James Baumgardner, PhD; Ahva Shahabi, PhD; Christopher Zacker, RPh, PhD; and Darius Lakdawalla, PhD
Both public and private payers have targeted cancer care as a prime source of healthcare savings. The testing of new payment models, such as the Oncology Care Model (OCM) by CMS and the Episode Payment Program demonstration by UnitedHealthcare, present 2 cases in point.1-3
Programs like these attempt to generate savings by changing provider incentives, without setting rules regarding how such savings should be achieved. The goal is to give providers the discretion to deliver high-quality, high-value care individualized for each patient. These incentives will have even greater impact if they are coupled with information that can help providers seek out and eliminate low-value care. To that end, the goal of our study was to identify the categories of cancer care that offered the greatest potential opportunities for savings within a treatment episode.
We explored this issue using tools drawn from the established literature on geographic variations in healthcare.4
When the cost of treating similar patients varies widely across geographic regions, efficiencies can be achieved by having high-spending regions emulate low-spending ones.5
The crux of the research problem is to define the concept of patient “similarity” and measure variation across regions that results from practice styles alone and not from variation in patient health.
Our study addresses that issue by considering patient and episode characteristics within an analysis that identifies the subcategories of spending (eg, chemotherapy, acute hospital inpatient care, imaging) most responsible for the interregional variation in total spending. Our particular focus was on spending per cancer care episode using the OCM’s definition of “episode,” because under the OCM, practices have a financial incentive to reduce total Medicare spending on their patients within these OCM-defined episodes.1
We examine interregional variation in spending per OCM-defined episode for 5 cancer types, representing a mix of solid and hematologic cancers that differ in prevalence and level of treatment innovation.
To reduce spending in OCM episodes, subcategories of spending that contribute most to interregional variation in standardized spending per episode may be the lowest-hanging fruit. An important caution, however, is that assessing the impact of differences in spending on patient outcomes is beyond the scope of the current study. Our goal is to flag for providers and health systems those categories of spending that contribute the most to differences in practice styles across regions. These categories ought to be viewed as the highest priorities for careful decision making about how to reach the appropriate trade-off between spending and outcomes.
We used data from the Surveillance, Epidemiology, and End Results Medicare (SEER-Medicare) database for 2006-2013. SEER-Medicare links data from the SEER program of the National Cancer Institute, which contains information from 18 cancer registries, with Medicare claims files. SEER-Medicare covers approximately 28% of the US population distributed geographically throughout the country. Our cost analysis included 2007 through 2013. The year 2006 was included for creating the Charlson Comorbidity Index (CCI) score of comorbidities for each episode, based on the patient’s prior-year claims. We excluded patients who spent time in Medicare Advantage (MA) because of incomplete treatment information and because MA patients were excluded from the OCM. The study was determined to be exempt from institutional review board oversight.
The patient cohort includes those with a new diagnosis of 1 of the 5 selected tumors between January 1, 2007, and December 31, 2011. Patients were required to be covered by Medicare parts A, B, and D from 12 months prior to the diagnosis through the end of data or date of death and to be treated with at least 1 chemotherapy drug in an outpatient setting that would trigger an episode within the OCM, including Part D chemotherapy claims. This restriction was chosen to allow us to define an episode in the same way as the OCM.
We identified a cohort of patients with 1 of the following 5 tumor types as the primary cancer: non–small cell lung cancer (NSCLC), advanced (stage III or IV) breast cancer (BC), renal cell carcinoma (RCC), multiple myeloma (MM), and chronic myeloid leukemia (CML). These tumor types were chosen to provide variety in prevalence in the population 65 years and older, ensure a mix of solid and hematological tumors, and offer variation in treatment patterns, innovation, and resource mix.
Episodes were defined as they are in the OCM. Either infusion or injection of outpatient chemotherapy or the filling of a prescription for Part D–covered chemotherapy triggered an episode. The OCM defines chemotherapy in a broad sense and includes antineoplastic drug therapies generally; for example, monoclonal antibody therapies.6
Likewise, we use the term chemotherapy
in the same broad sense as the OCM. Following the OCM definition, the episode ended either 6 months later or at death. As delineated
in the OCM, subsequent episodes for the same patient were allowed and began at first use of qualifying chemotherapy following the end of the previous episode.
