Caroline Carlin, PhD

 To compare the performance of methods to retrospectively attribute patients to provider systems by comparing the fraction attributed and the stability of assignment over time. 

Descriptive statistics are used to measure the fraction of patients attributed and stability of attribution from year to year. This study uses a panel of administrative claims data (2010-2011). Attribution rules were defined by unit of measure (count of physician visits, dollars paid), type of providers (primary care physicians [PCPs], all physicians), type of encounters (all visits, evaluation and management visits only), and level of concentration of care (majority, plurality). We created 32 retrospective attribution rules, spanning PCP-only rules, all-physician rules, hierarchical rules based on PCPs then all physicians, and lookback rules based on current-year PCP visits then prior-year experience.

 All methods exhibit a tradeoff between stability of attribution and fraction of the population attributed. This tradeoff is minimized when PCP-based rules are supplemented by a 1-year lookback when the current-year experience does not result in attribution. 

 We recommend using this lookback method when multiple years of data are available. In absence of multiple years of data, PCP-based rules maximize stability; hierarchical rules result in a greater fraction attributed with less loss of stability than simple all-provider rules.

The process of assigning patients to provider systems is a key component of many payment reform efforts. The majority of these payment systems rely on retrospective attribution of patients to provider systems, but there is no generally accepted method. We evaluated retrospective attribution rules, measuring the fraction attributed and the stability of attribution. 

The performance of attribution rules varies in terms of patients attributed and the stability of attribution. 

We recommend attribution rules based on primary care physician visits, with a look back to the prior year’s experience if the most recent year does not result in attribution.

Much of provider payment reform (bundled payments,

) requires patients to be linked to an individual provider or provider systems. Because prospective selection of a medical home is rare, many compensation methods rely on retrospective attribution of patients to providers. Retrospective attribution is the process of assigning responsibility for patients to an individual provider or provider systems after care is received. However, there is no generally accepted method of retrospective attribution, and methods are sometimes proprietary, making comparison of methods difficult.

 Attribution rules are typically defined by: 1) level of attribution (provider system, clinic location, individual physician), 2) provider types included (primary care physicians [PCPs], specialty physicians, nurse practitioners, physician assistants), 3) encounter type (any office visit, evaluation and management [E&M] visits), 4) unit of measure (visit counts, allowed charges), and 5) level of concentration of care required (majority, plurality).

For instance, the Medicare Shared Savings Program (MSSP) uses visits to PCPs (ie, general practice, family practice, internal medicine, and geriatric medicine physicians) to attribute patients to provider systems. In the MSSP, a provider system is responsible for a patient if the overall cost of all types of visits to PCPs was in that system. If a patient did not have PCP visits, then they use visits to all physicians as the attribution criterion.

 Both attribution and financial evaluation are retrospective. Specifically, after 2015 claims data were completed in 2016, Medicare used this 2015 experience to retrospectively attribute patients to provider systems and evaluate whether the year’s financial goals were achieved.

An ideal attribution rule balances 2 conflicting goals: maximizing the fraction of patients attributed and maximizing the stability of attribution to a particular provider system. An attribution rule assigning a large fraction of patients to a provider system makes payment reform relevant to a larger portion of the insured population. However, such rules may compromise stability. Conversely, attribution rules that consistently assign patients to the same provider system generate stable incentives for managing population health. An unstable attribution rule may assign patients to different provider systems over time due to variations in patterns of care rather than a true change in patient—provider affiliation; however, maximizing stability may require criteria so strict that the rule attributes few patients.

Our goal was to evaluate a variety of attribution rules by comparing the fraction attributed with the stability of assignment for each rule. Because we focused on shared savings contracts and other provider reimbursement methods administered at the level where reimbursement is negotiated, we attributed patients at the provider system level. We expected that more restrictive measures (PCPs only, E&M visits only) would exhibit greater stability of attribution over time and broader measures (all types of physicians, all types of visits) would attribute a greater fraction of the population but may be less stable.

There is a small body of literature evaluating attribution methods, much of it focused on attribution at the individual provider level. Mehrotra et al

 used commercial health plan data to compare 12 rules attributing patients to individual physicians and the subsequent categorization of physicians in cost categories. They found that the attribution rule could substantially affect the cost category assignment. Pham et al

 tested the fraction attributed and stability of attribution of Medicare patients to individual physicians; attribution rules were based on E&M visits and differed by concentration of care (majority, plurality), provider type (PCPs only, any provider), and a rule that allowed a patient to be attributed to several providers. The fraction attributed ranged from 79% to 94%, and stability of attribution ranged from 48% to 74%. In one of the few studies comparing attribution rules at the provider system level, Lewis et al

 compared prospective and retrospective application of the Medicare Shared Savings rule to Medicare beneficiaries. They found retrospective attribution more accurately identified patients who received care from an accountable care organization (ACO). A robustness check using claims from all types of physicians yielded the same results.

