Browse the corpus

Provider billing margin and cancer treatment selection: population based cohort study. OBJECTIVE: To estimate the association between the billing (profit) margin and clinical benefit of cancer treatments and use by oncologists. DESIGN: Retrospective population based cohort study SETTING: Population based sample using fee-for-service Medicare claims data. PARTICIPANTS: Medicare beneficiaries with an incident cancer diagnosis from 2015 to 2020; included all patients with cancer treatment indications for which the available treatment options varied in clinical benefit and billing margin. MAIN OUTCOMES AND MEASURES: The primary outcome was the cancer treatment each patient received among treatment options recommended by the National Comprehensive Cancer Network (NCCN). The characteristics of interest were provider billing margin (defined at the patient treatment level, using Medicare reimbursement rates corresponding to the patient's diagnosis date) and a proxy measure of clinical benefit (rank order of treatments based on Evidence Blocks scores developed by the NCCN, corresponding to the patient's diagnosis date). The association between treatment received, billing margin, and clinical benefit was modeled using a conditional logit model with inverse probability-of-treatment weights applied at the patient treatment level to control for patient and provider characteristics. RESULTS: Twelve cancer indications were examined comprising 19 397 individual patients. Provider billing margin ranged from $0 to $12 692 (£9440; €10 800) for each course of treatment. No association was found between a $100 increase in provider billing margin and the likelihood of treatment use (odds ratio 0.97, 95% confidence interval 0.91 to 1.03). Higher clinical benefit was associated with greater treatment use (1.62, 1.15 to 2.29). CONCLUSIONS: In this observational study of Medicare beneficiaries, selection of cancer treatments was associated with the treatment's clinical benefit but not billing margin. Therefore, it is unlikely that policy and price changes that affect the billing margin of cancer treatments would shift patterns of use.

Whether clinicians make patient care decisions based on financial considerations is controversial. Fee-for-service reimbursement—the dominant payment model among industrialized nations—gives healthcare providers a financial incentive to increase service use.1 Rationally, physicians might be expected to factor this financial incentive into their treatment decisions, balanced with other considerations such as clinical need and expediency. However, the professional ethos of healthcare providers—that patient interests should remain paramount—implies that revenue considerations should have no role at the bedside. Clinicians have been strongly resistant to the idea that treatment revenue could influence their clinical decision making.2 3

eed and expediency. However, the professional ethos of healthcare providers—that patient interests should remain paramount—implies that revenue considerations should have no role at the bedside. Clinicians have been strongly resistant to the idea that treatment revenue could influence their clinical decision making.2 3 The extent to which clinicians respond to billing incentives has direct relevance to policy. Within fee-for-service healthcare systems, physician fees are typically set by the central government through negotiations with providers. Policy makers must consider and balance the competing priorities of access, cost control, and care quality. To the extent that physicians are fee sensitive for a given service, relative overcompensation might lead to overuse, while undercompensation for clinically necessary services might negatively impact patient access and quality. Therefore, poorly constructed payment models could adversely affect patient care. On the healthcare system level, misaligned payment policy may adversely affect healthcare resource allocation and infrastructure. The healthcare system in the United States provides a cautionary example of this concern; providers have invested heavily in infrastructure for delivering highly compensated services such as cardiology and oncology,4 5 while undercompensated services such as mental health and obstetrics remain under-resourced.5 6 7

re. The healthcare system in the United States provides a cautionary example of this concern; providers have invested heavily in infrastructure for delivering highly compensated services such as cardiology and oncology,4 5 while undercompensated services such as mental health and obstetrics remain under-resourced.5 6 7 The association between reimbursement and clinical decision making is especially relevant to oncology. The high cost of cancer drugs may strain national healthcare budgets, presenting a strong rationale to policy makers to avoid incentivizing overuse. However, if cancer care providers are fee sensitive, then lower payment may result in lower uptake of beneficial new treatments. Oncology professional organizations have argued that inadequate compensation for delivering cancer drugs will harm access, leaving patients unable to receive some treatments.8 9

zing overuse. However, if cancer care providers are fee sensitive, then lower payment may result in lower uptake of beneficial new treatments. Oncology professional organizations have argued that inadequate compensation for delivering cancer drugs will harm access, leaving patients unable to receive some treatments.8 9 The US healthcare system provides a valuable example of evaluating provider fee responsiveness in delivering cancer care. In the US, providers receive fees for clinician administered drugs—such as intravenous or injectable chemotherapeutics—but not oral or prescription drugs. The value of this fee is proportional to drug price: more expensive drugs yield higher fees. Specifically, providers receive a fee of 6% of the average purchase price (average sales price) when treating patients with Medicare (the public payer covering all US adults beginning at age 65), resulting in a 6% profit margin on average. Among patients covered by commercial insurance companies, the margin varies among companies and providers, but is typically several fold higher than Medicare rates.10 11 These fees present oncology providers with the financial incentive to prescribe higher priced drugs. If providers respond to this incentive, it has the potential to hasten implementation of new, effective treatments (which are typically the most expensive), but also to further increase spending on cancer treatment, already $200 billion (£148 billion; €170 billion) each year in the US.12