Standardization of Claim Amounts
We converted every Medicare claim to a standardized amount to eliminate the effects of payment rate differentials across regions. We further adjusted claims for changes in Medicare prices over time, including the effects of the federal budget sequester beginning in April 2013. Our final measures of standardized spending are in 2013 (presequester) dollars. Because of these standardization steps, when we detect higher spending, it is due to greater utilization or greater use of higher-cost goods and services. For details on our sample and standardization methods, see the eMethods section in the eAppendix
(available at ajmc.com
Subcategories of Spending
We assigned every claim to 1 of 13 service subcategories following the definitions of Brooks et al,7
who designed the categories to be mutually exclusive and cancer-relevant. We modified the categories to incorporate Part D claims by adding Part D chemotherapy claims to the chemotherapy subcategory and by creating an additional subcategory for nonchemotherapy Part D (drug) claims. For every episode, we calculated both total spending and spending by subcategory.
We used the hospital referral regions (HRRs) defined in the Dartmouth Atlas of Health Care for our geographic regions in this study.8
The region for an episode was the patient’s HRR of residence. The HRRs were developed by the Dartmouth Atlas of Health Care to define regional healthcare markets based on referral patterns for tertiary care. They have been widely used to define geographic markets for studies of variation in use of medical services. Within the Medicare program, beneficiaries receive more than 80% of their care in their residence HRR.9
We created a measure of standardized spending per episode for each HRR for each cancer. Due to the potential for inaccurate inferences owing to sampling variation in HRRs with few episodes, we required at least 20 episodes for an HRR to be included in the analysis for a particular cancer. To adjust for patient and episode characteristics, we estimated a generalized linear model with a gamma distribution and log-link, with spending per episode as the dependent variable. The regression included the following covariates: patient age and age squared; CCI score and CCI score squared; number of prior episodes; indicators for male (except for the BC sample, which was restricted to female) patients, Medicaid eligibility, Medicare entitlement through disability, stage at diagnosis (except for the hematologic cancers, for which SEER staging does not apply), and death during the episode; and indicator variables for each HRR.
The estimates obtained from the regression were used to create a measure of standardized spending for each HRR by using the coefficient for the HRR’s indicator variable along with a standardized set of values for the covariates and their corresponding coefficient estimates. The standardization was defined at mean values of the covariates over the entire sample for the respective cancer types.
For each subcategory of spending, we used the same procedure to create a measure of standardized spending per episode for that subcategory for each HRR. Following the method of Newhouse and Garber, and using NSCLC as an example, we created a series of figures to graphically illustrate the variation in total standardized spending per episode across HRRs and to show the contribution of the respective subcategories to that interregional variation.10
Next, for each cancer, we developed a simple statistic that measured the contribution of each subcategory to the interregional variation in total spending per episode. The statistic, Sc
, is defined as
is the explained sum of squares from a multivariable least squares regression of HRR standardized total spending per episode on standardized spending of each of the 15 subcategories of spending; e2c
is the explained sum of squares from a multivariable least squares regression of HRR standardized total spending per episode on spending of each subcategory except for the category (c
) in question. (Within each HRR, we summed standardized subcategory spending per episode over all subcategories to arrive at standardized total spending per episode.) To the extent that the interregional variation in subcategory c
contributes to the interregional variation in total spending, e2c
will be small relative to e1
and the statistic
will be relatively large. It stands to reason that if dropping a single subcategory from the regression of total spending on the subcategories results in a large loss of explanatory power, then that subcategory must be important in explaining the variation in total spending.
presents summary statistics describing our data set and patient population. The numbers of episodes and patients, and the number of included HRRs that met our criterion of at least 20 episodes, reflect the differences in incidence of the 5 cancer types. Patients with BC had the lowest comorbidity index, on average, whereas patients with RCC had the highest comorbidity measure.
