As U.S. drug-pricing policy changes accelerate, policymakers need reliable ways to understand how reduced pharmaceutical revenues might affect R&D incentives, but the Congressional Budget Office’s (CBO) revised simulation model of new drug development remains too simplified and insufficiently validated to guide real-world decisions.
Summary
As the U.S. pricing policy landscape is evolving rapidly, introducing unprecedented changes to how pharmaceuticals are reimbursed and accessed, it is increasingly important to understand the consequences of these reforms on research and development (R&D) incentives. The Congressional Budget Office (CBO)’s model of new drug development is designed to help Congress assess how legislation that affects pharmaceutical industry revenues may influence the development of new drugs, alongside CBO’s estimates of the associated budgetary impacts.
In 2019, the CBO produced a high-level estimate of HR3’s (a bill that proposed a sweeping drug price setting policy) impact on drug innovation using a top-down approach relying on broad historical relationships between industry revenues and new drug development. In 2021, CBO changed their approach to a more detailed bottom-up economic simulation model, which they applied to the Build Back Better Act in 2021 and later the Inflation Reduction Act in 2022. The latest revisions to the CBO model have been shared through published presentation decks in late 2024 and early 2025.
While CBO has continued to make updates and improvements to its model, the model’s insights for informing policy decisions remain limited and unsupported. This demonstrates an important need to continue to conduct research to better understand drug development, but it also highlights that such an exercise should have limited utility in informing real world policy decisions. In this paper we review the most recent published methodology for CBO’s model and identify key limitations.
We have previously critiqued CBO’s drug development model. CBO has since made several updates to the model: (1) accounting for certain drug characteristics; (2) allowing expected revenues and costs to be correlated across development phases, and (3) using new data sources to estimate lifetime drug revenues. Despite these refinements, significant limitations remain. The model:
- Relies on a narrow conception of biopharmaceutical innovation that continues to exclude post-approval innovation;
- Incorrectly assumes that there is a single, homogeneous company representative of the entire industry;
- Assumes, unrealistically, that the number of drugs coming to market without the policy would be constant over time;
- Assumes that pharmaceutical companies make drug development decisions based on an oversimplified decision rule;
- Assumes product-level decisions, whereas decisions are often made at the portfolio-level, particularly in the early stages; and
- Produces headline estimates that are only averages and mask significant uncertainty.
As a result, the model remains a highly simplified and unrealistic representation of the pharmaceutical sector. The model’s headline estimates may understate the true impact of policies on innovation, while at the same time considerable uncertainty remains in both inputs and the model’s predictions. We therefore propose that CBO’s drug development model should be viewed as an academic exercise and should not be relied upon to inform policy decisions until much more refinement and validation has occurred.
What is the CBO simulation model of new drug development?
The CBO’s simulation model parameterises a structural model of decision-making in drug development as an alternative to directly using elasticity estimates from the literature that link changes in revenue to changes in innovation output. The model uses estimates of changes in expected future profits or development costs to estimate the percent change in the number of drug candidates entering the three stages of human clinical trials. By examining changes in decisions to enter at each stage, the model estimates when and by how much the number of new drugs entering the market will change in response to changes in anticipated revenue.
The model is based on a stylized representation of the pharmaceutical decision-making process. The approach is mathematically and computational sophisticated but requires significant simplification of the pharmaceutical decision-making process including the assumption of a single firm, which represents the entire marketplace. This firm is projected to continue development of a drug if expected returns exceed expected costs. Two mechanisms determine the observed change in the overall number of drugs developed. First is an immediate change in whether a potential drug candidate will enter one of the three phases of development based on its expected returns exceeding expected costs. Second are changes to the candidates available to enter a phase of development given changes to earlier phases of development. For example, if the model indicates that an illustrative Drug Candidate A fails to enter Phase II of development due to expected costs exceeding expected returns, Candidate A will then not be available for consideration of entering Phase III.
The model’s parameter values are derived from both estimation and calibration procedures. The model of the firm’s decision problem has several key inputs: estimates of success probabilities, expected development costs, expected time in development, financing costs, and expected returns. As CBO doesn’t observe all those values, the agency calibrates some parameters of the model and estimates others with restrictive parameterisations. Further, because CBO cannot observe the underlying joint distribution of expected costs and expected returns, they simulate it. The firm decides whether to enter simulated drugs into development. Simulated expected costs and expected returns for drugs that go into development are then compared with those that are observed.
CBO’s revisions to their simulation model
We published a critique of this simulation model in 2022, warning that policymakers should exercise caution in relying on the model’s findings for evaluating potential real-world policy changes. CBO subsequently made a number of revisions to the simulation model. At present there is no full, updated public working paper and associated model output to fully examine or determine the impact of these revisions. However, based on presentation decks shared in late 2024 and early 2025, there appear to be three major changes which only partially address criticisms of their original simulation model:
The model now takes into consideration certain drug characteristics.
The original simulation model did not reflect differences in drugs’ characteristics, producing a single aggregate estimate for all drugs. In the revised model, in deciding whether to move a drug candidate forward, the firm also considers select drug characteristics (including the number of competitors in the class, the route of administration, and whether the drug is a biologic), since these features can impact expected costs and returns.
