The economic case for scaling up health research and development: Lessons from the COVID-19 pandemic

Significance

The COVID-19 pandemic showed that a public research and development (R&D), push can accelerate discovery, yielding enormous health and economic benefits. To explore whether this feat could be repeated, we develop a model of R&D and apply it to Streptococcus A and Alzheimer’s disease. We find spending US$34 billion is optimal to develop vaccines against Streptococcus A. This spending is expected to generate US$2 trillion in benefits, corresponding to a return on investment of 23% per year for 30 y. For Alzheimer’s disease, we find optimal spending of US$5.7 trillion. The corresponding benefits are US$61 trillion with a return of 12.1%. We discuss policies to increase health R&D spending with the potential to unlock these massive global health benefits.

Abstract

In response to the COVID-19 pandemic, governments directly funded vaccine research and development (R&D), quickly leading to multiple effective vaccines and resulting in enormous health and economic benefits to society. We develop a simple economic model showing this feat could potentially be repeated for other health challenges. Based on inputs from the economic and medical literatures, the model yields estimates of optimal R&D spending on treatments and vaccines for known diseases. Taking a global and societal perspective, we estimate the social benefits of such spending and a corresponding rate of return. Applications to Streptococcus A vaccines and Alzheimer’s disease treatments demonstrate the potential of enhanced research and development funding to unlock massive global health and health-related benefits. We estimate that these benefits range from 2 to 60 trillion (2020 US$) and that the corresponding rates of return on R&D spending range from 12% to 23% per year for 30 y. We discuss the current shortfall in R&D spending and public policies that can move current funding closer to the optimal level.

The private sector is generally the dominant investor in disease prevention and treatment measures such as vaccines and pharmaceutical drugs. Its evaluations of health research and development (R&D) investments are usually based on expected commercial benefits and not the total benefits that would accrue to society at large. Insofar as vaccine and drug prices fall short of the incremental contributions to societal value, there is concern we are underinvesting in the discovery of new technologies for the protection and promotion of health. However, the COVID-19 pandemic demonstrated the enormous value of a different model with strong public funding of R&D and a guarantee of the market for final products. The billions of US$ spent developing COVID-19 vaccines created US$ trillions in value in the United States alone (1, 2). Moreover, COVID-19 vaccine development broke the vaccine development mold (3). Governments derisked private sector investment in development and production with large quantities of public funds. Cooperation between regulatory agencies and the private sector allowed parallel execution of clinical trials, compressing the typical 10-y vaccine development and regulatory process to 10 mo. Asking whether this feat can be repeated is natural. Could increased public funding and related political will yield innovative vaccines or therapies for other diseases that unlock large benefits?

To address this question, we develop a simple model of optimal R&D spending on new approaches to specific known diseases. The model provides insight into optimal global spending amounts, the amount of future harm from the disease that this spending could eliminate, and possible rates of return on R&D investment for vaccines and treatments. We apply the model first to COVID-19, finding that actual R&D spending was close to optimal and had very large rates of return. Encouraged by our model’s alignment with COVID-19 funding, we then examine the model’s implications for vaccines to prevent streptococcus A (Strep A) infection and for drugs to treat Alzheimer’s disease and related dementias (ADRDs). We find that optimal R&D spending on each is large and that the corresponding expected rates of return are substantial, well exceeding estimated rates of return for other productive uses of public funds.

Results

COVID-19.

Table 1 reports optimal R&D spending amounts. For more details on how we obtain these figures, please see Methods and SI Appendix. Our results imply that optimal spending on COVID-19 vaccines was US$9.84 billion, which would have funded 41 projects. (All calculated funding amounts are total amounts. For budgetary reasons spreading the spending over multiple years may be desirable.) Next, rates of return (RoR) for COVID-19 are calculated assuming a 2-y delay before harm reduction begins and assuming harm reduction is spread over a 10 y period. The rate of return on COVID-19 spending is estimated at 613% per year. If the probability that a vaccine project succeeds were lower, at 18%, funding 55 projects at a cost of US$13.2 billion would be optimal. In contrast, if a project had a 33% chance of succeeding, then optimal spending would fall to US$7.2 billion. If vaccines were less effective on average, reducing only 66% of COVID-19 harm, then optimal spending would rise to US$12.5 billion, funding 52 projects, and the RoR would fall to 561% per year. However, if successful vaccines reduced harm by 95%, then optimal spending would be US$8.6 billion, and the RoR would rise to 642% per year. Finally, if four approaches are needed to eliminate all the harm from COVID-19, then spending US$34.6 billion and funding 144 projects is optimal. (With more approaches, each project can address a smaller fraction of disease harm. Therefore, more projects need to be funded to get to the point where enough harm is diminished that funding additional projects is no longer cost-beneficial.). Fig. 1 plots the rate of return (right axis) of COVID-19 R&D as a function of possible probabilities of success. We find that the returns range from about 200% with a 1% probability of success to more than 800% with a 50% probability of success.

