recommended loosening the restrictions against gainsharing given appropriate safeguards for these concerns (MedPAC, 2007). CMS is planning two gainsharing-related demonstrations, the Medicare Hospital Gainsharing Demonstration and the Physician Hospital Collaboration Demonstration (Wilensky et al., 2006).
(4) Integrated Delivery Systems and Physician Group Practices. Existing integrated provider organizations are likely to have the greatest ability to assume responsibility for episodes of care because of the defined relationships between providers (Davis and Guterman, 2007; MedPAC, 2007a; MedPAC, 2007b; MedPAC, 2007c) . For example, integration was considered key to successful implementation of CABG bundling at Geisinger (Casale et al., 2007; Lee, 2007). Examples of integrated organizations with both hospitals and physicians include physician-led multispecialty group practices that also own hospital(s) (e.g., Mayo Clinic, Virginia Mason); hospitals that own physician groups (e.g., Intermountain Healthcare); or physician-hospital organizations (e.g., Advocate Health Partners; these have declined since late 1990s) (Cortese and Smoldt, 2007). However, there are several obstacles to attributing accountability for episodes of care primarily to integrated provider organizations. First, patients have the option to use services outside of the integrated provider organization, limiting control over the episode of care. Second, most physicians are organized in solo or small single-specialty practices, not integrated organizations or large groups (Pham and Ginsburg, 2007). Finally, Medicare currently does not recognize these integrated entities as a provider class eligible for payment (Davis and Guterman, 2007).
(5) Hospital medical staff. This model would assign accountability for acute care episodes to the entire medical staff of a hospital (holding the hospital accountable as well). In most proposals, the medical staff would be defined to comprise both hospital-based physicians such as pathologists and community-based physicians who see patients in the hospital. Since most medical staffs are not true legal organizations, they would have to form new legal entities in order to receive payment, including a performance-based bonus (Jencks and Dobson, 1985; Welch, 1989; Davis and Guterman, 2007) . This would essentially form a multispecialty group practice associated with a hospital (Fisher et al., 2006).
The concept of assigning accountability to a medical staff was initially tested in the context of payment reform after the initial Physician DRG concept - attribution to the attending physician - was defeated (Mitchell et al., 1987; Mitchell and Ellis, 1992) . These claims data analyses showed that paying a medical staff for physician services delivered during hospitalizations involved a more acceptable level of financial risk than paying individual physicians. The concept was then analyzed using claims data as a possible replacement for the Volume Performance Standard (Miller and Welch, 1992). The Volume Performance Standard was meant to control physician spending by linking annual fee schedule updates to the rate of increase in service volume. However, by measuring volume at the national level, the incentive for individual physicians was weak. Creating an equivalent arrangement at a smaller measurement unit, such as the hospital medical staff, would strengthen the incentive. Physician fees would be adjusted based on medical staff resource use in the prior year, so that different hospital staffs would have different payment rates. Neither of these studies explicitly examined quality measurement.
Elliott Fisher and colleagues have tested the feasibility of defining hospital medical staffs, which they call Accountable Care Organizations (ACOs), empirically using insurance claims data (Fisher et al., 2006; Bynum et al., 2007) . The ACOs were designed with the intent to hold them accountable for both quality and resource use. Beneficiaries were assigned to physicians and then through the physicians to hospitals based on service use in a defined time period. Beneficiaries were linked to the physician who was the generalist or medical subspecialist providing the plurality of their ambulatory care visits in a two-year period (the authors did not test how often the assigned physician changed over time). Physicians were assigned to a hospital based on the number of patients for whom they had submitted Part B claims or the number of hospital claims for which they were listed as attending or operating physician during hospitalization. If both of these values were zero, assignment was based on where the patients they treated were hospitalized. Using this method, 94 percent of physicians were assigned to a hospital. One-third of physicians bill at multiple hospitals, but typically provide the majority of their care at one hospital. On average, two-thirds of medical admissions and physician billing for patients were provided by the assigned hospital and medical staff (Bynum et al., 2007) . Tests were favorable for face, discriminant, and predictive validity of assignment (Bynum et al., 2007) . The advantages cited for using ACOs for accountability (compared to individual providers) include larger sample size (98 percent of physicians were assigned to ACOs serving more than 500 Medicare beneficiaries), broader scope of potential performance measures (e.g., measures of the fragmentation of care), and feasibility of including all contributing physicians within the measurement frame. The most important reason for using ACOs as the level of accountability, in the authors' view, is to establish accountability for local decisions about capacity, which drives utilization. In addition, hospitals and extended staffs would have greater incentive to invest in care management and coordination (Fisher et al., 2006). While developed with the intent of assigning a beneficiary to an ACO, the approach could also be used to assign episodes to ACOs and the research has been used as a basis for discussion of episodes by MedPAC in 2007 Commissioners meetings (MedPAC, 2007a; MedPAC, 2007b; MedPAC, 2007c) .
If the medical staff received payments, the organization would then need a process to allocate payment to individual physicians (Miller and Welch, 1992). One model could be contact capitation, where budgets were allocated to departments based on historical costs and then departments paid individual physicians based on productivity (Robinson, 1999). However, if the episode included a significant pre- or post-hospitalization window, it is possible that some physicians providing care would be geographically distant from the hospital (MedPAC, 2007c), which would require the development of alternative methods of payment allocation and performance accountability.
Since the hospital medical staff model uses "virtual" groups as the accountable entities, a significant barrier is the lack of integration between group members. In recent years, relations between physicians and hospitals have become increasingly strained (Fisher et al., 2006; Berenson et al., 2007) . This tension will likely be a significant barrier to holding hospitals and physicians jointly accountable for episodes of care (Pham and Ginsburg, 2007). On the other hand, it is possible that holding hospitals and medical staff jointly accountable for episodes of care could encourage physician-hospital collaboration.
(6) Other Virtual Groups. Some authors have raised the possibility of using other "virtual groups" defined by geographic areas or other characteristics (Davis and Guterman, 2007). No detailed proposals have been made, however.