Jesse Sussell, PhD; Jacqueline Vanderpuye-Orgle, PhD; Diana Vania, MSc; Hans-Peter Goertz, MPH; and Darius Lakdawalla, PhD
Insurers, healthcare providers, and others have noted that prices for novel oncology treatments are rising over time.1-3
Price per episode of treatment has risen steadily, and prior research suggests that although the efficacy of cancer therapies (as measured by gains in overall survival) is certainly improving over time, prices are rising even faster.4
If improved efficacy or effectiveness does not provide the dominant explanation, then what else might be contributing to the observed increase in prices?
One possibility is that firms today possess greater market power and use it to earn greater rewards for their drugs. If true, higher prices would coincide with higher revenues. At first blush, the magnitude of price growth in the oncology market may make growth in revenues seem self-evident. However, it becomes more difficult to assess when one considers the fact that newer drugs tend to treat fewer patients, as we document later. This decline in patients treated may be due to greater personalization of therapy, slower incremental progress that fails to force older drugs out of the market, or other factors. Regardless, rising prices alongside falling quantities make the trend in revenues an empirical question.
An empirically interesting context for this investigation is the market for targeted oncology therapies, first introduced in the late 1990s. Prior research has documented that targeted therapies have quickly come to dominate the market: For example, targeted therapies accounted for about two-thirds of all chemotherapy expenditures by 2011.5
In this study, we investigate trends in prices, quantities, and revenues for targeted oncology therapies. We aim to determine whether higher prices have coincided with higher revenues and rewards for innovation or whether, instead, price growth has coincided with flat or falling revenues. Analyzing a sample of targeted therapies intended to treat common tumor types, we estimate the growth in price per patient-year, the reduction in the average size of the annual patient base using each drug, and the resulting change in revenue. In auxiliary analysis, we also estimate the implied increase in costs of drug development per patient-year.
METHODS AND DATA
We focused on targeted cancer drugs launched since 1997, when the first targeted agent (monoclonal antibody) was approved by the FDA. We considered all drugs primarily aimed at extending survival and/or progression-free survival for patients with cancer and focused on the 6 most common tumor types: breast, colorectal, melanoma, non-Hodgkin lymphoma, non–small cell lung, and prostate.6
As some drugs are indicated for more than 1 tumor type, the unit of analysis for this study was therapy–tumor pair. To account for the fact that many drugs have more than 1 indication, and that approvals for new indications are frequently granted subsequent to the initial approval, we computed estimates for all possible therapy–tumor pairs, including all indications known at the time of this analysis. We estimated relative usage of a single agent for different indications by assuming usage proportional to relative disease incidence. After we dropped data for indications not of interest, our final sample consisted of 29 therapies and 33 therapy–tumor pairs (see eAppendix Table 1
[eAppendix available at ajmc.com
Using data for each of these therapy–tumor pairs, we defined 3 measures of interest: therapy price, the number of patients using each drug for each indication, and annual revenues. We estimated each of these quantities for each therapy–tumor pair of interest, for each year in our analysis sample (1997-2015), and then used panel regression to investigate how they changed with respect to drug launch year. In supplementary analysis presented in the eAppendix, we also estimate trends in per-patient research and development (R&D) cost.
Therapy Price, Patient Count, and Revenues
We estimated the total number of patients and the estimated price of treatment using the IQVIA National Sales Perspective (NSP) data set.7
The NSP reports nationally representative estimates of the total annual units of individual therapies distributed in the United States and of the total related revenues received by manufacturing firms. We combined these data with monthly dose and average length of treatment values specific to each therapy–tumor pair (obtained from FDA labels and clinical trials) to derive our price and patient count outcomes. Price was calculated as the cost for a full course of treatment (ie, the total revenue estimated to return to the manufacturer as a result of 1 patient being treated). Patient counts were imputed through analysis of the total number of therapy units distributed in combination with label information on dosing for a full course of treatment. Revenues were calculated within-year as the product of price and patient count. Revenues were calculated at the therapy, rather than therapy–tumor, level because this captures the total return on launching a new drug. Complete details on the construction of these measures are provided in the eAppendix.
