We identified all new drugs that received FDA approval between 1 January 2000 and 1 January 2022. The sample was then restricted to include only anticancer medicines, excluding those for supportive cancer care, diagnostic agents, and antiemetics. For these anticancer drugs, we identified all original and supplemental anticancer indication approvals until 1 January 2022. In our analyses, we only included drugs covered under Medicare Part B, given that no longitudinal price data are available for drugs covered under Medicare Part D from the Centers for Medicare and Medicaid Services (CMS). In general, Medicare Part B covers injectable cancer drugs (typically drugs that are administered at a hospital or doctor’s office), while Medicare Part D covers oral cancer drugs (typically self-administered drugs). Certain drugs with multiple routes of administration are covered by Medicare Part B and D, for example, everolimus.

For all identified cancer agents, we collected price data and information characterizing each drug’s characteristics, disease epidemiology, and market dynamics (electronic supplementary material [ESM] Table e1).

First, we characterized each drug’s innovativeness. Two reviewers assessed the novelty of the underlying drug target based on the World Health Organization’s Anatomical Therapeutic Chemical code. Drugs with novel targets were considered first-in-class, whereas those with known targets were considered next-in-class. Second, the University of Alabama’s drug database, ‘Drug Bank’, was accessed to determine each drug’s product type. Drugs were categorized as small molecules and others, which entail biological agents, antibody-drug conjugates, gene therapies, cell therapies, enzymes, and radionuclides. Third, we obtained information on the approval of companion biomarkers from FDA labels. Finally, we obtained data from FDA websites to determine if special FDA designations were associated with each drug, e.g. orphan designation, accelerated approval, fast track, priority review, and breakthrough therapy designation [32].

The clinical benefit of new cancer drugs was measured by their benefit in overall survival (OS), progression-free survival (PFS), and tumor response rates. We accessed FDA labels and ClinicalTrials.gov to collect data on OS and PFS hazard ratios and tumor response rates from randomized controlled trials (RCTs). Furthermore, we extracted median improvements in OS, PFS, and duration of response. The absolute median improvement in OS/PFS was calculated as the difference between median OS/PFS in the treatment relative to the control arm. The percentage improvement in OS/PFS was then calculated as the quotient of the absolute median OS/PFS improvement relative to the absolute median OS/PFS in the control arm. Although multiple analyses evaluated the association between the clinical benefit and launch prices of new drugs [2, 11,12,13,14, 16,17,18, 22], evidence scrutinizing the association between the clinical benefit and post-launch price changes of new drugs remains scarce [8]. Most European countries introduced regulations to limit drug price increases exceeding inflation and even reduce drug prices to control expenditure on new drugs. For instance, drug price increases are re-evaluated in Switzerland every 3 years, controlled by the government in England, and re-evaluated for drugs with new indications in Germany and France [33,34,35]. Given that over the study period there was no value-based pricing policy in the US that regulates launch prices and post-launch price changes, we hypothesized that there is no association between launch/post-launch prices and the drugs’ clinical benefit.

We obtained epidemiologic data for the US population in 2019 from the Global Burden of Disease study to describe the disease treated by each drug [36]. First, we collected disease incidence rates (per 100,000 US inhabitants) as a measure of disease rarity, and second, we collected disability-adjusted life-years (DALYs) per person as a measure of disease burden. DALYs are calculated as the sum of years lived with disability (YLD) and years of life lost (YLL). Therefore, DALYs not only capture the forgone lifetime but also the reduced quality of life that is caused by diseases. The disease-specific epidemiologic data were matched to each drug according to the treated disease specified in FDA labels.

Market dynamics were captured in two variables. We tracked the FDA approval of new supplemental indications for each drug. Given that these supplemental indications are often for non-orphan diseases supported by robust clinical trials with a relatively low clinical benefit (‘low-value indications’), we expect drug prices to decline following the introduction of new indications for the same drug [33, 37, 38].

