• Funding for research on the returns to public investment in R&D is available from a coalition of funders. Applications due April 30. See the full call here. This is part of the Pop Up Journal project that is a collaboration between the Alfred P. Sloan Foundation and the Abundance and Growth Fund at Coefficient Giving.

• The post below was originally published by me on the Abundance and Growth blog, where I write with the rest of the team at the Abundance and Growth Fund. I plan to cross-post posts I write there related to innovation, but I will also write about other topics there. If you want to follow our work, subscribe for updates.

• My colleague Jordan Dworkin has written two other pieces on that blog, which readers here might find interesting: Who Will Program Manage the Program Managers? and Is replication pro-progress or anti-risk?

• On to the post!

In 2019, Tyler Cowen and Patrick Collison wrote an article for The Atlantic titled We Need a New Science of Progress, which helped kick off the Progress Studies movement. In that article, they wrote:

For example, if our discoveries and inventions improve standards of living by 1 percent a year, children will by adulthood be 35 percent better off than their parents. If they improve livelihoods at 3 percent a year, those same children will grow up to be about 2.5 times better off. Whether viewed in terms of large or small improvements, progress matters a lot.

This is a common rhetorical strategy to justify why we ought to care about faster progress, and it’s one that I agree with: the magic of compounding means small differences in growth rates, sustained over time, add up.

But the case for caring about progress doesn’t rely only on the magic of compounding interest rates. Even small one-off changes in the rate of broadly shared progress are worth huge sums.

Thanks for reading The Abundance and Growth Blog! Subscribe for free to receive new posts.

Ballparking the value of a small boost to progress

Let’s illustrate this by estimating the value of a relatively small one-time increase in the rate of technological progress. Economists like myself generally assume increases in living standards are brought about by technological progress, so let’s consider the impact of a one-time increase in the rate of per-capita economic growth from 2% to 2.1% in the United States (2% is the long-run average). After one year of faster growth, assume we drop back down to the usual 2% per year. Going from 2% to 2.1% would be a 5% relative increase in the rate of growth. What would that be worth?

To ballpark it, consider that US GDP per capita is around $90,000. Growing by 2.1% instead of 2% means this increases by an extra 0.1%, or $90 per person. Shared over 340 million Americans, that works out to $30.6 billion.

But that’s only in the first year. Even though we’re imagining a one-off change in the rate of progress, that doesn’t mean the benefits are one and done. Technological progress is a cumulative process, where new technologies are built off older ones. If we grow by an extra 0.1%, in the next year we’ll be working off a slightly more advanced technological base from then on. So even if we drop back down to the historical average rate of 2% per year thereafter, we’ll still be 0.1% richer than we otherwise would have been. It’s like a runner who briefly sprints before reverting to their former pace. Even after they are no longer sprinting, they remain ahead of where they would have been, had they merely maintained a steady pace.

To keep things simple, let’s assume the current cohort of Americans enjoys an additional $90 per year for the rest of their working lives and not worry about discounting (we’ll come back to that). The average American is about 40 today, so if they retire at 65, that’s 25 years of earning another $90 per year. Ninety dollars a year, over 25 years, across 340 million people is $765 billion.

Another important benefit flowing from technological progress is health. In the USA, life expectancy has been on a long-run upward trajectory, increasing by about 0.1 years per year for several decades.1 If we think that is mostly driven by technological (in this case biomedical) progress, then an increase in the rate of progress should also have some health benefits. Just as we increased the rate of economic growth by 5% (going from 2 to 2.1%), let’s increase the annual gains to life expectancy by 5%: that means during our year of accelerated progress, we’ll add 0.105 years to life expectancy instead of 0.1. That’s an increase of life expectancy of 0.005 years, or 1.8 days.

To put a crass dollar value on that, let’s turn to the Value of a Statistical Life Year.2 The US Department of Health and Human Services has a range of estimates of the Value of a Statistical Life year, ranging from $282,000 to $1.5 million (see Table 3). The conservative lower bound will work fine for us. If we assume an extra year of progress yields an extra 0.005 years of life expectancy, this is valued at $1,410 per person. Across 340 million Americans, that’s $479 billion.

Add the long-run income and health benefits up and the upshot is this: the health and income benefits of a 5% one-time boost to the annual rate of technological progress are in the ballpark of $1.2 trillion.

Leaving the ballpark

The above calculation is illustrative, but takes a lot of shortcuts. It ignores changes in population. It assumes economic benefits are $90 per year, which is not correct if we assume growth is exponential. It ignores the impact on people not born today, who might still benefit from technological progress. And it doesn’t properly discount for benefits that arrive in the more distant future.

To try and capture all that, I’ve built a simple economic model that incorporates all this and more, and with the help of Claude code I’ve written a web tool you can play around with to value different kinds of interventions to the rate of progress. Under my default settings, this more realistic model also values the above policy change at $1.2 trillion. If you are interested in using this tool, read the explainer here. We’ve also put the underlying code up here as a commented python file, in case you want to build your own version. Just give the code to a frontier LLM and ask it to modify it to incorporate any changes you want to make.

What is an ocean but a multitude of drops?

The important takeaway from this exercise is that even tiny increases in broadly shared progress are almost inconceivably beneficial. Mechanically, this is because we’re assuming the kind of progress we care about is shared by many people over many years.3 The import of this fact can be hard to grasp intuitively. We’re all individuals, but this is a situation where focusing on an individual experience can give us the wrong intuitions. As individuals, if we think of what it would be like to be 0.1% richer and live an extra 1.8 days, it seems good but hardly something to get passionate about. But aggregated across the generations, the import is qualitatively different. The extra income cumulates to hundreds of billions; the extra days add up to more than a million years.

If we take the conclusions of this post seriously, then we should be willing to put a lot of effort into raising the rate of progress. And I see that as precisely what we, and the broader abundance and progress studies movements, are trying to do.

I’ve split this post into two to make it a bit easier to navigate. This is the second post. The first is here.

Special thanks to Nisha Austin for helping me put these posts together!

Titles Index

Titles are presented in random order. There might be additional authors on these papers - we’ve listed the associated job market candidate only.

1. Tenure and research trajectories by Giorgio Tripodi

2. Spatial Allocation of Inventors, Knowledge Diffusion and Growth by Furkan Kilic

3. Patents, Innovation, and Imitation in a North-South Model with Increasing Product Variety by Florence Ut Meng Ho

4. The Innovation Long-Run Risk Component by Fabio Franceschini

5. Import Competition, Innovation, and the Cost of Protectionism by Deniz Atalar

6. Financial Development and Endogenous Investment-Specific Technical Change by Daeeun Bae

7. The Origins of the Nitrogen Revolution by Christopher W. A. Sims

8. Knowledge Generality, Competition and Growth by Chenchuan Shi

9. Start-up Financing, Entry and Innovation by Charles Parry

10. The Role of Training for Technology Diffusion by Carolina Bussotti

11. Firm Scope and Innovation: The Role of Intangibles by Cagin Keskin

12. The value of conceptual knowledge by Benjamin Davies

13. Robust Technology Regulation by Andrew Koh

14. Technology and the Geography of Industrial Policy by Aditya Bhandari

15. Sexual Misconduct and Scientific Production by Manuela Collis

16. Worker Mobility and the Diffusion of Radical Technologies by Stephan Hobler

17. Patent challenge and generic entry by Xin Zhang

18. Innovation and Adaptation to Expanding Biological Threats by Shu-Chen Tsao

Titles, Abstracts, and Links to Papers

Tenure and research trajectories

Giorgio Tripodi

Tenure is a cornerstone of the US academic system, yet its relationship to faculty research trajectories remains poorly understood. Conceptually, tenure systems may act as a selection mechanism, screening in high-output researchers; a dynamic incentive mechanism, encouraging high output prior to tenure but low output after tenure; and a creative search mechanism, encouraging tenured individuals to undertake high-risk work. Here, we integrate data from seven different sources to trace US tenure-line faculty and their research outputs at a remarkable scale and scope, covering over 12,000 researchers across 15 disciplines. Our analysis reveals that faculty publication rates typically increase sharply during the tenure track and peak just before obtaining tenure. Post-tenure trends, however, vary across disciplines: In lab-based fields, such as biology and chemistry, research output typically remains high post-tenure, whereas in non-lab-based fields, such as mathematics and sociology, research output typically declines substantially post-tenure. Turning to creative search, faculty increasingly produce novel, high-risk research after securing tenure. However, this shift toward novelty and risk-taking comes with a decline in impact, with post-tenure research yielding fewer highly cited papers. Comparing outcomes across common career ages but different tenure years or comparing research trajectories in tenure-based and non-tenure-based research settings underscores that breaks in the research trajectories are sharply tied to the individual’s tenure year. Overall, these findings provide an empirical basis for understanding the tenure system, individual research trajectories, and the shape of scientific output.

Link

Spatial Allocation of Inventors, Knowledge Diffusion and Growth

Furkan Kilic

Where does innovation truly thrive? Inventive activity in the US is strikingly concentrated in a handful of hubs. This raises compelling questions: Does further agglomeration drive innovation, or could a more dispersed approach better leverage regional spillovers? To investigate, I exploit variation in patent citation lags across US states and develop a novel endogenous growth model with mobile inventors and workers. The model integrates an exogenous knowledge network that facilitates the dynamic exchange of ideas—laying the foundation for future inventions—between locations, revealing that inventors do not internalize how their location choice influences broader knowledge diffusion. These knowledge spillovers call for a targeted, place-based R&D subsidy to unlock latent innovation potential. Calibrating the model to data on inventor and worker allocations—and estimating the knowledge diffusion network from patent citations—I find that optimal policy would further concentrate inventors in established hubs, enhancing welfare by 1.8 percent in consumption-equivalent terms and boosting the economy’s long-run growth rate by 0.14 percentage points.

Link

Patents, Innovation, and Imitation in a North-South Model with Increasing Product Variety

Florence Ut Meng Ho

To understand the relationship between patent protection and technology transfer across countries and to equip policymakers with insights to balance innovation promotion and technology dissemination, this paper investigates the cross-country effects of patent protection on relative wages, innovation, technology transfer, and welfare in a North-South model with variety expansion. In this dynamic general equilibrium model, firms in the North perform innovative R&D, while firms in the South perform imitative R&D for technology transfer. Innovation occurs through the invention of new varieties of goods, and IPR protection is modeled in the form of patent breadth. We find that strengthening patent protection in the North permanently raises the relative wage between the North and the South, permanently decreases the rate of technology transfer, and temporarily increases the northern innovation rate. Conversely, strengthening patent protection in the South permanently reduces the relative wage, permanently increases the rate of technology transfer, and temporarily boosts the northern innovation rate. Calibrating this model to US-China data, our quantitative analysis reveals that when a country unilaterally strengthens patent protection, its domestic welfare increases. However, when both countries strengthen patent protection bilaterally, the South experiences a welfare gain while the North suffers a welfare loss.

Link

The Innovation Long-Run Risk Component

Fabio Franceschini

This paper provides robust empirical evidence that shocks to aggregate Research and Development (R&D) have persistent effects on macroeconomic dynamics and represent a significant risk for investors, as predicted by the “long-run risk” literature. The analysis focuses on a single variable, “effective R&D”, which captures the entire contribution of R&D to productivity growth, flexibly accounting for knowledge spillovers and product proliferation effects. Deviations of effective R&D from its equilibrium level can be empirically identified leveraging the error correction term in the cointegration relationship among R&D, total factor productivity, and the labor force. In US data, structural effective R&D shocks affect productivity and consumption growth rates beyond business cycle horizons and are associated with a significant risk premium in a cross section of stock and bond portfolios (around 2% annually), with cash-flow sensitivities proving a key determinant.

Link

Import Competition, Innovation, and the Cost of Protectionism

Deniz Atalar

How does import protectionism affect catch-up innovation and welfare? I develop a small open-economy model in which trade costs shape buyer–supplier relationships, and suppliers’ incentives to innovate. When trade costs rise, domestic buyers become more inclined to shift sourcing from foreign to domestic suppliers. The possibility of this shift triggers heterogeneous innovation responses among domestic suppliers: those suppliers that are less technologically advanced than their foreign competitors increase their innovation, while other, more advanced suppliers reduce it. I verify and quantify the buyers’ shift in sourcing as well as the suppliers’ heterogeneous innovation responses to this shift using novel Turkish firm–product-level data on firm-to-firm transactions, imports, production, and innovation expenditures. I build on these firm level responses to assess the impact of import protectionism on welfare. If domestic suppliers in the aggregate respond to a rise in trade costs by increasing innovation, the welfare losses of such a policy are partially mitigated. If, on the other hand, they respond by decreasing innovation, the welfare losses are amplified. Calibrating the model to Turkish microdata, I find that a 10 percent rise in trade costs in Turkiye triggers an increase in aggregate innovation, which mitigates roughly one-quarter of the welfare loss relative to a no-innovation benchmark.

Link

Financial Development and Endogenous Investment-Specific Technical Change

Daeeun Bae

I show that financial development is a key determinant of cross-country variation in the rate of investment-specific technical change. Using a large cross-country dataset, I document that countries with more developed financial markets exhibit higher rates of investment-specific technical change, and that investment goods production is more intensive in value added from high-R&D industries than consumption goods production. To explain these findings, I develop a multi-industry endogenous growth model with credit constraints on R&D expenditures. In the model, R&D drives productivity growth, and financial development disproportionately benefits the productivity growth of high-R&D industries because they are more dependent on external financing. Taken together with the different industrial composition of final goods production, financial development endogenously generates faster productivity growth in investment goods production. The quantitative analysis shows that this endogenous channel accounts for approximately 40% of the observed cross-country relationship between financial development and investment-specific technical change.

Link

The Origins of the Nitrogen Revolution

Christopher W. A. Sims

Many technologies raise productivity in locations constrained by their natural endowments yet diminish specialization across space. We show that the first commercial nitrogen fertilizers in history were one such “converging” technology. Leveraging natural variation in soil nitrogen deficiency and the sudden introduction of Peruvian guano and nitrates to 19th-century England, we provide two main empirical findings. First, locations specialized on the basis of their natural endowments before the introduction of fertilizer: nitrogen-deficient places devoted less land to nitrogen-intensive crops. Second, combining newly-digitized data and a difference-in-differences design, we show that these nitrogen-deficient places substantially reallocated toward nitrogen-intensive crops after fertilizer was introduced, indicating convergence across space. To quantify the welfare impact of this “converging” technology, we embed fertilizer into a quantitative spatial model of the English agricultural sector with realistic geography. The welfare gains from fertilizer were equivalent to two decades of annual productivity growth in agriculture. However, convergence implies a reduction in the gains from trade, which offsets up to 10% of these welfare gains under plausible trade cost regimes.

Link

Knowledge Generality, Competition and Growth

Chenchuan Shi

This paper studies how the generality of knowledge—its applicability across technologies and industries—shapes firms’ innovation strategies, market structure, and aggregate growth. I build an endogenous growth model in which firms choose between general and firm-specific R&D while competing for market leadership. General innovations enhance firms’ capacity to absorb and apply outside knowledge, creating spillovers within and across industries, whereas firm-specific innovations yield mainly private gains. The model predicts, and the data confirm, that (1) leaders favor firm-specific R&D while followers rely on general innovations to catch up, and (2) the gap in innovation generality between them follows a U-shaped pattern with market concentration. Leveraging variation in the enforceability of non-compete agreements across U.S. states, I provide empirical evidence consistent with the model’s spillover mechanisms. The findings point to a novel growth policy: encouraging general R&D, particularly among leading firms, can improve knowledge diffusion and sustain long-run growth.

Link

Start-up Financing, Entry and Innovation

Charles Parry

Venture capital (VC) is the key source of financing for high-growth start-ups, but with few alternatives, limited access can leave viable projects unfunded and constrain innovation. I develop and estimate an equilibrium model of the VC market to quantify these distortions in the US, explain cross-country differences in VC activity, and diagnose VC’s sectoral concentration. In the model, entrepreneurs and VCs meet in a frictional matching market and VCs endogenously stage capital injections over time to limit losses from hidden failure by entrepreneurs; however, reliance on follow-on funding exposes the start-up to premature closure if funding does not materialise. The model maps directly to observed funding histories, enabling estimation and policy counterfactuals. For US start-ups first funded in 2005–2015, my estimates suggest that 40% shut down despite having positive continuation value; with continued funding, half would reach an acquisition or IPO. I then estimate the model on UK microdata and find that financing conditions and acquisition opportunities, not project quality, drive US–UK differences; financing conditions account for two-thirds of the entry gap. Because UK start-ups struggle to reach late-stage rounds, retargeting existing support towards late-stage start-ups improves outcomes. Finally, the theory offers an explanation for VC’s concentration in software and services: frictions are least severe for short-horizon projects with ample acquisition opportunities. Absent frictions, the share of VC-backed software and services start-ups falls from 61% to 53%, offset by gains in science-based sectors.

Link

The Role of Training for Technology Diffusion

Carolina Bussotti

We study a dynamic model of new technology adoption in a labor market with search frictions, where worker training boosts the productivity gains from adopting the technology. Once trained, workers acquire general skills that can be used by any firm operating the new technology. As a result, firms can free-ride on the training investments of others, which inefficiently delays technology adoption and workforce training. We apply the model to the diffusion of Enterprise Resource Planning (ERP) systems in Portugal—a leading software technology currently used by 43% of European firms—whose adoption requires training workers in transferable skills. Using matched employer–employee data, we show that as the technology diffuses, firms train fewer workers upon adoption, consistent with the model’s free-riding mechanism. The calibrated model implies a 14% present-value loss in net output due to underinvestment in adoption and training. Finally, policy counterfactual analysis shows that training subsidies introduced at the onset of the diffusion process are about 10% more effective than those enacted ten years later.

Link

Firm Scope and Innovation: The Role of Intangibles

Cagin Keskin

Horizontal expansion through an increasing product portfolio lies at the core of modern endogenous growth literature. Yet evidence remains limited on how diversification across industries influences a firm’s trade-off between generating social surplus and capturing private returns. To investigate this, I categorize intangible assets by their spillovers: transferable intangibles (patents, software) generate social surplus, whereas embedded intangibles (organizational capital, brand value) primarily yield private returns. I document that diversified firms reallocate investment toward embedded intangibles, a strategic shift accompanied by declining markups and productivity, together with reduced innovation by their rivals. Motivated by this evidence, I extend a canonical endogenous-growth framework to endogenize firms’ allocation between transferable and embedded intangibles, allowing for both horizontal and vertical expansion. A key prediction of the model is that embedded intangibles are the primary driver of a firm’s ability to expand across industries, which also raises entry barriers for competitors and decreases social return rather than promoting long-run growth. Thus, a shift in innovative effort ultimately sacrifices economy-wide growth for firm-level market advantages, and quantitative analysis indicates that size-dependent taxes can substantially improve welfare.

Link

The value of conceptual knowledge

Benjamin Davies

We study the instrumental value of conceptual knowledge when making statistical decisions. Such knowledge tells agents how unknown, payoff-relevant states relate. It is distinct from the statistical knowledge gained from observing signals of those states. We formalize this distinction in a tractable framework used by economists and statisticians. Conceptual knowledge is valuable because it empowers agents to design more informative signals. It is more valuable when states are more “reducible”: when they can be explained with fewer common concepts. Its value is non-monotone in the number of signals and vanishes when agents have infinitely many signals. Agents who know more concepts can attain the same payoffs with fewer signals. This is especially true when states are highly reducible.

Link

Robust Technology Regulation

Andrew Koh

We analyze how uncertain technologies should be robustly regulated and how regulation should evolve with new information. An adaptive sandbox comprising a zero marginal tax up to an evolving quantity limit is (i) robust: it delivers optimal payoff guarantees when the agent’s learning process and/or preferences are chosen adversarially; (ii) dominant: it outperforms other robust and regular mechanisms across all agent learning processes and preferences; (iii) time-consistent: it is the only robust mechanism that can be implemented without commitment. Robustness is important: absent robust regulation, worst-case pay-offs can be arbitrarily poor and are induced by weak but growing optimism that encourages excessive risk-taking. Our results offer optimality foundations for existing policy and speak directly to current debates around managing emerging technologies.

Link

Technology and the Geography of Industrial Policy

Aditya Bhandari

Industrial policy around the world is increasingly targeting sectors heterogeneously across regions to account for sectoral agglomeration externalities. This paper provides a theoretical framework to study when such spatial targeting is optimal. I demonstrate that the optimal policy across regions depends on the underlying productivity structure: whether agglomeration affects Hicks-neutral (input-neutral) or Harrod-neutral (labor-augmenting) productivity, the two dominant formulations in spatial economics. When productivity is Hicks-neutral , optimal industrial subsidies increase with regional sector size. However, when productivity is Harrod-neutral, the optimal industrial subsidy is a constant ad-valorem wage subsidy across all regions. I apply this framework to manufacturing in England using data on 153 regions, where I estimate the productivity structure and the agglomeration elasticity. In the Hicks-neutral case, place-based policies raise welfare by 5.1% versus 2% for uniform subsidies. In the Harrod-neutral case, uniform subsidies raise welfare by 3.1%, while place-based ones reduce welfare by 0.7%. The estimation reveals predominantly Hicks-neutral technology, supporting place-based over uniform policies for manufacturing in England.

Link

Sexual Misconduct and Scientific Production

Manuela Collis

While sexual misconduct in the workplace has complex and lasting consequences for directly affected individuals, its broader organizational implications remain less well understood. Using a novel dataset of over 1,000 documented sexual misconduct cases across U.S. universities, I examine how these publicly reported incidents affect departmental scientific productivity. Using the benefit of hindsight, I record the year sexual misconduct occurs and the year it becomes public. I employ coarsened exact matching and a staggered difference-in-differences design to compare control departments with those that experienced subsequently publicized misconduct incidents. Sexual misconduct shows no discernible effect on departmental productivity when it occurs, but public reporting reduces publications by 0.1 per faculty member annually — equivalent to nine fewer publications over five years for a median department of 18 members. These findings reveal that organizational costs arise specifically from public disclosure rather than from the misconduct itself. This distinction between occurrence and disclosure effects suggests that protecting victims and maintaining productivity may require differentiated policy approaches as institutions navigate competing demands from legal frameworks, ethical obligations, and performance concerns. These dynamics help explain both why social pressures transform misconduct from HR concerns into strategic organizational challenges and why firms may prioritize confidentiality strategies.

Link

Worker Mobility and the Diffusion of Radical Technologies

Stephan Hobler

The spread of new, transformative technologies often relies on specialized knowledge among workers and managers. When the required expertise is scarce, human capital and worker mobility can become bottlenecks to technology diffusion. To study these dynamics, I develop a theory in which firms and workers accumulate technology-specific expertise through mutual learning, and worker mobility is subject to search frictions. I calibrate the model to the diffusion of predictive AI among U.S. firms, matching empirical patterns of technology adoption and worker flows using comprehensive microdata from LinkedIn. Worker mobility emerges as a key driver of diffusion. When a new technology is introduced, adoption is initially slow due to the lack of experienced workers and concentrated only among the most productive firm-worker pairs. Diffusion then accelerates through the poaching of workers from early adopters, generating an S-shaped diffusion curve. Aggregate output follows a J curve, with productivity gains taking time to materialize. A counterfactual matched to European-style labor markets with low mobility and long job tenures reveals a trade-off: the European economy delivers modestly higher steady-state output, but slows adoption by two-thirds and reduces welfare gains from the new technology by one-third—potentially contributing to transatlantic productivity gaps in technology-intensive sectors.