prescribe higher priced drugs. If providers respond to this incentive, it has the potential to hasten implementation of new, effective treatments (which are typically the most expensive), but also to further increase spending on cancer treatment, already $200 billion (£148 billion; €170 billion) each year in the US.12 Empirical data on whether oncologists respond to these fees are lacking. Several studies have found evidence of billing driven decision making in oncology services, including biopsy, surgery, and radiation.13 14 15 16 17 However, the system of price proportional drug fees currently used in the US is relatively recent (enacted in 2005), and studies evaluating the impact of this fee-for-service payment model on treatment decisions are lacking.18 The goal of this study was to assess the relative contributions of billing margin and clinical benefit in cancer treatment selection, and whether oncologists tend to prefer higher margin (more profitable) treatments over lower margin (less profitable) treatments.

We used a 100% sample of fee-for-service Medicare data from 2014 to 2020. Following previous work,19 we identified new cancer diagnoses by the occurrence of a cancer related diagnosis code in proximity to claims for cancer treatment, and no previous cancer diagnosis codes or treatment claims within a minimum one year period of continuous Medicare enrollment. Therefore, individual patients were observed only once. We identified cancer site (tissue of origin) using a claims based algorithm previously validated against SEER (surveillance, epidemiology, and end results) registry data.19 We included patients with an index diagnosis date from when the National Comprehensive Cancer Network (NCCN) first published NCCN Evidence Blocks scores for their cancer type (as early as 2015 for some cancers; see below) to the end of available claims data in 2020. Therefore, claims data for 2014 were used solely for baseline data for patients with 2015 index dates. We required age ≥66 years at diagnosis and continuous Medicare enrollment from one year before the index date through the outcome period. From this overall cohort, we then applied additional criteria as appropriate to identify smaller groups corresponding to each clinical scenario of interest. We included patients treated in all provider settings. During the study period, approximately 45% of Medicare insured patients with cancer received treatment in private physician offices, 45% in hospital affiliated outpatient practices, and 10% in outpatient practices affiliated with academic cancer centers.20

We included patients with an index diagnosis date from when the National Comprehensive Cancer Network (NCCN) first published NCCN Evidence Blocks scores for their cancer type (as early as 2015 for some cancers; see below) to the end of available claims data in 2020. Therefore, claims data for 2014 were used solely for baseline data for patients with 2015 index dates. We required age ≥66 years at diagnosis and continuous Medicare enrollment from one year before the index date through the outcome period. From this overall cohort, we then applied additional criteria as appropriate to identify smaller groups corresponding to each clinical scenario of interest. We included patients treated in all provider settings. During the study period, approximately 45% of Medicare insured patients with cancer received treatment in private physician offices, 45% in hospital affiliated outpatient practices, and 10% in outpatient practices affiliated with academic cancer centers.20 We obtained treatment options recommended by the NCCN, who produce the most widely used clinical practice guidelines in oncology. For each patient, we determined the systemic treatment regimen they started within 180 days of follow-up after diagnosis among the contemporary NCCN recommended regimens, following previous methods.20

d treatment options recommended by the NCCN, who produce the most widely used clinical practice guidelines in oncology. For each patient, we determined the systemic treatment regimen they started within 180 days of follow-up after diagnosis among the contemporary NCCN recommended regimens, following previous methods.20 We used the NCCN Evidence Blocks scores, an expert opinion based metric with high ease of use and validity,21 22 as our proxy measure for the clinical benefit of each treatment option. The NCCN Evidence Blocks grade each cancer treatment on five criteria: efficacy (the extent to which treatment improves survival and symptoms), safety (treatments with fewer side effects receiving higher scores), quality of evidence (the number and rigor of the supporting clinical trials), consistency of evidence (the degree to which clinical trials agree on the degree of benefit), and affordability (with less expensive treatments receiving higher scores). Each criterion is scored from one to five, with higher scores being more favorable. We abstracted Evidence Blocks scores for all cancer treatments recommended by the NCCN, as previously described.23 24 Because these guidelines are continuously updated to reflect new evidence, we repeated this abstraction every six months (1 January and 1 July) across the study period.

gher scores being more favorable. We abstracted Evidence Blocks scores for all cancer treatments recommended by the NCCN, as previously described.23 24 Because these guidelines are continuously updated to reflect new evidence, we repeated this abstraction every six months (1 January and 1 July) across the study period. For each treatment, we estimated the provider billing margin under Medicare payment rates, following previous work (for full details of methods, see supplementary appendix).23 24 We assumed billing margins of 6% of the average sales price for provider administered drugs, and no margin for prescription drugs. We included guideline concordant supportive care drugs (growth factors and anti-emetics). Margins for each treatment were re-estimated every six months to reflect manufacturer price changes and price changes driven by new generics or biosimilars.25 We used a multicohort design, including patients from each applicable cancer clinical scenario. From the many NCCN defined clinical scenarios (eg, adjuvant treatment, stage 3 melanoma), we included all those that reflect initial systemic treatment (including adjuvant or neoadjuvant) because subsequent line treatments are challenging to identify accurately in claims data, and all those that substantially varied in clinical benefit and billing margin among the recommended treatment options.