Without further information it is difficult to know how this impacts the model results, how much data are missing, and whether this is missing at random or biased towards particular characteristics. While some (limited) drug characteristics are accounted for, CBO did not make any adjustments for variation in pharmaceutical company characteristics, most notably firm size, which influences cost of capital, a major source of cost for long-term pharmaceutical investments.
The model now allows expected revenues and costs to be correlated across phases of development
The original model unrealistically assumed that signals about a drug candidate’s likelihood of success were independent across different phases of development.
The new model jointly estimates the firm’s decision-making at each stage of drug development, allowing for signals of a drug’s revenues and costs to be correlated across decision points. However, the CBO treats the “successful completion” outcome of drug development as exogenous; that is, once the firm decides to enter a phase of development, the firm is assumed not to be able to control whether that phase is successfully completed.
In addition, CBO predicts this exogenous success rate using a regression on a limited set of observed characteristics of a drug: therapeutic class (cancer, infection, and cardiovascular); whether it is a biologic; and route of administration (oral, and IV). They found that oral drugs, IV drugs and drugs to treat infection have higher probabilities of entry in all phases. While we will need to wait for the full paper to determine whether the success rates for this subset of therapeutic classes are reasonable based on other studies and industry opinion, it seems clear that the model may not adequately predict variation in the many disease areas omitted.
The model uses new sources of data to seed the model
Lifetime drug-level revenues are estimated by summing annual revenue data (from SSR Health). Where missing, they are predicted based on the limited set of a drug’s observed characteristics (e.g. therapeutic class and route of administration) and years since launch based on methodology from Wouters, McKee & Luyten (2020). Some details around the data are unclear, for example how extensive the missing data are. Of note, over the period covered in that study (2009-2018) 355 new drugs were approved but R&D expenditure was available for only 69 (18%). It is not at all clear why we should consider this a representative sample. Furthermore, it is claimed that “Data were mainly accessible for smaller firms, orphan drugs, products in certain therapeutic areas, first-in-class drugs, therapeutic agents that received accelerated approval, and products approved between 2014 and 2018.”. In addition, “Restricted availability of data in the public domain” was noted to be important, as well as the underlying assumptions in the cost calculations. Finally, there were two erratum indicating errors in the source data that need to be accounted for.
What remain the biggest problems with the CBO model?
It is difficult to determine the impact of these updates on the model, given that it has not been fully made public. However, despite these updates, at least six key limitations remain:
1. CBO relies on a narrow conception of biopharmaceutical innovation that continues to exclude post-approval innovation
The CBO analysis gives an incomplete picture of the losses associated with reduced biopharmaceutical innovation. The only measure of innovation that CBO considers is the annual number of new drugs coming to market. This is problematic for two reasons. First, it ignores indication expansion of existing drugs, and second, it may be a poor proxy for health impact.
The CBO model ignores the impact of changes in revenue on post-approval R&D that leads to new indication approvals of existing drugs, which represents a critical source of treatment advances for patients, particularly in areas such as oncology and rare disease. Considering the dominance of multi-indication drugs in some disease areas (around 65% in oncology), and the clinical significance of those extensions (with the majority targeting earlier stages of disease), this omission is significant.
Further, drugs do not all have the same health impact. A new curative therapy for a rare disease such as spinal muscular atrophy (SMA) may provide enormous health gains for each patient while another innovative treatment may help a much larger number of patients by preventing unnecessary heart attacks and hospital stays. A complete evaluation of any policy that produces declines in biopharmaceutical R&D must include estimates of the impact on the advances in science and human health. CBO explicitly notes that it is beyond the scope of their analysis to consider which types of drugs might be affected by a policy change and how the reduction in new drugs will affect health outcomes. However, without any analysis or consideration of the types of drugs or indication expansions that are likely to be lost, it is impossible to say which groups of patients stand to lose the most, or even whether the model’s output (number of new drug approvals) is a meaningful proxy for health impacts.
2. The CBO model incorrectly assumes that there is a single, homogeneous company representative of the entire industry
This mischaracterizes the complex and diverse innovation and investment ecosystem that exists today (see our HAS article for a characterisation of this ecosystem). Pharmaceutical companies differ in characteristics such as size, which affects capital costs, and other factors relevant for R&D decision-making.
3. CBO assumes, unrealistically, that the number of drugs coming to market without the policy would be constant over time
CBO measures the impact of drug pricing policies on innovation by the change in the number of new drugs coming to the market each year and assumes that this number would stay constant over time in the absence of the policy. However, there has been an upward trend in annual U.S. Food and Drug Administration (FDA) Center for Drug Evaluation and Research (CDER) approvals since at least 2005, and experts interviewed by OHE have suggested that developments such as genomics will contribute to an increasing rate of new molecular entity (NME) approvals.
