Policy makers will need a way to update bundled payments that reflects highly skewed spending growth of various care episodes

We used data from the 2003 and 2007 MarketScan Commercial Claims and Encounters Database from Truven Health Analytics (formerly Thomson Reuters and, before that, MedStat). The MarketScan data captured the claims of about 17.5 million commercially insured people in 2003 and, because of an increasing number of employers submitting claims, more than 31 million in 2007. About half of the insurance claims for this period were submitted by very large employers; the other half were submitted by insurance plans. Insurance plan data included claims from both large and small firms. For each of these years we selected people under age sixty-five who had drug coverage, were in a commercial plan with valid plan type, and were continuously enrolled for the full year (this requirement was to ensure that entire episodes were captured). We excluded people enrolled in capitated plans because their utilization data are frequently incomplete, which complicates episode formation and spending estimation. This process left 5.8 million and 11.1 million people in our samples for 2003 and 2007, respectively. We used three types of files for each year. First, the enrollment files contained information on patient demographics, enrollment periods, and type of health insurance coverage. Second, the event files contained inpatient, outpatient, and prescription drug claims, and they included dates and types of services, diagnosis ( International Classification of Diseases , Ninth Revision, Clinical Modification, or ICD-9-CM) and procedure ( Current Procedural Terminology , Fourth Edition, or CPT-4) codes, types of providers, and costs of services. Third, the episode files linked each event with a clearly defined episode of care and provided information on the type of episode, discussed below. These data included information on both reimbursements and charges. However, all analyses used the actual amount received by the provider, which included payments from both the insurer and the patient. We classified spending into episodes using Truven Health Analytics’ Medical Episode Grouper, which is widely available. 6 This commercial grouper employs a proprietary algorithm, based on clinical knowledge, which reviews claims and either assigns them to one of 569 possible episodes of care for different conditions or classifies them as “ungroupable.” Ungroupable claims were typically claims for prescription drugs or ancillary services such as laboratory or radiologic tests for patients who did not have a relevant episode for the service. We omitted these claims from our analysis. We further classified the episodes into five different types using the labels that the grouper assigned to each episode: acute, three types of chronic episodes (chronic maintenance, exacerbation of chronic condition, and chronic non-stratified), and well care. We combined the three chronic episode types into a single chronic category for reporting summary statistics and spending growth by disease, but we describe them separately when reporting spending growth by episode acuity. The lone well care episode, “encounter for preventive health services,” was considered separately. For each episode, we computed average spending per episode in 2003 and 2007 and the total count of episodes per capita, adjusted for age and sex. 7 We decomposed overall spending growth into growth in the number of episodes, changes in the mix of episodes, and growth in spending per episode, both at the disease level and for overall spending. We then examined the distribution of spending growth by episode acuity, calculating the contribution to overall spending growth for each type of episode. Finally, we examined the distribution of spending growth across disease categories, and we calculated the contribution of each disease to overall spending growth and, importantly, the heterogeneity in spending growth per episode. The formulas used for our decomposition are provided in the online Appendix . 8 To address the possibility that findings were unduly affected by outliers, we conducted a sensitivity analysis in which cost was top coded at the ninety-ninth percentile within each episode type. Analyses were conducted using two statistical software packages: SAS version 9.2 and Stata version 11. Our study had some limitations. First, our data set, although large, is a convenience sample from select employers and insurers—and thus is not representative of the private market. Our sample was restricted to people under age sixty-five who were continuously enrolled in noncapitated plans with drug coverage. These exclusions were necessary to ensure complete claims data on all members of the sample. We excluded 20 percent of spending from our analyses because we could not group those amounts to a particular episode. Although the distribution of spending growth across episodes was determined using one commercial grouper, there are several other episode groupers in use or development. However, we would expect spending growth across episodes to be skewed regardless of which grouper is used. Another limitation related to our inability to assess the appropriateness of episode spending growth or whether the appropriateness of spending growth per episode varied across episodes. Also, our use of fee-for-service spending growth as a basis for projecting episode trends assumes no behavioral changes on the part of providers in response to bundled payment incentives. Our results nonetheless illustrate the skewness in spending growth observed in the past; it remains to be seen how new payment models would (or should) affect that. Despite these limitations, this study is broadly relevant to payers and policy makers interested in operationalizing bundled payment.