Our study builds on and improves the literature in several ways. First, we attributed health plan members at the provider system level to support public and private shared savings contracts. Second, since these compensation methods are being adopted by private health plans, we expand the literature by assessing how attribution performs among commercially insured and managed Medicaid populations.

The study used data from a regional health plan headquartered in Minnesota, which is at the forefront of consolidation of providers into vertically integrated delivery systems (IDSs) capable of functioning as ACOs and of shared savings contracting with public and private payers.

 Yet, there remains significant exposure to other care settings, with 64% of the office visits in our data coming from providers practicing outside of an IDS setting. We believe that Minnesota is a mature shared savings environment that offers insights into the future of delivery of care in America.

We used claims and enrollment data from 2010 and 2011 for plan members (“patients”) aged 18 to 64 years. Patients were enrolled in commercial group, individual and family, and managed Medicaid products. The study was limited to adults younger than 65 years because Medicare paid on a primary basis for the plan’s Medicare Cost products, so we may not have observed all physician visits for these products.

Attribution measures were based on visits that occurred in a home or clinic setting or, for Medicaid only, the emergency department (as Medicaid patients may seek primary or specialty care at the emergency department in nonemergent settings

). Encounters were categorized as E&M visits (identified by the current procedural terminology [CPT] codes 99050, 99053-99057, 99058, 99201-99205, 99211-99215, 99241-99245, 99324-99328, 99334-99337, 99339-99345, 99347-99350, 99354, 99355, 99374, 99375, 99381-99387, 99391-99397) or non-E&M visits. We excluded most nonphysician encounters, although visits to nurse practitioners and physician assistants were retained with visits to physicians, categorized as primary or specialty care according to their area of practice. Providers specializing in general practice, internal medicine, pediatrics, family practice, or obstetrics-gynecology were categorized as primary care. Cost of care was defined as allowed charges, including both plan and member payments.

 shows sociodemographic characteristics of the population. Demographics (age, sex, plan type) were drawn from enrollment data. A categorical measure of health status (resource utilization band) and indicators of frailty and chronic conditions (diabetes, depression, high cholesterol, and high blood pressure) were computed by the Johns Hopkins Adjusted Clinical Groups (ACG) System

 based on diagnosis and pharmacy history. Neighborhood effects were based on 2010 US Census and 2011 5-year average American Community Survey data, within block (percent non-Hispanic whites), block group (levels of median income, education distribution), and census tract (urban, percent foreign-born).

Our attribution rules retrospectively assigned patients to a provider system based on the most recent calendar year of claims experience. These provider systems range from small practices to large IDSs with multiple clinic locations. We created provider systems by grouping providers within Tax Identification Numbers (TINs), additionally grouping TINs into the level at which the system contracts with the health plan. Each attribution rule is defined by: 1) provider type (any physician, PCPs only), 2) encounter type (any visits, E&M visits only), 3) unit of measure (number of visits, allowed charges for visits), and 4) required concentration of care (majority, plurality with a minimum fraction of 35%). Criteria 1 and 2 define the scope of data needed, criterion 3 defines the metric used, and criterion 4 defines the decision rule. Together, these criteria define 16 attribution rules. For example, a patient could be attributed to the provider system from which they receive the majority of E&M visits with a PCP. In the case of a tie, attribution was made to the provider system providing the greater fraction of the alternate unit. For example, if an attribution rule was tied at 50% of visits in 2 provider systems, the tie was broken by allowed charges.

We tested 2 additional sets of 8 attribution rules that increase the fraction attributed. These rules first used a PCP-only method and, for those not initially attributed, a supplementary method. In the hierarchical method, if a member was not initially attributed, the analogous all-physician method was used. In the lookback method, if a member was not initially attributed (2011), we looked back to the member’s experience in the prior year (2010) to see if they had coverage and would be attributed in the prior year. If so, this method carried that attribution forward to 2011. Altogether, we evaluated 32 attribution rules.

 displays the results for the 32 rules, showing fraction attributed and stability of attribution, with details available in the eAppendix [

]. Fraction attributed is simply the percentage of patients attributed. Stability is the percentage of patients attributed in the current and prior years that stayed in the same system in both years. Figure 1 plots the fraction of patients attributed against the stability of attribution for the all-physician (squares), PCP (diamonds), hierarchical (triangles), and lookback (circles) attribution rules, including enrollees with both partial and full years of coverage. eAppendix Figure 1 shows that while enrollees with partial years have a lower fraction attributed, the relative performance of the attribution rules is consistent between partial-year and full-year enrollees. The type of visit used for attribution is indicated by the size of the symbol, with larger symbols indicating the use of all visits and smaller symbols indicating the data were restricted to E&M visits. If the attribution units were cost (allowed charges), the outline of the symbol matched the color of the fill; black outlines indicate that the count of visits was used. Where the decision rule was based on majority of cost or counts, the symbol is filled with a darker shade; lighter fill indicates a plurality rule. Figure 1 shows that the lookback rules clearly dominate in both fraction attributed and stability. For these rules, the choice of cost versus count or majority versus plurality makes little difference in their performance. Restricting our attribution process to E&M visits only results in a minor reduction in the fraction attributed for lookback methods.