Furthermore, the time-constant baseline that CBO assumes is 44 new drugs per year, set equal to the average number of annual CDER approvals in the period 2015-19. While this average includes new biologics license application (BLA) approvals by the CDER, it does not include approvals by the FDA’s Center for Biologics Evaluation and Research (CBER) for products such as vaccines and cell and gene therapies
It is critical that CBO’s assumed baseline reflects reality. If new drug approvals are expected to increase over time in the absence of the policy, as the evidence suggests, then erroneously assuming a constant baseline will underestimate the absolute numbers of medicines being lost to a modelled policy. A 10% drop from the 2011 level of 30 CDER approvals is equivalent to 3 drugs, but a 10% drop from the 2018 level of 59 is almost double that.
4. CBO assumes that pharmaceutical companies make drug development decisions based on an oversimplified decision rule
The revised CBO model now accounts for the relationship between phases of clinical development, but there are still significant shortcomings in how decisions are modelled. In CBO’s modelling, a drug candidate will advance to the next phase of development as long as expected returns exceed expected costs for that individual drug. This means that the level of expected net returns does not influence the probability of the drug being progressed, conditional on the returns being positive. The policy could lower expected net returns to almost zero but as long as they remain above zero, it will have no impact on the probability of the drug being progressed into the next stage of development.
In reality, drug developers often require returns to be some positive multiple of costs to make the high risk of investment worthwhile, suggesting that CBO underestimates the impact of the policy on development decisions. For example, in CBO’s model, two drugs with expected net returns of 0.01% and 15% are equally likely to advance into the next stage of clinical development. Yet, a positive return of 0.1% is hardly attractive for investment when the stock market is earning 7% or when a new green energy investment could garner a 10% return.
5. The model assumes product-level decisions, whereas decisions are often made at the portfolio-level, particularly in the early stages
A further limitation of the simple decision rule used in the modelling is that it abstracts from how pharmaceutical R&D decisions are made in practice. Firms rarely assess the viability of individual drug candidates in isolation; instead, development choices are typically taken with reference to the company’s wider R&D portfolio and, indirectly, the portfolios of competitors. Risk and return are actively managed across portfolios by spreading investment across therapeutic areas, molecule types, and modes of action, and some products form part of broader platform technologies rather than standalone assets. The model also omits the substantial role of non-pharmaceutical investors, most notably venture capital, in financing early-stage biomedical innovation – an issue we examine in more detail in our recent article in Health Affairs Scholar.
These simplifications matter because reductions in expected revenue are unlikely to lead to a proportional reduction in R&D across all programmes. Instead, firms and investors may respond by rebalancing portfolios towards projects with lower development risk and more predictable returns. For example, a company facing tighter revenue expectations may choose to prioritise late-stage assets in common chronic conditions, where clinical pathways and reimbursement outcomes are well understood, over early-stage programmes in areas such as neurodegenerative disease or rare cancers, where scientific uncertainty, trial failure rates, and development costs are higher but the potential health gains are substantial. As a result, the disincentive effects of reduced revenues may fall disproportionately on high-risk, high-return areas that typically correspond to the greatest unmet clinical need. This spillover effect is not captured in the unrealistic model which assumes investment decisions are made at the product level and in isolation.
6. Headline estimates that can be generated from such a model are only averages and mask significant uncertainty
There is significant uncertainty around both the values of key inputs used in the model (which are calibrated or estimated) and around the projected impact of the policy. For example, based on applications of previous versions of the model we find that even small differences in key inputs, such as success rates in early clinical development, can have relatively large downstream effects and therefore meaningfully impact the number of new drugs coming to market (Cookson and Hitch, 2022). Policymakers should therefore exercise caution in relying on CBO’s headline point estimates, given the uncertainty in their estimation. Future work should test the sensitivity of estimates to changes in assumptions, with ranges to reflect the level of uncertainty.
What’s next for the CBO model
The CBO outlined short-term steps to refine its drug development model, including incorporating expected competition within therapeutic classes, comparing results from the revised November 2024 model with earlier versions, and running new simulations to assess the effects of other policies on drug development (such as the impact of NIH cuts and changes in FDA review times, for which initial results were released in July 2025). Looking longer term, the CBO has called for further research to address two major gaps, which we identified in our final point above: first, the lack of indication-level analysis, which limits understanding of how policies affect firms’ decisions about pursuing and sequencing new indications; and second, the absence of health outcomes modelling, with the CBO acknowledging that its current approach does not capture the value of innovation or how policy-driven changes in drug development affect the mix, effectiveness, and availability of treatments, patient outcomes, and broader population health.
Conclusion
The most recent updates to CBO’s innovation simulation model have partially addressed some of the concerns that have been raised with the model. However, the model remains an oversimplification that does not adequately represent the reality of biopharmaceutical investing, leading to inaccurate predictions of the impact on innovation of new policies. Moreover, CBO’s own calls for additional research raise questions as to whether the model’s outputs (i.e., number of new drug approvals) are a reasonable proxy for outcomes needed by policymakers, such as the size and nature of changes to health outcomes.
Our previous conclusion therefore holds: the simulation model may be of academic interest, but it cannot be reliably used to inform policymaking, at least not without significant qualifying information. This includes a better accounting of the uncertainty around any point estimates derived from the model and a stronger acknowledgment of the possible inadequacy of the model’s current outputs in capturing the magnitude and nature of policies’ impacts on human health.