We then monitored the number of new competitors entering the market for each drug. We used two alternative measures of new competitors. First, we counted the number of new cancer drug indications receiving FDA approval within the same disease during each quarter. This represents a broad measure of competition in the market of anticancer drugs (variable: new competitors [broad]). Second, we counted the number of new anticancer drug indications receiving FDA approval within the same disease, in the same line of therapy, for the same treatment setting, with the same biomarker during each quarter. This represents a narrow measure of competition in the market of anticancer drugs (variable: new competitors [narrow]). The narrow measure of competition might be more reflective of the underlying market dynamics in the cancer drug market, given that each drug and indication often fills a distinct therapeutic niche that is defined by the therapeutic setting (neoadjuvant vs. adjuvant vs. metastatic), line of therapy (first-line vs. second-line vs. advanced-line), and biomarker profile (for example, differentiated by driver mutations for non-small cell lung cancer (NSCLC): KRAS vs. EGFR vs. ALK vs. BRAF vs. MET vs. ROS1 vs. HER2 [ERBB2] vs. NTRK). For both measures of competition, we included the market entry of all new drug indications with FDA approval regardless of insurance states, e.g. we included drug indications covered under Part B and D.

Drug prices were calculated according to a methodology that has been described in prior articles [2, 16, 28, 39]. First, we accessed the CMS' quarterly average sales price (ASP) data files to obtain drug pricing data from 2005 to 2023. For each drug, we then calculated monthly treatment costs based on the dosing regimen defined in FDA labels for the average US patient with a body weight of 70 kg and a body surface area of 1.7 m2 [2, 14, 16, 17, 39]. As a result, these treatment costs only include the drug price and do not consider any additional charges for doctor’s fees, delivery expenses, administrative fees, or supportive care that may be necessary for the treatment of cancer patients.

We used descriptive statistics to describe the sample’s baseline characteristics, and then conducted random-effects regression models to evaluate the association between post-launch price changes and collected variables. Random-effects regressions were performed to examine the association between time-varying and time-invariant variables on drug prices. The use of random-effects rather than fixed-effects models was confirmed by performing the Hausman test (χ2 = 12.47, p = 0.0861) and the Lagrange Multiplier (LM) test (\(\overline^\) = 42,858.22, p < 0.001). All models account for drug-level clustered standard errors to adjust for heteroscedasticity and autocorrelation. Drug prices, disease incidence, and disease prevalence were transformed with the natural logarithm to account for their skewed distribution.

For all models, the dependent variable (\(_\)) is the inflation-adjusted log-price for each drug (\(d\)). First, we evaluated the association between each independent variable and launch as well as post-launch drug price changes in a series of separate univariate regression analyses. In these models, each independent time-invariant variable (\(_\)) was included alongside an interaction term between the time-invariant variable and the time since launch (\(_\)) (Eq. 1). We defined drug launch as the first time a drug’s price was listed in CMS files. Coefficients of the independent variable (\(_\)) can be interpreted as the association between the independent variable of interest and launch prices. The coefficient of the interaction term (\(_\)) can be interpreted as the association between the independent variable of interest and post-launch price changes. Product type, innovativeness, companion biomarkers, special FDA designations, disease incidence, and DALYs per person were included as time-independent variables. Across all models, \(_\) represents the drug-specific error and \(_\) represents the idiosyncratic error.

$$_=\alpha +__+__+___+_+_$$

(1)

Among models with time-varying variables, the variable of interest (\(_\)) was the only independent variable and its coefficient can be interpreted as the post-launch price change (Eq. 2). The number of new competitors and new supplemental indications were included as time-varying variables.

$$_=\alpha +__+__+_+_$$

(2)

Thereafter, a multivariate regression model was conducted (Eq. 3). Special FDA designations, except for the orphan designation, were excluded as they are often granted concurrently.

$$_=\alpha +__+\sum_^__+___)+\sum_^(__)+_+_$$

(3)

Sensitivity analysis was conducted using two-step fixed-effects, rather than random-effects, regression models. First, we constructed a fixed-effects panel regression including all time-varying variables. Based on this model, we predicted log-prices at launch. Thereafter, we conducted an ordinary least squares (OLS) regression including all time-invariant variables on predicted log-prices at launch.

Coherent with previous studies, we examined the cancer drug market within drug classes based on the correlation of prices [24, 29,30,31, 40]. The relationship between nine drug classes with a total of 25 injectable cancer agents was analyzed and visualized using a Pearson correlation matrix. Stigler and Sherwin (1985) suggest that price movements between two products can be used to define the extent of a market [41]. A positive correlation coefficient close to 1 indicates that two products are competing in the same market, whereas a low correlation coefficient suggests that the two products are competing in separate markets [42]. We tested for causality between each drug pair’s logarithmic first difference of prices using the Granger causality test [43].

Data were stored in Microsoft Excel (Microsoft Corporation, Redmond, WA, USA) and were analyzed using Stata software, version 14.2 (StataCorp LLC, College Station, TX, USA). Two-tailed p-values <0.05 were considered significant. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines where applicable [44].