Geographic Variation in Physicians’ Responses to a Reimbursement Change

Although much research has revealed U.S. geographic variation in the intensity of health care utilization and the level of Medicare spending,¹ such variation in response to Medicare policy changes has received much less attention. This limitation has become more important in the face of the myriad Medicare-reimbursement changes included in the Patient Protection and Affordable Care Act (ACA). We studied the variation in geographic response to a major reform of Medicare’s reimbursement system for physician-administered drugs (Part B), the vast majority of which are chemotherapy agents.

On January 1, 2005, Medicare instituted an average sales price (ASP) payment system for physician- administered drugs, setting reimbursement at the national average of manufacturers’ sales prices from two earlier quarters plus a 6% margin. The goal was to remedy a system that, as was widely acknowledged, overpaid for many drugs. As compared with Medicare’s previous system of reimbursement based on the average wholesale price (AWP) — a list price that was often unrelated to purchase prices — the new ASP system reduced profit margins substantially for many chemotherapy-drug manufacturers. In one extreme case, that of paclitaxel, a drug commonly used to treat lung cancer, standardized monthly reimbursements decreased by a factor of 10 when the ASP system took effect. Physicians responded to this change by increasing the rate of chemotherapy treatment for patients with lung cancer, with the increase concentrated in treatment given by office-based oncologists. ² Rates of chemotherapy treatment increased by more than 10%, or about 2 percentage points, within 30 days after a diagnosis of lung cancer and by almost 4 percentage points within 180 days.

Building on these results, we examined differences among states in how physicians responded to this change in reimbursement. To capture patients across all 50 states and the District of Columbia, we relied on Medicare claims. Our cohort of 878,923 patients was derived from the group of Medicare beneficiaries who had at least one fee-for-service claim in a physician’s office or outpatient-hospital setting, who had received a diagnosis of lung cancer (International Classification of Diseases, 9th revision [ICD-9] codes 162.0 through 162.9) between 2003 and 2005, and who had no claim with such a diagnosis in the previous 12 months. For these beneficiaries, we obtained all Medicare claims from both institutional (e.g., inpatient and outpatient hospital) and noninstitutional (e.g., physician’s office) settings from 2002 to 2006. We used standard algorithms to infer a diagnosis of lung cancer and exclude patients with incorrectly entered diag- nostic codes or who were undergoing evaluation for lung abnormalities that were subsequently deemed benign.²

We analyzed treatment patterns by month of diagnosis to discern whether the January 2005 switch to ASP-based reimbursement generated sharp changes in treatment overall or by state. We studied patients whose cancer was diagnosed up to 11 months before or 11 months after the reimbursement change. We excluded cases diagnosed in December 2005 because of what appeared to be incomplete claims data.

In estimating the probability of beneficiaries’ undergoing chemotherapy, we adjusted for modest differences in the mix of patient characteristics across states by controlling for beneficiaries’ age and its square, sex, race, coexisting conditions, and the presence of metastases. Metastasis was defined by either the presence of two or more noninstitutional claims, separated by at least 28 and no more than 365 days, with a secondary cancer diagnosis or one institutional claim with such a secondary diagnosis. Although this algorithm is likely to miss many beneficiaries with secondary cancer, it has been shown to have reasonable specificity.³ We controlled directly for beneficiaries’ state of residence to capture fixed differences in the supply of oncologists or in the setting of cancer treatment, and we included an interaction between the state of residence and whether the date of diagnosis was in the period after the reform took effect to assess differences in response to the payment change.

Oncologists’ response to the payment change varied markedly across states (see graph), with some states increasing treatment rates by more than 4 percentage points within 30 days after diagnosis and a few actually reducing treatment rates. A small part of the variation is random, as shown by the upper limit of the 95% confidence interval in the graph, but the great bulk of it is real; we can reject the null hypothesis that the change in chemotherapy treatment was the same across states (P<0.001).

In the graph, we also pay specific attention to the states involved in the Surveillance, Epidemiology, and End Results (SEER) program. Because of its richness of clinical detail, the SEER database has been used in a wide variety of cancer studies. Of undoubted value for studies of cancer epidemiology, it has also been linked to Medicare claims and is used regularly to study responses to Medicare policy.⁴ Our results raise questions about SEER’s value for this latter purpose. More than 50% of the national cohort in SEER areas are in states in the bottom three deciles of the response to the reimbursement change (appearing on the right side of the graph) and less than 10% are in the top decile. Even after regression adjustment for patient characteristics, we can reject the hypothesis that the changes observed nationally are the same as those observed in either claims from SEER states or SEER–Medicare itself (P<0.05). In short, in the case of this particular Medicare-payment change, SEER–Medicare includes a disproportionate number of states with smaller responses to the payment change.

We draw two conclusions from these results. First, to the longstanding question of why physicians practice medicine differently in different geographic areas, we can add the question of why their responses to a change in reimbursement are so geographically clustered. It may not be surprising that a physician with large, unpaid debts for, say, education loans might respond to a fee cut differently from a physician who is near retirement and has paid off any education loans, has paid for his children’s college education, and has no mortgage on his residence — but physicians of both types are found in every state. Even more puzzling is why oncologists in Minnesota responded by increasing chemotherapy rates much more than those in California or why oncologists in New Hampshire and Connecticut responded by substantially increasing chemotherapy rates, while those in Rhode Island responded by increasing them only slightly and those in Massachusetts responded by decreasing them, albeit slightly.

Chemotherapy drugs are sold in a national market. Although prices vary with practice size, it seems unlikely that practice-size differences alone could fully explain the varied responses among the states. Understanding this variation is of increasing policy importance. Although there has been considerable discussion of how the changes that the ACA makes in Medicare reimbursement might affect Medicare spending and Medicare beneficiaries on average, there has been little to no explicit recognition that the effects may vary geographically.⁵

Second, despite SEER’s wealth of clinical detail, its lack of national representativeness is a drawback to using SEER–Medicare data to evaluate national changes to the financing and or- ganization of cancer care. Despite SEER–Medicare’s many virtues, more attention should be paid to the limited geographic scope of these data.

Standardized Change in 30-Day Chemotherapy Rates by State.

The height of each bar represents the standardized change in the probability of administering chemotherapy after the ASP payment system took effect, relative to the change in South Carolina. T bars indicate the upper limit of the 95% confidence interval for this change. South Carolina was chosen as the reference category because its (unadjusted) change is roughly equal to the national change; its adjusted change is 0.017. We adjusted for patient age and its square, sex, and race; the Deyo–Charlson comorbidity score; metastases; and calendar month when treatment was initiated. Standard errors are adjusted to allow for an arbitrary covariance structure at the state level. The F value for the null hypothesis that the state changes are jointly equal to zero is 38279, meaning that we can reject the null hypothesis at any standard level of significance.