Descriptive statistics for our sample are shown in Exhibit 1 . The age and sex distribution of our samples remained relatively constant over time. Overall, the grouper allocated 78 percent of spending into 558 types of episodes: 361 acute, 196 chronic, 9 and one well care. The grouper was unable to allocate 23.7 percent and 21.1 percent of spending to episodes in 2003 and 2007, respectively. The final sample included total annual spending of $18.9 billion and $42.1 billion in 2003 and 2007, respectively, on inpatient services, outpatient services, and prescription drugs. Between 2003 and 2007, spending per capita increased by 15 percent, from approximately $3,253 to $3,739, holding age and sex constant. This increase was lower than the national private spending growth rate of 21 percent, which includes growth due to the aging of the population. 10 Exhibit 2 displays the decomposition of per capita spending growth into changes in spending per episode, episode mix, a shift from lower-cost to higher-cost episodes or the reverse, and the number of episodes per beneficiary. The increase in spending per episode was the key driver of spending growth, accounting for 73 percent of the total. The number of episodes per person increased as well, from 10.39 to 10.94, accounting for a 36 percent increase in spending growth. The episode mix was slightly tilted to less expensive episodes over time, although not greatly. Spending growth was highly skewed, with 10 percent of episodes accounting for 82.5 percent of total spending growth. Spending on the 10 percent of episodes with the smallest contribution to cost growth actually decreased12.9 percent between 2003 and 2007, contributing −16 percent to overall spending growth. Among episodes with at least 1,000 observations per year, spending per episode rose 323 percent in the fastest-growth episode and fell 75.5 percent in the slowest. Exhibit 3 displays the contribution to spending growth of the major episode types. Chronic episodes accounted for 43.5 percent of baseline spending and 40.6 percent of spending growth. Acute episodes accounted for 53.1 percent of baseline spending and 46.2 percent of spending growth. The remaining 13.2 percent of spending growth was attributable to well care. We found much heterogeneity in spending growth per episode across episodes and within each episode type as well. Among acute episodes with at least 1,000 observations per year, spending per episode rose 323 percent in the fastest-growth episode, disorders of bilirubin excretion, and fell 75.5 percent in the slowest-growth episode, bacterial meningitis. Among chronic episodes with at least 1,000 observations per year, spending per episode rose 190 percent in the fastest-growth episode, other maternal conditions affecting newborns, and fell 44.1 percent in the slowest-growth episode, hepatitis G. Exhibit 4 displays our decomposition of cost increases for the ten episodes with the largest and the ten with the smallest contributions to per capita spending growth. The lone well care episode was the single largest contributor to per capita spending growth, accounting for more than twice as much spending growth—13.2 percent—as the next-fastest-growth episode. Other top contributors included breast cancer, contributing 5.3 percent; osteoarthritis of the lower back, 4.5 percent; and type 2 diabetes, 4.4 percent. In contrast, reductions in spending on angina pectoris and acute myocardial infarction actually slowed per capita spending growth by 5.9 percent and 2.3 percent, respectively. There was also considerable heterogeneity in spending growth per episode across the ten episodes with the largest and the ten episodes with the smallest contributions to per capita spending growth. Among the top ten episodes, spending per episode rose 48.4 percent in the two fastest-growth episodes, colon cancer and multiple sclerosis, and fell 23.9 percent in the slowest-growth episode, renal failure. Among the ten episodes with the smallest contribution to overall spending growth, spending per episode rose 19.8 percent in the fastest-growth episode, benign breast neoplasm, and fell 39.1 percent in the slowest-growth episode, sleep disorders. The robustness of our results was maintained when we dampened the effect of outliers by top coding spending at the ninety-ninth percentile within each episode type. Sensitivity analysis results are presented in more detail in the Appendix . 8