Link

Patent challenge and generic entry

Xin Zhang

Pharmaceutical innovation depends on strong primary patents that allow originators to recoup R&D costs. However, drug companies often engage in ever-greening that prolongs patent protection by filing follow-on patents with little therapeutic gain. We study a policy lever that works with market forces to screen out weak follow-on patents: the Hatch-Waxman Act, which incentivizes challenges to ever-greening patents by granting the first successful challenger a period of marketing exclusivity. We investigate how the length of first-filer exclusivity shapes generic firms’ incentives to initiate challenges, which can curb the extra monopoly protection created by ever-greening while preserving incentives for genuine discovery and protecting consumer welfare through earlier generic entry. Using a two-stage structural model that endogenizes challenge and entry decisions, we estimate the fixed costs of generic entry with moment inequalities. We find that the current 180-day exclusivity raises challenge rates by about 4 percentage points. Extending exclusivity primarily activates challenges in markets that would otherwise go unchallenged: a two-year exclusivity increases the challenge rate to 15.38%. Effective exclusivity is highly heterogeneous across therapeutic classes: reaching a 20% challenge rate requires roughly two years for antimicrobials but less than one year for genitourinary drugs.

Link

Innovation and Adaptation to Expanding Biological Threats

Shu-Chen Tsao

Developed countries often possess the capacity to innovate technologies that mitigate global biological threats, such as vaccines for infectious diseases, but innovate little when not directly exposed. This paper develops a spatial dynamic game to study endogenous innovation incentives as biological threats expand into developed countries. In the model, all countries can control threats locally, while only a few can innovate. I decompose how two externalities—threat diffusion and technology spillovers—and their interaction shape strategic innovation incentives. Moreover, I show that these analytical results can be empirically estimated using a tractable GMM framework. Using evidence of dengue-transmitting mosquitoes expanding into the U.S., I estimate that endogenous U.S. vaccine innovation could reduce dengue cases in the Americas by 54% relative to a no-innovation scenario and assess the resulting global welfare implications. Finally, I show that greater exposure may fail to spur innovation when sustained eradication through local control is optimal for a class of biological threats.

Link

• The Institute for Progress is hosting the third iteration of its free online PhD short-course: The Economics of Ideas, Science, and Innovation. If you’ve completed the equivalent of a first year economics PhD, apply by January 9. I’m giving the lecture on the returns to R&D!

• I have a new blog! It’s a group blog by the team at The Abundance and Growth Fund (where I’m the program director). My plan is to return to writing more often in 2026, with the hope that I will finally have time now that the Abundance and Growth Fund has finished hiring. But reflecting changes in my job, I plan to write about more than just innovation this year. My tentative plan is to cross-post things I write about innovation here, but not the rest.

  The Abundance and Growth Blog

  All things progress and growth from the Abundance and Growth Fund team at Coefficient Giving.

  By Matt Clancy

• Speaking of writing that’s about more than innovation, check out my first post over at the Abundance and Growth Blog: 28 Thoughts About Abundance and Growth.

• Also on the Abundance and Growth Blog, we are going to publish a series on job market papers related to abundance and growth in early 2026. Send your economics job market papers on housing, energy, infrastructure, clinical trials, state capacity, and high skilled immigration to abundanceandgrowth@coefficienctgiving.org! And subscribe if you want those posts in your inbox.

• The Alfred P. Sloan Foundation is hiring for an Economics Program Associate.

• Merry Christmas Eve to those who celebrate!

In this special annual edition of What’s New Under the Sun, we have a big bundle of the titles, abstracts, and links to innovation-related PhD job market papers from 2025. Some were sent my way following my request in the last newsletter, but to find the majority of these, I manually looked at the titles of ~1,300 job market papers. Even so, I’m sure I missed some great papers. If that’s you, email me and I’ll add you to the posts.

I’ve split this post into two to make it a bit easier to navigate. This is the first post.

Special thanks to Nisha Austin for helping me put these posts together!

Titles Index

Titles are presented in random order. There might be additional authors on these papers - we’ve listed the associated job market candidate only.

1. Specialization by design: the unequal geographic effects of modular product design by Vishan Gandhi Nigam

2. Technology M&A and Knowledge Diffusion by Zili Yang

3. Venture Capital Contracts and Heterogeneous Innovation by Yucheng Wong

4. Trade, Research Productivity, and Growth: A Dynamic General Equilibrium Approach by Yaming Chang

5. The Impact of Intellectual Property Rights on Innovation and Follow-on Development: Evidence From the Bayh-Dole Act by Tallon Howie

6. Innovation through Recombination by Songyuan Teng

7. Technological Change and the Market for Books, 1450-1550 by Qiyi Charlotte Zhao

8. Factories of Ideas? Big Business and the Golden Age of American Innovation by Pier Paolo Creanza

9. Solving Problems of Unknown Difficulty by Nicholas Wu

10. Quantifying Knowledge Spillovers Using Firm and Product Dynamics by Mohamad Adhami

11. Beyond the Lab: The Effect of PhD Programs on Innovation by Manfredi Aliberti

12. The Prestige-Testability Tradeoff in Science by Kurtis A. Hingl

13. Ownership Structure and Economic Growth by Koki Okumura

14. Startup Acquisitions and Innovation in the Biopharmaceutical Industry by Hayley Wabiszewski

15. Extreme Heat and Directed Innovation by Enjie (Jack) Ma

16. How Does Industry Shape Academic Science? Evidence from “Million Dollar Plants” by Hongyuan Xia

17. How Network Hiring by Entrepreneurs Shapes Firm Formation and Performance by John F. Bonney

18. Science, Startups, and the Problem of Value Capture: Thin Acquisition Markets, Weak Outside Options by Roger Masclans

Titles, Abstracts, and Links to Papers

1. Specialization by design: the unequal geographic effects of modular product design

Vishan Gandhi Nigam

I show that modular design – a revolution in how firms organize innovation – concentrates industrial production in large countries. Modular products follow common rules called design platforms, and thus can share inputs while remaining customized for local needs. Combining six new datasets on global automotive design and trade, within-firm event studies of platform rollouts and mergers, and a model with scale economies in shared input production, I find that design platforms reshape global trade in two phases. First, platform-sharing across destinations increases trade and enables within-firm specialization; for instance, poor countries export engines for affordable car segments. Second, platform-sharing across product segments creates winner-take-all supply chains in which a firm’s largest markets produce most inputs. In both phases, design platforms expand the scope of home-market effects, concentrating input production in large markets for shared platforms rather than individual products. By quantifying the model, I show modular design has unequal effects across countries and shapes the returns to industrial policy: in particular, universal platforms (expected by 2030 for EVs) double American and Chinese input production shares, reduce production by over 80% in most smaller countries, and imply that U.S. input tariffs on China (but not on third countries) have larger reshoring effects.

Link

2. Technology M&A and Knowledge Diffusion

Zili Yang

This paper examines how technology mergers and acquisitions (tech M&As) affect the diffusion of target firms’ pre-acquisition innovations in the United States. Using US patent and M&A data from 1980 to 2021. This study employs a difference-in-differences approach comparing successful acquisitions with exogenously failed deals; it finds that tech M&As significantly increase external diffusion of targets’ technologies, as measured by patent citations, with effects concentrated within the acquirer’s industry. Tech M&As do not diminish young firms’ ability to cite and build upon acquired targets’ patents, contradicting concerns about innovation foreclosure. To interpret these findings and quantify aggregate implications, I develop an idea flow model where firms improve productivity by choosing innovation intensity based on potential targets’ technologies, with acquisitions affecting both the innovation step size in learning from targets and the cost of accessing them. The model, calibrated to the empirical estimates and US innovation data, reveals that doubling the 2015 tech M&A rate would increase annual productivity growth by five hundredths of a percentage point, with the diffusion channel contributing 40% of this increase. Surprisingly, relaxing restrictions on post-acquisition knowledge appropriation yields negligible growth effects: reduced spillovers from acquired targets are offset by increased innovation using independent technologies as acquisition values rise. These findings underscore the importance of incorporating diffusion effects and general equilibrium forces into antitrust policy for tech M&As.

Link

3. Venture Capital Contracts and Heterogeneous Innovation

Yucheng Wong

This paper studies how venture capital (VC) reshapes startups’ innovation choices by insuring against default risk, and explores the macroeconomic implications of this mechanism. I develop a dynamic general equilibrium model in which startups choose between conservative (low-risk, low-return) and aggressive (high-risk, high-return) innovation while endogenously selecting their financing mode. Debt financing features state uncontingent repayments and exposes startups to default. By contrast, VC financing is a state-contingent dynamic contract with one-sided limited commitment from startup. Evidence from a new dataset linking VC deals, balance sheets, and patents supports the model predictions: VC-backed startups begin with higher leverage, show greater post-financing profit dispersion, and generate more high-quality patents. Calibrated to financing and innovation data, eliminating VC reduces the aggregate output by 5 percent and the mass of large firms by 11 percent, despite only 0.2 percent of startups ever receiving VC.

Link

4. Trade, Research Productivity, and Growth: A Dynamic General Equilibrium Approach

Yaming Chang

Exporting exposes firms to foreign buyers and rivals, providing knowledge that makes innovation more effective. How much does this mechanism raise innovation efficiency, and how does it affect long-run growth? I develop and estimate an endogenous growth model in which foreign-market exposure enhances firms’ innovation efficiency, the exporting–innovation efficiency (EIE) channel, while broader product scope dilutes managerial attention and lowers marginal returns. The model predicts that innovation efficiency rises with export intensity but declines with product scope and firm size. Using firm-level data on Chinese manufacturers, I document patterns consistent with these predictions and estimate that export exposure increases the effective knowledge available for innovation by about 9% at the product level. Although firm-level gains are modest, the aggregate effect is economically meaningful: eliminating the EIE channel reduces the long-run growth rate by about 1.3 percentage points (around 10 percent of observed growth). With the EIE channel active, trade liberalization disproportionately benefits firms with broad export portfolios, boosting their innovation efficiency and growth and accelerating the exit of low-productivity firms. The result is a more right-skewed productivity distribution and greater market concentration.

Link

5. The Impact of Intellectual Property Rights on Innovation and Follow-on Development: Evidence From the Bayh-Dole Act

Tallon Howie

Intellectual property rights (IPR) are thought to promote innovation, but inhibit follow-on development by restricting the dissemination of new ideas across firms. Exploiting a change in patent policy for U.S. government-sponsored inventions (1980 Bayh-Dole Act), I find that strengthening IPR increases innovation, follow-on development, and cross-firm diffusion of ideas. I use these estimates to calibrate a general equilibrium growth model featuring a two-stage R&D process and endogenous knowledge diffusion. IPR raises R&D intensity by strengthening appropriation and increases knowledge diffusion by enabling innovators to shield licensees from competition. A quantitative analysis suggests that IPR increase aggregate welfare, although a compulsory licensing policy sometimes better balances innovation incentives with broader diffusion. These results highlight trade-offs in patent policy and the government’s ability to leverage R&D funding to promote dissemination and development of ideas.

Link

6. Innovation through Recombination

Songyuan Teng

New ideas often recombine existing ones; this insight is emphasized in recent economic growth theories, but evidence on its empirical relevance is scarce. This paper takes combinatorial growth to measurement by studying the pharmaceutical industry, where the distinction between novelty (discovering new building blocks) and recombination (assembling building blocks into products) is transparent. I uncover the substantial and rising importance of recombination, the firm life-cycle from knowledge accumulation to recombination, and the value premia for novelty. Motivated by these facts, I develop a theory of firm dynamics that distinguishes firm knowledge stocks from product portfolios. Innovation operates along two distinct yet intertwined margins: novel innovation expands knowledge, while combinatorial innovation deploys that knowledge to create new products. The calibrated model captures salient empirical patterns, implies sustained growth through rising recombination, and highlights sharp policy trade-offs: subsidizing novelty boosts short-run growth, while subsidizing recombination raises long-run growth with heterogeneous effects across firms.

Link

7. Technological Change and the Market for Books, 1450-1550

Qiyi Charlotte Zhao

Conventional views consider printing a cost-reducing technology. This paper examines unusually granular product- and firm-level data and proposes a new framework for understanding this technology’s economic impact. I show that relative to its predecessor, manuscript production (i.e., hand-copying), printing introduced new incentives and constraints that altered both the product’s nature and the market’s structure. First, printing’s business model encouraged the production of shorter and simpler books targeting a poorer and less educated audience. Second, its cost structure led to product differentiation and prolific trade rather than direct competition and localized production, making available a greater variety of products offering diverse information and perspectives. Rather than making medieval books cheaper, printing’s core contribution to economic development might lie in fostering popular demand for literacy through a variety of simpler products that lacked an economic basis under manuscript production.

Link

8. Factories of Ideas? Big Business and the Golden Age of American Innovation

Pier Paolo Creanza

This paper studies the Great Merger Wave (GMW) of 1895–1904—the largest consolidation event in U.S. history—to identify how Big Business affected American innovation. Between 1880 and 1940, the U.S. experienced a golden age of breakthrough discoveries in chemistry, electronics, and telecommunications that established its technological leadership. Using newly constructed data linking firms, patents, and inventors, I show that consolidation substantially increased innovation. Among firms already innovating before the GMW, consolidation led to an increase of 6 patents and 0.6 breakthroughs per year—roughly four-fold and six-fold increases, respectively. Firms with no prior patents were more likely to begin innovating. The establishment of corporate R&D laboratories served as a key mechanism driving these gains. Building a matched inventor–firm panel, I show that lab-owning firms enjoyed a productivity premium not due to inventor sorting, robust within size and technology classes. To assess whether firm-level effects translated into broader technological progress, I examine total patenting within technological domains. Overall, the GMW increased breakthroughs by 13% between 1905 and 1940, with the largest gains in science-based fields (30% increase).

Link

9. Solving Problems of Unknown Difficulty

Nicholas Wu

This paper studies how uncertainty about problem difficulty shapes problem-solving strategies. I develop a dynamic model where an agent solves a problem by brainstorming approaches of unknown quality and allocating a fixed effort budget among them. Success arrives from spending effort pursuing good approaches, at a rate determined by the unknown problem difficulty. The agent balances costly exploration (expanding the set of approaches) with exploitation (pursuing existing approaches). Failures could signal either a bad idea or a hard problem, and this ambiguity generates novel dynamics: optimal search alternates between trying new approaches and revisiting previously abandoned ones. I then examine a principal–agent environment, where moral hazard arises on the intensive margin: how the agent explores. Dynamic commitment leads contracts to frontload incentives, which can be counteracted by the presence of learning. The framework reflects scientific discovery, product development, and other creative work, providing insights into innovation and organizational design.

Link

10. Quantifying Knowledge Spillovers Using Firm and Product Dynamics

Mohamad Adhami

Knowledge spillovers are a common rationale for government support of innovation, yet evidence on their magnitude remains limited. In this paper, I quantify the wedge that spillovers create between social and private rates of return to innovation. To do so, I build a novel semi-endogenous growth model featuring multiproduct firms and endogenous exit of products. In equilibrium, product exit exhibits negative selection and is preceded by a gradual decline in market share, consistent with facts I document using barcode-level data. Through the lens of the model, these dynamics of product exit are informative about spillovers: by accelerating growth in the quality of new products, stronger spillovers increase the rate at which incumbent products lose market share and exit. Since comprehensive datasets track firms rather than products, I leverage the model to infer the wedge created by spillovers from data on firm exit by age. Across U.S. private nonfarm employer businesses, I infer spillovers that drive a 16 percentage point wedge between the social and private rates of return to innovation.

Link

11. Beyond the Lab: The Effect of PhD Programs on Innovation

Manfredi Aliberti

This paper estimates the causal impact of PhD programs, designed to train individuals to advance the frontier of knowledge, on innovation. I exploit the centrally planned and staggered rollout of doctoral programs across Italian universities and construct a new dataset linking program openings to local patenting activity. The introduction of PhD programs increased patenting by 21% between 1986 and 2001. Using admission exam scores in a regression discontinuity design, I show that about 22% of this effect is direct and driven by the increased patenting of program graduates, while most of the remainder reflects spillovers to local firms. These findings indicate that PhD programs stimulate technological progress both by increasing graduates’ inventive output and by strengthening the surrounding innovation ecosystem. A cost–benefit analysis based on patent valuations suggests that the social return to these programs exceeds total costs by at least 46%. Finally, I estimate that PhD programs raised Italy’s GDP by 0.6% to 4.7% over the same period.

Link

12. The Prestige-Testability Tradeoff in Science

Kurtis A. Hingl

Where ideas are difficult to test directly, does the scientific community rely more on prestige markers to evaluate them? In this paper, I adopt the cultural evolutionary concept of “prestige,” translate it into economics through a simple reputation model, and propose this hypothesis of a prestige-testability tradeoff: scientific fields that are less testable rely more on prestige markers, manifesting a higher concentration. I present empirical evidence of this prestige-testability tradeoff in two ways. Firstly, in bibliographic data of the corpus of scientific research from 1900 to 2015, I find that the concentration of author prestige markers—citations and h-indexes—is consistently negatively associated with a straightforward measure of testability—the incidence of the word “test” in the titles—across nineteen fields and across subfields within each field. Secondly, I use the occurrence of a paradigm shift toward more testability in the mid-1990s as an event study: the “credibility revolution” in microeconomics. Though not truly exogenous, this paradigm shift reflects a testability shock that is suitably uncovered by a staggered event-study design. I find that the credibility revolution administers a leveling effect on its adopters, based on various citation metrics and share of papers in top-five journals: authors below-median pre-adoption on these prestige markers see clear and persistent increases in their prestige markers, while their above-median peers do not, which I interpret as evidence for the prestige-testability tradeoff. I argue that this prestige-testability tradeoff framework is an important lens for viewing the organization of science, an important factor in a number of science policy decisions, and likely a feature of other social learning environments.

Link

13. Ownership Structure and Economic Growth

Koki Okumura

This paper examines how the rise of common ownership affects economic growth and social welfare. We develop an endogenous growth model that incorporates three inter-firm networks: ownership, product-market rivalry, and innovation. In the model, a large number of oligopolistic firms make forward-looking R&D investment decisions, internalizing externalities on commonly owned firms arising from product-market competition and technological spillovers. We estimate the model using data on over 700 publicly traded U.S. firms with patents. Our counterfactual analysis shows that the observed increase in common ownership between 1999 and 2017 reduced the annual growth rate by 0.12 percentage points and social welfare by 0.6%. This finding suggests that, under common ownership, the internalization of the negative externality from innovation that reduces competitors’ market shares dominates the internalization of the positive externality associated with technological spillovers.

Link

14. Startup Acquisitions and Innovation in the Biopharmaceutical Industry

Hayley Wabiszewski

Regulators have expressed growing concern that acquisitions of biotechnology startups by big pharmaceutical firms may stifle innovation by removing potential competitors. This paper quantifies the dynamic equilibrium effects of such acquisitions on innovation, entry, and market structure in the biopharmaceutical industry. I construct a novel project-level dataset linking comprehensive pharmaceutical R&D data with acquisition and sales information from 2000 to 2018, which enables tracking of projects across phases of development. Descriptive evidence shows that acquired projects have lower transition rates than non-acquired startups in early phases but higher rates in later stages, with the pattern differing significantly between oncology and non-oncology markets. To separate selection on project quality into acquisition from the causal effects of economies and diseconomies of scale, I estimate a dynamic oligopoly model with endogenous drug development, startup acquisitions, and entry decisions by startups and big firms. Firms select on unobserved project quality at each phase of R&D and into acquisition, allowing higher quality projects to reach later phases of R&D and for both positive and negative selection into acquisition. The model recovers both the average treatment effect of acquisition and the average treatment effect on the treated. Counterfactual simulations of an acquisition ban show that project approvals would rise by 8–9% in small and medium non-oncology markets but fall by 5–9% in large nononcology and oncology markets. The results highlight that regulators should adopt market-size- and therapeutic-category-specific policies when evaluating or limiting startup acquisitions.

Link

15. Extreme Heat and Directed Innovation

Enjie (Jack) Ma

Can directed innovation mitigate climate damages? I provide systematic evidence outside agriculture that firms adapt to extreme heat through directed technological change. Linking firm-level production data to patent records for nine EU countries (2000–2020), I establish three results. First, extreme heat acts as a labor-biased productivity shock: labor-intensive firms experience larger losses and lose market share to capital-intensive rivals. Second, firms shift toward capital and redirect innovation toward labor-saving technologies, especially in heat-exposed, labor-intensive industries. Third, this endogenous innovation response is economically significant—labor-saving patents filed in response to heat offset 26 percent of aggregate productivity losses over the period. Overall, the results show that innovation is not merely a driver of growth but also an active margin of climate adaptation.

Link

16. How Does Industry Shape Academic Science? Evidence from “Million Dollar Plants”

Hongyuan Xia

Firms rely on academic science and actively participate in the production of scientific knowledge. However, the impact of industry on academic science remains unclear. This study utilizes the site selection decisions of “Million Dollar Plants” (MDPs) to estimate the causal effects of industry on academic science. I compare the responses of scientists in counties that successfully attracted MDPs (”winners”) with those in counties that narrowly missed out on these MDPs (”runners-up”). The arrival of an MDP in a “winner” county shifts research of local scientists toward topics relevant to the firm, but not at the expense of either the quantity or quality of their work. This shift in research direction is not primarily driven by direct funding or collaboration. Instead, it occurs immediately after the announcement but before the physical establishment of these plants and is more likely to affect scientists without prior experience in commercialization. These findings indicate that scientists are refocusing their attention toward more applied and firm-relevant research.

Link

17. How Network Hiring by Entrepreneurs Shapes Firm Formation and Performance

John F. Bonney

Many entrepreneurs rely on their personal networks to hire their first employees. How important is this practice for the formation and performance of new firms? I study this question using Norwegian administrative data that allow me to link entrepreneurs to their firms, employees, and former coworkers. To identify causal effects, I develop an instrumental variables framework that jointly models entry and network hiring, allowing for endogenous selection on both margins. The results reveal three main findings. First, each ex-coworker hired in the firm’s first year raises annual revenues in the following four years by over $250K and crowds in other hires, without reducing average productivity. Second, without the ability to hire ex-coworkers, a quarter of network-hiring entrepreneurs would not have started their firms at all. Third, counterfactual simulations show that, compared to entry subsidies, networks enable entry of entrepreneurs who create substantially more jobs, survive longer, and achieve higher value added per worker. Interpreted through the lens of a simple model, the data suggest that private information about coworker quality is a key driver of network hiring. Taken together, the results show that access to human capital through networks is an important determinant of entrepreneurial entry and success.

Link

18. Science, Startups, and the Problem of Value Capture: Thin Acquisition Markets, Weak Outside Options

Roger Masclans

Startups commercializing science-based innovations are crucial for tackling pressing challenges, yet, in critical sectors such as energy, industrials, and materials, entrepreneurial activity remains limited. This paper investigates whether weak value capture at exit constrains these ventures. I estimate value creation and capture in startup acquisitions by combining acquisition prices with acquirer stock returns, adjusting for market noise to isolate the economic signal attributable to the acquisition. Science-based startups capture 46 cents per dollar of acquisition-induced surplus, compared to 61 cents for non-science startups—a 24% penalty. Conversely, they create 20% more joint surplus, consistent with continued entry despite the capture penalty. To explain these patterns, I examine a central mechanism: the structure of a startup’s exit conditions. I argue that science-based startups face thinner, more concentrated acquisition markets and limited ability to scale independently, features that weaken the startup’s bargaining power. Indeed, I find that science-based startups face up to 40% fewer potential acquirers, who are 53% larger on average, and that their value capture is more sensitive to acquirer concentration. Concentrated markets have a dual effect: large incumbents enable greater surplus creation, but also shift bargaining power away from startups, allowing acquirers to ex tract most of the gains from innovation. Finally, I find that the capture penalty diminishes when startups can scale commercialization independently. The results suggest that constrained exit environments limit returns to science-based entrepreneurship, highlighting the importance of competitive acquisition markets, markets for technologies, and alternative commercialization pathways in incentivizing upstream innovation.

Link

Innovation Job Market Papers Continues in Part 2!

Still no new post, but a few quick updates.