3 melanoma), we included all those that reflect initial systemic treatment (including adjuvant or neoadjuvant) because subsequent line treatments are challenging to identify accurately in claims data, and all those that substantially varied in clinical benefit and billing margin among the recommended treatment options. Substantial variation in clinical benefit was defined as at least one treatment having an efficacy score that was higher, and safety, quality of evidence, and consistency of evidence scores that were all at least as high as other recommended treatments, following previous work.20 We did not use the affordability score in assessing clinical benefit.23 Substantial variation in billing margin was defined as a difference of ≥$2000 or a difference that was at least twofold in relative terms and ≥$100 in dollar terms between the treatments with the highest and lowest clinical benefit within the scenario. The relatively small minimum dollar value difference (≥$100) was informed by the literature on drug industry payments to physicians, which has found that even small dollar amounts can sway prescribing.26 27 28 We excluded scenarios that would not be identifiable within claims data (eg, defined by clinical criteria not available in claims) or had insufficient sample size (figure S1 and methods in supplementary appendix).

ustry payments to physicians, which has found that even small dollar amounts can sway prescribing.26 27 28 We excluded scenarios that would not be identifiable within claims data (eg, defined by clinical criteria not available in claims) or had insufficient sample size (figure S1 and methods in supplementary appendix). Patient characteristics included age, sex, low income subsidy status29; race or ethnicity (using the Research Triangle Institute race classification); rural residence; region; median ZIP level income and percentage below poverty from the American Community Survey 2008-12; a frailty index30; and National Cancer Institute comorbidity index.31 We assigned the primary oncologist for each patient as the physician with a Medicare specialty code related to oncology who had the plurality of evaluation and management claims in proximity to the index date.32 Provider characteristics included specialty and gender (National Plan and Provider Enumeration System), medical school graduation year (Centers for Medicare and Medicaid and Services doctors and clinicians file), and clinical volume (number of assigned patients within the dataset, determined for each cancer type separately).

er characteristics included specialty and gender (National Plan and Provider Enumeration System), medical school graduation year (Centers for Medicare and Medicaid and Services doctors and clinicians file), and clinical volume (number of assigned patients within the dataset, determined for each cancer type separately). We assessed the association between provider billing margin and treatment selection using the McFadden conditional logit model.33 In this model structure, the dependent variable is the selection of a given treatment (among the alternative treatments), and the independent variables are treatment level characteristics. Our modeled independent variables were billing margin (in $100 increments) and clinical benefit (rank order of treatments based on NCCN Evidence Blocks scores, scaled so that for each scenario and time point the treatment with the highest NCCN Evidence Blocks scores has clinical benefit=1 and the lowest scores clinical benefit=0; eg, within each scenario, the highest ranking treatment at a given time point has a clinical benefit score of 1; for details, see table S1 in supplementary appendix). Billing margin and clinical benefit are time varying with respect to each patient’s index diagnosis date. To account for patient and provider level characteristics, we estimated inverse probability-of-treatment weights on the level of the treatment choice (using all variables described above in patient and provider characteristics), stabilized by the overall prevalence of each treatment within our sample and winsorized (setting maximum or minimum values, or both, to reduce the influence of outliers) at the 5% and 95% percentiles. We conducted a complete case analysis because the missingness for covariates was very low.

d provider characteristics), stabilized by the overall prevalence of each treatment within our sample and winsorized (setting maximum or minimum values, or both, to reduce the influence of outliers) at the 5% and 95% percentiles. We conducted a complete case analysis because the missingness for covariates was very low. Because the conditional logit model requires a common referent choice (in this study’s context, a cancer treatment) and all included cancer treatment options were unique to the clinical scenario, we estimated the model separately within each scenario. To achieve effect estimates aggregated across scenarios, we conducted meta-analyses (for billing margin and clinical benefit separately), treating each scenario as a separate study. Owing to observed heterogeneity, we used random effects estimates as the preferred interpretation of the overall effects of interest.

rio. To achieve effect estimates aggregated across scenarios, we conducted meta-analyses (for billing margin and clinical benefit separately), treating each scenario as a separate study. Owing to observed heterogeneity, we used random effects estimates as the preferred interpretation of the overall effects of interest. We conducted several sensitivity analyses with the following modifications to our primary model: dropped inverse probability-of-treatment weights from the model, in line with our theory that patient and provider characteristics are not confounders of the margin-selection association (figure S1 in supplementary appendix); assumed more conservative use of supportive care drugs in estimating billing margin (as opposed to more liberal use in the primary model) because these drugs may comprise a substantial portion of the billing margin for some treatments34; used a weighted average (as opposed to rank ordered) formulation of the NCCN Evidence Blocks scores as the measure of clinical benefit; included an outlier scenario (metastatic pancreas); and included only scenarios that had a >$2000 difference in billing margin among recommended treatments. Medicare analyses were performed using R version 3.5.2. Meta-analysis was conducted using the metagen package.