shows summary statistics for spending per episode and the distribution of spending across subcategories of services. The least expensive episodes, on average, occurred for BC ($20,887) and the most expensive for MM ($52,489). Chemotherapy was the largest category of spending across all cancer types, ranging from 25.9% of total spending for BC to 67.8% for CML.
illustrates the contribution of selected subcategories to the variation in total spending per episode across regions for NSCLC, from lowest to highest spending region in terms of standardized spending per OCM-defined episode (see eAppendix Figure
for additional detail). For economy of presentation, we developed the statistics shown in Table 3
to summarize the contribution of each subcategory to the inter-regional variation in total spending for each cancer type. Because we have standardized for differences in Medicare prices (payment rates) and for observable patient and episode characteristics, differences in our reported measure of standardized spending represent differences in utilization for a patient with the same, “average,” characteristics.
The first panel of the Figure depicts the differences across HRRs in total standardized spending per episode for NSCLC. The interquartile range is $7281. The ratio of spending at the 80th to the 20th percentile is 1.3.
Chemotherapy was the largest contributor to the interregional variation in total spending, as shown by the general size and progression of bars in the Figure, which correspond to the relatively high value of the contribution statistic (0.2896) in Table 3. Acute hospital care, which includes both facility and inpatient physician services components, was also a large contributor. Notably, imaging contributed little to the interregional differences, as illustrated by the roughly flat overall trend of the bars in the Figure, which translates to the low value for imaging (0.0178) in Table 3.
Table 3 displays the contribution to interregional variation for the 5 respective cancers and for all spending subcategories. The memorandum lines provide 2 measures of the size of the variation in total standardized spending across regions.
Chemotherapy spending was the largest contributor to interregional variation in total episode spending across regions for 4 of the 5 cancers, and especially for RCC and MM. Spending on acute hospital care was the next largest contributor for NSCLC and MM but was less important for RCC and CML. The category of “unspecified outpatient hospital and other Part B” spending was the second largest contributor to variation for RCC and CML. That category consists primarily of facility charges for outpatient hospital use.
Nonchemotherapy Part D (drug) spending was a substantial contributor to interregional spending variation for advanced BC. Radiation therapy was a larger contributor to interregional cost variation than inpatient care for both RCC and CML, but it played an insignificant role for other cancer types studied.
Imaging, nonchemotherapy Part B drugs, physician evaluation and management services, and diagnostics were among the negligible contributors (defined here as <0.02) to the interregional variation in total spending.
Our summary statistics in Table 2 are consistent with the findings of Rocque et al, who found in a sample of Southeastern US cancer centers that chemotherapy is the highest spending category, followed by inpatient hospital care.11
Consistent with a Dartmouth interpretation, the subcategories with higher values in Table 3 may be the most fruitful in yielding savings, given apparent differences in practice styles or professional opinion about appropriate total spending. A higher value in Table 3 indicates a greater contribution of that subcategory to the variation across regions in total standardized spending within OCM-defined episodes.
The OCM definition of an episode appears to play an important role in explaining the high contribution of chemotherapy spending listed in Table 3. Because the OCM defines an episode as the 6-month period commencing with chemotherapy, OCM-defined episodes are necessarily “chemo-centric.” Our results might appear to contradict those of Brooks et al,7
who concluded that inpatient care was a more substantial driver of interregional variation than chemotherapy, but that difference likely occurs because the OCM episodes of care begin at a conceptually different index date. Although both episodes last 6 months, those in the Brooks study began at cancer diagnosis, whereas in the OCM, episodes commence with chemotherapy treatment. In a different analysis, Wang et al concluded that radiotherapy was the largest driver of regional cost differences for prostate cancer, but this study also used an episode definition (2 months prior to 12 months after diagnosis) different from that of the OCM.12
Although less substantial than the contribution of chemotherapy spending, the importance of acute inpatient hospital care is not surprising, given the earlier finding of Brooks et al.7
Consistent with their conclusions, our results suggest that use of inpatient care has a substantial effect on real spending differences across areas. Our results are consistent with the longstanding view that avoiding admissions can reduce cost, but as demonstrated by our results for RCC and CML, acute inpatient hospital care is not a uniformly strong contributor for all cancers. Importantly, the study by Brooks et al does not contradict our findings in that it does not include RCC or CML.