Table 1.

Optimal R&D funding and social rates of return for COVID-19 vaccines, Group A Strep vaccines, and ADRDs treatments

	Projects funded	Optimal spending (US$)	Social surplus (US$)	Internal rate of return
COVID-19
Baseline calibration
	41	9.84 billion	30.6 trillion	613%
Parameter sensitivity
Success probability = 18%	55	13.2 billion	30.6 trillion	550%
Success probability = 33%	30	7.2 billion	30.6 trillion	687%
Harm reduction = 95%	36	8.6 billion	30.6 trillion	642%
Harm reduction = 66%	52	12.5 billion	30.6 trillion	561%
Approaches = 4	144	34.6 billion	30.6 trillion	382%
Group A Streptococcus
Baseline calibration
	226	33.9 billion	1.85 trillion	23%
Parameter sensitivity
Harm reduction = 70%	108	16.2 billion	1.87 trillion	29.4%
Success probability = 5%	554	83.1 billion	1.79 trillion	16.4%
Total Strep A harm 2×	252	37.8 billion	3.74 trillion	28.1%
Approaches = 4	396	59.4 billion	1.82 trillion	18.8%
ADRDs
Baseline calibration
	49,083	5.7 trillion	61 trillion	12.1%
Parameter sensitivity
Harm = US$200 trillion	61,557	7.1 trillion	171 trillion	17.1%
Approaches = 11	118,888	13.7 trillion	49 trillion	6.8%
Success probability = 1%	74,999	8.7 trillion	56 trillion	9.4%
Success probability = 3.4%	32,874	3.8 trillion	63 trillion	14.7%
Harm reduction = 11.9%	7,074	0.8 trillion	67 trillion	26.5%
Harm reduction = 21.24%	4,335	0.5 trillion	67 trillion	30.8%
Harm reduction = 1.97%	30,525	3.5 trillion	64 trillion	15.2%
Approaches = 11, Success probability = 1%	172,601	19.9 trillion	37 trillion	4.5%

Note: For Strep A and ADRDs, internal rate of return is calculated assuming a 10 y delay before harm reduction begins and assuming harm reduction is spread out evenly over 30 y. For COVID-19, we use a 2-y delay and assume harm reduction is spread over a 10 y period. All monetary values are in 2020 US$.

Fig. 1.

Internal rate of return vs. probability of success note: Solid line—Strep A, dotted Line—Alzheimer’s, dashed line—COVID-19. COVID-19 R&D returns on the right axis.

Our calculations suggest that global spending on COVID-19 vaccine development was near optimal. Estimates of public funding total US$9.2 billion with US$5.5 billion to US$6.6 billion for vaccines (4, 5). However, at 126, the world may have funded too many distinct vaccine projects. That funding choice could have been optimal, however, if the expected probability of success were lower than our calibration, or uncertainty about which approaches would work required diversification across approaches. In either case, under optimal or actual COVID-19 vaccine R&D funding, the rate of return on this R&D was enormous as these vaccines averted a tremendous number of deaths and a large amount of lost economic output. Because our calculations come close to matching actual COVID-19 R&D spending, we take these results as supportive evidence for the accuracy of our model and apply the model to two additional diseases with greater confidence.

Other authors argue that COVID-19 pandemic spending was inadequate (6, 7). Our results differ from these analyses as we focus solely on R&D costs. These authors argue that pandemic spending was inadequate because funding manufacturing capacity in advance for multiple vaccine candidates even before knowing whether these candidates would be approved (6) and supporting supply chain resilience and health system strengthening to prepare for the next pandemic (7) would have been optimal. We do not disagree with these authors. This additional spending is likely optimal to prepare for and respond to a pandemic. However, our goal in this paper is not to calculate total optimal pandemic spending but to analyze optimal R&D spending for an endemic disease and to analyze the extent to which the COVID-19 pandemic can shed light on this issue.

Group A Streptococcus.

Optimal funding is US$33.9 billion, enough to fund 226 R&D projects. [We have previously published similar results for Strep A (8).] This investment generates US$1.85 trillion in benefits to society (social surplus). The rate of return on this investment is 23% per year for 30 y. For Strep A and ADRDs, we calculate rates of return using a 10 y delay and assume harm reduction is spread out over 30 y. It is important to note that our optimal spending number will likely result in multiple successful vaccines or treatments, with subsequent successes improving on past successes by, for example, increasing efficacy or treating different strains in different populations. In this sense, optimal spending should be thought as spending to exhaust the value of Strep A vaccine R&D, not spending needed to guarantee at least one vaccine with high probability.