We conducted several sensitivity analyses with the following modifications to our primary model: dropped inverse probability-of-treatment weights from the model, in line with our theory that patient and provider characteristics are not confounders of the margin-selection association (figure S1 in supplementary appendix); assumed more conservative use of supportive care drugs in estimating billing margin (as opposed to more liberal use in the primary model) because these drugs may comprise a substantial portion of the billing margin for some treatments34; used a weighted average (as opposed to rank ordered) formulation of the NCCN Evidence Blocks scores as the measure of clinical benefit; included an outlier scenario (metastatic pancreas); and included only scenarios that had a >$2000 difference in billing margin among recommended treatments. Medicare analyses were performed using R version 3.5.2. Meta-analysis was conducted using the metagen package. Patients and the public were not directly involved in the design of this study. However, interactions with patients informed the motivation for this study; we have found that patients are not aware of the compensation policy for provider administered drugs, and upon learning about it, patients express deep concern about the potential for this system to incentivize higher cost care, which would result in higher out-of-pocket costs to them. Unpublished findings from this study have been presented to the leadership of patient advocacy groups.

We obtained treatment options recommended by the NCCN, who produce the most widely used clinical practice guidelines in oncology. For each patient, we determined the systemic treatment regimen they started within 180 days of follow-up after diagnosis among the contemporary NCCN recommended regimens, following previous methods.20 We used the NCCN Evidence Blocks scores, an expert opinion based metric with high ease of use and validity,21 22 as our proxy measure for the clinical benefit of each treatment option. The NCCN Evidence Blocks grade each cancer treatment on five criteria: efficacy (the extent to which treatment improves survival and symptoms), safety (treatments with fewer side effects receiving higher scores), quality of evidence (the number and rigor of the supporting clinical trials), consistency of evidence (the degree to which clinical trials agree on the degree of benefit), and affordability (with less expensive treatments receiving higher scores). Each criterion is scored from one to five, with higher scores being more favorable. We abstracted Evidence Blocks scores for all cancer treatments recommended by the NCCN, as previously described.23 24 Because these guidelines are continuously updated to reflect new evidence, we repeated this abstraction every six months (1 January and 1 July) across the study period.

gher scores being more favorable. We abstracted Evidence Blocks scores for all cancer treatments recommended by the NCCN, as previously described.23 24 Because these guidelines are continuously updated to reflect new evidence, we repeated this abstraction every six months (1 January and 1 July) across the study period. For each treatment, we estimated the provider billing margin under Medicare payment rates, following previous work (for full details of methods, see supplementary appendix).23 24 We assumed billing margins of 6% of the average sales price for provider administered drugs, and no margin for prescription drugs. We included guideline concordant supportive care drugs (growth factors and anti-emetics). Margins for each treatment were re-estimated every six months to reflect manufacturer price changes and price changes driven by new generics or biosimilars.25

We used a multicohort design, including patients from each applicable cancer clinical scenario. From the many NCCN defined clinical scenarios (eg, adjuvant treatment, stage 3 melanoma), we included all those that reflect initial systemic treatment (including adjuvant or neoadjuvant) because subsequent line treatments are challenging to identify accurately in claims data, and all those that substantially varied in clinical benefit and billing margin among the recommended treatment options. Substantial variation in clinical benefit was defined as at least one treatment having an efficacy score that was higher, and safety, quality of evidence, and consistency of evidence scores that were all at least as high as other recommended treatments, following previous work.20 We did not use the affordability score in assessing clinical benefit.23 Substantial variation in billing margin was defined as a difference of ≥$2000 or a difference that was at least twofold in relative terms and ≥$100 in dollar terms between the treatments with the highest and lowest clinical benefit within the scenario. The relatively small minimum dollar value difference (≥$100) was informed by the literature on drug industry payments to physicians, which has found that even small dollar amounts can sway prescribing.26 27 28 We excluded scenarios that would not be identifiable within claims data (eg, defined by clinical criteria not available in claims) or had insufficient sample size (figure S1 and methods in supplementary appendix).