Episode-based bundled payment systems have gained traction in the public and private sectors. Yet many important questions must be addressed before episode-based payment can become a sustainable solution to rapid spending growth. 2 , 11 – 13 Whether or not bundled payment can control spending growth depends in part on the extent to which spending growth is driven by increases in cost per episode as opposed to increases in episode volume. Spending growth that is attributable to increased number of episodes will not be materially curtailed by episode-based payment; spending growth due to costs per episode is more likely to be affected. On this first point, our news is generally good. We found that spending per episode—including price increases, not episode volume—was the main driver of spending growth, accounting for 73.1 percent of overall spending growth. 14 The success of episode-based payment models in curbing the rate of spending growth, as opposed to just providing a one-time reduction in level of spending, will rely greatly on the ability to update payment rates over time. Initial episode-based spending models may rely on the fee-for-service system as a reference when updating rates. However, an alternative updating approach will be needed as the fee-for-service system shrinks and becomes less representative of the prevailing costs of care. Importantly, uniform updates based on historical spending growth will be appropriate only if spending growth per episode is relatively evenly distributed across episodes. Our findings on this second point are more problematic. Our analysis found spending growth to be highly skewed across episodes: 10 percent of episodes accounted for more than 80 percent of spending growth, and within-episode spending growth ranged from −75 percent to 323 percent across episodes. This skewness has important implications for bundled payment mechanisms. Most notably, it implies that reimbursement rates cannot be updated uniformly across episodes. Some episodes, such as encounter for preventive health services and malignant breast neoplasm for females, experienced rapid growth. Others, such as angina pectoris and acute myocardial infarction, did not. In this context, uniformly updating payment rates would result in overpayment for angina and acute myocardial infarction care and underpayment for breast cancer care and preventive health services. Our finding that the encounter for preventive health services topped the spending growth list warrants further investigation into specific clinical drivers of spending in primary care. The marked growth in well-care spending concurrent with an impressive decline in spending on acute exacerbations of chronic conditions raises the question of whether the two are related—in other words, whether there might be a causal relationship between preventive care delivered in a well care episode and a decline in spending on acute exacerbations of chronic illness. In designing a system to differentially update reimbursement rates, it is critical to understand why spending per episode rises at different rates for different episodes. It might be the case, for example, that during our short time window, spending growth reflected only differential increases in the price per service (for example, inpatient or outpatient services or prescription drugs) within episodes and that these differential increases would reverse themselves over time. However, most research on spending growth indicates that a large share of health care spending increases must be due to the diffusion of medical technology. If so, updates for different episodes should be based on some sense of where technology is driving changes in practice patterns.

Episode-based payment models constitute a promising advance in the evolution of payment reform; implementing and realizing cost containment from these systems, however, will be challenging. The marked skewness in spending growth demonstrated in this study indicates that reimbursement rates should not be updated uniformly across episodes of care. In the short term, as payers experiment with a limited number of episodes, individual updates for each episode may be appropriate. As the breadth of covered episodes expands, though, more systematic approaches to updating payment rates will likely be required. Furthermore, whereas episode reimbursement rate updates can help control spending growth, they do nothing to control the number of episodes or those episodes’ clinical appropriateness. Over time, moving from updates based on historical cost growth, which may reinforce overuse, to updates based on the appropriate use of evidence-based best practices would be ideal. In the future, it will be important to decide whether increases in episode spending are justified by improved health or whether they reflect unnecessary use. Coupling payment with quality measures may be helpful. However, given time lags, it will also be important to have a system that assesses the need to increase reimbursements to allow the diffusion of valuable new medical technologies and other innovations. Developing such a system will require the integration of greater clinical knowledge into payment update systems. Yet as episode-based payments become a larger part of our future payment system, we must begin the discussion of how such payments can be updated to allow the appropriate diffusion of medical technology.