If a vaccine were to reduce harm by 70% (instead of 30%), optimal spending falls to US$16.2 billion, benefits change very little, and the RoR on this investment rises to 29.4%. When R&D projects are assumed to be less successful (5% chance versus 15%), more projects should be funded because more projects are needed to reduce expected harm to the point where future projects are no longer beneficial. RoR, in this low success probability scenario, is 16.4%. Fig. 1 plots the RoR as a function of the possible probabilities of success. We find that the RoR is less than 10% if the probability of success is only 1% but more than 30% when the probability of success approaches 50%.

Underestimating harm from Strep A implies underestimating optimal spending. We examine sensitivity to doubling Strep A harm. Here, optimal spending rises; however, it rises by a factor smaller than two. Under this scenario, benefits to society double. The return to investment rises to 28.1%.

Finally, if four approaches are needed to address all harm associated with Strep A, optimal spending rises to US$60 billion. Social surplus differs little from the baseline case. The return to investment falls but remains large at 18.8% per year.

Current vaccine R&D efforts are approximately 10% of what we find to be optimal. A recent survey finds only 23 distinct projects currently in progress (9).

Alzheimer’s Disease.

For ADRDs R&D, optimal spending is US$5.7 trillion, which, while very large on its face, pales compared with the estimated total harm of ADRDs (US$76 trillion). Funding of this magnitude can support 49,083 individual projects to find an effective treatment. We assume the number of possible projects to fund is unlimited and the supranational organization has access to the optimal level of funding. However, if the number of projects to fund or financial resources do not exceed the optimal amount, then funding all available projects and spending the organization’s entire budget would be optimal.

Our finding that optimal spending is much larger for ADRDs than for Strep A is due to three factors: the (much) larger estimated harm from ADRDs, the lower success probability of ADRDs R&D, and the smaller fraction of harm each success alleviates. This level of spending would prevent the bulk of projected ADRDs harm, leading to social benefits of US$61 trillion. These benefits yield an expected return of 12.1% per year.

We examine the sensitivity of our results to varying parameters from our baseline calibration. If true harm caused by ADRDs is valued at US$200 trillion, optimal spending would be US$7.1 trillion and the rate of return on this investment rises to 17.1%. Similarly, if using 11 approaches to treat all ADRDs harm is required, spending rises to US$13.7 trillion with an RoR of 6.8%.

Under a lower success probability (1% vs. 2%), harm would fall more slowly; therefore, funding more projects is optimal, and optimal funding increases to US$8.7 trillion with an RoR of 9.4%. If instead each project has a higher probability of success at 3.4%, then optimal spending is US$3.8 trillion dollars with an RoR of 14.7%. Fig. 1 plots the RoR as a function of the probability of success. We find that returns are relatively small, around 10% when the probability of success is only 1% but grow to more than 30% when the probability of success is close to 50%.

As the calibration section explains, harm reduction depends on the survival probability and how long we assume an effective treatment would push back ADRDs onset. In a large push-back (3 y), high survival probability scenario, the fraction of harm reduction is 11.96% and optimal spending on ADRDs R&D is US$816 billion with an RoR of 26.5% per year. In a large push-back, low survival probability scenario, the fraction of harm reduced per success is 21.3%, optimal spending is US$500 billion, and the corresponding rate of return is 30.8%. In the low push-back, low survival probability calibration, optimal spending is US$3.5 trillion with an RoR of 15.2%.

Finally, when 11 approaches are needed to address ADRDs, with each project having only a 1% success probability, optimal spending rises to US$19.9 trillion as funding a very large number of projects becomes necessary.

Our results imply that the current level of Alzheimer’s disease R&D funding is woefully inadequate. Since 1995, total private expenditures on R&D at the clinical stage were estimated to be US$42.5 billion (10). Moreover, the US National Institute for Health budget for Alzheimer’s research is only $1.8 billion annually (11).

Discussion

Optimal spending on R&D for vaccines and treatments is very large. These values are estimated to be on the order of a trillion US$ for ADRDs and on the order of tens of billions of US$ for Strep A. However, and more importantly, the benefits are orders of magnitude larger. In the case of Strep A R&D, benefits range from US$1.8 trillion to US$3.7 trillion. Returns on investment range from 19% to 29% per year for 30 y. Moreover, as the results section describes, current spending falls substantially short of optimal spending. The current level of Strep A vaccine R&D is about 10% of what we estimate to be optimal; for Alzheimer’s disease research, the annual US NIH budget is only 0.03% of our estimate of the optimal total amount.