Patient characteristics included age, sex, low income subsidy status29; race or ethnicity (using the Research Triangle Institute race classification); rural residence; region; median ZIP level income and percentage below poverty from the American Community Survey 2008-12; a frailty index30; and National Cancer Institute comorbidity index.31 We assigned the primary oncologist for each patient as the physician with a Medicare specialty code related to oncology who had the plurality of evaluation and management claims in proximity to the index date.32 Provider characteristics included specialty and gender (National Plan and Provider Enumeration System), medical school graduation year (Centers for Medicare and Medicaid and Services doctors and clinicians file), and clinical volume (number of assigned patients within the dataset, determined for each cancer type separately).

We assessed the association between provider billing margin and treatment selection using the McFadden conditional logit model.33 In this model structure, the dependent variable is the selection of a given treatment (among the alternative treatments), and the independent variables are treatment level characteristics. Our modeled independent variables were billing margin (in $100 increments) and clinical benefit (rank order of treatments based on NCCN Evidence Blocks scores, scaled so that for each scenario and time point the treatment with the highest NCCN Evidence Blocks scores has clinical benefit=1 and the lowest scores clinical benefit=0; eg, within each scenario, the highest ranking treatment at a given time point has a clinical benefit score of 1; for details, see table S1 in supplementary appendix). Billing margin and clinical benefit are time varying with respect to each patient’s index diagnosis date. To account for patient and provider level characteristics, we estimated inverse probability-of-treatment weights on the level of the treatment choice (using all variables described above in patient and provider characteristics), stabilized by the overall prevalence of each treatment within our sample and winsorized (setting maximum or minimum values, or both, to reduce the influence of outliers) at the 5% and 95% percentiles. We conducted a complete case analysis because the missingness for covariates was very low.

We conducted several sensitivity analyses with the following modifications to our primary model: dropped inverse probability-of-treatment weights from the model, in line with our theory that patient and provider characteristics are not confounders of the margin-selection association (figure S1 in supplementary appendix); assumed more conservative use of supportive care drugs in estimating billing margin (as opposed to more liberal use in the primary model) because these drugs may comprise a substantial portion of the billing margin for some treatments34; used a weighted average (as opposed to rank ordered) formulation of the NCCN Evidence Blocks scores as the measure of clinical benefit; included an outlier scenario (metastatic pancreas); and included only scenarios that had a >$2000 difference in billing margin among recommended treatments.

Patients and the public were not directly involved in the design of this study. However, interactions with patients informed the motivation for this study; we have found that patients are not aware of the compensation policy for provider administered drugs, and upon learning about it, patients express deep concern about the potential for this system to incentivize higher cost care, which would result in higher out-of-pocket costs to them. Unpublished findings from this study have been presented to the leadership of patient advocacy groups.

There were 12 clinical scenarios eligible for analysis (figure S2 in supplementary appendix). We included 27 076 patients across all 12 scenarios (figure S3, table S2 in supplementary appendix). Among eligible patients, the median age was 73, 59.7% were women, 77.4% were white, and 20.1% received low income subsidies (table 1). Clinical scenario groups ranged in size from multiple myeloma, non-transplant eligible (n=4771, 17.6%) to metastatic HER2 (human epidermal growth factor receptor 2) positive breast cancer (n=212, 0.8%). Patient characteristics within clinical scenarios are available in table S3 in the supplementary appendix. Of the 27 076 patients, 19 397 (71.6%) received systemic treatment during the study period with a regimen that was NCCN recommended and were therefore eligible for inclusion in the statistical analysis. This subset of patients had similar characteristics to the overall cohort (table 1). Cohort characteristics 1=£0.74, €0.85. Data are numbers (%) unless stated otherwise. Characteristics are shown for overall cohort and analytic subset that received guideline concordant systemic treatment. AIAN=American Indian and Alaska Native; EGFR=epidermal growth factor receptor; HER2=human epidermal growth factor receptor 2; IQR=interquartile range; NCI=National Cancer Institute.

1=£0.74, €0.85. Data are numbers (%) unless stated otherwise. Characteristics are shown for overall cohort and analytic subset that received guideline concordant systemic treatment. AIAN=American Indian and Alaska Native; EGFR=epidermal growth factor receptor; HER2=human epidermal growth factor receptor 2; IQR=interquartile range; NCI=National Cancer Institute. Across all 12 clinical scenarios, a total of 147 treatments were NCCN recommended for any portion of the study period and were received by at least one patient (table S4 in supplementary appendix). Of the 147 treatments, 30 (20.4%) were newly recommended during the study period. Seventy treatments (47.6%) were received by at least 20 patients. There was a high degree of variation in the provider billing margin and the clinical benefit of the treatments received by patients, within and between scenarios (fig 1, table S5 in supplementary appendix). Across all patients, the median billing margin of the received treatment was $917 (interquartile range $225-1855), and the median clinical benefit was 0.71 (0.50-0.92). The scenario with the highest median billing margin was adjuvant treatment for HER2 positive breast cancer ($4293, $3920-5751), and the scenario with the highest median clinical benefit was metastatic pancreas (1, interquartile range 0.8-1).