Of course, it may be possible that only a single year of experience is available with which to determine attribution, precluding the lookback method. In this event, the PCP rules provide more stability relative to the alternatives, although they result in a smaller fraction of patients attributed. In contrast, the hierarchical rules provide a larger fraction attributed than the PCP rules, at the cost of some stability of attribution. As for the lookback rules, the choice of cost versus count or majority versus plurality makes little difference to either the PCP rules or the hierarchical rules.

Although we also find a relatively larger fraction of the patients attributed using all-visit rules rather than PCP rules, this is at a greater cost of stability relative to the performance of both the PCP and hierarchical rules. These rules also exhibit the greatest sensitivity to the other criteria, including the choice of cost versus count or majority versus plurality.

Both provider systems and insurers may be interested in the fraction of allowed charges for attributed patients (sum of plan and patient payments) rather than the fraction of patients attributed. 

 compares, for all attribution rules, the fraction of patients in the attributed population (top panel; previously displayed on the horizontal axis of Figure 1) with the fraction of allowed charges for the attributed population (bottom panel). We found that the attributed population generates a very large fraction of the allowed charges for all methods, with the lookback and hierarchical methods capturing 95% or more of allowed charges, regardless of the other methodological criteria.

Most of the meaningful differences in both percent of patients attributed and charges from attributed patients are driven by the class of attribution rule chosen, although the difference in percent of charges from attributed patients is not meaningfully different between the lookback and hierarchical classes. There are 2 additional factors beyond class of attribution rule that drive meaningful differences: 1) The hierarchical methods are sensitive to whether all visits or only E&M visits are used; this is most meaningful in the fraction of patients attributed. Using all visits generates a larger fraction of attributed patients and allowed charges. 2) The all-provider methods are sensitive to the choice of plurality versus majority rules, with the greater instability across all-provider methods seen in percent of charges from attributed patients relative to percent of patients attributed. All-provider plurality rules have a larger fraction of patients and allowed charges in the attributed population than do all-provider majority rules. Given that the other 3 classes of attribution methods rarely capture fewer than 90% of allowed charges in the attributed population, the instability of this metric for all-provider rules is the most striking result in Figure 2.

Health status of patients may substantially change the performance of an attribution method. Related work

 using a PCP-based method shows that a patient with a more complex health status is more likely to be attributed, but also more likely to switch attributed providers from year to year, possibly due to primary care and specialty relationships spanning multiple provider systems. To assess the performance of the attribution rules by health status, we repeated the calculation of fraction attributed and stability of attribution for patients by health status.

We computed categorical measures of health status (resource utilization band) and indicators of chronic conditions (diabetes, depression, high cholesterol, and high blood pressure) with the Johns Hopkins ACG System,

 based on diagnosis and pharmacy history. With these indicators, we measured health status in 2 ways: by the presence or absence of at least 1 chronic condition and whether or not the ACG system categorizes the member as high/very high risk (“complex health status”) relative to lower-risk categories.

Whether segmenting the population by the presence of chronic conditions (eAppendix Figure 2) or by health status complexity (eAppendix Figure 3), these tables show a higher fraction attributed for sicker patients across all attribution rules. In contrast, the tables show attribution stability is higher for patients with 1 or more chronic conditions across all attribution rules relative to those with no chronic conditions (Figure 3), but relatively consistent when comparing those with high complexity in health status to those with low complexity in health status (Figure 4). Thus, we see a marked difference between the impact of chronic conditions (typically managed in a primary care setting) and complex health status (a mix of chronic and acute conditions) on stability of attribution.

This study compared the performance of retrospective attribution rules used to assign patients to provider systems. We evaluated rules based on the fraction of patients attributed and stability of attribution. Results showed that the lookback method, especially when all visits were used to compute the measure, clearly dominated all alternatives. If 2 years of data are not available to compute attribution, the preference for breadth of attribution or stability of attribution will drive the decision to hierarchical methods or PCP methods, respectively. One would expect provider systems to favor the more stable PCP rules and the insurer to favor the hierarchical rules providing broader accountability to the provider system.