To validate our findings regarding trends over time in the average number of patients, we separately analyzed patient counts in an independent data set, the Medicare Current Beneficiary Survey (MCBS). The MCBS is a survey of a representative sample of Medicare beneficiaries and contains data on diagnosis and drug utilization.8
For each of the therapy–tumor pairs of interest, we counted the number of individuals in the MCBS sample who (1) had an indication for that tumor type and (2) reported using that therapy. We then used sampling weights to inflate these counts to the population level. We note that the MCBS results are estimates of counts for the Medicare population only and should not be construed as separate estimates for the aggregate US population. We repeated this process for the individual years between 1997 and 2012. (At the time of this analysis, 2012 was the most recent year for which MCBS data were available.) Finally, we compared trends over time in average patient counts as estimated in the IQVIA and MCBS analyses.
We sought to determine whether each of our main outcomes (therapy price, number of patients, and annual revenues) was correlated with therapy launch year. To do this, we fit a series of regression models using the above measures as outcomes and therapy launch period as the key independent variable. We converted therapy launch year into a categorical variable with 3 groups: drugs launched between 1997 and 2002 (reference category), drugs launched between 2003 and 2009, and drugs launched between 2010 and 2015. We present results from 2 sets of regressions. In the first, we regress each of our outcomes of interest on this categorical variable corresponding to launch time period. In the second, we include an additional regressor (years post launch) to control for life cycle trends in the price and quantity of a drug following its market entry. We report trends in average regression-adjusted price and quantity and then report movement over time in the entire distribution of regression-adjusted annual revenues per therapy. We conducted sensitivity analyses involving (1) adding covariates to the base-case model, (2) using alternative values for key parameters related to the cost of R&D, and (3) asserting a linear relationship between launch year and outcomes, rather than the period fixed-effects structure described above. These are discussed in detail in the eAppendix.
We begin with an analysis of the first study outcome, episode treatment price. The fact that newer oncology products are increasingly costly has been extensively documented in the literature.9,10
This trend is also readily apparent in our data on targeted agents. Figure 1
presents the results of a series of regressions that use therapy price as the dependent variable and launch year as the key independent variable and that follow the specifications defined previously.
We find a statistically significant positive correlation between price and launch year. The episode treatment cost for drugs launched between 2003 and 2009 was, on average, $23,000 greater than that for drugs launched between 1997 and 2002. The difference was even greater for drugs launched between 2010 and 2015—a statistically significant average difference in episode treatment cost of about $43,000.
This study sought to examine whether price growth has coincided with revenue growth. Thus, trends in annual average patient counts play a crucial role.
presents the results of the regressions for the patient count outcome, first without and then with the control for time since launch. Figure 2 demonstrates a strong negative relationship between launch year and average patient count. Relative to therapies launched in the early period, the dummy models suggest that therapies launched in the middle period were used by, on average, 28,000 to 35,000 fewer patients annually, whereas therapies launched in the late period were used by 33,000 to 44,000 fewer patients. Detailed time-series plots of patient counts for individual therapy–tumor pairs, and for average values within launch period, are presented in eAppendix Figures 1-4
To confirm these results, we conducted a separate analysis of patient counts by therapy–tumor pair in the independent MCBS data set. Comparative results are presented in Figure 3
Note that the IQVIA data set covers the entire US population, whereas MCBS covers Medicare patients only. Each is designed to be nationally representative for its particular sample frame.
There is clear evidence that annual patient populations are smaller for more recently launched drugs: In the main analysis using IQVIA data, the average patient count fell from 48,520 per drug for drugs launched in the early period to 4781 per drug for drugs launched in the late period, a decline of 90%. A decline of similar magnitude (85%) is observed in the Medicare data.
The reduction in quantity seems to have offset growth in price. The entire distribution of annual revenues has fallen over time. We use the regression-adjusted (ie, predicted) revenues from our regression model of revenues as a function of years since launch and time period. We also aggregate this up to the therapy level to eliminate the possibility that newer drugs spawn more indications and thus artificially lower revenues per tumor type. This permits uniform comparisons over time that account for the way in which revenue evolves over the life cycle of a drug. Figure 4
(A and B) presents the distribution of regression-adjusted annual revenues (at the therapy level) for each of the 3 launch periods; the difference between the 2 panels is that Figure 4B removes a single influential outlier—bevacizumab—from the data set. Both panels show that the distribution of regression-adjusted annual revenue has shifted left over time. In both cases, the most recent distribution ranges from $250 million to $500 million, whereas the earliest period shows a distribution from about $250 million to more than $900 million (all values are reported in 2015 US$). The sole difference between the distributions lies in the middle period. In the full sample, the main mass of the distribution lies between those of the early and late launch periods, but significant right skew is present. This long right tail is caused by the presence of a single drug, the blockbuster bevacizumab. Annual revenues for that drug (limited to the 6 indications of interest to this study) routinely exceeded $1 billion, in part because it was approved for more than 1 of those tumor types.
presents the shifts in the distributions at key percentile points. Because bevacizumab is such an outlier, the Table accurately describes the distributions in both panels of Figure 4 (ie, the bevacizumab data points lie beyond the 90th percentile).