artile range $225-1855), and the median clinical benefit was 0.71 (0.50-0.92). The scenario with the highest median billing margin was adjuvant treatment for HER2 positive breast cancer ($4293, $3920-5751), and the scenario with the highest median clinical benefit was metastatic pancreas (1, interquartile range 0.8-1). Distribution of provider billing margin and clinical benefit of treatment received by each patient. $1 (£0.74; €0.85) added to all billing margin values to allow for representation on log scale. Clinical benefit measured using National Comprehensive Cancer Network Evidence Blocks scores. EGFR=epidermal growth factor receptor; HER2=human epidermal growth factor receptor 2 In the primary analysis, there was no association between treatment billing margin and the likelihood of use (odds ratio 0.97, 95% confidence interval 0.91 to 1.03; P=0.3; fig 2, upper panel). There was a positive and statistically significant association between clinical benefit (eg, NCCN Evidence Blocks scores) and the likelihood of use (1.62, 1.15 to 2.29; P=0.006; fig 2, lower panel). These results were unchanged in all sensitivity analyses (figures S4-S7 and table S6 in supplementary appendix).

panel). There was a positive and statistically significant association between clinical benefit (eg, NCCN Evidence Blocks scores) and the likelihood of use (1.62, 1.15 to 2.29; P=0.006; fig 2, lower panel). These results were unchanged in all sensitivity analyses (figures S4-S7 and table S6 in supplementary appendix). Overall association of treatment selection with provider billing margin and clinical benefit. Meta-analysed results across clinical scenarios for provider billing margin (upper panel; random effects P=0.3) and clinical benefit (lower panel; random effects P=0.006). Scenario of metastatic HER2 positive breast cancer is not included in clinical benefit meta-analysis because after removal of rarely selected treatments, no variation was observed in clinical benefit among remaining treatments. EGFR=epidermal growth factor receptor; HER2=human epidermal growth factor receptor 2

In this study of cancer treatment selection, we found no evidence that oncologists used more profitable treatments over less profitable treatments. A positive association was observed between treatment use and clinical benefit; the most effective, guideline preferred treatments were used more often. However, after controlling for clinical benefit, we found no evidence of an association between billing margin and treatment selection. This result was robust to several sensitivity analyses. Whether observations were weighted to balance patient characteristics did not substantially affect results, suggesting that these characteristics (age, comorbidity, etc) are not major confounders of the billing margin treatment selection association. One clinical scenario (metastatic pancreas) was challenging to model owing to collinearity of margin and clinical benefit, which resulted in unexpectedly large estimates; however, summary results did not change regardless of whether this scenario was included or excluded. In all sensitivity analyses, there remained a strong, positive association between clinical benefit and treatment selection, while the association with billing margin was null.

resulted in unexpectedly large estimates; however, summary results did not change regardless of whether this scenario was included or excluded. In all sensitivity analyses, there remained a strong, positive association between clinical benefit and treatment selection, while the association with billing margin was null. Our findings are unexpected in the context of previous research on physician responsiveness to financial incentives. Previous studies have consistently found that reimbursement affects clinical decision making in oncology18; however, only one previous study examined whether billing incentives influence chemotherapy selection.35 This study found evidence of an effect, but was limited to one cancer type (breast) and studied the time period before the current US Medicare reimbursement model (with the 6% fee) was enacted in 2005. Before this change, physician fees for cancer drugs were far higher,13 36 37 and whether the current, relatively modest fees influence treatment selection has not been studied. Therefore, one possible explanation for our unexpected findings is that the 6% profit margin is too small to affect treatment decisions.

2005. Before this change, physician fees for cancer drugs were far higher,13 36 37 and whether the current, relatively modest fees influence treatment selection has not been studied. Therefore, one possible explanation for our unexpected findings is that the 6% profit margin is too small to affect treatment decisions. Although we found no association between margin and treatment selection in the aggregate, we cannot exclude the possibility that some smaller subsets of providers may be margin responsive. Previous research found heterogeneity among providers in billing responsive practice patterns. Use of high margin drugs (including those with low clinical value)38 39 is generally higher in physician offices—where oncologists commonly receive compensation proportional to their billed services—compared with hospitals.40 41 42 43 44 45 For-profit and provider owned practices are additional provider categories in which the use of low value, high margin services may be increased.14 43 46 Therefore, although our overall estimate is null, further research is warranted to assess heterogeneity among provider subgroups.