First, I have new research paper out in Science called “What if the NIH had been 40% smaller?”, coauthored with the great Pierre Azoulay, Danielle Li, and Bhaven Sampat. Here’s an abstract:

Calls to reduce the National Institutes of Health (NIH) budget by 40% have surfaced amid broader federal spending debates. To gauge the potential effect of such a policy, we analyze 27 years of NIH extramural awards and their peer review scores to identify the bottom 40% of grants that would have been eliminated from 1980 to 2007, had this policy been implemented historically. We then identify new molecular entity drugs, approved in the 21st century, that are linked to these at-risk grants. Fourteen drugs have patents that directly acknowledge research support from at-risk grants, and more than 50% cite at least one research publication funded by at-risk grants. We find no evidence that grants which are linked to at-risk grants are of lower quality than grants that are not linked. These findings illustrate the foundational role of NIH support in enabling the science that underpins most drug discoveries.

Check it out. It’s only a few pages and I think it reads pretty easy! One reason I was excited about this project is that it’s very transparent. A non-specialist can very easily understand what we’re doing.

Second, for the last two years I have been doing an annual roundup of innovation job market papers. I would like to do this again. Please help me out by sending me any and all innovation-related PhD job market papers. Send me your paper. Send me your student’s paper. Send me your friend’s paper. Send me a cool paper you saw online or at a seminar. Later this Fall I’ll compile them into a post. Email me at matt@newthingsunderthesun.com.

Third and finally, posting has been light (almost non-existent) since I took on leading the Abundance and Growth Fund at Open Philanthropy, but we are nearly done hiring people to join the team, and I hope to have more time to write once everyone has started. In the meantime, I have been using lynkmi for awhile now to keep track of interesting research. It’s a website for sharing and tagging links. If you want to see links to research I think might be good fodder for New Things Under the Sun, go here.

Thanks everyone,

Matt

Hi Everyone,

In March, I became the (first interim, later permanent) head of Open Philanthropy’s Abundance and Growth Fund, a new $120+ million program to accelerate economic growth and reduce the cost of living. Now we’re looking to hire 2-4 people to join me and Jordan Dworkin (who joined last year to work on innovation policy), to do strategic grantmaking and research across a wide range of areas. In addition to innovation policy, we’re exploring supporting work related to housing, energy, infrastructure, state capacity, healthcare and clinical trials, economic dynamism, high-skilled immigration, or building abundance/progress studies communities.

To do that well, we need help! We’re planning to hire two different kinds of role, which we’re calling the specialist track and generalist track. The specialist track will focus on one or at most two areas (we are especially interested in finding someone to lead our housing work), while the generalist track will have a broader portfolio. We also expect people on the specialist track will come in with more experience (5+ years), then the generalist track (2-3 years). And we’re especially interested in folks with experience in US federal policy making.

If that sounds interesting to you, apply by July 27 for full consideration. The full job descriptions are below:

• Specialist track

• Generalist track

And if you aren’t personally interested, but know someone who might be a good fit, you can earn $5k if your referral results in a hire.

Finally, for all the rest of you New Things Under the Sun readers: posting has been extremely light since I took on this new role. The main reason for that is that there is a lot to do, and we’re understaffed. Once this team is in place, my expectation is that I’ll be able to return to writing more often, which I’m looking forward to.

Best,

Matt

Subscribe now

Consistency in the Returns to R&D

The post Government Funding for R&D and Productivity Growth reviewed a few recent papers that try to empirically measure the productivity impact of government spending on R&D. Those papers qualitatively found similar results, but it was hard to assess whether the magnitudes from each approach were consistent with each other, because they each used different methods and reported results in different ways. Now, Fieldhouse and Mertens (2025) makes some welcome progress on this question. From the updated article, the article now concludes:

So… what is the return to government funded R&D?

To answer this, recall that Jones and Summers (2021), the paper I mentioned at the outset of this article, argued via a thought experiment that a dollar of R&D on average is worth several dollars in benefits. Another way to express the benefits of R&D is via a social rate of return, which can be understood as the interest rate you would need to be offered on a conventional investment to be indifferent between it and getting the returns from R&D. By this measure, Jones and Summers argue as a baseline that the return on R&D averages 67% (much higher than the return on most investments!).

At first glance, it seems like Fieldhouse and Mertens (2023) find answers that are dramatically different—they estimate social returns to nondefense public R&D in the range of 140-210%, compared to the 67% from Jones and Summers. But in a follow-up paper (Fieldhouse and Mertens 2025), they show these answers are not that difficult to reconcile if you keep in mind a few key points.

First, the 67% estimate from Jones and Summers is an average over all R&D spending, including both government and private sector R&D. The Fieldhouse and Mertens (2023) estimate is not an overall average—it's specifically for nondefense government R&D.

To see how these different estimates can be reconciled, we need to consider the composition of R&D spending. Over the post-war period, private sector R&D has accounted for roughly 54% of all R&D spending, defense R&D for about 26%, and nondefense government R&D for approximately 20%. If we use the estimate that private sector R&D generates a return of about 55% (which they pull from another paper, Bloom et al., 2013, discussed in more detail here), and defense R&D generates a return of about 25% (which is consistent with the findings in Fieldhouse and Mertens 2023), and nondefense R&D generates a return of 175% (the midpoint of their 140-210% range), then the weighted average return would be:

54% × 55% + 26% × 25% + 20% × 175% = 71%

This is very close to the 67% estimated by Jones and Summers (2021).

What about Dyévre (2024)? I think his answers are also consistent with the conclusion that public funding for R&D is actually higher than the levels implied by Jones and Summers (2021). Here’s a simply argument to see why.

As a benchmark, suppose that 100% of annual economic growth is driven by (100% of) annual R&D. This is basically the assumption made in Jones and Summers (2021). Let’s suppose there is a constant relationship between R&D spending and growth, just as a benchmark. If that’s true, then we should expect a 1% increase in annual R&D to generate a 1% increase in annual growth. Annual GDP per capita growth in the USA has been about 1.8% per year since the 1950s, so a 1% increase in the growth rate is 0.018%. Do Dyévre’s results match this benchmark?

As noted previously, government R&D has averaged only about 45% of total R&D, so increasing overall R&D by 1% would require increasing government spending by more than 1% - specifically, 2.2%. Recall that Dyévre finds that a 1% increase in government R&D funding generates roughly a 0.024% increase in productivity after five years, so a 2.2% increase should lead to a 0.0528% increase in productivity.

In other words, Dyévre’s empirical approach finds that an increase in government spending equivalent to 1% of total R&D spending tends to lead to a 0.0528% increase in productivity growth; that’s a lot more than the 0.018% benchmark that I think is consistent with Jones and Summers (2021). One would need to do some careful work to be sure, but it also seems broadly consistent with the Fieldhouse and Mertens (2023) result which found the return to non-defense R&D was much higher than the overall average for R&D.

Read the whole post

Changes in the Composition of Government R&D

In my post Frequently Asked Questions About US Government Funding for R&D, I noted that government spending on R&D has grown more slowly than overall government spending, private sector spending on R&D, and US GDP. But looking only at aggregate funding for R&D hides some interesting dynamics; not all kinds of R&D have seen their level of government support fall, relative to GDP. In this update, I note:

There have been substantial changes in the composition of federal spending on R&D over time. Federal support for basic and applied research has been roughly constant as a share of GDP since the 1970s, at roughly $0.38 for every $100 of GDP. Meanwhile, support for development has fallen from a high of $1.50 in every $100 of GDP during the height of the space race to roughly parity with support for basic and applied research.

Read the whole post

A big retraction

To close, one of the biggest papers in the economics of innovation of the last year was Toner-Rodgers (2024), which purported to describe an experiment conducted by an unnamed materials science company on the effects of AI on R&D. The paper had a lot of interesting implications, and I wrote about it in the articles Prediction Technologies and Innovation and Do Prediction Technologies Help Novices or Experts More?, and to a more minor degree in the post What if we could automate innovation? On May 16, 2025, MIT announced it had conducted an internal confidential review, and concluded the paper should be withdrawn from public discourse. We don’t know exactly what the story here is, but MIT economists Daron Acemoglu and David Autor stated “we have no confidence in the provenance, reliability or validity of the data and in the veracity of the research.”

New Thing Under the Sun is a living literature review, and I aim for articles to reflect the current state of our knowledge. I think the current state of knowledge no longer includes the findings in Toner-Rodgers (2024), so I’ve dropped them from the review. This has had the following effects on the updated articles:

• Prediction Technologies and Innovation looked at evidence that access to a better prediction technology (which included everything from gene maps to software to AI) can inadvertently retard the pace of knowledge growth, by encouraging researchers to narrow their research agenda to domains where the new prediction technology can be useful. Papers looking at early prediction algorithms in structural biology, and gene maps illustrated that this concern about research narrowing can indeed happen. Somewhat in contrast, Toner-Rodgers had found that access to AI tools increased the novelty of materials discovered (which is not exactly the same thing as the question about narrowing). Without Toner-Rodgers (2024), the article is now a bit more internally consistent, in that all the papers discussed illustrate a narrowing effect of new prediction technologies. But the article no longer has any discussion of modern AI tools.

• Do Prediction Technologies Help Novices or Experts More? argued that prediction technologies can benefit experts or novices more, depending on what kinds of sub-problems they help solve. Most of the papers discussed found evidence for both dynamics; in some situations a prediction technology (like a satellite map or a genome association study) helps novices more, and in others, experts. Again, somewhat in contrast, Toner-Rodgers (2024) found pretty unambiguous evidence that experts benefitted from access to the AI tool. Indeed, researchers who were the least productive in the pre-AI era didn’t benefit at all. Without Toner-Rodgers (2024), I think we are kind of back where we started, in asserting that prediction technologies in general can help either experts or novices more, but now the post doesn’t discuss any articles about the effects of AI on this process. I predict we won’t have to wait too long until that changes though.

• What if we could automate innovation? is mostly a discussion of one theoretical model of innovation, which assumes innovation is comprised of many different tasks. In the model, when a subset of tasks get automated, researchers reallocate their effort away from that task and towards the ones that are not (yet) automated. The post discusses Toner-Rodgers (2024) as an illustration of this phenomenon, but the theoretical argument does not depend on it (and was written well before the paper).

So a general theme is that I haven’t dramatically revised my views on the effects of prediction technologies on innovation. We now know less than we thought about the impact of AI on innovation, but I think what we knew from this paper was always pretty provisional, given how fast AI changes. It was always more like the impact of a specific technology, at a specific time. I suspect we’ll soon have lots of new papers on these topics.

On a personal note, I’m sorry I got taken in by this paper and signal-boosted it. As with all the papers I write about, I read it carefully. At the time, I thought it was almost too good to be true - I had never seen a more comprehensive set of data - but not actually too good to be true. It now seems like I was wrong.

Should I tighten up my standards so this doesn’t happen again? I ultimately think the tradeoffs for any feasible changes I could make to avoid this kind of error would hurt New Things Under the Sun more than help it. The premise of this project is that I publish faster and with less hedging than would often be the norm in academia, but what I publish is more provisional. It’s a living literature review, with the expectation that findings will be revised as we learn more.

Should academia tighten its standards? I am in favor of a regime where the level of resources devoted to assessing a paper’s veracity is tailored to the likely impact of the research. In such a regime it is easy to do preliminary research, which ensures we get lots of shots on goal, but also we can be (more) confident in influential results. In general, I don’t think we currently invest enough in ensuring the rigor of high impact papers (we should make replication the norm there, for example). But I would not advocate for increasing the level of oversight in a preprint like Toner-Rodger (2024).

Until Next Time

Thanks for reading! If you want to chat about this post or innovation, don't hesitate to reach out at matt@newthingsunderthesun.com.

• Innovation Growth Lab is hiring for multiple roles. Applications are due May 1. Learn more here.

• The Institute for Progress is hiring a Fellow / Senior Fellow to work on Metascience. Applications are due May 11. More information here.

On to the post!

Subscribe now

This post was jointly written by me and Caroline Fry, assistant professor at the University of Hawai’i at Manoa.

This article will be updated as the state of the academic literature evolves; you can read the latest version here.

The US government plays a significant role in supporting research and development (R&D) in developing countries. The US, for example, funds international collaborations, fellowships and training programs, research infrastructure, direct grants to international researchers, and incentives for US scientists to address global development challenges. Cuts to programs like this are likely to have a significant negative impact on this kind of research.

In this post we make three points. First, the social return on investment of R&D in developing countries may be extraordinarily high. Second, a significant portion of this return flows back to the US. Third, US support is a large share of developing country R&D, but a very small share of US federal spending.

R&D in Developing Countries Probably Has Very High Social ROI

As a starting point, let us consider the returns to R&D in high income countries. As discussed in this post, a variety of evidence is consistent with every dollar of domestic R&D spending by the US government generating $2-5 of benefits via its impact on economic growth. If we try to put a dollar value on other benefits of technological progress, such as improved health, then we might plausibly double this social ROI.

We have some good reasons to believe that supporting R&D in developing countries could be just as high yielding as domestic R&D. It’s sometimes assumed that developing countries don’t need domestic science and research capabilities, because they can benefit from R&D investments made in high income countries. But that isn’t always the case. Sometimes the research conducted in one country simply isn’t applicable in others. As we discussed in our earlier post When research over there isn’t useful here, diseases, environments, and even social systems vary across countries. This suggests there is scope for R&D in developing countries to generate benefits that won’t happen by default as a result of research done in high income countries.

At the same time, there are also reasons to think the returns to R&D in developing countries might actually be higher than domestic R&D. The burden of knowledge hypothesis, for example, proposes that innovation becomes more challenging as a field grows more mature, because unsolved problems require increasingly more detailed knowledge to solve. The kinds of R&D that is done less often in high income countries like the USA, such as research on neglected diseases and agricultural systems in developing countries, might be less mature, and hence have less of a “burden of knowledge.”

Empirically, we have a few papers consistent with the idea that US funding for R&D in developing countries has a high ROI, even relative to the tall standards for domestic R&D. Alston, Pardey, and Rao (2021) focus on estimating the ROI of agricultural R&D in developing countries, both performed by the network of Consultative Group for International Agricultural Research (CGIAR) international research centers (which are supported in part by the US government), and by developing country public sectors. Across 200+ studies, the median value for the ROI of R&D at CGIAR was around $9.5 in benefits (mostly via increased agricultural production) for every dollar in spending. In a different analysis, they estimate how much agricultural production would have been lost over 1960-2015 if the productivity of agriculture in developing countries had been stuck at the level of 1961, since these are the benefits that are plausibly sacrificed without agricultural R&D. If they divide the value of this lost production by the total cost of agricultural R&D performed by CGIAR and developing country governments, this implies the developing world gained $13.70 in agricultural production for every dollar invested in R&D. Whichever approach one takes to estimate the ROI of agricultural R&D in developing countries, the results compare favorably to the average of around $2-5 that studies have found for the marginal dollar of US R&D.

While Alston, Pardey, and Rao (2021) estimate the historic returns of agricultural R&D in developing countries, Rosegrant et al. (2023) look forward. They consider various scenarios around how increased spending on agricultural R&D in the developing world translates into yield gains, and then use economic models of the agricultural economy to estimate the value of these gains. This exercise finds even larger gains, with every dollar of R&D generating $29-$35 in benefits.

Some other studies also find high returns in other domains. Fry and Ganguli (2024) look at the impact of the AIDS International Research Training Program run by NIH, which provides training at top US institutions to LMIC scientists.1 Ganguli and Fry show that when program graduates of the program return to their home institutions, their peers, working the same department, also appear to benefit. Ganguli and Fry’s model implies this training program induces new scientific publications at a cost of around $10,000 per publication. This is potentially orders of magnitude cheaper than the cost per publication in the USA.2

Finally, a 2024 report by Policy Cures estimated that a set of medical advances for the treatment of neglected diseases saved 598mn years of life (across millions of people who did not die prematurely) between 2000 and 2024. Policy Cures estimates the cost of R&D for all research related to the neglected diseases was $97.9bn between 1994 and 2022. If we value a lost year of life at $20,000 (which they argue is reasonable, though note it is far less than values used in the USA), then this implies a social ROI of roughly $122 in benefits (over 2000-2024) for every dollar spent on R&D for neglected diseases (over 1994-2022). And they also note we can expect far more lives to be saved in the future from these drugs. That said, it is important to note that while this research targeted developing country problems, a large portion of this research was conducted in high income countries. Still, it is suggestive of the kinds of social returns that are possible with the kind of R&D that would likely be prioritized by developing countries.

US Self-Interest and R&D in Developing Countries

Of course, the problem with R&D has always been not that it has low returns, but that most of the benefits flow to people who aren’t the R&D performer. That’s probably even more true for R&D performed abroad: a large share of the benefits of US funded R&D performed in developing countries do not directly flow to US taxpayers. Nonetheless, it’s very likely the US captures at least some fraction of the benefits. Given that those benefits may be extremely large, capturing even a small fraction of them can mean supporting developing country science is a good investment even when only considering US interests. Here we lay out some of the channels through which the US benefits from R&D in developing countries.

To start, it has been well documented that R&D is characterized by large “knowledge spillovers,” wherein knowledge discovered by one R&D investment is useful to actors who did not perform the R&D.3 Scientific knowledge and technological inventions discovered in other countries can be adapted for US use. While the US will probably capture a smaller share of spillovers compared to domestic R&D, this may be offset by a higher social ROI.

Moreover, support for R&D in developing countries can also generate benefits not achievable by funding strictly domestic R&D. For example, HIV/AIDs and Covid-19 illustrate how diseases originating anywhere can become global pandemics. Yet local research capacity is often the first line of defense. Fry (2023),4 for example, shows that researchers in countries where Ebola outbreaks actually happen are much more likely to switch toward working on that topic (often in connection with foreign researchers), compared to similar scientists working in unaffected countries. More generally, several of the mechanisms deployed by the US to support developing country R&D facilitate the formation of international collaborations.5 This facilitates US researchers’ access to data and expertise not available domestically.6

Finally, we can think of support for developing country R&D as a form of humanitarian aid rather than purely as a form of R&D. Humanitarian aid can be justified with a variety of self-interested reasons (in addition to pure altruism which is itself a laudable objective). Global prosperity may reduce conflicts that affect US interests, or grow markets for US export. Aid is also a diplomatic tool for soft power and political influence. Indeed, this can be particularly true of international science, which tends to involve international collaboration.7 Finally, to the extent it promotes economic development, support for science can mitigate the need for altruistic humanitarian aid in the future.

The US Role in Developing Country R&D

Cuts to aid from the US will likely have a dramatic and negative impact on developing country R&D. The full extent of this support is difficult to quantify, due to the decentralized nature of funding and the diverse channels through which it is distributed across agencies and initiatives, but to get a sense of the scale of this funding that is at risk, we can look to the numerous US agencies that contribute to supporting science in developing countries. To begin, USAID allocated around $220 million to R&D in 2022, of which around 11 percent was likely directly granted to researchers in developing countries (based on the average rate of local funding at USAID reported in early 2025; this does not include funding provided to local researchers via grants to US institutions). Meanwhile, the Fogarty International Center (FIC) at the NIH, which supports global health research and capacity building, had a budget of approximately $95 million in 2023. It’s not clear how this compares to overall funding for research, but in 2022, US public funding for neglected disease research accounted for $1.918 billion, representing more than 76% of the world's investment in neglected disease research, though not all of this is performed in developing countries.

Recent US changes have already impacted global science investments. For instance, USAID funds 21 US-based Feed the Future Innovation Labs, which conduct research domestically while also partnering with universities in developing countries for fieldwork, training, and dissemination. This initiative supports the kind of agricultural research that we noted above has particularly high social returns (see Dalton and Fuglie 2022 for a discussion of the returns to this particular program). These initiatives were put on hold earlier this year, and its unclear how many of them have resumed operations. Another example of the impact of shifts in funding priorities include the termination of the USAID Strategic Partnerships for Enhancing Research and Knowledge (SPARK) program that began in 2024. SPARK, and its predecessor PEER, represented unique mechanisms for directly funding LMIC-based scientists, fostering local research careers, and addressing regionally relevant challenges. The reduction or delay in such initiatives signals potential challenges for sustaining scientific progress in these regions.

To close, there are some profound asymmetries in play here. While US support for R&D in developing countries is a large share of total developing country R&D, it actually represents only a tiny fraction of overall US science and development expenditures. USAID’s total budget in 2022 was $74 billion, meaning its R&D support accounted for less than 0.3% of total agency spending. Similarly, the Fogarty International Center’s $95 million budget constituted just 0.2% of the NIH’s total $47.7 billion appropriation in 2023.

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.

But before we get into that, some announcements.

Most importantly, my job is (temporarily) changing in ways that affect New Things Under the Sun. Open Philanthropy (my employer) has launched a new Abundance and Growth Fund, which will spend $120mn over the next three years to accelerate economic growth and boost scientific and technological progress while lowering the cost of living. I’ll be the interim lead of this program while we conduct a hiring round.

During that time, I will probably write a bit less, and less of what I do write will be for New Things Under the Sun. Instead, I plan to write more about specific ways to accelerate economic growth and boost scientific/technological progress, drawing more on the internal research I and others do at Open Philanthropy. I think this state of affairs will last a few months, and then I’ll reassess.

Other Announcements:

• Speaking of other writing, I have a new piece at Notes in Progress titled The Value of Technological Progress: Evidence from the Life of Matt Clancy. It dropped today; check it out!

• The International Conference on the Science of Science is accepting submissions for the 2025 conference this summer from now until March 21. More here.

• The Experiment Foundation is seeking researchers to collaborate on evaluating its Science Angel program. Applications due March 28. More here.

• As noted above, Open Philanthropy is launching the Abundance and Growth Fund, and seeking a Senior Program Officer to lead its strategy and grantmaking. Refer a strong candidate and get a $5k reward if Open Phil hires them. Learn more and apply here by March 31st.

• Study the impact of AI on the science and research ecosystem with an AI early career fellowship! There are a few funding organizations, including UKRI, the Alfred P. Sloan Foundation, and the Social Sciences and Humanities Research Council (SSHRC) of Canada. Applications due April 10. More details here.

• The Metascience Alliance, a new cross-sector network fostering collaboration to better understand and improve the research system, is looking for your feedback! Read more here.

Now for some updates to the living literature review!

Subscribe now

Starbucks and Innovation

The post Urban social infrastructure and innovation looks at some evidence that cities serve as a place where the intermixing of people and their ideas facilitates innovation. I’ve added to this post two recent papers that look at the influence of Starbucks(!) on innovation. The updated post picks up after discussing a paper on the impact of bars on innovation during the prohibition era.

Today, it’s not so obvious that bars would serve the same third place function as they did in prohibition. Instead, coffee shops might fill that role.

Starbucks provides a particularly useful case study of modern third places. At the time it was founded, coffee shops as a place to spend time and meet other people were not very common, and Starbucks saw part of its value proposition to customers to be providing this. The company's expansion from Seattle provides a natural experiment: the first Starbucks opened in Seattle in 1971, but it wasn't until 1993 that the company opened its first East Coast location in Washington DC.

Two recent papers leverage this rollout of Starbucks locations to assess how this kind of social infrastructure affects innovation. Andrews and Lensing (2024) look at the effect of getting a local Starbucks on patents, while Choi, Guzman, and Small (2024) examine impacts on the rates of entrepreneurship. Both papers want to compare metrics of innovation in localities that have more Starbucks to localities with fewer, but face the same issue: Starbucks doesn't randomly choose where to open new stores. The company probably tries to open stores in neighborhoods that are either doing well, or are forecast to be doing well - exactly the kind of places that might have gotten more innovative anyway, even without a Starbucks.

Each paper tackles this challenge differently. Andrews and Lensing exploit the fact that Starbucks expanded geographically outward from Seattle. The basic idea is that the further you are from Seattle, the fewer Starbucks you are likely to have, but this effect diminishes over time. In some analyses, they use distance from Seattle to predict how many Starbucks a place would be expected to have in a given year, and then use those predictions to assess the impact of Starbucks on patenting. The key assumption is there aren't other factors that affected patenting and spread out concentrically from Seattle over a few decades; just Starbucks. As a bit of evidence that this theory is correct, they show in one analysis that if you try to predict patenting using distance from Seattle (and time) in the twenty years *before* Starbucks was founded, there's no obvious pattern; you only see distance from Seattle over time mattering during the era of the Starbucks rollout.