We note that our general conclusion, that health R&D has very high potential returns and that globally we appear to be underfunding this R&D, holds here for diseases that differ greatly in terms of attributes and treatment approaches. Consequently, our conclusions are likely quite general. In fact, one intriguing use of our model would be to calculate optimal R&D for the prevention and treatment for many different diseases and find those diseases with the largest gap between actual and optimal spending. Such an analysis is beyond the scope of the current article. However, we underscore that our model does not make the case for public spending for all health R&D. For example, our analysis would not justify spending on a disease with very little prospect for effective treatment or low population harm (though it may be argued for on ethical grounds).

Returns to the health R&D projects that we analyze are large even compared with other social interventions that receive considerable support. For example, estimates of the return to more years of education range from 9% to 10% per year (12, 13). Our results call upon policymakers to promote expanded funding for the development of Strep A vaccines and ADRDs treatments, mirroring calls to increase funding for pandemic preparation (14) and the public push to defeat COVID-19.

In our model, a single supranational organization carries out funding. This is a useful device—what economists call the social planner’s problem—to find the optimal allocation of resources. However, in practice, various actors make R&D decisions: pharmaceutical companies, national health institutes and ministries, and private investors and foundations. Typically, the allocation of resources from the decentralized decisions of these various actors would not reach the optimal allocation of resources. Therefore, public policy is needed to move society closer to the optimum. We now discuss reasons why the current situation is unlikely to reach the optimum and some available funding mechanisms to increase health R&D. We acknowledge the “sticker shock” that some of our optimal spending numbers may evoke. We do not imagine that our proposed financing mechanisms will get us all the way to the optimal levels; we simply point out that starting from a suboptimal level, an increase in R&D spending (reasonably allocated) will increase societal welfare.

Economic analyses suggest that private sector development of treatments for health conditions will fall short. First, some valuable research and development, like basic research, is hard to patent and therefore is unlikely to provide an adequate return on investment (8). Examples include developing an animal model that can better predict the effects of candidate ADRDs treatments and cataloging the Strep A variants that circulate the globe. Second, given that an individual R&D project is quite risky, the private sector is likely to find that the resulting returns are insufficiently large to justify the high levels of risk. Even in our most optimistic scenarios, a typical project will fail more than 80% of the time, resulting in total loss of investment. In contrast, the large portfolio of many projects we consider here greatly reduces the risk of not developing a single viable product (8).

To fill this development gap, the public sector has multiple potential policy options. The most basic is direct funding of vaccine R&D projects. (Nongovernmental donors may also have a role. However, given the scale of funding, government involvement will likely be required to reach optimal levels.) On a large scale, this approach reduces the risk of R&D by diversifying across many projects. To raise funds for this investment, governments can increase taxes, reduce other spending, or use debt finance. In terms of reallocating spending, we note the proposed, 2024 US defense budget is US$842 billion (15) and health spending in the United States alone was US$4.5 trillion (16) in 2022. Utilizing some of this spending for R&D for preventative health measures seems prudent. Debt finance, when possible, allows for a better temporal alignment of project costs and benefits as any health R&D project is likely to return benefits years in the future. Debt allows a government to borrow money now and pay it back later when benefits from the R&D begin. Moreover, advanced economies can currently borrow at interest rates substantially lower than our estimated annual returns to R&D investment (8). For example, since 2005, real (inflation-adjusted) interest rates have been below 2.5% in the United States (17). Nontraditional debt finance that conditions on outcomes of the R&D funding may be an additional way to supplement R&D funding (18).

Instead of direct funding, a government could support a large bond fund for private sector investment into health R&D, enabling private investors to pool investments into many R&D projects simultaneously (8, 19). The diversified fund would have lower risk than each individual project. Successful projects would yield profits that provide a return to bond holders. The government could guarantee the principal on such a bond investment or provide preferential tax treatment to promote such a fund. This approach reduces the risk of vaccine or drug development and provides a role for both public and private sectors.

Speeding up the regulatory process, while monitoring for adverse effects, would also increase R&D. The cost of capital—the private sector’s required return to invest in the pharmaceutical industry— is estimated conservatively at 8% (8, 20). Even at this lower bound, a 2-y regulatory delay implies that expected profits from an investment would have to be 16% higher to justify investment. Reducing the time from investment to market would increase the number of projects the private sector finds viable, raising R&D expenditures. To this point, research finds that a 3 mo regulatory delay leads (nontrivially) to one fewer drug developed per development category (21).

As our estimates are for global spending, international cooperation is needed to move the world substantially toward globally optimal levels. This is especially true as countries can “free ride” off the R&D efforts of others. In response to COVID-19, the international community has noted the importance of international cooperation on pandemic-related R&D. The World Bank Pandemic Fund and the Coalition for Epidemic Preparedness Innovations’ “100 Days Mission” are initiatives to increase international cooperation for health R&D. Our results support looking beyond pandemics, to identify the endemic diseases that should be treated with similar urgency.