ared with hospitals.40 41 42 43 44 45 For-profit and provider owned practices are additional provider categories in which the use of low value, high margin services may be increased.14 43 46 Therefore, although our overall estimate is null, further research is warranted to assess heterogeneity among provider subgroups. Although this single study should not be taken to suggest that financial incentives never affect clinical decision making, our results indicate that the financial incentives in fee-for-service billing may not play a large part in the selection of cancer drugs. Concerns have been raised that fee-for-service billing might lead to overuse of expensive treatments because it presents providers with the financial incentive to do so.37 47 48 Our findings suggest that this may not occur to a large extent, and that the clinical benefit of cancer drugs remains the primary factor in treatment selection. In many instances, treatments with low reimbursement remained heavily favored by clinicians. For example, we observed that cisplatin remained the most used radiosensitizing agent for locally advanced head and neck cancer despite its very low margin and higher margin (but less beneficial) agents such as cetuximab being available. In other instances, treatments with high reimbursement were used most frequently, but these treatments also typically had the greatest clinical benefit. For example, for metastatic HER2 positive breast cancer, the treatments pertuzumab plus trastuzumab plus either docetaxel or paclitaxel were the most frequently used and had the highest billing margin as well as the highest clinical benefit.

uently, but these treatments also typically had the greatest clinical benefit. For example, for metastatic HER2 positive breast cancer, the treatments pertuzumab plus trastuzumab plus either docetaxel or paclitaxel were the most frequently used and had the highest billing margin as well as the highest clinical benefit. Fee-for-service billing also presents a potential concern about underuse of low price treatments. If low reimbursement for clinically necessary drugs makes them financially less attractive for providers, then patient access could suffer. Our finding that oncologists use the clinically appropriate treatments, even when they are less profitable, suggests this concern may be less likely to occur as well. This finding has implications for payment policy. Policy makers in healthcare systems that do not pay fee-for-service for cancer drugs may be reassured that the absence of such fees is unlikely to reduce care quality; drug fees do not appear to be necessary for the adoption of new, effective treatments. In the US context, physician groups have opposed the Inflation Reduction Act—a recent law that will negotiate lower prices on some high cost drugs and consequently reduce physician fees for administering them—on the basis that lower physician fees will harm patient access.9 49 In the context of our findings, the Act’s drug price negotiations appear less likely to negatively affect care quality while still achieving their primary objective of substantially lowering drug prices.

ly reduce physician fees for administering them—on the basis that lower physician fees will harm patient access.9 49 In the context of our findings, the Act’s drug price negotiations appear less likely to negatively affect care quality while still achieving their primary objective of substantially lowering drug prices. This study’s conclusions are supported by several strengths, including a large, nationally representative dataset, a focus on clinical settings with high variation in billing margin among treatment options (these would be the most affected by margin seeking behavior, adding confidence to our null result), and a cohort of cancer scenarios with broad representation of liquid and solid tumors, and adjuvant and metastatic settings. We conducted a detailed cohort creation process to ensure the accuracy of the study sample. Most importantly, we used detailed and previously published methods for measuring treatment cost and clinical benefit,23 24 and applied these to create a longitudinal analytic framework which compared each patient’s condition with the guideline recommended treatment options—and the treatments’ benefits and costs.20

le. Most importantly, we used detailed and previously published methods for measuring treatment cost and clinical benefit,23 24 and applied these to create a longitudinal analytic framework which compared each patient’s condition with the guideline recommended treatment options—and the treatments’ benefits and costs.20 This study has limitations resulting from its observational design. Identifying incident cancers and inferring stage using claims likely results in some degree of misclassification. This study can only evaluate the association between billing margin and use, and cannot infer causality. We used a proxy measure of clinical benefit (recommendations in guidelines by NCCN), which is imperfect but is the only source of treatment recommendations in oncology with the necessary breadth and timely incorporation of new clinical evidence. Our observation that use is strongly associated with our clinical benefit measure (an expected finding) and that this same proxy measure has produced the expected use patterns in previous studies20 supports its usefulness in characterizing cancer treatment patterns at scale and across time. Our conclusions are limited to the US setting and the fee-for-service Medicare population, and it is possible that results would be different in other payer settings, such as commercial insurance, where drug billing margins are far higher,10 11 50 or Medicare Advantage, where plans may control the use of high cost cancer drugs.51 52 53

re limited to the US setting and the fee-for-service Medicare population, and it is possible that results would be different in other payer settings, such as commercial insurance, where drug billing margins are far higher,10 11 50 or Medicare Advantage, where plans may control the use of high cost cancer drugs.51 52 53 We did not analyze treatment patterns within provider subgroups, such as for-profit networks or privately owned clinics, which might have greater financial motivation in treatment decisions.40 54 We assumed that providers would receive no margin for oral or prescription cancer drugs, which is not the case within practices with integrated dispensing pharmacies; however, previous work suggests that pharmacy integration may not significantly affect prescribing patterns.55 We also assumed the historical 6% in estimating billing margin, although margins have been closer to 4.3% under budgetary “sequestration” policies enacted in 201356; this is unlikely to have a major impact on treatment selection because the change in margin would be proportional across treatment options. Additionally, our estimated billing margin does not reflect the reality that margin is heterogeneous across providers: the margin is the difference between provider reimbursement and drug purchase price, and purchase price varies across providers.48 Our analysis focused on treatment selection among patients receiving treatment; we did not evaluate the impact of billing incentives on the decision of whether to offer treatment, and previous work suggests that billing incentives might affect this decision.13 15 45 The statistical model did not account for potential correlation of treatment choices between physicians or practices; accounting for this could result in wider confidence intervals, which could conceivably lead to the association of clinical benefit and treatment selection no longer being statistically significant.