Outpatient procedures, especially in a hospital outpatient setting, are probably more likely sources for potential savings than suggested by the “outpatient procedures” line in Table 3. This is because the category “unclassified hospital outpatient and other Part B services” may capture some of the hospital outpatient facility charges that are coupled with the utilization of outpatient procedures. The latter category contributes more to total spending variation for all 5 cancer types and comprises mainly unspecified hospital outpatient facility charges, which may be incurred for performing outpatient procedures.
The potential spending reductions on nonchemotherapy Part D drugs for advanced BC and on radiation therapy for RCC and CML may be less promising than they appear in Table 3. Unlike other contributors to variation, these particular results hinged on single high-spending HRRs. Removal of the single highest spending HRR for other Part D drugs in BC changed the contribution statistic from 0.209 to 0.013, consistent with the negligible contribution seen for other cancers. Removal of the single highest spending HRR for radiation therapy reduced the contribution from 0.040 to 0.015 for RCC and from 0.067 to 0.019 for CML, suggesting that interregional differences of opinion on appropriate use of radiation therapy in RCC and CML play a limited role in total spending variation. Perhaps a review of utilization in these subcategories is appropriate, but only single HRRs in the sample are the reason for review.
Just as noteworthy as items that appear to be fruitful targets for cost cutting are subcategories that are unimportant contributors to interregional total spending variation. Although imaging and nonchemotherapy Part B drugs are suggested as areas for savings in the literature,13
those subcategories were negligible contributors to interregional variation in spending within OCM-defined episodes.
Although a subcategory can be a nontrivial fraction of overall spending (Table 2), it may be a negligible contributor to interregional variation in total spending. For example, for NSCLC and advanced BC, Table 3 indicates that radiation therapy is not a substantial contributor to interregional differences in total episode spending despite accounting for 5% or more of total spending (Table 2). This suggests that radiation therapy is not a fruitful target for OCM savings for those cancer types because lower spending on radiation therapy is not the reason that low total spending regions have low spending.
Importantly, this analysis does not include patient outcomes data. Thus, we cannot determine if differences in utilization of particular services across regions result in differences in important outcomes.
Additional questions for future research exist but remain beyond the scope of this study. When more data are available to compare OCM-participating and nonparticipating practices, researchers can further evaluate the OCM’s effectiveness, but they will need to take into consideration the fact that participation is voluntary. In addition, if sufficient data were available from MA plans, it would be interesting to compare MA plans with both OCM-participating and nonparticipating fee-for-service practices. Another potential evaluation is a more broadly targeted alternative payment model than the OCM—for example, one that considers episodes not necessarily confined to chemotherapeutic treatments. Finally, as more treatment innovations occur, our findings may need to be revisited.
We took a Dartmouth-style approach to identify opportunities to reduce costs within OCM-defined episodes by quantifying how different subcategories of care contribute to observed differences in standardized total episode spending across regions. Given these results, while also recognizing that higher spending does not necessarily imply less efficiency if outcomes are better, OCM-participating practices may wish to evaluate their treatment styles within the high-priority categories identified here.
This study used the linked SEER-Medicare database. The interpretation and reporting of these data are the sole responsibility of the authors. The authors acknowledge the efforts of the National Cancer Institute; the Office of Research, Development and Information, CMS; Information Management Services, Inc; and the SEER Program tumor registries in the creation of the SEER-Medicare database.
The authors thank Sarah Green and Seanna Vine for excellent programming support, and Sarah Beers for medical coding. In addition, they thank Amitabh Chandra, Mark Linthicum, Michelle Brauer, Caroline Huber, Joanna MacEwan, and Lara Yoon for their advice and assistance.