We have found the scale of current health R&D to be suboptimal, despite the strong case our results make for supporting such R&D. Indeed, there have been prominent calls for the government to take a more active role in drug development (22). Yet, the private sector remains the dominant source of funding to bring new drugs to market. Various potential reasons explain this state of affairs. First, politicians may be reluctant to support such a large change to the status quo. This reluctance is known as loss aversion or status quo bias in politics (23). Lobbying by pharmaceutical companies may also limit the expansion of the public sector into drug development (24). Finally, given the presence of asymmetric information in drug discovery (25), i.e., researchers knowing more about the likelihood their projects will succeed than the government, some form of private capital may be advantageous to screen for the most promising projects.

Given these limitations on the ability of governments to achieve optimal R&D, we see a strong role for public–private partnerships in enhancing global R&D. A recent example is the successful development of a malaria vaccine resulting from a partnership among PATH, the Bill & Melinda Gates Foundation, and GSK (26). We hope that this success will encourage nonprofits to fund health R&D through regulatory approval. This route is especially important to address diseases with burdens mostly in lower income countries. Treatments for these diseases are unlikely to attract much private R&D capital. Advance market commitments (27) may help to fill in this gap along with increased donations to GAVI, especially for initiatives to promote vaccine R&D in addition to vaccine delivery.

Our work has limitations. While we calculate social surplus, we do not analyze the pricing of treatments. Thus, we cannot predict how the surplus will be divided between reduced harm to patients and profits to manufacturers. This issue is important because if treatments are not affordable, or funds are not available to make them affordable, they are unlikely to reach all who need them, resulting in a smaller-than-predicted reduction in disease harm. Relatedly, our model could overstate spending in cases where scale-up of existing treatment options is a more cost-effective strategy. We note, however, the paucity of alternative treatments for ADRDs and the mixed evidence on their cost effectiveness (28, 29). In the case of Strep A, if increased access to penicillin could more cheaply alleviate harm than developing vaccines, optimal R&D spending on vaccine development would be lower. Studies have shown benefit–cost ratios of almost 40 for antibiotic scale-up (30). However, our estimates imply even larger benefit–cost ratios for a Strep A vaccine. Additionally, we note that Strep A harm in our model is calibrated using the experience of Australia, a country where antibiotics are widely available, implying that scaling up penicillin availability alone is insufficient to address Strep A harm.

Additionally, our model of optimal R&D investment abstracts from some more realistic aspects of the R&D process. One important departure is a dynamic aspect, where one can choose additional funding after observing past successes and failures. While we are confident that our main conclusion—that the returns to health R&D are large—is robust to adding dynamic features, some implications of a dynamic model would differ from our static model.

For one, funding decisions would be conditioned on past outcomes. Optimal funding would be reduced when projects succeed more than expected and would be increased when projects succeed less than expected. In the context of our static model, we view the optimal spending amounts as giving the correct spending levels on average. Another important feature of a dynamic model of R&D is the ability to learn from past successes or failures. One may want to increase funding to the most promising approaches or perhaps increase funding to the approaches that are found to be more difficult and therefore need more funding to succeed. In the case of developing a Strep A vaccine—where the baseline probability of success is quite high—we view learning as less important. With Alzheimer’s disease, given the difficulty of finding successful treatments, this learning effect could be quite important. Finally, we note that in the case of learning, time to successful treatment may be slower than our assumption of 10 y if spreading out funding over many years is important to allow for learning.

In our model, we consider funding R&D for one disease at a time. In reality, funding is allocated across multiple pathogens simultaneously. Modeling these features could have some potentially interesting implications. For example, one may want to fund approaches (e.g., mRNA vaccines) if these approaches were promising for multiple diseases. The cost of funding an R&D project may also be higher than simply the project cost as one should account for the opportunity cost of not funding projects for other diseases. In this case, our results indicate spending on overall health R&D is too small, but our disease-specific estimates of spending would be too large.

We omit the cost of manufacturing and delivery that would be necessary to realize these benefits. However, overall benefits of R&D spending are measured in trillions of dollars and therefore reasonable manufacturing and delivery costs would reduce measures of benefits and rates of return, but only slightly. Indeed, in other work, we found the cost of manufacturing and delivering a prospective Strep A vaccine to be between 2% and 3% of the total benefits of such a vaccine (8). However, our model would need to be modified for therapies with administration costs that are significant fractions of the costs of development (e.g., multimillion-dollar-per-case gene therapies).