Meanwhile, Choi and coauthors take a different approach, and compare census tracts that got a Starbucks to similar ones that didn't. In some cases, they specifically compare tracts that got a Starbucks to ones that were supposed to get one, but where the deal fell through for idiosyncratic reasons (like zoning board rejections). The idea here is that these pairs of census tracts were similarly attractive to Starbucks, but only some actually got one.

Under their preferred statistical models, both papers find substantial effects. Andrews and Lensing find that when the growth rate of new Starbucks establishments in a county increases by 10%, patenting in that county increases by 3-4%. Choi and coauthors find getting a first Starbucks in your census tract leads to 5-15% more startups in that tract over the next seven years.

Those are big effects - do we really believe they're the result of complicated (for the time) coffee orders? Both papers present additional evidence to try and shore up our confidence. For example, Choi and coauthors find particularly large effects - nearly 30% more startups - in an analysis of locations opened through a partnership between Starbucks and Magic Johnson. This partnership specifically targeted low-income minority neighborhoods that Starbucks probably would not normally have entered. That suggests the benefits of getting a Starbucks aren't merely about the company's ability to forecast which neighborhoods are about to take off.

There's also evidence that these effects really are driven by people meeting and sharing ideas at Starbucks. Andrews and Lensing find patents are more likely to cite other patents from the same state after Starbucks arrives, consistent with more local exchange of ideas. Choi and coauthors find you get similar effects for other chains that try to be third places, such as Caribou Coffee, but not for chains like Dunkin' Donuts that serve coffee but aren't set up for extended seating. They also find no effect from small Starbucks locations (like the ones inside Target stores) that don't facilitate interaction.

Lastly, another reason for the big effect might be that it's not all about connections made at Starbucks. Andrews and Lensing find that when Starbucks enters a market, it tends to predict additional coffee shops opening up too. So it might be that the increase in interactions happens at coffee shops in general, and that Starbucks is just a catalyst to jumpstarting a coffee scene. That's consistent with another result from Choi and coauthors: there is no impact on entrepreneurship when Starbucks enters a census tract that already had a coffee shop.

Read the whole post

A New Way to Measure Unusual Combinations of Research Ideas

The post The best new ideas combine disparate old ideas looks at a bunch of evidence that papers which forge unexpected connections between different ideas tend to be higher impact, as measured in different ways. In this literature, it’s always a challenge to find ways to measure the “ideas” that papers are combining. A variety of approaches have been used over time. In this update, I add discussion of a new one.

… Yu and Romero (2024) … look at the datasets used by papers. The data used in a paper represents another concrete way to track what ideas and evidence it draws upon.

Yu and Romero study over 30,000 papers that use datasets from the Interuniversity Consortium for Political and Social Research (ICPSR), a major repository of social science data. Most papers (71%) use just one dataset, but those that do combine datasets in unusual ways tend to receive more citations. For each standard deviation increase in how atypical a paper's combination of datasets is, papers receive 18.4% more citations over three years. The effect is similar over five and ten year windows. These papers are also more likely to be discussed outside of academia: papers with more atypical combinations of datasets are significantly more likely to be mentioned on Wikipedia and Twitter.

Read the whole post

Twitter and Citations to Economics Research

The post Twitter and the spread of academic knowledge looks at a variety of studies that try to assess whether tweeting about papers raises awareness of research by other academics. One claim I made in that post was that if tweeting about a paper boosts citations on the order of 10%, which several studies suggest might be the case, then existing studies were too small to detect that effect with high confidence. A new paper supports that claim.

Sofer (2024) offers the largest study to date on this question, examining nearly 3,000 National Bureau of Economic Research (NBER) working papers published between 2015 and 2018. The NBER working paper series is a prestigious platform where economists share research before formal peer-reviewed publication. Usefully for Sofer, all NBER working papers receive a tweet from the NBER communications office, regardless of author preferences.

Similar to Chan et al. (2023) [discussed more in the main post], Sofer exploits the quasi-random timing of tweets to identify variation in tweeting that isn’t related to paper quality. Rather than just looking at day of the week effects [as Chan did], Sofer develops two instrumental variables: a "Twitter Attention Index" that measures economists' activity levels on Twitter at different hours of the day and days of the week, and a "News Pressure Index" based on the Wall Street Journal's front page "What's News" section, which indicates when major economic news events might be diverting attention from academic content. He shows that when the NBER tweets about a paper during hours when economists are highly active on Twitter (typically weekday mornings around 9 AM), that paper receives significantly more engagement. Conversely, when a paper is tweeted on days with major economic news events, it receives less attention. Sofer uses the timing of tweets to predict twitter activity, with no reference to a paper’s underlying quality, and then looks to see if these predicted tweets are correlated with subsequent citations.

Sofer finds different effects across different time horizons. In the first year after publication, increased Twitter visibility doesn't significantly affect academic citations. However, by the fourth year after publication, papers with higher Twitter visibility (as predicted by the time of day NBER tweeted about it) receive about 1.44 more citations – an effect that's (finally) statistically significant at the 5% level.

Read the rest of the post

Measuring Disruption

The post Science is getting harder looks at a number of different indicators that I interpret as indicating it’s getting more challenging to make scientific discoveries with the same impact as the past. One of these indicators is the consolidation-disruption (CD) index. There has been some debate about how this indicator has changed over time and how to interpret those changes, and I’ve updated the post to reflect the evolution of the debate.

Note however, that my target reader for New Things Under the Sun is a non-specialist interested in what the current state of the evidence says, not so much the history of how we got to our current understanding. Accordingly, I give the debate short-shrift in the update. However, if you want to see how I’ve covered this update here was the original coverage and here was a major update. I will update this post again as the research evolves.

To see the basic idea of the CD index, suppose we want to see how disruptive is some particular paper x. To compute paper x’s CD index, we would identify all the papers that cite paper x or the papers x cites itself. We would then look to see if the papers that cite x also tend to cite x’s citations, or if they cite x alone. If every paper citing paper x also cites x’s own references, then paper x has the minimum CD index score of -1. If some papers cite x and no papers cite any of paper x’s references, paper x has the maximum CD index score of +1. The intuition here is that if paper x overturned old ideas and made them obsolete, then we shouldn’t see people continuing to cite older work, at least in the same narrow research area. But if paper x is a mere incremental development, then future papers continue to cite older work alongside it.

Park, Leahey, and Funk (2023) compute the CD index for a variety of different datasets of academic publications, encompassing many millions of papers. They find the average disruption index for a paper has declined substantially since the 1940s. But we need to be cautious about interpreting a declining CD index score as evidence that the average paper has, in fact, become more incremental and less disruptive over time. This is because changing citation practices can also impact a paper’s disruption score. In particular, all else equal, it can be shown that disruption scores will tend to decline if papers make more citations, and if they cite older papers. Both have happened (see the second appendix for more discussion), so we need a way to try and separate out the decline in the disruption index that we can attribute to changing citation norms, and the share that we can attribute to a decline in actual disruption.

This paper generated a lot of debate, and in a 2025 response, Park, Leahey, and Funk perform an analysis that tries to separate out the influence of changing citation practices and disruption. The key idea is to generate 10 placebo citation networks, where citations are not informative about the extent to which a paper is disruptive, but which do exhibit changing citation practices. They can then see if the actual decline in disruption exceeds the decline observed in the placebo networks. To do that, they randomly shuffle the actual citations papers make to other papers. So instead of paper y citing paper x, they redirect the citation so that paper y now cites some other paper z, where z is some random paper published in the same year as x. This kind of reshuffling preserves the tendency over time of papers to cite more references and to increasingly cite older papers. But since the citations are now random, they should not convey any real information about actual disruption. These placebo networks will exhibit a decline in disruption, for the reasons noted above related to changing citation practices, but we can see if the actual decline in disruption exceeds this or not.

In the following figure, drawing on 25 million papers from the web of science, Park, Leahey, and Funk (2025) look at the average disruption of a paper over time, controlling for the disruption score of the same paper in the placebo networks. For example, a paper that cites a lot of papers, or a lot of older papers, might be more likely to have a low disruption score, purely as a consequence of this citation strategy. But that will be reflected in the disruption score of that paper in these placebo networks (which preserve the growing tendency to cite more and older papers). As indicated in the figure, actual papers still exhibit a decline in disruption relative to these placebos. While changing citation norms definitely contribute a lot to the decline in the disruption index, Park, Leahey, and Funk still see a decline when we try to control for those.

(I think it’s also worth pointing out that this entire simulation approach takes it for granted that papers will increasingly cite more and older work, and looks for a decline in disruption above and beyond what we would expect purely from that fact. But as argued elsewhere, the shift towards citing older work is itself potentially evidence that something is changing in science.)

Read the rest of the post

Until Next Time

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.

Table of Contents

1. How much does the US federal government spend on R&D?

2. What kind of R&D does the government fund?

3. What’s a reasonable guess at the average return on investment (ROI) of R&D?

4. Has the average ROI of R&D spending changed over time?

5. What’s a reasonable guess at the marginal, rather than average, ROI of government R&D?

6. Is government support for R&D self-financing?

7. Can private sector research substitute for government research?

8. If R&D is so good, why isn’t our government already doing even more of it?

9. If R&D is so good, why does science have so many well publicized problems?

10. What’s your preferred policy?

To make this document more skimmable, after each question I’ve written a very short high level response in a quote block (aiming for the length of a tweet). A more detailed response is given below each quote block.

If you have additional questions you would like to see answered, please email me!

How much does the US Federal Government spend on R&D?

The US government spent ~$160bn on R&D in 2022, about 2.6% of government spending. As a share of GDP or federal spending, federal support for R&D has declined for decades. We spend more on R&D in absolute terms than other big R&D producing countries, but are middling as a share of GDP. The private sector currently spends about 4x as much.

More details:

In 2022, the federal government spent about $160bn on R&D. This is about 0.6% of GDP. For comparison, the private sector spent about $673bn on R&D in that year, which is about 2.6% of GDP. In inflation-adjusted terms (pictured below), federal R&D spending has risen substantially since 1953.

To put this into context, it’s useful to consider it relative to three other benchmarks that have also increased over this period. In the following figure, I look at government support for R&D relative to private sector R&D spending, total government expenditures, and GDP. As the figure makes clear, while government support for R&D has grown in inflation-adjusted terms, that growth has failed to keep pace with the growth of these other benchmarks.

Federal support for R&D was extremely elevated during the space race - in 1964, we were spending more than $2 on R&D for every $1 of private sector R&D and for every $20 of federal spending - but this fell rapidly. By 1990, the Cold War was over and the federal government was spending about $0.75 on R&D for every $1 of private sector R&D, about $0.95 out of every $20 of government spending, and about $1.03 out of every $100 of GDP. By 2022, this has declined even further, to $0.24 for every $1 in private sector spending, $0.52 out of every $20 in federal spending, and $0.61 out of every $100 in GDP.

We can also compare the US to other countries for more context. The US government leads the world in the absolute level of R&D it supports. The following figure gives government spending totals for the 8 countries or regions that account for the largest share of global R&D spending in 2021 (amounts adjusted for purchasing power parity).

However, as a share of GDP, the US is more middling among major R&D performers.

In 2021, the Federal spending on R&D was about 0.7% of GDP. In contrast, the South Korean government spent slightly more than 1.1% of GDP on R&D, which are levels the US has not been at since the late 1980s. The German government also spent a significantly larger share on R&D, at over 0.9% of GDP.

What kind of R&D does the government fund?

About 75% of R&D spending is split roughly evenly between the Department of Defense and Department of Health and Human Services (mostly NIH). About 75% of the rest is split between the Department of Energy, NASA, and the National Science Foundation.

More details:

In 2022 (see table RD-12), about half of R&D was devoted to experimental development (the majority at the Department of Defense), and the other half to basic and applied research. Focusing on the research side (see table SRD-7), about 45% of research funding was spent on the life sciences, the majority at the National Institutes of Health. The rest is largely split up between the Department of Energy, Department of Defense, NASA, and the National Science Foundation. The biggest fields of study, after life science, are engineering and the physical sciences.

What’s a reasonable guess at the average return on investment (ROI) of R&D?

Averaged across government, the private sector, and other non-profits, the ROI is probably about $5.50 for every dollar of R&D, if you focus only on GDP. If you put a dollar value on other benefits of R&D, I think $11 for every dollar is reasonable.

More details:

While it is very challenging to estimate the benefits from a specific R&D program (although many have tried, as I will discuss), it is actually relatively straightforward to estimate the average ROI of research in general. The fundamental argument is a pretty simple one. If we believe that economic growth is derived from technological progress and technological progress is derived from spending on R&D and investment costs associated with building new technologies, then we can compute the average return on investment of R&D by dividing the value of growth by the cost of R&D and investment. You have to do some standard accounting adjustments to account for the fact that costs and benefits occur over different points in time, but Summers and Jones (2021) walks through how to do this calculation.1 If you assume a constant share of GDP devoted to R&D and investment generates constant exponential growth, which has been very roughly true for decades, then Jones and Summers (2021) shows the average benefits-to-costs ratio is given by this formula:

Benefits-cost-ratio = g / (s * r)

Where g is per capita economic growth, s is the share of GDP spent on generating technological progress, and r is the interest rate.

We will focus on the USA. In that case, g has averaged 1.8% since the 1950s and a pretty standard value for the interest rate r in economics is 5%. To estimate s, let’s add together the share of GDP devoted to R&D, which has averaged 2.5%, and the share of GDP devoted to embodying new ideas in technology. Since the embodiment of growth enhancing ideas into technologies is overwhelmingly performed by the private sector, Jones and Summers suggest that net private sector investment is a reasonable proxy for this value. That has averaged 4.0% of GDP since the 1950s. Adding together R&D and private net investment, we get 6.5% as the share of GDP devoted to realizing technological progress.2

Taken together, we have 1.8% / (6.5% * 5%) = $5.50.

However, this formula only captures benefits that are reflected in GDP per capita. Nordhaus (2005) argues health gains range from 59%-126% of the value of income gains over the 20th century. I think we should err on the upper end of this range, because health is not the only benefit of technology not captured by GDP.3 Accordingly, I think it’s reasonable to assume the total benefits of technological progress are double the purely monetary value and average out to about $11 per $1 spent.

Has the average ROI of R&D spending changed over time?

There is some suggestive evidence that the ROI of all R&D has declined since the early 2000s, though it still remains the case that a dollar of R&D probably generates several dollars worth of economic growth benefits. It seems probable this will reverse as a consequence of maturing AI, but it’s hard to say how much.

More details:

We can use the previously discussed Jones and Summers (2021) formula to see if the ROI of R&D has changed; this will be only suggestive though, because there are various lags between when R&D is performed and when growth is affected. In the figure below, I compute the benefit cost ratio using their formula, but instead of using the long-run average value for g (the per capita growth rate) and s (the share of GDP invested in technological progress), I use the average over only the preceding 15 years. This lets us see if the ROI of R&D has changed.

It has; most notably, it has declined noticeably since the early 2000s.

There has been a decline in the average ROI of R&D from a peak of nearly $8 in 1998 to roughly $4.25 in the most recent year (or $8.50 if we count non-monetary benefits). Mechanically, this decline is driven by the growth rate of GDP per capita falling, rather than an increase in spending on technological progress. Indeed, the latter has also been declining, as net private sector investment falls, but the rate of GDP per capita growth decline has been more rapid over this period.

(Note that despite the recent decline, the data here remains consistent with every dollar spent on R&D generating substantial net benefits.)

Looking to the future, it seems likely to me the declining ROI of R&D will reverse as AI technology continues to mature and diffuse. It’s difficult to say how much though. Assuming we do invent AI that can perform most of the cognitive work of R&D reliably, we can turn to economic theory about what happens when some parts of the innovation process become automated. These theories generally tell us that when some tasks are automated the long-run impact depends on other potential bottlenecks. For the application of AI on innovation, we may be bottlenecked on the supply of data, the pace of iterative real-world testing, and the diffusion of new technologies, to take three examples.4 But we’re in a bit of uncharted territory here, and it’s hard to know how significant these bottlenecks will turn out to be.

What’s a reasonable guess at the marginal, rather than average, ROI of government R&D?

This is very challenging to estimate, but a variety of research points to an additional dollar of government sponsored R&D generating $2-$5 in benefits via economic growth.

More details:

If you’re unfamiliar with the term, the ROI of the marginal dollar is the ROI of the last dollar spent. It’s harder to estimate, but more informative about the impacts of an increase or decrease of R&D spending. In general, we would expect the marginal ROI of R&D to be lower than the average, since we probably fund the most promising R&D first and the least promising R&D last. Assuming we have at least some ability to differentiate promising and unpromising research, increased spending will have a lower ROI than average, since the additional funding will be spent on less promising projects.

Meanwhile, government R&D might be different than the average for all R&D if, for example, the government on average selects better or worse R&D projects, or if increases in government funding crowd out or crowd in private sector R&D.

We have a few different papers that look at the impact of marginal government R&D funding for specific programs. Azoulay et al. (2019) focuses on NIH research, estimating that $1 in additional research funding from the NIH (on the margin) is associated with roughly $2 to $3 in additional pharmaceutical sales for newly discovered drugs. This is an underestimate of the total value created, since it is restricted to the pharma sector, for drugs that get patented (not all do).

Some other papers (discussed in this post) investigate programs like the SBIR program (which largely funds R&D by small firms). These studies compare firms that barely win one of these grants to ones that barely lose, again zeroing in on the impact of government R&D support for marginal firms. One of those papers5 estimated that a €1.6mn grant generally yielded €2mn in value to the winning firm after a few years, using a conservative valuation of the resulting patents. But that is only the value to the firm receiving the grant, and another paper about these programs6 found spillovers to other firms increased this value by 2-4x. That implies the marginal SBIR grant generates $2.50 to $5 for every dollar.

A different approach to this question is to compare the impacts of marginal versus average returns to R&D. Park, Lee, and Kim (2015) also looks at NIH research funding, comparing the publication outcomes of research funded under its normal budget to research funded by an unexpected budget increase (which was part of the stimulus to combat the 2008 financial crisis). Depending on how you measure outcomes, the marginal research (which was funded by the budget increase) had between 7% and 43% worse outcomes than normally funded research. Applying that to my estimate that the average returns to R&D are roughly $5.50 for every dollar in spending implies we should get $3.14 to $5.12 in benefits for a dollar of additional R&D spending.

Finally, both Dyèvre (2024) and Fieldhouse and Mertens (2024)7 both look at the overall productivity effects of US government R&D over the last 70 years. Both of these are primarily based on the effects of marginal changes to R&D; when agencies increase or decrease their R&D budgets. Those effects are quite large and consistent in magnitude with the average estimates implied by Jones and Summers (2021).

Now; this kind of empirical work is really tricky and the above results should be taken with some humility. However, taken together, we have a few program specific studies consistent with a marginal ROI of government R&D spanning $2 to $5.12, against an overall average for all R&D of roughly $5.50, as implied by the Summers and Jones (2021) calculations. This is consistent with the marginal ROI of government research being 36-93% of the average ROI of all R&D.

(Note again that none of these estimates put a dollar value on the benefits of innovation not captured by GDP; my preference is to double these estimates to account for these other benefits of technology, but reasonable people can disagree.)

Is government support for R&D self-financing?

Possibly. Some studies indicate government sponsored research generates enough future taxes, via its impact on growing the economy, to entirely pay for itself. But other studies suggest otherwise. Further study is probably a good idea.

More details:

If government support for R&D has a sufficiently large impact on economic growth, then it is possible that R&D spending may eventually increase the tax base enough to pay for its own costs. I don’t think our evidence on this is strong enough for me to insist this is the case; but I think it quite possibly true and worth further study (see Gullo et al. 2025 for a much more thorough discussion of related issues).

To get the ball rolling though, consider an argument for and against this possibility.

The case against is relatively straightforward. In the last question, I pointed to a variety of studies that indicated the marginal dollar of R&D generated $2-$5.12 in economic benefits. Given US taxes have averaged roughly 17% of GDP, these would indicate $1 of R&D generates about $0.34-$0.87 in future tax revenue, which is insufficient to finance itself.

The case in favor comes from Dyèvre (2023) and Fieldhouse and Mertens (2024), the most comprehensive studies of US government funding for R&D. To start, observe that Dyèvre (2024) finds that a 1% increase in government R&D spending is associated with a 0.025% increase in productivity after 5 years. In 2022, a 1% increase in government R&D spending is about $2bn. We can use this tool by the Congressional Budget Office to estimate the ten-year budget impact of a 0.025% increase in the productivity growth rate that takes place in the fifth year of a ten year budget. The tool implies this additional $2bn in R&D spending would reduce the deficit (over ten years) by $7bn! And when Fieldhouse and Mertens (2024) model the long-run impact of R&D, they too note “[these] estimates also suggest that government funding of nondefense R&D is self-financing from the perspective of the federal budget, at least in the long run.”

Can private sector research substitute for government research?

The private sector probably leaves some R&D avenues under-explored, given the extent to which R&D benefits other firms. This is supported by data on the differences between private and publicly supported research, and by the growth impacts of government funded R&D.

More details:

Even as government funding for R&D has declined as a share of GDP, private sector funding has increased by a more than offsetting amount, such that total R&D as a share of GDP has increased. This is good; I think government funded R&D is most valuable in an environment where the private sector is also doing extensive R&D.

Even so, R&D is characterized by a free-rider problem. New knowledge tends to have wide application, and it tends to be hard to prevent people from using those ideas. The existence of these “knowledge spillovers” is a classic reason why there may be underinvestment in R&D. When the private sector decides how much R&D to do, it weighs the costs it bears against the benefits it expects to receive - not the benefits all firms expect to receive. But this argument also implies the federal government is uniquely positioned to benefit from funding R&D, since it can capture these spillovers via the tax system. When a firm invests in R&D, a big chunk (possibly the majority) of the benefits flow to other US firms. A private investor doesn’t see any upside from that, but a government investor does.8

The size of these knowledge spillovers appear to be substantial (see my post Knowledge Spillovers are a Big Deal for an overview of some evidence on its scale), and tools that firms use to try and mitigate them (like patents) don’t work that well, especially for some kinds of R&D. As a consequence the private sector doesn’t invest in a lot of research that would be valuable to society, but for which they can’t capture enough of the benefits to make it profitable.

You can see this in a few places in the data.

First, the government has tended to specialize in research that has more spillovers than the private sector, suggesting those kinds of research are affected most adversely by these issues. To take one example, about one third of government R&D is characterized as “basic” research (rather than applied research or development), but only 7% of private sector research is (see Table RD-3). Basic research is associated with scientific research, and a variety of papers have found this kind of research is particularly valuable for innovation.9

Second, more specific to government funded research, Dyèvre (2023) also looks at the patents held by government agencies and finds they differ from the typical private sector patent in a variety of ways. For example, they’re more likely to be cited by a wide range of patent technology classes, more likely to cite the scientific literature, and more likely to anticipate new categories of technology.

Third, as we’ve seen, studies that look at the impact of increases in government spending on national productivity tend to find it has large positive effects. That’s consistent with there being a lot of valuable lines of research out there that were not being adequately explored prior to the government expansion.

If R&D is so good, why isn’t our government already doing even more of it?

My guess: even though R&D has a high ROI on average, it’s hard to make the case for any specific program. Even expert peer reviewers often disagree on what is good research, and tangible benefits arrive after a long delay and are often too diffuse to easily pin back to any particular R&D program.

More details:

The same factors that make it difficult to estimate the ROI of R&D make it difficult for people to attribute tangible benefits arising from an R&D program to a specific program. First, evaluating the quality of research is very challenging, even for domain experts. The post What Does Peer Review Know? reviews a variety of studies that document the relationship between expert peer review scores and subsequent impact is very noisy (though it is positive).