45 The statistical model did not account for potential correlation of treatment choices between physicians or practices; accounting for this could result in wider confidence intervals, which could conceivably lead to the association of clinical benefit and treatment selection no longer being statistically significant. Among Medicare beneficiaries starting systemic, pharmacological cancer treatment, the clinical characteristics of cancer treatments appear to be more important determinants of treatment selection than reimbursement considerations. Differences in billing margin among treatment options do not appear to influence which treatment patients receive, whereas treatments with greater clinical benefit were consistently used more often than those with lower benefit. In the United States, compensation for provider administered cancer drugs is volume based and proportional to drug price, creating the financial incentive to use more expensive drugs Whether cancer care providers respond to this incentive by selecting more profitable treatments is unknown This study suggests that oncologists were not more likely to choose more profitable cancer treatments over less profitable treatments Oncologists were more likely to choose treatments with higher clinical benefit scores These findings suggest that policy changes that affect drug price or provider reimbursement may be unlikely to affect use patterns

Our findings are unexpected in the context of previous research on physician responsiveness to financial incentives. Previous studies have consistently found that reimbursement affects clinical decision making in oncology18; however, only one previous study examined whether billing incentives influence chemotherapy selection.35 This study found evidence of an effect, but was limited to one cancer type (breast) and studied the time period before the current US Medicare reimbursement model (with the 6% fee) was enacted in 2005. Before this change, physician fees for cancer drugs were far higher,13 36 37 and whether the current, relatively modest fees influence treatment selection has not been studied. Therefore, one possible explanation for our unexpected findings is that the 6% profit margin is too small to affect treatment decisions.

Although this single study should not be taken to suggest that financial incentives never affect clinical decision making, our results indicate that the financial incentives in fee-for-service billing may not play a large part in the selection of cancer drugs. Concerns have been raised that fee-for-service billing might lead to overuse of expensive treatments because it presents providers with the financial incentive to do so.37 47 48 Our findings suggest that this may not occur to a large extent, and that the clinical benefit of cancer drugs remains the primary factor in treatment selection. In many instances, treatments with low reimbursement remained heavily favored by clinicians. For example, we observed that cisplatin remained the most used radiosensitizing agent for locally advanced head and neck cancer despite its very low margin and higher margin (but less beneficial) agents such as cetuximab being available. In other instances, treatments with high reimbursement were used most frequently, but these treatments also typically had the greatest clinical benefit. For example, for metastatic HER2 positive breast cancer, the treatments pertuzumab plus trastuzumab plus either docetaxel or paclitaxel were the most frequently used and had the highest billing margin as well as the highest clinical benefit.

This study’s conclusions are supported by several strengths, including a large, nationally representative dataset, a focus on clinical settings with high variation in billing margin among treatment options (these would be the most affected by margin seeking behavior, adding confidence to our null result), and a cohort of cancer scenarios with broad representation of liquid and solid tumors, and adjuvant and metastatic settings. We conducted a detailed cohort creation process to ensure the accuracy of the study sample. Most importantly, we used detailed and previously published methods for measuring treatment cost and clinical benefit,23 24 and applied these to create a longitudinal analytic framework which compared each patient’s condition with the guideline recommended treatment options—and the treatments’ benefits and costs.20

Among Medicare beneficiaries starting systemic, pharmacological cancer treatment, the clinical characteristics of cancer treatments appear to be more important determinants of treatment selection than reimbursement considerations. Differences in billing margin among treatment options do not appear to influence which treatment patients receive, whereas treatments with greater clinical benefit were consistently used more often than those with lower benefit. In the United States, compensation for provider administered cancer drugs is volume based and proportional to drug price, creating the financial incentive to use more expensive drugs Whether cancer care providers respond to this incentive by selecting more profitable treatments is unknown This study suggests that oncologists were not more likely to choose more profitable cancer treatments over less profitable treatments Oncologists were more likely to choose treatments with higher clinical benefit scores These findings suggest that policy changes that affect drug price or provider reimbursement may be unlikely to affect use patterns

In the United States, compensation for provider administered cancer drugs is volume based and proportional to drug price, creating the financial incentive to use more expensive drugs Whether cancer care providers respond to this incentive by selecting more profitable treatments is unknown

This study suggests that oncologists were not more likely to choose more profitable cancer treatments over less profitable treatments Oncologists were more likely to choose treatments with higher clinical benefit scores These findings suggest that policy changes that affect drug price or provider reimbursement may be unlikely to affect use patterns