We use market prices to value disease harm. In this case, the same health outcome will be costlier in a high-income country than in a low-income country. This assumption is relevant from the point of view of a national policymaker deciding how much to spend on treating a disease within the country, but we acknowledge the ethical concerns of attributing higher costs to the same health outcome based on national context (8). Assigning equal dollar values to health outcomes irrespective of income could address this issue and would likely increase the value of optimal R&D for diseases with large burdens in low-income countries. Additionally, we measure harm alleviated by the treatment as harm no longer experienced from the specific disease but we may potentially overestimate the benefits of treatment when being cured of one disease makes someone more likely to experience another.

We note that while we have calibrated the success probability of a health R&D project based on prior data and on expert opinion and provide sensitivity analysis of our assumptions, the true probability of success is unknown and depends on the likelihood of engineering and scientific advances.

While our optimal spending numbers initially appear large, the correct lens by which to view them is relative to their expected benefits. Indeed, one reviewer likens our result regarding the scale-up of health R&D to the Manhattan project in terms of scope and urgency. The COVID-19 vaccine R&D effort was a similarly extraordinary situation—but not from the point of view of achieving globally optimal R&D levels. The pandemic demonstrated that a large public investment in health technologies can unlock enormous net benefits and yield extremely high returns. The tools are now available to extend this knowledge to many illnesses and pathogens that cause substantial illness across the globe and to end the world’s potentially substantial underinvestment in health-promoting technologies.

Methods

In this paper, we generalize a framework applied to COVID-19 vaccines (31) which we have also previously applied to Strep A vaccines (8). We focus here on vaccines and drug treatments, but the model is general and could apply to other medical interventions like medical devices, behavior modification, or institutional reforms, e.g., universal health care.

In our model, a supranational organization considers a table of disease-specific R&D projects to fund: the table columns correspond to different approaches to reduce disease harm (e.g., for COVID-19 vaccines, mRNA and an inactivated virus could be different approaches). The entries under each column correspond to different projects under each approach (e.g., under the mRNA vaccine approach, the Moderna project, the Pfizer-BioNTech project, etc., could be the different projects). Projects are applied in the sense that they have the potential to lead to a treatment for a specific disease; we do not consider more basic science intended to have more diffuse applications. The organization’s goal is to choose projects to fund from the table to maximize the expected benefit of this funding minus its cost. The benefit of the funding is the amount of disease harm reduced, and the cost of the R&D funding is the number of funded projects times the (assumed) constant per project cost.

The organization is risk-neutral regarding outcomes. However, when funding a large number of projects is optimal, the variance in potential outcomes is small. We also assume the number of possible projects to fund is unlimited. As projects are ideas, no natural limit exists. However, if the optimal number of projects exceeds the total number of potential projects, then funding all potential projects would be optimal. Similarly, if optimal funding exceeds the amount of funding available to the organization, then investing all available funding would be optimal.

Simplified Innovation Model.

Because considering every funding combination is computationally infeasible, we must make some mathematically convenient assumptions to find the optimal solution.

• Approach Success Probability: With probability 1 − p_n, no project under approach n will ever produce a successful intervention or treatment. This probability of success is independent across approaches.

  • oThis assumption regarding the approaches allows for correlation in the failure probability of R&D projects. For example, if abnormal buildup of amyloid in the brain is not an underlying cause of Alzheimer’s disease, treatments that slow the buildup of amyloid cannot prevent disease progression. Similarly, having never produced a commercial vaccine before, assuming some chance that mRNA technology would not produce a successful vaccine seemed prudent in 2020.
  • oThe approaches create a correlation in the success probability of each project. Therefore, diversifying across approaches to reduce the overall risk of failure becomes desirable.

• Project Success Probability: Conditional on the approach succeeding, the probability of success of any project j under that approach is p_j|n, which is independent of the probability of success of any other project. This probability is the same for all projects under approach n. We view independence as a reasonable baseline assumption. Various choices in R&D projects that different scientists make will lead to success or failure independent of the choices of other scientists. However, allowing for correlation in the probability of success across projects under the same approach is mathematically feasible.

• Harm Addressed from Approaches: Each approach can address at most Δ_n of the total harm.

  • oThe Δ_n sum to less than or equal to 1.
  • oHarm is partitioned, i.e., approaches 2, …, N cannot address the first fraction Δ₁ of harm; approaches 1,3, …, N cannot address the second fraction Δ₂ of harm; and so on.
  • oOur assumption here, that the approaches partition harm, facilitates finding a solution to the model. However, we also believe it helps to capture some important properties of the diseases we analyze. For example, because ADRDs vary, multiple approaches will likely be needed to address all the harm they cause. Similarly, the large variety of Strep A strains suggests that different approaches will be needed to vaccinate against all prevalent strains. Finally, even with COVID-19, some vaccine approaches led to vaccines that were less expensive to manufacture and easier to distribute, suggesting that multiple approaches were helpful in ensuring that harm was alleviated in lower-income countries.