Eventually, R&D may yield unambiguous benefits in the form of new technologies that give us new capabilities, but by this point it can be hard to attribute these back to underlying R&D. Partially this is due to the extent of R&D spillovers and partially due to the long lags between when R&D is conducted and when technologies are developed. These problems are probably more acute for the kinds of R&D the government is best suited to pursue, such as basic science. In the post, How long does it take to go from science to technology? I summarize a variety of evidence pointing to 20-year lags as being reasonable for basic science (though note this is only a minority of government funded R&D).

Given all that, I think it is somewhat remarkable that we do, in fact, spend more than one hundred billion of dollars per year on R&D! At the same time, it’s not surprising that the share of federal spending devoted to R&D has been steadily declining for decades.

If R&D is so good, why does science have so many well publicized problems?

Science is done by people, and things done by people usually have problems. But two things are true. Despite its problems, the ROI of R&D is high. And we can and should try to make the ROI of R&D even higher.

I plan to write a lot more about ways to improve R&D efficiency this year, so watch this space.

What’s Your Preferred Policy?

I would prefer to increase (and certainly not decrease) government support for R&D. That said, I would be disappointed if we increased R&D but left our current innovation ecosystem fundamentally unchanged. The returns to R&D are high; but I think the returns to R&D about doing R&D are even higher.

More details:

The case for increasing government support research for R&D seems pretty clear to me. A variety of research suggests the marginal R&D of government funded R&D is high. Meanwhile, the government is well positioned to be the primary investor in American innovation, since it has claims on a broad array of the benefits of R&D, via the tax system (it’s possible government R&D is even self-financing). Besides, increases would hardly be unprecedented. Current levels of R&D support are below international and historic domestic benchmarks.

That said, it is challenging to know just how much R&D should. There simply isn’t good evidence on how the ROI of R&D changes over very large swings in R&D support. But it seems likely that the ROI would decline as we spend more. I doubt we could increase R&D 10x and still have it earn a good ROI, but it would be great to try much smaller increases.

But I don’t want to leave the impression that I think the US innovation ecosystem is beyond reproach, and that we should just write it a blank check. Instead, I think we should try to increase the productivity of the innovation sector, just as we try to increase the productivity of other sectors of the economy. To do that, we need to invest in technological progress in innovation itself.

Here’s a simple illustrative example. Suppose we assume that a dollar of R&D generates $5.50 worth of economic growth, via our ability to more productively use the factors of production. Let’s assume R&D about R&D has a similar efficacy: every dollar invested in meta-research generates $5.50 worth of increased research output, via our ability to more productively use the inputs to innovation and R&D. If each of those dollars worth of research output itself generates $5.50 in economic growth, then a dollar invested in meta-research ultimately generates 5.50 x 5.50 = $30.25 in economic growth.

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.

What’s the return on government funding for research?

There are a few places in the academic literature you can look to for insight. Jones and Summers (2021) uses a hypothetical thought experiment to make the case that, on average, every dollar of R&D spent probably generates several dollars in benefits via its long-run impact on economic growth (see What are the returns to R&D? for more discussion). But that result applies only to R&D in general, government and non-government, bundled together. Is government funding above or below this average? This approach can’t say. Moreover, while I find it a compelling thought experiment, at some point we probably want to check the results against data.

On the other extreme, there are many papers that have looked at specific government programs to fund research – see the posts An example of high returns to publicly funded R&D and More science leads to more innovation for discussion of some of this work. While the work described there is grounded in careful empirical analysis, we may worry that the results from any specific government program are not representative of the impact of government research support more broadly.

Against this background, two recent papers – Fieldhouse and Mertens (2023) and Dyèvre (2024) – make progress in tackling the question of how federally funded R&D at large affects productivity growth. Fieldhouse and Mertens tackle this problem in a conceptually straightforward way. Taking the US as a whole, what has historically happened to US productivity after the federal government increases its spending on R&D? Does it go up? When and by how much?

Subscribe now

Using data on US productivity and defense and non-defense R&D, for the period 1947 to 2019, they find that, in general, a 1% increase in R&D “capital stock”1 by the federal government tends to be followed 8-15 years later by a 0.2% increase in national productivity.2 In the figure below, you can see the average correlation between productivity in the following 15 years and an increase in government defense and non-defense R&D. The red line is very simple; just the average correlation between R&D and productivity growth. The blue line is the link after we try to remove the influence on productivity of a bunch of other non-R&D factors: the capital utilization rate, lagged productivity, lagged R&D values, defense spending news, and a measure of stock market returns for various innovation sectors (designed to measure expectations about innovation). As you can see, these adjustments don’t actually much matter. Interestingly, non-defense R&D seems to have a stronger but more delayed impact on productivity, compared to defense R&D.

It can’t be that simple, can it? Well no. One problem is that changes in R&D don’t happen randomly. Changes in R&D spending reflect other things going on that might also impact productivity. Fieldhouse and Mertens worry in particular that changes in R&D may reflect changes in underlying national economic conditions. For example, it might be that R&D is treated like a form of economic stimulus: maybe we increase R&D spending when the economy slips into a recession. If R&D only changes when the economy slips into a recession, we might then confound the impact of a recession on productivity with the impact of R&D on productivity. On the other hand, sometimes we change R&D for reasons that are unlikely to be confounded with changes in the economy. For example, a geopolitical rivalry between the US and the USSR led to a massive expansion in R&D for space technology.

This latter type of R&D change seems like the kind that could actually be informative about how R&D affects productivity. To focus on this kind of R&D change, Fieldhouse and Mertens draw inspiration from a famous 1989 paper by Romer and Romer that tried to understand the impact of monetary policy. Even more so than government R&D spending, interest rate policy is adjusted in response to economic changes, so Romer and Romer faced a similar problem as Fieldhouse and Mertens. Their solution was to use their best judgment to identify cases in history where interest rates were changed for reasons other than fluctuations in economic output and study monetary policy in those cases. Fieldhouse and Mertens take a similar approach: they look at the five agencies most responsible for R&D, and identify 218 cases when these agencies’ R&D budgets undergo a particularly large revision (up more than 5% in a year, or down more than 2.5%). For each of these cases, they then examine primary sources to try and understand why the R&D budget was changed in that year. Using their judgment, they then split this sample of major R&D changes into two categories: changes motivated by changing economic conditions, and the rest. The figures and results above are what tends to happen to productivity after a change in R&D spending that isn’t caused by fluctuating economic conditions.

(How good is their judgment in classifying these R&D policy changes? For what it’s worth, it turns out that all this classification work doesn’t actually matter much for their overall results. They get similar results if they just use all the data.)

One potential concern is that, in addition to changing in response to economic conditions, R&D might be changed because policymakers anticipate R&D is likely to be particularly impactful (or not). For example, the launch of Sputnik might have communicated to US policymakers not only that they were falling behind technologically, but also that an increase in R&D spending for space was likely to bear fruit – after all, the Soviets had demonstrated certain feats in space were feasible. If changes in R&D are responding mostly to changes in “technological opportunity,” then this kind of result might overstate the benefits of R&D in general, though they try to control for that with stock market valuations of innovative sectors. But if you want to evaluate their judgments, the paper includes a 630 page(!) “narrative appendix” which contains the primary sources and classification of each R&D policy change. (As an aside, this seems like it could be a useful dataset on what drives changes in R&D spending!)

The connection between R&D and productivity is further strengthened by a set of other results that suggest changes in R&D spending also anticipate changes in various proxies for technological change. Increases in R&D appropriations are associated with later increases in the number of new PhDs, researchers, and the number of new scientific and technological books.

One bottom line of this paper is that, over the last 70 years, if you look at swings in US government funding for R&D, increases tend to be followed by more PhDs, more researchers, more books on technology, and, eventually higher productivity growth, again across the whole US. The advantage of focusing on productivity in the entire United States is that doing so casts a very wide net; you’ll pick up gains from R&D wherever they happen nationally. That’s important, because as I’ve written here, knowledge spillovers are a big deal in innovation. The gains to unanticipated beneficiaries of R&D often exceed the gains to the intended target. But the disadvantage of focusing on the productivity of the US as a whole is that you end up with a relatively small number of observations (about 70 years and one country).

Another way to tackle this problem is to focus on individual firms- there’s only one America, but there are lots of firms. If publicly funded R&D is improving national productivity growth, then we should also be able to observe an effect on the individual firms that make up the economy. Dyèvre (2024) takes this approach.3 By studying individual firms, Dyèvre is able to draw on a bit over 7,000 data points (different firms in different years). Instead of focusing on how overall US productivity changes in response to changes in overall R&D, he wants to leverage this data to test how an individual firm’s productivity changes after an increase in government R&D that is relevant to that firm.

What is relevant government R&D? Well, for a firm working on solar panels, research funded by the Department of Energy is probably more relevant than research funded by the National Institute of Health. And within the Department of Energy, research conducted during years when solar panel research was a priority is probably more relevant to a solar panel firm than research conducted during years when the Department of Energy’s focus was primarily on other things. But how to identify what kinds of research each agency is working on in a given year at scale?

Dyèvre uses the patents firms and government agencies hold as a way to assess the overlap of their research. For example, a firm working on solar panels is likely to hold patents classified in patent category H02S: generation of electric power by conversion of infrared radiation, visible light or ultraviolet light, e.g. using photovoltaic [pv] modules. The Department of Energy also gets patents for its inventions, and the years in which its patents are classified as belonging to patent category H02S are also the years in which it is more likely to be doing R&D relevant to that firm. Using variation in what kinds of technologies get patented across different agencies in different years, Dyèvre can trace out what happens to the productivity of energy/life science/aeronautics/etc. firms when the government increases its R&D spending respectively on energy/life sciences/aeronautics/etc.

Compared to simply looking at total government R&D and its impact on total national productivity, we get a lot more data and variation, at the cost of adding a bit of noise to our data, both because patenting does not perfectly reflect what firms are working on,4 and because sometimes firms may benefit (possibly indirectly) from R&D on topics different from their own. With this data, Dyèvre estimates that a 1% increase in government R&D is associated with roughly a 0.023-0.025% increase in productivity after 5 years.5 At the end of this post, I’ll try to assess whether this is the kind of level we should be impressed by.

One other paper lends a bit more support to the general conclusions of Fieldhouse and Mertens (2023). Policy analysts typically assume that the economic effects of R&D differ depending on whether the spending is for defense or nondefense purposes (see CBO 2018, for example). Fieldhouse and Mertens (2023) also separately analyze the impact of defense and non-defense government R&D on productivity, and generally find results supporting that distinction (non-defense R&D has a greater impact on productivity). This finding is also echoed in a third paper, Moretti, Steinwender, and Van Reenen (2025), which focuses more directly on government funded defense R&D.

Moretti and coauthors essentially create estimates of industry-specific spending on defense R&D across OECD countries (as well as firm-specific defense R&D contracts for French firms) and look to see what happens after changes in defense-spending on R&D specific to your industry or firm. They estimate that a permanent 1% increase in annual defense R&D spending in a particular industry is associated with an 0.08% increase in productivity growth in that industry. The way Moretti and coauthors set up their analysis, we should interpret this increase as the long-run increase.6 This is smaller than the long-run effects of non-defense R&D found by Fieldhouse and Mertens, but larger than the long-run effects of defense R&D they find (which were essentially nil).

One potential explanation for this intermediate result (stronger effects for defense R&D, but weaker effects for non-defense R&D than in Fieldhouse and Mertens) is that Moretti, Steinwender, and Van Reenen’s estimates of defense R&D spending might in fact be picking up a mix of defense and non-defense R&D. The paper does not separate out the two types of R&D spending, and they note across their countries the average correlation between national defense and non-defense R&D spending is 0.29. So it’s possible their predicted defense R&D is also correlated with unobserved non-defense R&D (which they’re not specifically controlling for), which would lead to productivity increases if we assumed the Fieldhouse and Mertens result is correct.

In the end we have three empirical papers that try to assess the broad effects of government support for R&D. One looks at the overall impact of government supported R&D on US productivity, another at the impact of specific slices of that R&D on firms that are likely to be most directly impacted. Both approaches find R&D generates significant productivity impacts. One of these papers (Fieldhouse and Mertens 2023) also separately analyzes defense and non-defense spending, finding the productivity effects of defense-related R&D are lower than non-defense ones. A third paper looking specifically at defense R&D (but for an international sample, not just the USA) also finds long-run effects of defense R&D that are lower than the long-run effects of non-defense R&D identified by Fieldhouse and Mertens.

What’s the ROI though?

So to respond to the question that kicked things off: what is the return to government funded R&D?

Let’s focus on Dyèvre (2024), which I think is easiest to interpret. Recall, he finds that a 1% increase in government R&D funding generates roughly a 0.024% increase in productivity after five years. Is that good?

As a benchmark, suppose that 100% of annual economic growth is driven by (100% of) annual R&D. This is basically the assumption made in Jones and Summers (2021), the paper I mentioned at the outset of this article, which argued that every dollar of R&D generates several dollars of GDP over the long run. Let’s suppose there is a constant relationship between R&D spending and growth, just as a benchmark. If that’s true, then we should expect a 1% increase in annual R&D to generate a 1% increase in annual growth.

That’s broadly consistent with Dyèvre’s results. Annual GDP per capita growth in the USA has been about 1.8% per year since the 1950s, so a 0.024% increase in productivity (as Dyèvre finds) is equivalent to a 1.3% increase in the annual rate of growth. That implies we’re actually getting more than 1% increase in annual growth for a 1% increase in R&D, especially given that government R&D is only a minority of overall US R&D.

To be clear, I wouldn’t take this exercise too seriously - there are a lot of subtleties you should adjust for to do this kind of exercise properly. But as a back of the envelope calculation, I think it’s quite consistent with the results of Jones and Summers (2021), which argued that we should expect a dollar of R&D to generate several dollars of GDP.

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.

If you want to read more, the following posts were mentioned above:

• What are the returns to R&D?

• An example of high returns to publicly funded R&D

• More science leads to more innovation

• Knowledge Spillovers are a Big Deal

• The Size of Firms and the Nature of Innovation

• Bigger Firms Have Different Incentives

• Can we learn about innovation from patent data?

• The NBER is again running an Innovation Research Boot Camp this summer, with space for 25 PhD students (or recent PhD graduates). Apply by February 6. More here.

• The Washington Center for Equitable Growth is requesting proposals for policy-relevant research that promotes competition and supports workers in an era of AI innovation. Applications due February 10. More here.

Email me to suggest an announcement for the next newsletter. On to the post!

This article will be updated as the state of the academic literature evolves; you can read the latest version here. You can listen to this post above, or via most podcast apps here.

[May 2025: One of the articles discussed here has been retracted. See the updated article here.]

Which kind of inventor (or scientist) is going to benefit more from artificial intelligence: novices or experts? In theory, it can go either way.

Here’s a simple model. Assume you need to solve a bunch of sub-problems to discover or invent something. Novices aren’t very good at solving any sub-problems, while experts are good at solving the sub-problems related to their expertise. Finally, we’ll assume AI can help solve some of these sub-problems.

Suppose AI is good at solving the same sub-problems that experts are good at. In that case, AI doesn’t help experts very much, since it just helps them do what they already knew how to do. But novices using AI might become just as effective as experts. In this case, novices benefit much more from AI than experts, since they now operate on a level playing field with them.

Alternatively, suppose AI is good at solving sub-problems that experts are not good at. In that case, both novices and experts benefit, but experts might benefit a lot more. In this scenario, the experts use the AI to help solve sub-problems they are bad at, and continue using their own expertise to solve the things AI is bad at. Novices can also benefit from AI, but continue to struggle on the sub-problems that AI is bad at. It’s possible the gap between them and experts actually widens.

There are all sorts of middle cases as well, where AI helps solve some of the sub-problems experts are good at, but maybe not all of them. In that case, who benefits most may depend on how important the remaining problems are.

AI is not the only kind of tool people use to help them solve sub-problems they are facing. But the same logic applies to other “prediction technologies” as well.

Literal Maps

Let’s start with Nagaraj (2022), which illustrates the above in unusually concrete terms. We’ll be focusing on mining companies that are trying to find new deposits of gold (itself a common metaphor for the inventive process!). A key prediction technology for this activity is the aerial photo, since overhead photos help mining companies predict where gold deposits will be by revealing associated geological features, such as faults and lineaments. It had long been possible to acquire images like this by paying for a plane to survey some land, but it was costly. Then, beginning in 1972 the Landsat program began to systematically photograph the Earth from space and make these images available through mail-order for a modest fee. Nagaraj looks at how the sudden availability of satellite imagery affected gold-mining by new firms (novices, in the context of this post), and by incumbents (the “experts”).

Satellite mapping began in 1972 and continued for the next decade, which means some areas got mapped before others. Who got mapped when doesn’t seem to have been driven by gold potential; there was also some randomness in how useful the maps were based on the extent of cloud cover when an image was taken. Nagaraj’s main analysis compares the annual number of new gold discoveries on a particular tract before and after it was mapped (or mapped when cloud cover was low). The figure below documents the average difference in new gold discoveries made between tracts that receive a low-cloud photo and those that don’t, with the low-cloud image becoming available at year 0. We can see that prior to the availability of a good image, there’s no systematic difference in the rates of new gold discoveries between different tracks, and then after some tracts get good images they start finding gold deposits at a higher rate than the ones that lack a good image (though it takes several years to see an effect).

So who benefits more from satellite maps: the novices or the experts? As we noted at the outset, it should depend on whether there are important sub-problems associated with making a gold discovery that expertise helps with, and which maps do not help with.

Nagaraj focuses on the sub-problems of gold mining associated with institutional and regulatory barriers: environmental regulations, labor regulations, legal institutions, etc. These are the kinds of barriers that we would expect incumbent firms to have some experience dealing with, but for which satellite imagery doesn’t help much. Accordingly, we might expect that access to good satellite images helps novices with discoveries when these other barriers are low, but to help experts when they are high. And that’s basically what Nagaraj finds.

To measure the strength of these other sub-problems, Nagaraj relies on a 2014 survey of 4200 industry managers on the costs of exploration arising from these kinds of institutional and regulatory barriers. He labels the 25% of land tracts belonging to the lowest cost territories as “low cost”, the middle 50% of land tracts as “medium cost”, and the top 25% as “high cost.”

In all cases, senior firms (our experts) benefit from mapping. Across all three types of land tract (low, medium, and high cost) they make more discoveries in a tract after it has been mapped, as compared to unmapped tracts. That said, the benefits of mapping shrink as other costs rise - they see the largest increase in discoveries for the low cost areas, and the smallest increase in discoveries in the high cost areas.

On the other hand, Junior firms (our novices) only benefit from mapping when costs are low or medium. Among high cost land tracts, satellite mapping does not seem to help these novice firms at all. So on high cost land tracts, this prediction technology widens the gap between experts and novices. On the low and medium cost lands though, things are a bit more nuanced. When a region gets mapped, the absolute number of new discoveries made by senior firms increases by more than the absolute number of new discoveries made by junior firms. But junior firms make fewer discoveries overall, so that in proportional terms, they benefit more. The net effect is that the share of discoveries made by novices rises after mapping, concentrated in low and medium cost land tracts.

Metaphorical Maps

We find similar results when we turn from literal maps to metaphorical ones. Kao (2023) and Tranchero (2023) both examine the impact of genome mapping studies on innovation in drug discovery. Genome mapping studies scan the genes of people with specific diseases to identify all gene mutations that turn up more often than would be expected by chance. Since many drugs work by targeting specific genes that are causally related to diseases, identifying genes strongly associated with specific conditions helps drug companies prioritize potential drug targets and predict which drug candidates are more likely to succeed.

Kao (2023) looks at 168 systematic cancer mapping efforts published in prestigious journals. Each of these studies focused on a particular type of cancer; for example, a study might sequence a set of ovarian cancer tumors to identify all genes that tend to exhibit mutations in ovarian cancer cells. Tranchero’s study is instead focused on all diseases, and his genome-mapping data is based on 1,259 genome-wide association studies, which try to identify the genes that disproportionately exhibit mutations in people with a given disease. As discussed in Prediction Technologies and Innovation, both studies find that innovative effort around specific gene-disease pairs increases after associations between a disease and gene are discovered (relative to the level of innovative effort around gene-disease pairs with no associations found). Kao measures innovative effort with the number of phase II clinical trials focusing on a particular disease and gene, while Tranchero measures innovative effort with the number of new patent applications associated with a particular disease and gene.

Mapping studies provide one very particular type of information: they tell you that there is an unusually strong statistical association between a gene mutation and a drug. That’s useful to know, but it can sometimes mislead, because it turns out that strong associations between a gene and a disease are often spurious and not meaningful. Tranchero has an interesting way to identify this: sometimes a small study identifies an association between a gene and a disease, but then a bigger study comes along later and doesn’t find the same association. That suggests the first study just identified a false positive and there is no actual causal relationship between the gene and a disease. This happens surprisingly often - Tranchero finds that about 85% of disease-gene associations do not replicate in subsequent studies. So while a statistical association between a gene and a disease is suggestive, an ability to discern which associations make sense and are most likely to represent true causal relationships is a valuable complement to what genome mapping studies tell you.

It isn’t obvious which associations are meaningful and which are spurious. As noted in Prediction Technologies and Innovation, firms often seem to get tricked by these false positives; there is an increase in patent applications for gene-disease pairs when an association is found, even if that association later turns out to be spurious (by Tranchero’s definition). However, Tranchero also shows that firms who have some expertise related to the gene don’t get fooled by these false positives. To measure that, he looks at the publications associated with firms and labels firms who have published articles related to a particular gene to have “gene knowledge.” We can imagine that these firms have a better understanding of the role a gene might play, and hence have a better ability to discern meaningful and spurious associations. And it turns out these firms are more likely to file patent applications when a genome wide association study finds an association that is not subsequently overturned, but they are not more likely to file applications for associations that will turn out to be spurious.

Drug discovery requires solving many other sub-problems besides deciding if there is a link between a gene and a disease. For example, running clinical trials involves solving a lot of logistical and other sub-problems. Moreover, the kinds of challenges faced vary across different kinds of trials, and firms will have expertise related to running different kinds of trials. Firms whose expertise is most closely related to the other problems associated with running a clinical trial on a particular disease-gene pair should benefit more from learning about an association.

Kao has one proxy for this. She finds that after an association between a cancer and a gene is discovered, there is a larger increase in clinical trials for drugs that had been previously tested in similar settings (for example, for other diseases associated with the gene in question) than for totally novel drug applications. That’s consistent with there being a bigger response to mapping data from firms with expertise in the kinds of clinical trials you need to run if you take an association seriously. Again; experts can benefit more from a prediction technology if there are important other sub-problems that the prediction technology doesn’t help with, but expertise does.

In contrast, we can also predict that firms with expertise very closely related to the knowledge disclosed by mapping studies will benefit less from them. That seems to be the case. For example, for each company in her dataset, Kao looks to see if they have previously conducted phase II trials associated with a particular cancer, and if they have published an above-average number of articles about genome mapping. She takes that as a proxy that this firm is more likely to have private mapping information about associations related to that cancer. If private mapping data and public mapping data are close substitutes, this might be a good example of what happens when a prediction technology solves the sub-problem that an expert is good at. And here, rather than help the experts, she finds the firms with more publications about mapping respond to publications about associations between this particular cancer and a gene only about 40% as strongly as firms that are less likely to have private mapping information about the cancer and gene.

Judgement and AI Tools

For a final case study, let’s turn to a paper looking at the impact of AI on discovery in material sciences, circa 2022. Toner-Rodgers (2024) looks at a large materials science company, which in mid-2022 began rolling out access to an AI tool to randomly selected research teams. This tool allowed researchers to specify desirable properties for new compounds, and predicted which kinds of compounds might have those properties. As discussed more in the post Prediction Technologies and Innovation, Toner-Rodgers found teams with access to the tool discovered substantially more new materials, measured in a variety of different ways.