Provider systems may be interested in the stability of their attributed patient population as a fraction of their total patient population rather than as a fraction of the insurer’s enrolled population. We examine this in detail in our eAppendix, finding that results are driven by size of the provider system (IDS, large physician-only system, or small physician-only system). The fraction of patients served by a system that are attributed to that system—and the fraction of revenue generated by those attributed patients&mdash;declines markedly with system size. Smaller systems see more turnover in their attributed population from year to year. Even the largest vertically integrated systems see significant “leakage,” as only 47% to 61% of their attributed patients’ revenue is received by the system.

This study has a key limitation: an external validity threat. National trends in provider concentration, value-based contracting, and healthcare utilization may be different than those in Minnesota. However, we believe Minnesota is on the leading edge in the trends of provider concentration and value-based contracting, so our results have relevance to national trends.

Our results allow us to make a clear recommendation of attribution rules, based on the fraction attributed and stability of the attribution. We recommend lookback rules using all PCP visits for attribution when 2 years of population history are available; we recommend a lookback method when multiple years of data are available. When there is a single year of history, the contractual parties must negotiate to determine whether the greater stability of all-visit PCP rules or the greater fraction attributed under all-visit hierarchical rules is the determining factor.&ensp;

The authors thank Ravi Narayanan, Beth Lindholm, Megan Savage, and Era Kim (Medica Research Institute) for their help in processing the claims data, and Jon B. Christianson (University of Minnesota) and Scott Ode (Medica Research Institute) for their helpful comments regarding this manuscript. We are grateful for the substantive input from reviewers; this input has resulted in meaningful improvements in our presentation.

 Medica Research Institute (LH, CC), Minnetonka, MN.

 Medica Research Institute Intramural Award Program grant.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (CC); acquisition of data (CC); analysis and interpretation of data (LH, CC); drafting of the manuscript (LH, CC); critical revision of the manuscript for important intellectual content (LH, CC); statistical analysis (LH, CC); obtaining funding (CC); and supervision (CC).

 Lucas Higuera, MA or Caroline Carlin, Medica Research Institute, 401 Carlson Pkwy, Mail Route CW105, Minnetonka, MN 55305. E-mail: higue003@umn.edu. 

1. Ridgely MS, de Vries D, Bozic KJ, Hussey PS. Bundled payment fails to gain a foothold in California: the experience of the IHA bundled payment demonstration. 

2014;33(8):1345-1352. doi: 10.1377/hlthaff.2014.0114.

2. Cromwell J, Trisolini MG, Pope GC, Mitchell JB, Greenwald LM, eds. 

Pay for Performance in Health Care: Methods and Approaches.

 Research Triangle Park, NC: RTI Press; 2011. https://www.rti.org/sites/default/files/resources/bk-0002-1103-mitchell.pdf. Accessed May 15, 2017.

3. Lewis VA, McClurg AB, Smith J, Fisher ES, Bynum JPW. Attributing patients to accountable care organizations: performance year approach aligns stakeholders’ interests. 

2013;32(3):587-595. doi: 10.1377/hlthaff.2012.0489.

4. Pantely SE. Whose patient is it? patient attribution in ACOs. Milliman website. http://us.milliman.com/uploadedFiles/insight/healthreform/whose-patient-is-it.pdf. Published January 19, 2011. Accessed May 15, 2017.

5. Chernew ME, Mechanic RE, Landon BE, Safran DG. Private-payer innovation in Massachusetts: the ‘alternative quality contract’. 

2011;30(1):51-61. doi: 10.1377/hlthaff.2010.0980.

6. Bailit MH, Hughes C, Burns M, Freedman D. Shared savings payment arrangements in health care: six case studies. The Commonwealth Fund website. http://www.commonwealthfund.org/publications/fund-reports/2012/aug/shared-savings-payment-arrangements. Published August 20, 2012. Accessed May 15, 2017.

7. Mehrotra A, Adams JL, Thomas JW, McGlynn EA. The effect of different attribution rules on individual physician cost profiles. 

2010;152(10):649-654. doi: 10.7326/0003-4819-152-10-201005180-00005.

8. Medicare Shared Savings Program: shared savings and losses and assignment methodology: specifications. CMS website. https://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/sharedsavingsprogram/Downloads/Shared-Savings-Losses-Assignment-Spec-v2.pdf. Published December 2014. Accessed May 15, 2017.

9. Pham HH, Schrag D, O’Malley AS, Wu B, Bach PB. Care patterns in Medicare and their implications for pay for performance. 

Articles

The authors compare methods of retrospectively attributing patients to provider systems by the fraction of patients attributed and the stability of attribution over time.

Caroline Carlin, PhD

Articles

A Comparison of Retrospective Attribution Rules