• Harm Addressed from Project Successes: Each successful project j, under approach n, reduces the remaining expected disease harm by a fraction δ_n. This fraction can vary by approach but is constant for all projects under approach n. Implicitly, we assume here that successes lack spillover benefits to other diseases and that benefits of successes are always net positive, i.e., adverse effects do not outweigh treatment benefits.

In our model, the organization pays a one-time R&D cost and realizes a one-time benefit. In practice, R&D decisions take place over time and benefits are realized over time. We incorporate this reality into our calculations in two ways. First, we estimate total disease harm for multiple years into the future and calculate total harm as the present discounted value of this harm. Second, to estimate rates of return, we assume that benefits from R&D investment begin several years after R&D costs are paid and that these benefits accrue evenly over multiple periods. The exact delay before benefits begin and the window of time over which benefits spread out vary by disease.

Model Solution.

The funding organization considers the available projects and calculates the expected benefit of funding each project (hereafter referred to as the benefit) as a product of the expected amount of harm remaining from the disease, the fraction of harm the new project’s success would alleviate, and the probability that the newly funded project will successfully produce an approved vaccine or drug treatment.

Mathematically, the expected benefit equals $p_n\ast H\ast {\mathrm{\Delta }}_n\ast E\left(1-{\mathrm{\Psi }}_n\right)$$\ast p_{\mathrm{j}|\mathrm{n}}\ast {\delta }_n$, where H is total harm from the disease and $E\left(1-{\mathrm{\Psi }}_n\right)=$ ${\sum }_{s=0}^S{\left(1-{\delta }_n\right)}^{s}P\left(s\right)$, where S is the number of projects of approach n that have already been funded and P(s) is the probability of getting s successes from a binomial distribution with success probability $p_{j|n}$ and number of Bernoulli trials that is equal to S.

The organization funds the highest-benefit project if the benefit of that project exceeds its cost. It repeats these calculations with the remaining projects. The organization then funds the next highest-benefit project if the benefit of that project exceeds its cost. Importantly, the benefit of funding a project falls as more projects are funded, because some of those projects are likely to succeed and reduce the remaining harm to address. The organization continues in this manner repeating these calculations until the benefit of funding the next project is less than its cost. At this point, we have found the optimal number of projects to fund, and by multiplying by the per-project cost, we calculate the optimal amount of funding.

Fig. 2 illustrates the model calibrated for Strep A. The orange line (MC, marginal cost) represents the cost of funding a project. It is constant and does not depend on the number of projects funded. By contrast, the blue line (MB, marginal benefit) represents the benefit of funding the next highest-benefit project given the number of previously funded projects. This line slopes downward because the more projects that have been funded the more likely a successful vaccine will be developed. As a result, we expect that less harm will remain from Strep A, thereby making an additional project less beneficial. While the MB line is above the MC line, the organization should continue to fund projects. The organization should continue funding projects until it reaches the 227th project (the point of intersection). At this point, harm from Strep A is expected to have been reduced to such a degree that funding additional projects is no longer cost beneficial (8).

Fig. 2.

Incremental costs and benefits vs. projects funded for Strep A R&D. Note: Figure y axis is a log scale.

Fig. 2 allows us to calculate social surplus (total expected benefits minus total cost) as the area of the triangle formed by the y axis, the MC line, and the MB line. We use this social surplus as an input into our calculations of R&D rates of return. (Note, these returns differ from commercial rates of return, which depend on the prices private companies charge.)

Based on Fig. 2, we can analyze the comparative statics of our model: how changes in the various parameters will affect optimal funding. First, all else being equal, the more harm caused by the disease the larger optimal funding will be. Note that as harm increases the MB curve shifts to the right and the optimal number of funded projects increases. However, a change in the success probability has an ambiguous effect on the optimal number of projects. On the one hand, a higher probability of success raises the benefit of a project and shifts out the MB line. On the other hand, each funded project is more likely to succeed and therefore the slope of the MB line becomes more negative. We therefore cannot from the diagram alone determine the effect on the optimal number of funded projects. Similarly, changing our parameter capturing the fraction of harm reduced from a success has an ambiguous effect on the optimal number of projects funded. Therefore, we rely on our model calibration and numerical calculations to determine the effect of changing these parameters.

Calibration.

COVID-19.