But Toner-Rodgers has data on 1,018 scientists spread across 221 different research teams (this is apparently a very big company). That lets him see which kinds of scientists most benefitted from access to the tool.

He starts by computing how many new materials each scientist tends to discover per year, before they had access to AI. He groups scientists into percentile buckets (the 10% least productive, then the 10% next most productive, then the 10% next most productive and so on). Then, he compares the number of new materials being discovered by scientists with access to the new AI tool, to the number being discovered by scientists in the same productivity bucket, but who don’t have access to the tool. The results are in the figure below. The horizontal axis is the the productivity decile; lower numbers indicate scientists who produced fewer new materials per year, prior to the introduction of AI.

If we think of the lower-productivity scientists as being analogous to the novices of this post and the higher-productivity scientists as being analogous to experts, then this is a very clear case of an AI tool helping the experts more than the novices. Indeed, we can’t statistically reject the hypothesis that the 30% of scientists with the lowest productivity don’t benefit at all from the AI tool; they do just as well as their similar colleagues without the tool. Meanwhile, scientists who were in the top 30% of productivity prior to the tool’s introduction produce 60-80% more new materials with AI than similar colleagues without it.1

If access to AI widens the gap between “novices” and “experts” (not really the right terms for the above figure, but it’s what we’re using in this post), then the discussion at the beginning of this post suggested that may be because experts are good at solving sub-problems that the AI does not help solve. Fortunately for Toner-Rodgers, scientists at this organization keep logs of their activities, which he has access to. A log will include the time a scientist began working on a task, how long they spent on it, and a text description of the task; scientists make 7.8 entries per week on average. These logs provide unusually detailed information on the kinds of sub-problems that scientists face in this organization.

They aren’t standardized, but Toner-Rodgers uses Claude-3.5 to label work activities with one of four categories: idea generation, evaluation, testing, and “other.” Here, “idea generation” refers to activities related to coming up with designs for novel compounds, “evaluation” refers to activities that help determine which designs to test (for example, by doing initial theoretical calculations or reviewing the literature), and “testing” involves running tests on proposed ideas to assess them for desirable properties.

Toner-Rodgers then runs a statistical model to try and infer each scientists’ skill at the idea generation and evaluation steps. The basic idea here is to use their work logs to figure out on which research projects a scientist was doing work on idea generation and on which projects they were doing work on evaluation (or they might be doing both). Toner-Rodgers assumes the probability a project succeeds (in this case, a new material is ultimately added to the firm’s internal database) is higher when the people assigned to idea generation have a higher skill for that task, and the people assigned to evaluation have a higher skill for that task. For example, suppose we have two hypothetical scientists, Alice and Bob. When Alice works on idea generation and Bob on evaluation, we notice that projects tend to succeed, but when Bob works on idea generation and Alice works on evaluation, projects tend to fail. When Alice and Bob each do both tasks, their success rate is somewhere in the middle. That set of facts lets us infer Alice is good at idea generation, but not evaluation, and Bob is the reverse. Toner-Rodgers runs a statistical model that infers a scientists’ skill using a similar idea.

With these inferred skill levels, he then can ask who benefits most from using an AI tool: people good at idea generation or idea evaluation? The AI tool seems most likely to help at the idea generation stage, since it is designed to propose compounds that are likely to meet certain requirements. So we would predict that people good at idea evaluation benefit more from using the tool than people good at idea generation. And that’s basically what Toner-Rodgers finds: people skilled at each task benefit, but people skilled at idea evaluation enjoy about 4x the benefits of people good at idea generation.

Indeed Toner-Rodgers finds that, over time, the advantage of experts actually grows. This seems to be because after working with the AI tool, scientists begin to reallocate more and more of their time towards idea evaluation tasks, and away from idea generation tasks (see the figure below).

Echoes of Automation

To return to the beginning of our post, we asked which kind of inventor/scientist is more likely to benefit from artificial intelligence: novices or experts?

In theory it could go either way, since it depended on the kinds of sub-problems that AI helped with and whether those overlapped substantially with the kinds of sub-problems that experts have typically had advantages over novices at. So we turn to empirics related to various prediction technologies and find… it can sort of go either way.

In gold-mining, at least in proportional terms, novices benefit more when a prediction technology helps them identify new deposits, so long as other barriers to exploration (regulatory and institutional, for example) are small. Or the prediction technology can disproportionately help experts, so long as those other barriers are high, since expert firms are better at dealing with those barriers. In drug discovery, gene maps erode a lot of the advantages of experts who have private mapping data, but not the advantages of experts who have different kinds of expertise (for example, related to running clinical trials or the genes in question). For materials science researchers, novices in general don’t seem to benefit much from the tool, but there is still some variation in how much it helps different kinds of experts. Experts whose main strength is generating ideas benefit from the tool much less than experts skilled at evaluating ideas.

So in this case I think we have a remarkable agreement of theory and empirics. Who benefits most from a new prediction technology? It depends!2

That’s a bit too nihilistic about what we’ve learned though I think. We have in fact made some progress: the impact of future AI on scientists and inventors will depend on your assumptions about what future AI will do well and how much that overlaps with existing expertise. My own view is that the current AI paradigm is going to keep getting better at identifying useful patterns in data, whether the protein data bank or the corpus of scientific journal articles. So to the extent an expert’s work is primarily based on mining patterns from data, AI is potentially going to erode their advantages over novices. I think this would be bad news for me if I made my living purely from writing New Things Under the Sun!

On the other hand, as your expertise gets further away from pulling patterns out of data, maybe experts will begin to benefit more from AI than novices. Maybe there isn’t a lot of data for the AI to work with - this could be the case in fields where tacit knowledge is important, or maybe on the cutting edge where experiments are generating knowledge but nothing has been written up yet. Alternatively, maybe pulling patterns from data is just a starting point for a lot of other work; in materials science, it suggests compounds, but you still need to test them.

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.

If you want to read more, the following posts were mentioned above:

• Prediction Technologies and Innovation

• When the robots take your job

• Gaétan de Rassenfosse has launched a new living literature review, The Patentist, about intellectual property rights. Check it out!

• Graduate students and employees at qualifying organizations, consider a summer internship at Open Philanthropy! Applications are due January 12.

• The National Bureau of Economic Research is offering research fellowships to study innovation and productivity policies. Applications due January 16.

• The Research Transparency and Reproducibility Training (RT2), hosted by the Berkeley Initiative for Transparency in the Social Sciences, will be May 21-23, in Berkeley, CA. Learn more / apply here. Applications due January 19.

• The Metascience 2025 conference will be in London over June 30-July 2 and has an open call for proposals until February 7, 2025. I plan to be there!

• Apply for a studentship to do a PhD with the London School of Economics and CERN to study the socio-economic impacts of big science - think Large Hadron Collider or International Space Station. Applications will be considered on a rolling basis until May 22.

Email me to suggest an announcement for the next newsletter. On to the post!

This article will be updated as the state of the academic literature evolves; you can read the latest version here. You can listen to this post above, or via most podcast apps here.

[May 2025: One of the papers discussed here has been retracted. See the updated version of the post here.]

Some inventions and discoveries make the inventive process itself more efficient. One such class of invention is the prediction technology. These can take a lot of forms. AI is one example of a technology that can help scientists and inventors make better predictions about what is worth trying as a candidate solution to a problem, but as we’ll see, there are many other kinds of prediction technology as well.

Prediction technologies seem obviously valuable to scientists and inventors: discovery is all about stepping into the unknown, and a good prediction technology can make it more likely the trip into the unknown will be a productive one. But people often worry that they can have a negative effect if they focus too much inventive effort onto the things the prediction technology helps most with.

Streetlights and Rationality

The most famous example of this phenomenon is the proverbial story of the drunk man struggling to find some keys he dropped. It’s night time in the parable, and he’s looking for his keys under the streetlight, but not having any luck. When asked if this is where he dropped them, he says “no, I dropped them over in that field; but this is where the light is.” In the same way, prediction technologies might bias innovators to look where it’s convenient, not necessarily where it’s most important to look.

The parable doesn’t make much sense if you assume the person looking for their keys is more clear-headed though; if they know they dropped their keys far from the light, they should know not to search near the light. Similarly, it’s hard to tell a story where a clear-sighted inventor is worse off by getting access to a prediction technology. Such an inventor can just assess whether it’s worth using the prediction technology, if it means the potential solution won’t be very significant.

Hoelzemann et al. (2024) point out a way that prediction technologies could make us worse off though, even if scientists/inventors are clear-headed and rational. The key idea is that the discoveries made by scientists and inventors tend to be public, and so scientists/inventors learn from each other.1 In such a world, collective innovation happens more quickly if individuals pursue different lines of research. For example, if I see a rival’s line of research fails, then I know not to go down that road myself. If it succeeds, I can copy that approach.

In this kind of process of collective discovery, a prediction technology can actually make us worse off if it leads everyone to focus on the same small number of research lines. Hoelzemann and coauthors first show this with theory; if you don’t care about the knowledge spillovers to your peers, then it can be rational for everyone to follow the same promising research line, rather than trying different approaches (which might seem less promising individually).

Hoelzemann and coauthors also run an experiment where participants must choose an unknown payout from five different options. Participants play two rounds with each payoff, so if someone finds a high payoff in the first round, everyone learns it and can opt to choose that payout in the second round. Hoelzemann and coauthors show that if you don’t tell participants anything, then they will tend to all guess at different payouts in the first stage, which helps them learn which payouts are best in the second round. But if you tell participants about an intermediate payout - not the best, but not the worst - then many participants are more likely to simply select that payout (which is for sure not bad, even if it’s not great). Since more people pick the same thing, they find the group is less likely to discover the best payout and hence less likely to receive the best payout in the second round. Indeed, in the experiment, participants are usually worse off with information about an intermediate payout than when they don’t know anything at all.

This is a nice demonstration of a theoretical downside to prediction technologies. Let’s now turn to some studies of how real-world prediction technologies affected innovation.

Structural Biology in 2003

To start, let’s consider Kim (2023) who looks at the impact of a new software program in structural biology. This is a field where scientists typically try to figure out the 3D structure of proteins by interpreting data from x-ray diffraction patterns. Kim studies the 2003 introduction of a software program called Phaser that was useful in inferring the 3D structure of proteins.

A key limitation of Phaser is that it only helps you find the structure of proteins that are closely related to proteins whose structure we already know. If there are no known structures “nearby” in the space of possible proteins, the program isn’t helpful. But maybe those are precisely the proteins we want to know more about?

Kim looks at how the research focus of structural biologists changed after Phaser was introduced. To do that, she partitions the universe of proteins (both those with known and unknown structures) into tens of thousands of clusters, where all the proteins in a given cluster have similar amino acid sequences (and hence, are more likely to have similar structures).2 She then restricts her sample of clusters to those where the proteins were known to exist before 1998 (when her sample begins), and which contain at least one human protein. In some of these clusters, we additionally knew the structure of some of their proteins by 1998, and in others we did not. Kim looks to see how scientific interest in proteins in these different kinds of clusters changed after Phaser came along and made it easier to discover proteins in clusters where we already knew the structure of some other proteins in the cluster.

She finds that clusters with known structures saw a 7% increase in the number of solved structures, relative to clusters without known structures, following the introduction of Phaser. In other words, this prediction tool did disproportionately increase effort on proteins that were “under the street light.” Worryingly, she also finds some evidence that the program increases work on proteins that are less important. Clusters with known structures see relatively larger increases in the number of protein structures that are solved but which don’t get published in academic journals, don’t get cited by patents, and don’t seem to merit annotation about their functions from experts in another database. Kim takes these as proxy evidence for a relative increase in work on less important protein structures.

Gene-Disease Association Maps

Tranchero (2023) and Kao (2023) also document the strong differential effects of prediction technologies on innovation, this time in the context of drug discovery. Both examine the impact of genome mapping studies that establish correlations between diseases and mutations in certain genes (i.e., “people with this disease tend to have mutations in this gene”). Both look at how clinical trials focused on specific diseases and genes evolve following genome mapping studies; Kao focuses on cancers, and Tranchero on disease more generally.

A key part of both datasets is built from genome mapping efforts. Kao looks at 168 systematic cancer mapping efforts published in prestigious journals. These studies were cancer-specific; for example, a study might sequence a set of ovarian cancer tumors to identify all genes that tend to exhibit mutations in ovarian cancer cells. Tranchero’s study is based on 1,259 genome-wide association studies, which try to identify the genes that disproportionately exhibit mutations in people with a given disease, compared to people without it. Drug companies benefit from knowing about these associations, because they can develop drugs that target different genetic targets.

Both Kao and Trancheo then look to see how innovation is affected for treatments that target a specific gene, to treat a specific disease. In the figure below, we’re focusing either on the number of phase II clinical trials associated with a particular gene-cancer pair (Kao 2023, left) or the number of patent applications associated with drugs for a particular gene-disease pair (Tranchero 2023, right). Both figures compare the average number of clinical trials/patent applications for a gene-disease pair that eventually has an association discovered (at time zero in the figure) to the average number of clinical trials/patent applications for gene-disease pairs that do not have a discovered association. We can see that after associations are discovered, there’s a relative increase in innovation efforts based on that gene-disease combination.

Once again, we can ask whether this prediction technology pulls innovators away from important lines of research.

One problem with genome association studies is that sometimes they misidentify chance associations as meaningful associations. Tranchero has an interesting way to identify this: sometimes a small study identifies an association between a gene and a disease, but then a bigger study comes along later and doesn’t find the same association. That suggests the first study just identified a false positive and there is no actual meaningful relationship between the gene and a disease. This happens surprisingly often - Tranchero finds that about 85% of disease-gene associations do not replicate in subsequent studies. Tranchero uses this fact to identify a bunch of gene-disease associations that subsequently are revealed to be likely false positives. He shows the number of patent applications associated with a gene-disease pair does increase significantly, even if the pair is subsequently revealed to be a false positive. Patent applications based on false positives do not tend to generate valuable patents by a variety of metrics: they’re not highly cited, and they’re not associated with above-average patent value.

It’s not all bad though. While a false disease-gene association increases related patent applications by about 0.07-0.15 per year, a true association increases applications by substantially more (0.5-0.7 per year), and is associated with more valuable subsequent patents. Meanwhile, Kao finds that clinical trials initiated after cancer-gene associations are discovered are a bit more likely to generate positive outcomes (for example, a statistically significant increase in survival probability).

Material Science Machine Learning

As a final prediction technology, let’s consider machine learning. Toner-Rodgers (2024) looks at a large materials science company, which in mid-2022 began rolling out access to an AI tool to randomly selected research teams. This tool allowed researchers to specify desirable properties for new compounds, and predicted which kinds of compounds might have those properties. Toner-Rodgers can measure the impact of this tool on innovation in a few ways. First, the company has an internal database of materials it believes are viable for product use; new additions to this database are one way he tracks discoveries. Second, he can look at patent filings (see this post for a discussion of how well patents do and do not track innovation outcomes). Third, he can look at new products sold by the company that incorporate newly discovered materials. Below, we can see how the rate of discovery across these different measures changes when teams get access to the AI tool, as compared to teams that lack access.

These are pretty gigantic effects; teams with an AI tool are discovering 20-50% more new materials than their peers without access to this prediction technology!

To look at the kinds of new materials developed, Toner-Rodgers uses a few different approaches. First, he uses a measure of how similar the crystal structure of a new material is to the structures in existing datasets.3 Second, he looks at the text of filed patents. He assumes patents with language that is more dissimilar to existing patents, or which introduce a large share of new technical terms are more novel (an approach used in other literatures). Third, he looks at the share of new materials that are also new product lines, rather than improvements to existing ones (which most new materials are). And finally - he just surveys the researchers to ask whether they think AI helps them make more novel materials. Across each of these measures, he finds teams that have access to the AI tool produce more novel materials than teams that didn’t. For instance, 73% of surveyed respondents thought the tool generated more novel designs than other methods.

The size of these effects was meaningful, if not revolutionary. To take one illustrative example, imagine you line up patents by how similar they are to existing patents (based on using similar uncommon words). Designate the far left 0%, where you have the patents that are most unlike existing patents, in their word choices, and designate the far right 100%, where you have the patents that are the most similar to existing patents, again in terms of word choices. On average, we would put the patents of teams using AI at the 42% position, and the teams not using AI at 48%.

Taken together then, access to an AI tool has a pretty dramatic effect on the number of new materials discovered, and appears to direct research effort towards materials that are more, rather than less, distinctive, at least as compared to previous materials.

Streetlights and Floodlights

To sum up, we’ve got four studies of prediction technologies (software in structural biology, genome-disease maps, and an AI tool for material science), all of which study the effects of prediction technologies on innovation in slightly different ways. None of these studies look directly at the kind of herding concerns that we kicked off this post with. But at least some of them do show that prediction technologies can certainly affect research choices in a way that could lead to a net decline in the diversity of research approaches. The phaser software program and genome-disease association maps both lead to significant increases in research of the kind the prediction technology makes more predictable. And in both cases, it’s also plausible that the increased research action was of the mediocre type that Hoelzemann et al. (2024) warns could lead to an overall decline in the rate of collective invention. The proteins that became more common to study seem less valuable, and a large share of the disease-gene associations that were investigated are probably false positives. The results for AI have a quite different flavor, but we’ll come to those in a minute.

But before we do that, it’s worth pausing for a second to emphasize that the potential downsides of streetlight effects or the narrowing of research approaches aren’t unique to prediction technologies. That’s because prediction technologies are not the only thing that helps us predict which lines of research are likely to be most promising. Another thing that does that is knowledge, in general. Indeed, I have another post, Science as a map of unfamiliar terrain, that uses a prediction technology - a map - as a metaphor for understanding the impact of science on innovation.

Hoelzemann et al. (2024) is the paper we kicked off with, that used theory and an experiment to highlight how knowledge could lower the rate of collective invention. They close their paper with a real world illustration of how the streetlight effect can happen simply as a consequence of normal research. In this case, they focus on research to uncover the genetic roots of different human diseases. They aim to show that when there is an early discovery of a gene weakly related to a particular disease, that leads to less exploration than if no gene had been found at all, which ultimately delays the discovery of genes strongly related to a given disease. To do that, they look at over 4,000 diseases listed in the DisGeNET dataset, which compiles all scientific publications that link human diseases to their genetic causes. DisGeNET scores the strength of an association between a gene and a disease, so over the period 1980-2019, Hoelzemann and coauthors look to see what is the strongest gene-disease association discovered in the first 10% of papers published about a disease.

In the figure below, they look at the average number new genes added to the dataset, per paper, before and after a disease-gene association of medium strength is discovered,4 as compared to other diseases. After year 0, which is when a medium-strength gene-disease association is found, the number of new genes studied (per paper) drops, consistent with a narrowing of research strategies after scientists see that one approach is at least somewhat promising.

Hoelzemann and coauthors argue that this reduction in exploratory research can slow collective innovation. And indeed, they find diseases which discover a gene-disease association of moderate strength actually take longer to discover very strong gene-disease associations than diseases that find only weak gene-disease associations. In their dataset, a disease which initially discovers only a weak gene-disease association will typically discover a strong gene-disease association two years earlier than a disease which initially discovers a medium strength gene-disease association!

So it’s actually anything that might make some research lines disproportionately more attractive that can lead to inefficient research herding: prediction technologies yes, but also discoveries, and presumably other things like changing costs of some kinds of research relative to others. The pressure to reduce exploratory research when a field makes progress is probably just another unfortunate property of the natural world, like Winter. You can’t stop winter, and you can’t stop the fact that discovery and progress make some research lines more attractive and hence broad-based exploration less attractive. Instead, the solution here seems to be setting up incentives that counteract these tendencies by rewarding exploration directly; one way to do that is to reward novelty itself, for example by giving disproportionate credit to people who are first to make a discovery. We in fact do this, though that has its own drawbacks (see the post Publish-or-perish and the quality of science).

Alternatively, you can “just” invent prediction technologies that cover a very broad range of research lines. There’s an interesting contrast between Kim (2023) and Toner-Rodgers (2024). Both papers are about using software to predict molecular structure. However, Kim (2023) finds access to a prediction technology reduced the novelty of scientist efforts - more protein structures got published that were similar to existing proteins. Toner-Rodgers (2024) instead finds access to a prediction technology increased the novelty of research efforts - more materials were created that were dissimilar to existing materials. But I don’t think there’s really any contradiction here. People flock to the light offered by these prediction technologies, but different technologies throw out different amounts of light. We can imagine Kim (2023)’s technology is like a lonely streetlight, only illuminating protein structures that are near to others we already know, while Toner-Rodgers’ technology is a gigantic set of floodlights that illuminate a whole field.5

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.

If you want to read more, the following posts were mentioned above:

• Knowledge spillovers are a big deal

• Can we learn about innovation from patent data?

• Science as a map of unfamiliar terrain

• Publish-or-perish and the quality of science

In this special annual edition of What’s New Under the Sun, we have a big bundle of the titles, abstracts, and links to innovation-related PhD job market papers from 2024. Some were sent my way following my request in the last newsletter, but to find the majority of these, I looked at the titles of job market papers from graduating PhDs at ~175 economics, business, and other departments. Even so, I’m sure I missed some great papers. If that’s you, email me and I’ll add you to the posts.

I’ve split this post into three to make it a bit easier to navigate. This is the third post. The first post is here, and the second is here. Subscribers should receive all three in their email.

Subscribe now

Titles Index

Titles are presented in random order. There might be additional authors on these papers - I’ve listed the associated job market candidate only.

41. Technology Choice, Spillovers, and the Concentration of R&D by Todd Lensman

42. The Impact of Wage Differentials on R&D Offshoring and Invention Value by Divya Sebastian

43. Technology Rivalry and Resilience Under Trade Disruptions: The Case of Semiconductor Foundries by Weiting Miao

44. Open Source Software Policy in Industry Equilibrium by Jeff Gortmaker

45. Infrastructural Shocks and University Success in the United States by Jian Qi

46. Teams and Text: Collaborative Innovation in the Knowledge Space by Joseph Emmens

47. Green Innovation Under Pressure by Xintong Li

48. Not on Terra FIRRMA: Foreign Investment in US Startups and Innovation by Fiona Paine

49. The Geography of Innovation in the United States by Weiliang Tan

50. Where do toxic innovations come from? A study of competition and innovation in the US shale oil and gas industry by Divya Saxena

51. Awakening Latent Human Capital: The Opening-Up and Entrepreneurship in 19th-Century China by Li Duan

52. Industry Shakeouts after an Innovation Breakthrough by Xiaoyang Li

53. Are Societal Promises in Science and Technology Substantiated? A Study of Value Expressions in Patents by Sergio Pelaez

54. A New Order? Digital Disruption and Entrepreneurial Opportunities by Javad Shamsi

55. Disasters, Capital, and Productivity by Tarikua Erda

Titles and Abstracts

41. Technology Choice, Spillovers, and the Concentration of R&D

Todd Lensman

The direction of innovation shapes both current technologies and future innovation opportunities, as firms acquire expertise and create public knowledge through discovery. But how do firms choose which technologies to develop, and why might they fail to exploit new technological paradigms? I study these questions in a tractable new model of directed innovation and firm dynamics, highlighting a novel connection between market structure, the direction of innovation, and economic growth: Expertise in a current technology gives incumbents a comparative advantage at innovating it relative to entrants, who instead favor a new technology with higher growth potential. Each firm’s innovation decisions influence others through knowledge spillovers, which can inefficiently delay or prevent the emergence of the new technology. Concentrating R&D resources in a small number of firms can exacerbate this problem by amplifying the influence of incumbents, even though it accelerates growth in the absence of a technology choice. I provide empirical evidence for the theory using data on firm patenting and R&D expenditures, and I apply it to explain the historical development of mRNA vaccines.