Annual adjusted revenues for the median drug have fallen from about $580 million to $287 million, a decline of about 50%. There is a decline of roughly 40% at the 25th percentile and nearly 60% at the 75th percentile. The only region of increase occurs at the 10th percentile, where revenues increased from the early to middle period, only to fall back down in the final period.
Results from the sensitivity analyses are presented and explained in the eAppendix.
The rising cost of novel oncology therapies has been a source of great controversy in recent years.11-13
Our analysis confirms that prices of oncology drugs are indeed rising rapidly. We show that the number of patients taking each drug has dropped substantially over the same period of time. As a result, revenues have fallen at every point in the distribution, after accounting for life cycle growth in revenues over years since launch. This suggests that price growth is unlikely to have resulted from greater pricing power, at least within this market segment. Profit-maximizing firms with more pricing power would never willingly make decisions that lead to lower revenues for each drug launched. One exception to this point might occur if costs of drug discovery or production have fallen significantly. Although the extant academic literature on the costs of drug discovery remains controversial, all of it points to rising costs.14
We know of no academic publications on trends in the costs of oncology drug production; more research is called for in this area.
This study has several important limitations. The IQVIA NSP data set, which provided much of the source data for this study, contained “restrictions” (partially missing values) for some therapies in some years; we recoded these as missing. This step will not affect the results of the study, as long as the circumstance of missing data is not correlated with our outcomes of interest. Also, our measure of the cost of individual therapies contains only the component of total cost that returns to pharmaceutical manufacturers; it does not incorporate markups by wholesalers or hospitals, nor does it include any confidential rebates paid by manufacturers to purchasers. In addition, we focus on only treatments for the 6 most common tumor types; as such, our results have limited generalizability to other forms of cancer. Finally, to estimate therapy prices, we assume that average treatment duration is equal to the duration indicated on the drug label. In practice, individual patients’ duration of therapy may be longer or shorter than is suggested by the label because of factors such as mortality, discontinuation, and extended treatment at the discretion of the physician.
Previous research has suggested that we should be skeptical of the notion that prices are rising solely because effectiveness is rising.4
Our study also casts doubt on an explanation for price growth that relies solely on rising R&D costs: If this were so, firms would necessarily respond by launching drugs capable of earning higher revenues. Our findings suggest instead a relationship between price growth and average patient counts, although the precise nature of this relationship is not fully clear. One possible explanation for declining patient counts is relatively slow growth in effectiveness over time; this would improve the ability of older drugs to remain on the market. The consequences of this longevity would be reduced market share and reduced revenue for newer drugs. A second possible hypothesis for declining patient counts would be increased competition: All else being equal, an increase in the number of drugs approved for a given tumor type would lead to a decline in the average number of patients per drug. This explanation, however, would also suggest an increase in price competition, which is inconsistent with the observed data.
A final possible explanation for the trends we observe is growth in the development of drugs that target patients with specific biomarkers (sometimes referred to as personalized medicine) within the targeted oncology market. By design, these drugs target subsets of the total population of patients with the indicated cancer. Shrinking patient counts might be in part the result of more personalized therapies that treat narrower indications. For example, trastuzumab is indicated for human epidermal growth factor receptor-2–overexpressing breast cancer, which comprises only 15% to 20% of invasive breast cancer cases.15
Mechanically, personalization leads to lower revenue when we hold prices constant. From another perspective, this is a cost to society of personalization: Without significant offsetting price growth, personalized therapies generate lower revenues and returns to innovators; this reduction in returns may reduce the rate of drug discovery in the long run.16
More research is needed on these and other hypotheses for the causes of drug price growth and the observed decline in average patient counts.
Despite acknowledged limitations, this study provides surprising new data on declining patient populations treated by targeted cancer agents. This pattern is likely an important and, to our knowledge, previously undescribed factor that lies behind trends in revenues and rewards from innovation in oncology.