We calibrate global harm from COVID-19 at US$34 trillion. Previous estimates for the United States calculate harm at US$16 trillion (1). We obtain our estimate by applying the methodology of that previous estimate to the entire world. (Calculations are available upon request from the authors.) The probability that a COVID-19 vaccine R&D project succeeds is 24% based on data from the Regulatory Affairs Professionals Society (32). The society tracked 126 vaccine projects, of which 30 gained approval. We also use a lower estimate of 18% based on World Health Organization data (33). We calculate a harm reduction parameter of 83% as the average efficacies of the Pfizer and Moderna vaccines (95%), the Johnson & Johnson vaccine (66%), and the AstraZeneca vaccine (76%) (34). (We consider lower efficacy estimates—perhaps due to the emergence of variants—in our sensitivity analysis.) We take the cost of developing a successful vaccine, inclusive of failures, to be 2020 US$1 billion (35, 36). Finally, we consider only one approach to addressing COVID-19—the vaccine approach—and sensitivity to four vaccine approaches: inactive/attenuated virus, viral subunits, viral vector, and nucleic acid (DNA/RNA) (37). After consulting with industry experts, we calibrate the probability that an approach can succeed to 90%. (Our industry experts argued that all these approaches have scientific validity and that with enough time, resources, and funding they are likely to produce at least one successful product.) Finally, given the uncertainty of our parameter choices, we examine sensitivity in the results section and find that our qualitative conclusions are robust to a wide range of values.

Group A Streptococcus.

We allow for two approaches to Strep A vaccine development: the M-protein and a generic other approach (38). Again, both have a 90% chance of success. A single project has a 15% chance of leading to an approved vaccine (8, 39, 40). The fraction of Strep A harm that a single approved vaccine could eliminate is 30% (8, 41). We calibrate the cost per project the same as for the COVID-19 vaccine.

Using estimates from a study for Australia, we find total global harm caused by Strep A to be US$2.1 trillion. This estimate is obtained by extrapolating estimated harm to the non-Indigenous Australian population to high- and upper-middle-income countries and estimated harm to the Indigenous Australian population to lower-middle- and low-income countries (41).

While we do not claim to equate the experience of a low-income country population to that of an Indigenous population of a high-income country, several points support our use of the Australian data to estimate global harm. These estimates examine the various manifestations of Strep A and provide a monetary cost of Strep A disease. Moreover, the Indigenous population experiences Strep A like lower-income country populations do. As this study (41) points out, the Australian Indigenous population has incidence rates of Strep A–induced disease similar to those of lower-middle-income countries. Other work has shown the rate of rheumatic heart disease (a serious complication from repeated Strep infections) for Indigenous Australians is between that of South-Central Asia and Sub-Saharan Africa (42). Finally, using these Australian estimates makes our results robust to the alternative of scaling up antibiotic use. The robustness stems from the fact that Australia has high antibiotic availability, implying that our extrapolation gives us a potential estimate of remaining harm after a global scale-up of access to antibiotics.

Alzheimer’s Disease.

We calibrate harm from ADRDs using the present discounted value of the estimated global economic burden of Alzheimer’s disease over the next 30 y (43). The resulting value is US$76 trillion. We calibrate the number of R&D approaches to three based on a review of approaches to develop ADRDs treatments (44). This survey divides approaches to treating ADRDs into three categories: biologic, small molecule symptom reducing, and small molecule disease-modifying. We consider sensitivity to 11 approaches, using the survey’s subclassifications based on mechanism of action.

Each approach has a high probability of success, equal to 90%, but the probability of an individual project succeeding is low. Clinical trials for ADRDs treatments have had low success rates: as of 2020, 196 Alzheimer’s disease trials had failed and four had resulted in approval from the US Food and Drug Administration (45). These results imply a 2% success probability for individual projects.

The high probability of success for each approach but low probability of success for an individual project is not a contradiction. Our assumption of a 90% probability that the approach will succeed means only that if we did not limit funding and funded a very large number of projects, we would have a 90% chance of seeing a success. This is akin to saying that even if buying a lottery ticket gives an individual little chance to win, someone will likely win when enough tickets are sold.

The fraction of harm a successful project alleviates is calculated under the assumption that a success will push back Alzheimer’s disease onset by 3 mo, based on the lower range of estimates of the effect of Donepezil, a successful Alzheimer’s treatment (46, 47). A 3-mo time discount factor and a 3 mo survival probability capture this benefit. (Intuitively, if the effect of a treatment is to push back disease harm for a given amount of time, but that harm will eventually be experienced if the patient survives, then the value of that treatment can be measured by how much the patient discounts the future and how likely she is to survive to experience that harm.) This benefit calculation leads to a harm reduction from a successful treatment equal to 1.05%. Finally, to calculate c, the project cost, we use data from DiMasi et al. (48). Their estimates imply a cost of developing a successful treatment, inclusive of failures, of 2020 US$5.77 billion after adjusting for inflation and expected growth of R&D costs.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

There are no data underlying this work.