Link

42. The Impact of Wage Differentials on R&D Offshoring and Invention Value

Divya Sebastian

Over the past forty years, the international organization of R&D has globalized, as evidenced by an increased reliance of US firms on overseas inventors. However, there is little evidence of the nature and direction of inventions produced overseas as compared to those at home. Using a simple framework of location choice of R&D, this paper provides evidence for an offshoring penalty and considers which firms can overcome the offshoring penalty. Amid restrictive immigration policies to hire overseas talent in the US, the paper also shows timely evidence of how the relative cost of inventing at home can change the nature of inventions produced at home and overseas. Using a sample of patents from publicly traded US firms that invent both at home and overseas, I test the predictions from the framework empirically.

Link

43. Technology Rivalry and Resilience Under Trade Disruptions: The Case of Semiconductor Foundries

Weiting Miao

This paper studies the impact of industrial policies on technology competition and consumer welfare amid rising global trade disruption risks. Distilling key empirical features from novel data on the semiconductor foundry industry, I develop and estimate a dynamic oligopoly model that integrates step-by-step innovation, trade disruption risk, and industrial policies. While distortions from market power and technological externalities justify subsidies, their optimal levels depend on the magnitude of trade disruption risk: when the risk is low, the optimal subsidy rate remains low, as the welfare benefits are distributed globally, but the costs are borne exclusively by the subsidizing government. My quantitative model shows that a 35% trade disruption risk makes the 25% investment subsidy under the US CHIPS Act optimal, resulting in a 6% welfare improvement for the U.S. The paper also analyzes the CHIPS Act’s restrictions on investments in rival countries, intended to secure technological leadership against their firms. Its efficacy depends on the strength of technology spillover restrictions and the scale of the rival home market secured for rival firms.

Link

44. Open Source Software Policy in Industry Equilibrium

Jeff Gortmaker

Open source software (OSS) is a form of public knowledge widely provided and relied on by the private sector. To study the effects of growing government involvement in this critical public good, I build a new empirical model where high-tech firms choose software inputs and developer labor in competitive equilibrium. For estimation, I create a new dataset of OSS and in-house investment for the global web development industry, where software choices are directly observable. I simulate counterfactuals to assess the global impact of China tightening its recent internet restrictions on crossborder OSS collaboration or increasing its financial support for domestic OSS. I find that stricter restrictions do little to boost domestic OSS investment. Instead, lost spillovers raise web development costs in China by $2 per dollar of disincentive and $7 globally. Heightened subsidies prove more effective at increasing domestic investment and cut global costs by $11 per dollar of subsidy—tripling if the US responds in kind.

Link

45. Infrastructural Shocks and University Success in the United States

Jian Qi

What factors lead some universities to flourish while others falter? Using newly digitized data on historical college enrollments, the spatial distribution of World War II military surplus, and institutional-level research output, I leverage quasi-random variation in access to military property surpluses from the postwar demobilization to estimate the impact of infrastructural inputs on enrollment and other institutional performance metrics. Constructing a balanced panel of colleges spanning each decade from the 1930s to the 2010s, I find that exposure to military real property surplus enabled U.S. institutions, particularly public universities, to expand enrollments in the postwar years and boost research output in the long run. Individual-level analyses further reveal that successful universities generate positive spillovers within the local labor market. These findings underscore the enduring benefits of investing in higher education infrastructure.

Link

46. Teams and Text: Collaborative Innovation in the Knowledge Space

Joseph Emmens

In this paper, I study the impact of an expanding scientific and technological frontier on team innovations. To do so, I present a novel framework that integrates inventor teams and their patent texts. I model collaboration directly through a Bayesian model of Natural Language Processing. Applied to patent text data, this model builds a map of inventors, teams, and research fields, referred to as the knowledge space. Trained on over 400,000 U.S. patents from the USPTO PatentsView database, this framework allows me to tackle unanswered questions on how teams create new knowledge. Specifically, I investigate the effect of prior work on a team’s ability to produce a breakthrough—an innovation that sparks a new and successful research field. Leveraging high-dimensional patent text data, I back out two new measures: breakthrough patents and a team’s knowledge field, the set of research fields accessible to the team. I combine this with data on premature inventor deaths as a quasi-natural experiment. This identifies how team innovations change as they pivot to more or less advanced research fields. The framework unifies key elements of collaboration. Teams build on existing knowledge, and prior work both supports and obstructs innovation. I show that teams generate more breakthroughs when building on enough prior work to incorporate valuable knowledge, but not so much as to stifle novelty.

Link

47. Green Innovation Under Pressure

Xintong Li

As the impacts of climate change intensify, firms must overcome the technical challenges of emission reduction while coping with damage from frequent disasters. This paper provides novel evidence on how firms adjust the pace and direction of their innovation in response to costly physical disasters. Leveraging granular data on the joint spatial distribution of climate hazards and economic activity across the U.S., we do not find firms exposed to acute physical risks cut their R&D expenditure after disaster shocks. Instead, our patent analysis reveals a subsequent shift in these firms’ innovation efforts toward green technologies. To examine the underlying mechanisms, we extend the directed technical change framework by incorporating regulatory incentives and firms’ learning about future risks. Our findings demonstrate that post-disaster recovery, when the advantage of “dirty” vintages weakens, presents a unique window of opportunity for policies to accelerate the low-carbon transition while enhancing climate resilience.

Link

48. Not on Terra FIRRMA: Foreign Investment in US Startups and Innovation

Fiona Paine

Countries have increasingly been using economic policies to further geopolitical and national security goals. Thus far, economists have focused on studying tariffs and subsidies despite a broader range of economic tools actually being implemented. How costly are these other policies and what are their effects on capital markets, investment, and the economy more broadly? In this paper, I examine a 2018 U.S. law (FIRRMA), which expanded the government’s ability to review and block transactions on national security grounds to include venture capital (VC) investments by foreign investors. I use the passing of FIRRMA, its differential impact on specific VC industries, and the role of Chinese investors in U.S. venture capital to study whether foreign investment screening impacts capital supply. I find that FIRRMA had a negative effect on capital supply in impacted industries due to two factors: 1) the specialization of VC investing (such that the substitution of outside capital into impacted industries is low) and 2) networks in VC investing (there are spillovers to domestic syndication partners in impacted industries). I further find that the change in capital supply is costly, leading to lower innovation by startups. I introduce a novel way of measuring innovation early in the life of a startup using text from startup websites. I use this measure to show there is a selection effect where VCs give first round funding to less innovative startups after FIRRMA. Finally, in a case study of the biotechnology industry, I show that impacted startups suspend drug projects at higher rates, and in particular their risky projects.

Link

49. The Geography of Innovation in the United States

Weiliang Tan

A defining trend in U.S. innovation is its increasing geographic concentration, exemplified by the growth of high-tech clusters like Silicon Valley. What factors drive this increasing spatial concentration, and what are its implications for regional and aggregate growth? Using comprehensive data on patents, firms, and inventors from 1976 to 2018, I find that innovation became more concentrated in high-skill cities only after 1990, with the sudden rise of information and communication technologies (ICT) playing two distinct roles in this process. First, there was a compositional shift in innovation towards ICT, which is colocated with ICT production and concentrated in high-skill cities. Second, firms that were initially concentrated in high-skill cities produced more non-ICT patents likely due to spillovers from ICT innovation and ICT enabled reductions in communication costs, which allowed these firms to expand production to lower-cost regions and enhanced the profitability of new ideas. Worker migration to high-skill cities amplified the effects of these mechanisms, intensifying the spatial concentration of innovation. To better understand the mechanics of innovation across space and its consequences for macroeconomic growth, I develop a model of spatial growth with endogenous and directed innovation, technology diffusion, and worker mobility. The model provides an analytical characterization of the spatial direction of innovation on the transition path and how its steady-state distribution across space determines long-run aggregate growth.

Link

50. Where do toxic innovations come from? A study of competition and innovation in the US shale oil and gas industry

Divya Saxena

I study the effect of competition on the nature of innovation, particularly innovations that generate negative environmental externalities. I argue that competitive pressures drive firms to undertake innovations that have the potential to enhance yields but also pose significant risks to the environment. I examine the effect of competition on the toxicity of innovations in the design of chemical cocktails (exploratory formulae) used in the production of shale oil and gas in the US between 2011 and 2016. Utilizing the negative oil price shock of 2014 as an instrument to competition for firms in the industry, I show that competition is positively associated with using exploratory formulae that have toxic chemicals, especially by poorly performing firms. In exploring the origins of such toxic innovations, I show that poor-performing firms act as conduits for experimentation with toxic chemicals by suppliers under the conditions of high competition. In discussing the productivity implications, I show that the productivity gains from using toxic exploratory formulae are positive but short-term. I highlight how changes in competitive intensities can impact the nature of innovation generated by firms which can be detrimental to environmental health.

Link

51. Awakening Latent Human Capital: The Opening-Up and Entrepreneurship in 19th-Century China

Li Duan

This study exploits a special historical case-openings of treaty ports in 19th-century China- to examine how upper-tail human capital, quantified via book creation, impacted modernization when facing external pressures. Employing a prefecture-level panel dataset from 1840 to 1904, the study establishes book density, indicative of knowledge endowment, as a significant and positive predictor of modern firm entry following the opening of treaty ports. To understand the mechanism, a critical aspect lies in understanding the Civil Service Examination (keju), an indigenous institution that historically dominated talent accumulation and allocation in China. By integrating data with keju, we find that exposure to Western influence mobilized the segment of upper-tail human capital at the bottom or outside of the keju system into entrepreneurship. This paper illustrates the dynamics between indigenous institutions and external pressures.

Link

52. Industry Shakeouts after an Innovation Breakthrough

Xiaoyang Li

Conventional wisdom suggests that after a technological breakthrough, the number of active firms first surges, and then sharply declines, in what is known as a “shakeout”. This paper challenges that notion with new empirical evidence from across the U.S. economy, revealing that shakeouts are the exception, not the rule. I develop a statistical strategy to detect breakthroughs by isolating sustained anomalies in net firm entry rates, offering a robust alternative to narrative-driven approaches that can be applied to all industries. The results of this strategy, which reliably align with well-documented breakthroughs and remain consistent across various validation tests, uncover a novel trend: the number of entry-driven breakthroughs has been declining over time. The variability and frequent absence of shakeouts across breakthrough industries are consistent with breakthroughs primarily occurring in industries with low returns to scale and with modest learning curves, shifting the narrative on the nature of innovation over the past forty years in the U.S.

Link

53. Are Societal Promises in Science and Technology Substantiated? A Study of Value Expressions in Patents

Sergio Pelaez

This study explores the relationship between value expressions (VEs) in patent documents—narrative statements that describe societal or commercial problems, solutions, or promises—and the technological orientation of patents in artificial intelligence and nanotechnology. Using generative and discriminative large language models (LLMs), we categorize public value expressions (PVEs), private value expressions (PRIVEs), and combined expressions (PVE-PRIVEs) from 175,730 U.S. patents (2005-2023), totaling 7.1 million sentences. While all patents have a technological orientation based on their technical features, only some are considered socially oriented—those whose technological capabilities align with societal challenges as defined by the UN's Sustainable Development Goals (SDGs). Using a technology-based mapping of patents to SDGs provided by LexisNexis Intellectual Property Solutions, we analyze how narrative VEs correlate with this objective technological orientation. Our findings reveal that patents containing PVEs are more likely to be socially oriented based on their core technical features, while those with PRIVEs show the opposite pattern. Notably, patents articulating both public and private values (PVE-PRIVEs) demonstrate the strongest alignment with social orientation. These results challenge the notion that value-based statements in patents are merely rhetorical, suggesting instead that they meaningfully reflect or signal an invention's potential societal impact. This study contributes to the ongoing discussion on the role of value-based statements in science, technology, and innovation (STI) and highlights the potential of VEs as useful elements in responsible innovation frameworks.

Link

54. A New Order? Digital Disruption and Entrepreneurial Opportunities

Javad Shamsi

Does the rise of digital marketplaces primarily benefit large incumbent firms or facilitate the entry of entrepreneurs, including those from minority backgrounds? This paper studies the growth of food delivery applications in the UK—UberEats and Deliveroo—and their impacts on local restaurants. To study this, I construct a novel dataset that measures the staggered spatial expansion of these apps and I employ a dynamic differencein-differences framework. I find that app entry increases local restaurant counts (by 35%) and employment (by 12%) over four years and does not crowd out dine-in expenditures. This increase is driven by the entry of small and independent businesses, with ethnic minority entrepreneurs gaining disproportionately from lower entry costs and reduced dependence on prime locations. This democratization in entrepreneurship fosters greater diversity in cuisine offerings, enhancing consumer choice.

Link

55. Disasters, Capital, and Productivity

Tarikua Erda

The fixed nature of physical capital could delay adjustment to rapidly occurring climate shocks. Prior theoretical work predicts this would amplify economic damages from climate change as global warming intensifies disaster risk. Using rich confidential microdata from the US Census Bureau, I establish novel causal evidence on how manufacturing firms readjust their capital, particularly their machinery, in response to large, federally declared floods. While floods degrade capital, surviving plants replace capital and see higher productivity because they upgrade technology as they rebuild. I document substantial reallocation of second-hand capital across plants following floods, notably from low-productivity exiters toward well-performing young plants. Ultimately, capital adjusts relatively quickly, and is reallocated toward better use following flooding, boosting aggregate productivity. These outcomes stem from the expanded credit access that federal disaster spending creates. My findings reveal a new channel through which government disaster spending revives disaster-hit economies and underscore its crucial importance in a warming world.

Link

Thanks for reading! Back to our normal posting next time!

In this special annual edition of What’s New Under the Sun, we have a big bundle of the titles, abstracts, and links to innovation-related PhD job market papers from 2024. Some were sent my way following my request in the last newsletter, but to find the majority of these, I looked at the titles of job market papers from graduating PhDs at ~175 economics, business, and other departments. Even so, I’m sure I missed some great papers. If that’s you, email me and I’ll add you to the posts.

I’ve split this post into three to make it a bit easier to navigate. This is the second post. The first post is here, and the last is here. Subscribers should receive all three in their email.

Subscribe now

Titles Index

Titles are presented in random order. There might be additional authors on these papers - I’ve listed the associated job market candidate only.

21. Innovation Spurred: Evidence from South Korea's Big R&D Push by Luis F. Jaramillo

22. Firewall for Innovation by Jie Zhou

23. Unveiling the Hidden Risk Premium of Knowledge-Intensive Firms: Introducing the HKR Factor by Pedro Vallocci

24. Shifting Tastes, Advancing Technologies: A New Perspective on Income Inequality by Mila Markevych

25. How Patient is Venture Capital? by Namrata Narain

26. The Long-run Impacts of Ancient Chinese Civil Exams on Contemporary Local Innovation by Chenxi Tang

27. Exploring the Role of Social Media in the Diffusion of Economic Research by Yotam Sofer

28. Shifting Gears: Environmental Regulation in the Car Industry and Technological Change Among Suppliers by Johannes Gessner

29. Commercializing Contrarian Ideas: Evidence from AI Contests by Luca Gius

30. R&D Expenditures, Productivity, and Wage Inequality: Evidence from an R&D Tax Credit by Amirhossein Tavakoli

31. Innovation and Adoption in the presence of distortions by Álvaro Pinzón

32. Commission Fee Structure and Innovation in Digital Platforms by Byoungmin Yu

33. Customer capital and firm innovation by Duong Dang

34. Lock-In and Productive Innovations: Implications for Firm-to-Firm Innovation Pass-Through by Lucía Casal

35. Tariffs and Innovation in a Schumpeterian Economy with North-South Technology Transfer by Florence Ut Meng Ho

36. Mergers and Mismatches in the Labor Market for Creativity by Ke Shi

37. Combining Complements: Theory and Evidence from Cancer Treatment Innovation by Rebekah Dix

38. Lobbying, Innovation and Aggregate Productivity by Nasir Hossein Dad

39. From Research to Development: How Globalization Shapes Corporate Innovation by Chan Kim

40. Machine Versus Muscle, Bot Versus Brain: Effects of Artificial Intelligence on Heterogeneous Skill Groups by Wenjia Cao

Titles and Abstracts

21. Innovation Spurred: Evidence from South Korea's Big R&D Push

Luis F. Jaramillo

We study how South Korea’s first “mission-oriented” R&D program, implemented between 1992 and 2001, shaped innovation and economic outcomes. Using new textual data and a language model to identify targeted and control technological classes, we exploit the fact that some of the planned research projects were not implemented because of budget shocks. We use a local projections event study to compare the outcomes of targeted technological classes to those of control classes. Despite the absence of differential trends before the program, by ten years after the extension of program support, future-citation-weighed patenting output in the targeted classes doubled and real exports tripled relative to the control technology classes. These results stand when we study cross-country evidence. Technological classes with less concentrated patenting output before the program drive our results. Using market-based patent valuations, we find that the program’s benefits exceeded its costs by over a factor of three. Our findings suggest that technology policy was central to South Korea’s transition to a knowledge-intensive economy.

Link

22. Firewall for Innovation

Jie Zhou

Do protectionist policies foster domestic growth and innovation in the digital economy, and if so, how? This paper investigates the impact of the Great Firewall (GFW) in China – the world's largest system of internet regulation – on the development of domestic mobile apps. By blocking foreign apps at times determined mostly by political considerations, the GFW prompted a 30% user base expansion for Chinese substitute apps (identified through their baseline text descriptions). Monthly data on these apps’ underlying technologies, extracted from their compiled source code, reveal that Chinese substitute apps accelerated their innovation efforts, with in-house development increasing by 14% two years after the blockage. This technological progress spilled over broadly post-blockage, as both domestic and foreign apps adopted more Chinese technologies. I further show that increased access to data was one important driver. Chinese apps requested more types of sensitive data and were more likely to share user data access with outside firms after their foreign substitutes were blocked. These increased types of user data generate innovation; quasi-random variation in the introduction of new data access raises in-house technology development. Finally, using data-sharing networks between app developers, I show that in-house development also increased at the firms that user data was shared with. In summary, protectionist policies brought about through China's GFW boosted its app industry, potentially contributing to China's leadership role in this fast-growing industry.

Link

23. Unveiling the Hidden Risk Premium of Knowledge-Intensive Firms: Introducing the HKR Factor

Pedro Vallocci

This paper introduces a novel measure of knowledge capital risk derived from textual analysis of firms' 10-K filings. Using cosine similarity to R&D-related terms, I construct a High-Knowledge Risk (HKR) factor that captures firms' exposure to uncertainty from R&D outcomes. I demonstrate that the HKR factor offers significant explanatory power for cross-sectional equity returns, even after controlling for established equity pricing factors. This explanatory power persists in out-of-sample tests. Portfolios formed on HKR exhibit a substantial return premium, with high-HKR equities outperforming low-HKR equities over a 14-year period. A trading strategy based on the HKR factor also displays the highest Sharpe ratio among compared factors, indicating an attractive risk-adjusted return. These findings suggest that knowledge capital risk represents a distinct source of priced risk not fully captured by traditional asset pricing models and highlight the importance of knowledge capital risk in asset pricing.

Link

24. Shifting Tastes, Advancing Technologies: A New Perspective on Income Inequality

Mila Markevych

This paper examines how changing consumer demand affects income inequality in the context of technological change in the US. I develop a general equilibrium structural transformation model that incorporates time-varying demand shifters –Temporal Demand Growth Factors (TGFs). The estimates of TGFs reveal significant heterogeneity in demand patterns across goods and households. Counterfactual analysis shows that TGF-driven demand effects substantially moderate the rise in income inequality due to technological change. In the absence of these demand effects, the increase in income inequality between 1989 and 2021 is 73% larger. Changes in demand particularly benefit workers in less productive and more labour intensive non-routine manual and routine cognitive sectors, consistent with Baumol’s cost disease. The reallocation of economic activity towards sectors with lower productivity growth, driven by changes in demand, is associated with more equitable income distribution, suggesting that demand driven slowdown in productivity growth is not necessarily detrimental to our economic wellbeing.

Link

25. How Patient is Venture Capital?

Namrata Narain

Venture capital is a major source of finance for innovation in the U.S. economy, but how successful has it been in financing long-term innovation? I evaluate the allocation of venture capital to technologies, exploiting a new technology-level measure of the expected time between investment and innovation. Using natural language processing, I assign innovation funding between 1980-2022 from the following sources to technologies: venture capital firms, public companies, and six government programs and agencies. Venture capital is disproportionately allocated to technologies with short lags between investment and innovation; this allocation resembles that by public companies and is shorter-term relative to the distribution of commercial value. I show that the desire to raise follow-on funds (“fundraising pressure’’) leads venture capital fund managers to invest in innovation with short lags.

Link

26. The Long-run Impacts of Ancient Chinese Civil Exams on Contemporary Local Innovation

Chenxi Tang

China was historically heavily influenced by civil exams (keju), with the exams focused on recruiting the best academic individuals to serve in government. This paper explores the relationship between prefectures with historically more top scorers on the national exam (jinshi) between 1371 and 1905 and contemporary innovation, measured as the number of top scientists and engineers and the number of patents and their quality, finding a strong positive relationship. Results are robust to using an instrumental variable strategy that measures the minimum river distance from each prefecture to the nearest pine and bamboo forest. A doubling of the number of historically top scorers on the national exam leads to a 33% increase in the number of top scientists and engineers and a 92% increase in the number of patents. Investments in military equipment and telegraph construction play crucial roles sustaining the long-term effects of China’s civil exams.

Link

27. Exploring the Role of Social Media in the Diffusion of Economic Research

Yotam Sofer

For more than a decade, social media have become a key channel for knowledge dissemination used by scientists as a whole, and economists in particular. However, their role in the diffusion of knowledge is understudied. This article investigates the role of social media visibility of working papers on diffusion outcomes. While previous studies focused on the diffusion of STEM research, this article explores the diffusion of economic research. To do so, a data set of all NBER working papers published between 2015-2018, covering their social media mentions, as well as bibliometric and altmetric indicators, is used. To estimate the causal effect of social media visibility on diffusion, an instrumental variable approach, leveraging quasi-random variation in social media posting policy of the NBER's communication office, is employed. The results indicate heterogeneity in the role social media play in the diffusion of economic research. Increased social media visibility of working papers positively affects the likelihood and the extent to which research is diffused to public discourse (measured by blogs and news mentions), within the first year from publication, as well as within the scientific community (measured by academic citations), four years post-publication. No effect on citations in policy documents was found. Lastly, the likelihood to publish a working paper in a peer-reviewed journal is found to be unrelated to social media visibility of the working paper. The results of this article provide evidence for the role social media play in the diffusion of economic knowledge.

Link

28. Shifting Gears: Environmental Regulation in the Car Industry and Technological Change Among Suppliers

Johannes Gessner

Decarbonizing industries to mitigate climate change requires technological change. Innovation by suppliers can play a crucial role in the technological transition, particularly when suppliers have expertise in zero-emission technologies. In this paper, I study the effect of environmental regulation in a downstream industry on the innovation outcomes of suppliers in the context of the European CO2 emission standard for passenger cars. I construct a novel data set that links administrative data on car manufacturer compliance to supplier patent data using information on automotive supply chains. To identify the causal effect of changes in the stringency of the emission standard, I leverage the heterogeneous exposure of automotive suppliers to changes in the composition of the European car market in the aftermath of the 2015 Volkswagen diesel scandal. Exposure to more stringent environmental regulation increases innovation for zero-emission vehicle technologies among existing suppliers. In addition, the likelihood that car manufacturers form new supply chain links to firms with expertise in technologies to reduce vehicle emissions increases in response to more stringent environmental regulation. These results suggest that environmental regulation induces economically significant technology spillovers to the regulated firms.

Link

29. Commercializing Contrarian Ideas: Evidence from AI Contests

Luca Gius

This work builds on the notion that some ideas are not only distinctive, but also contrarian: pursued by few, and met with open skepticism by most. Such skepticism hinders contrarians from attracting resources but allows them to experiment openly without fear of immediate imitation. To investigate this dynamic, I leverage hundreds of Artificial Intelligence contests where researchers either employ popular, state-of-the-art methods or pursue alternative, contrarian approaches. These contests act as sudden “moments of clarity” that reveal which methods perform best. They allow researchers to attract resources for commercialization, but also potentially alert competitors of the opportunity. Through a difference-in-differences design, I find that when contrarian contestants win, they are up to 7× more likely to found startups, and their startups attract up to 3× higher venture capital valuations. This cannot be simply explained by the fact that contrarians are more likely to achieve breakthroughs; much of the advantage arises by winning close races. Instead, while close victories validate contrarian entrepreneurs and allow them to attract resources, mainstream researchers tend to adopt contrarian methods only when they conclusively outperform traditional approaches. This reluctance to build on contrarian ideas allows contrarians to leverage public demonstrations to attract resources without provoking immediate competition.

Link

30. R&D Expenditures, Productivity, and Wage Inequality: Evidence from an R&D Tax Credit

Amirhossein Tavakoli

This paper examines the impact of R&D tax credits on between-firm and within-firm wage inequality. Leveraging a regression kink design and matched employer-employee tax records, I estimate a large and statistically significant increase in R&D expenditures. The results show that R&D-intensive firms respond to tax credits with substantial increases in R&D expenditures, leading to significant gains in profitability, productivity, and wages while non-R&D-intensive firms show minimal changes. These gains disproportionately benefit high-skill, older, and long-tenured workers, exacerbating wage inequality between and within firms. High-skill workers experience the largest earnings gains, with a 10 percent increase in generosity of the tax credits leading to a 1.6 percent rise in their annual earnings. In contrast, low-skill workers see no significant changes. These findings provide evidence of rent-sharing mechanisms and highlight the role of R&D tax credits in contributing to wage inequality.

Link

31. Innovation and Adoption in the presence of distortions

Álvaro Pinzón

Factor misallocation significantly contributes to the productivity gap between developed and developing economies. This paper examines the dynamic misallocation of R&D resources in Colombia’s manufacturing sector using a unique dataset on firms with over 10 employees, capturing comprehensive firm-level details on productivity, R&D, and market distortions, allowing for a precise assessment of innovation dynamics in a developing economy. I document distortions and the link between productivity, R&D activity, and innovation. I develop and calibrate a model to the Colombian manufacturing industry, where firms face market distortions and invest in R&D to improve productivity. The results show that reducing distortions to U.S. levels leads to a 64% increase in aggregate TFP, with 34% of this growth driven by R&D reallocation and improved selection. Notably, R&D resources shift from low- to high-productivity firms, even as overall innovation rates remain almost constant.

Link

32. Commission Fee Structure and Innovation in Digital Platforms

Byoungmin Yu

This paper quantifies the welfare effects of regulating commission fees in digital platforms, focusing on third-party app developers’ innovation and pricing decisions. I employ a comprehensive dataset of music apps within the Apple iOS store in the United States from October 2018 to February 2024 to estimate app users’ demand and app developers’ cost parameters. The paper reveals key findings with three policy counterfactual simulations where I sequentially solve for optimal innovation and pricing decisions. First, a cap on commission fees promotes innovative efforts by third-party app developers and improves social welfare. Second, when the platform adds a unit fee scheme under the fee cap, developers partly pass unit fees on to app users by increasing in-app purchase prices. Third, a hypothetical buy-out of a streaming app by the platform leads to a significant decrease in the innovative efforts and market share of the acquired app. Notably, welfare analysis without quality adjustment is predicted to underestimate the impact of fee cap on social welfare by 0.91% - 2.06% points compared to the full-stage model estimates. This research highlights the importance of considering quality changes along with price fluctuations when evaluating regulatory intervention in digital platforms.

Link

33. Customer capital and firm innovation

Duong Dang

This paper studies the role of customer capital in driving firm innovation decisions and the resulting effects on aggregate productivity and concentration. I develop a step-by-step innovation model where households form deep habits in consumption. These habits form customer capital for firms: firms can decrease prices and increase production to build customer capital and raise future profits, at a potential loss to current profits. As the strength of habits increase, leader firms face higher and more inelastic demand while followers face lower demand. I show how these movements in demand result in an increase in innovation by leader firms relative to follower firms, leading to greater productivity dispersion and concentration. I find evidence for this effect in data on U.S. public firms: in sectors where outputs are more heavily consumed by older households—those with stronger habits—the most productive firms increase their R&D investment relative to others. I discipline the strength of habits in the model base on micro estimates of household evolution of consumption. I then use the model to quantify the effects of changes in aggregate customer capital arising from aging demographics. The model suggests that the shift toward older households between 1980 and 2019 accounts for 10%-35% of the observed trends in rising revenue productivity dispersion among firms, increasing market concentration, and higher aggregate markups. The model also highlights how customer capital influences the effectiveness of innovation policies: with customer capital, innovation subsidies have a significantly larger impact on concentration and markups—around two to three times greater than in an environment without customer capital.

Link

34. Lock-In and Productive Innovations: Implications for Firm-to-Firm Innovation Pass-Through

Lucía Casal

Firms innovate to improve efficiency and reduce their costs of production (productive innovations) and to increase customer dependency by making products harder to substitute (lock-in innovations). In this paper, I quantitatively study the macroeconomic implications of lock-in innovations for aggregate productivity and market power. I develop a theoretical framework that allows firms to invest in lock-in innovations by reducing product substitutability, while also nesting standard macroeconomic models of productive innovations. A key prediction of the model is that productive innovations by suppliers increase customer firms’ sales by lowering input costs, while lock-in innovations decrease customer firms’ sales by allowing suppliers to charge higher prices for products that are harder to substitute. I use this theoretical insight to identify the nature of innovation in the data and calibrate the model to the U.S. economy. Informed by the observed changes in the response of customer firms’ sales to their suppliers’ innovations, I find that 37% of innovations are lock-in, and that their incidence has doubled in recent decades, especially for high markup firms. Moreover, had the incidence of lock-in innovations remained at pre-2000 levels, observed aggregate productivity would have been 3% higher, median markups would have stayed at pre-2000 levels, and markup dispersion would have been 9% lower.

Link

35. Tariffs and Innovation in a Schumpeterian Economy with North-South Technology Transfer

Florence Ut Meng Ho

This paper develops a North-South quality-ladder model with northern innovative R&D, southern adaptive R&D and imitative R&D to analyze the effects of tariffs on innovation, technology transfer, relative wage and welfare. We find that increasing southern tariff decreases the relative wage between the North and the South permanently, increases the technology transfer rate permanently and decreases the northern innovation rate temporarily. In contrast, increasing northern tariff increases the relative wage permanently, decreases the technology transfer rate permanently and either increases or decreases the northern innovation rate, depending on the size of the North-South labor ratio. Moreover, we calibrate this model to the US-China data to perform a quantitative analysis. We find that imposing tariff in the home country yields welfare gain in itself and yields welfare loss in the foreign country. The numerical results are consistent with the analytical policy implications.

Link

36. Mergers and Mismatches in the Labor Market for Creativity

Ke Shi

This paper introduces a novel empirical framework to assess the impact of ownership consolidation on labor markets, addressing growing concerns about labor market power. I develop a two-sided matching model tailored to the creative labor force, a segment characterized by strong worker-firm compatibility. Applying this model to a major merger in the U.S. publishing industry, I leverage rich text data to analyze its effects on the author labor market. Counterfactual merger simulations reveal a trade-off between efficiency gains, creative misalignment, and redistributive effects. While the merger alleviated capacity constraints, post-merger integration resulted in significant creative misalignment between authors and publishers. The merger also triggered substantial value transfers from competing publishers and authors to the merged entity, with established authors bearing the heaviest losses. Notably, the merger's anticompetitive effects manifested primarily in labor markets rather than consumer markets. This research extends merger evaluation beyond consumer impact, providing a framework for analyzing the broader consequences of mergers on labor markets characterized by worker-firm complementarities.

Link

37. Combining Complements: Theory and Evidence from Cancer Treatment Innovation

Rebekah Dix

Innovations often combine several components to achieve outcomes greater than the “sum of the parts.” We argue that such combination innovations can introduce an understudied inefficiency—a positive market expansion externality that benefits the owners of the components. We demonstrate the importance of this externality in the market for pharmaceutical cancer treatments, where drug combination therapies have proven highly effective. Using data on clinical trial investments, we document several facts consistent with inefficiently low private innovation: firms are less likely than publicly funded researchers to trial combinations, firms are less likely to trial combinations including other firms’ drugs than those including their own drugs, and firms often wait to trial combinations including other firms’ drugs until those drugs experience generic entry. Using microdata on drug prices and utilization, we quantify the externalities that arise from new combinations and find that the market expansion externality often dominates the standard negative business stealing externality, suggesting too little innovation in combination therapies. As a result, firms may have incentives to free ride off others’ innovation, which we analyze with a dynamic structural model of innovation decisions. We use the model to design cost-effective policies that advance combination innovation. Redirecting publicly funded innovation toward combinations with high predicted market expansion or consumer surplus spillovers minimizes crowd out of private investments, increasing the rate of combination innovation and total welfare while remaining budget neutral.

Link

38. Lobbying, Innovation and Aggregate Productivity

Nasir Hossein Dad

We study the impact of firms lobbying activities on innovation and aggregate productivity in the United States. We build a quantitative model where firms make decisions about lobbying and R&D investments to grow. Lobbying can either complement R&D by increasing its returns or substitute for R&D as an alternative way to boost profits, making the net effect theoretically ambiguous. To determine which effect dominates on average, we use firm-level lobbying data and a shift-share instrumental variable strategy to estimate the causal effect of lobbying on R&D expenditure. We find that lobbying significantly reduces R&D expenditure at the firm level. We calibrate the model to the U.S. economy and find that eliminating lobbying would lead to a 3.5% increase in aggregate productivity. The gains are primarily driven by improvement in firm-level productivity distribution, through an increase in firm-level innovation. We then use the model to evaluate the impacts of U.S. Senator Elizabeth Warren’s proposal to tax lobbying progressively and find that such a policy could increase aggregate productivity in the U.S. by 1.57%.

Link

39. From Research to Development: How Globalization Shapes Corporate Innovation

Chan Kim

I show that globalization has shifted U.S. corporate innovation from scientific research to commercial development. Analyzing data from publicly traded firms, I find evidence that substantial tariff reductions at export destinations following the “Uruguay Round'' led U.S. firms focus on a narrower range of technologies, reducing their emphasis on scientific research. To explain these findings, I develop a multi-product firm model that distinguishes between research and development. Globalization—modeled as an expanded market size—reallocates profits toward products for which firms hold a competitive advantage. Consequently, firms increasingly focus their innovation efforts on core products, favoring development as it more effectively targets them. The model embeds a crucial welfare trade-off: while a greater focus on development increases high-productivity products, a stronger emphasis on research enhances the overall innovation efficiency of the economy through knowledge spillovers. Calibration to U.S. manufacturing firms shows that these innovation decisions double the productivity gains from globalization but reduce welfare gains. The welfare-maximizing policy suggests that research subsidies should exceed development subsidies, particularly after globalization, to offset the decline in research share.

Link

40. Machine Versus Muscle, Bot Versus Brain: Effects of Artificial Intelligence on Heterogeneous Skill Groups

Wenjia Cao

This paper studies effects of artificial intelligence (AI) on employment and wages for heterogeneous skill groups in the U.S. by introducing and analyzing a task-based framework. I first categorize labor into four skill groups based on skill specializations: (1) abstract and AI-intensive; (2) abstract-intensive but not yet AI-related; (3) routine-intensive; and (4) manual-intensive. The demand for AI skills is then measured by matching phrases for AI-developing skills to descriptions of online job postings. I document a consistent upward trend in the share of AI postings for the high-skilled AI-complement group during my sampling period, 2012-21. There is a strong growth in both employment and wages for abstract and AI-intensive occupations associated with an increasing demand for AI skills, while abstract but not-yet-AI occupations have much smaller growth. Middle-skilled occupations experience wage declines associated with an increase in the standard deviation of the intensity that AI-developing skills are required for job tasks. Employment and wage gaps between abstract and AI-intensive occupations and other skill groups widen as the labor market favors workers with AI skills, consistent with my theoretical model's implications. I also discuss whether AI is possibly a general-purpose technology.

Link

Innovation Job Market Papers 2024 continues here.

In this special annual edition of What’s New Under the Sun, we have a big bundle of the titles, abstracts, and links to innovation-related PhD job market papers from 2024. Some were sent my way following my request in the last newsletter, but to find the majority of these, I looked at the titles of job market papers from graduating PhDs at ~175 economics, business, and other departments. Even so, I’m sure I missed some great papers. If that’s you, email me and I’ll add you to the posts.

I’ve split this post into three to make it a bit easier to navigate. This is the first post. The second post is here, and the third is here. Subscribers should receive all three in their email.

Subscribe now

But first, a few announcements.

• Two new living literature reviews have recently launched, with support from Open Philanthropy. Check them out and subscribe if you are interested in the topic:

  • Migration Literature Review by Lauren Gilbert, on all things immigration. See her first post on the academic literature spawned by The Mariel Boatlift.

  • Governing with AI by Justin Bullock, a review on the impact of AI on governance. See his first post on the history of AI in bureaucracy.

  • These join other living literature reviews on social collapse, the research to policy pipeline, AI in science, scaling in human societies, and interdisciplinarity!

• The National Bureau of Economic Research is seeking research proposals related to investments in early career researchers for a forthcoming conference. More information here.

• Finally, for all PhD students interested in innovation, note the Eighth Annual Wharton Innovation Doctoral Symposium is accepting proposals now. More information here.

Email me to suggest an announcement for the next newsletter. Now for the job market papers!

Titles Index

Titles are presented in random order. There might be additional authors on these papers - I’ve listed the associated job market candidate only.

1. Self-Selection and the Diminishing Returns of Research by Lorenz K.F. Ekerdt

2. A structural model of mentorship in startup accelerators: Matching, learning, and value creation by Mohaddeseh Heydari Nejad

3. Is green technological change skill biased? by Maren Holthe Hedne

4. Knowledge and Firm Growth in Space by Jack Liang

5. Innovation, Financial Frictions, and Hysteresis Effects of Monetary Policy by Ozgen Ozturk

6. Monetary Policy, Price of Risk, and Growth by Anindo Sarkar

7. How Do Robot Subsidies Affect Aggregate Productivity and Firm Dispersion? Theory and Evidence from China by Runhong Ma

8. Regulation-Driven Innovations: A Textual Analysis of U.S. Patents and Federal Regulations by Zhoudan Xie

9. Exporting State-Promoted Technologies and the Direction of Global Innovation: Evidence from 5G Standardization by Myeongwan Kim

10. Patent Protection in Developing Economies: The Role of Market Power and Technology Access by Weili Chen

11. Early Mentors for Exceptional Students by Ian Calaway

12. Minding Your Business or Your Child? Motherhood and the Entrepreneurship Gap byValentina Rutigliano

13. The Effect of Transport Infrastructure on Innovation: The Role of Market Access in the English Railway Boom by Giorgio Ravalli

14. Knowledge is (Market) Power by Jinglun Yao

15. Amenity-Biased Technical Change by Gerard Maideu-Morera

16. Asymmetric Information and Digital Technology Adoption: Evidence from Senegal by Deivy Houeix

17. Innovation and Technological Mismatch: Experimental Evidence from Improved Seed by Sergio Puerto

18. Human Capital and Growth: The Role of High-Skill Labor Concentration by Julio Brandao-Roll

19. Production Outsourcing and Innovation: Evidence from China’s Pharmaceutical Industry by Shi Gu

20. Learning from Multinationals by Minyoung Song

Titles and Abstracts

1. Self-Selection and the Diminishing Returns of Research

Lorenz K.F. Ekerdt

The downward historical trend of research productivity has been used to suggest that there are severe permanent diminishing returns of knowledge production. We argue that a substantial portion of the declining research productivity is a transitory phenomenon caused by self-selection in researchers' ability and the expansion of the research sector. To quantify these transitory diminishing returns, we develop a model of self-selected researcher supply and estimate it using data on the labor force share and earnings distribution of researchers. Our results suggest that the average ability of researchers has fallen dramatically. We then use our findings to revisit the estimation of the knowledge production function and its resulting prediction on long-run economic growth. We find that switching from an accounting framework without considering self-selection to one with nearly doubles the long-run growth rate of per capita income predicted by semi-endogenous growth models.

Link

2. A structural model of mentorship in startup accelerators: Matching, learning, and value creation

Mohaddeseh Heydari Nejad

Entrepreneurial success depends on reducing uncertainty about the quality of ideas and selecting effective strategies to bring the idea to market. Mentorship plays a critical role in this process. In this paper, I examine how mentorship improves entrepreneurial outcomes within the Creative Destruction Lab (CDL), a global mentorship-driven startup accelerator, through two channels: the direct effect of improving startup quality and the screening effect of identifying high-quality startups. Using mentorship interaction data from CDL, I apply machine learning algorithms to generate quantifiable measures of mentors' advice. I propose and estimate a structural model of mentorship, where the dynamics of quality accumulation are influenced by both the direct effect of mentors' advice and the screening effect from mentors' learning. I find that mentorship generates value through both direct and screening effects, with significant spillovers of quality signals between mentors. This model enables a counterfactual analysis, quantifying the value added by mentors when they actively shape the strategic direction of startups, compared to a more passive role where they support the execution of the entrepreneurs' original plans. The counterfactual analysis shows that entrepreneurs benefit from mentors' strategic guidance, with significant heterogeneity across sectors. In emerging sectors like quantum, mentors' strategic input has minimal impact, especially early on, suggesting that a more passive mentorship approach may be more beneficial. In these sectors, screening gains grow over time as mentors accumulate information and provide guidance that better reflects the true quality of the startups. These results offer important managerial implications for the design of intermediaries, such as accelerators that provide mentorship, suggesting that guidance approaches should be tailored to the specific needs and developmental stages of each sector.

Link

3. Is green technological change skill biased?

Maren Holthe Hedne

Technologies that limit greenhouse gas emissions have become a key component of modern technological change, but little is known about their impact on labour market inequality. This paper studies whether and when green technology increases the demand for workers with and without higher education. Green technologies differ from technologies previously shown to be skill biased because their main function is to improve clean relative to dirty production, rather than augment or replace types of labour. Using linked employer-employee data and a novel shift-share instrument to account for endogenous technology choices, I show that the skill-bias of green technologies is highly heterogeneous. Specifically, I find that firms in manufacturing increase their share of high-skilled workers in response to an imported green technology shock, while firms in most other sectors (such as construction) decrease their skill ratio. The political feasibility of climate policies depends on the winners and losers of the green transition. My results imply that the green transition is not inherently skill-biased — instead, the labour demand and income inequality effects of green technologies vary significantly depending on the sector where the green transition occurs.

Link

4. Knowledge and Firm Growth in Space

Jack Liang

I study how knowledge facilitates firm-level growth across space. Using US Census microdata, I identify establishments specialized in knowledge production for other establishments of the firm. Firm growth is disproportionately concentrated near these establishments, suggesting that within-firm geographic frictions inhibit the perfect replication of knowledge across production units in space. I show that these specialized knowledge establishments relate to increased learning and adoption of firm knowledge in the firm’s production establishments. These findings motivate a dynamic model of firm growth via the accumulation of knowledge in which firms jointly determine where to locate their production of knowledge and output. The firm’s problem is dynamic, combinatorial, and features many continuous state and choice variables, so I solve it using a novel computational algorithm. I estimate the model and use it to understand the effects of geographic shocks. Counterfactual analysis demonstrates that firms’ knowledge investment decisions amplify the welfare effects of local productivity shocks by 21% and propagate these effects to other regions.

Link

5. Innovation, Financial Frictions, and Hysteresis Effects of Monetary Policy

Ozgen Ozturk

This paper examines the link between access to external finance and the long-term impact of monetary policy on productivity growth. By leveraging loan-level data merged with firm-level balance sheet information, we show that firms' R&D expenditures decline after a monetary tightening, with heterogeneous responses. Firms that lack access to external finance for funding R&D activities experience sharper cuts in R&D spending compared to those with better access. Within an endogenous growth model with nominal rigidities and financial frictions, we interpret this pattern as access to external finance enables firms to sustain innovation during periods of monetary tightening. Our model findings suggest that these short-term impact of monetary policy on R&D investment can have long-lasting effects on productivity, as current R&D efforts drive future productivity growth. Additionally, we show that when firms are provided with the financial flexibility to borrow to finance innovation activities, and monetary policy targets the output gap, it is possible to stabilise output without inducing hysteresis effects.

Link

6. Monetary Policy, Price of Risk, and Growth

Anindo Sarkar

We uncover a novel channel by which monetary policy affects the economy’s supply side through affecting risk premia. In this channel, monetary policy affects the effective risk aversion, that is, the price of risk in the economy. This impacts equilibrium R&D investments and, eventually, TFP growth. Using an asset pricing model, we construct measures of the price of risk shocks and show that increases in the price of risk decrease aggregate R&D. We then quantify the contribution of our channel to the overall R&D and TFP growth response to monetary policy shocks by constructing an endogenous growth model with time-varying risk aversion. Using the model, we find that the price of risk channel accounts for 20% of reaction of R&D and 33% of the reaction of TFP growth to unanticipated monetary Policy.

Link

7. How Do Robot Subsidies Affect Aggregate Productivity and Firm Dispersion? Theory and Evidence from China

Runhong Ma

This study examines the effects of robot subsidies in China’s manufacturing sector. Exploiting differences in the timing of the subsidy implementation across municipalities, I find the introduction of a robot subsidy has heterogeneous impacts across firms of different scale. Although the subsidy results in a 13 percent increase in applications for robot patents, the facilitated access to robotics leads to a 14 percent reduction in new firm's entry in the manufacturing sector, along with a significant increase in turnovers of bigger industrial enterprises. Using a stylised model, I show that the interaction between financial frictions and endogenous automation helps reconcile the empirical findings: ex-ante capital misallocation causes a uniform subsidy to disproportionately benefit firms with better access to capital. The distortion creates an efficiency trade-off: while a subsidy can enhance overall automation, it also exacerbates automation dispersion, which reduces efficiency. To quantify the net efficiency impact of these competing forces, I embed this mechanism into a dynamic heterogeneous firm model, calibrated to match key features of the Chinese industrial sector. The model indicates that a robot subsidy of 20% narrows the gap between mean and optimal automation levels by 22 percentage points, while raising automation dispersion by 49 percentage points. This leads to a 1.2 percent increase in aggregate output, along with a 2.4 percent decline in total factor productivity.

Link

8. Regulation-Driven Innovations: A Textual Analysis of U.S. Patents and Federal Regulations

Zhoudan Xie

Some innovations are developed to comply with or circumvent legal and regulatory requirements. While these regulation-driven innovations can generate societal benefits, they may also incur unintended economic costs. This paper explores this unique type of innovation and examines its relationship with firm dynamics, creative destruction, and economic growth. I present a simple Schumpeterian model demonstrating how regulation-driven innovations can serve as a strategy for firms to achieve higher growth, deter competitors, and reduce the rate of creative destruction. Guided by the model’s implications, I identify regulation-driven innovations from U.S. patents issued between 1976 and 2020 by measuring the degree of alignment between patents and federal regulations. I construct this measure by estimating textual similarities between patent documents and regulatory texts using natural language processing techniques. Linking the measure with patent- and firm-level data, I find that innovation-regulation alignment is positively associated with the economic value of patents and the growth in size and market power of innovating firms. At the aggregate level, however, the static gains for innovating firms fail to offset the dynamic social costs from reduced reallocation and competition.

Link

9. Exporting State-Promoted Technologies and the Direction of Global Innovation: Evidence from 5G Standardization

Myeongwan Kim

Standardization ensures compatibility but potentially shapes innovation by locking in certain technologies. Unlike other countries, the Chinese government coordinates its firms to advance specific domestic technologies in international standard-setting organizations (SSOs). This paper studies the impact of this policy on global innovation in 5G. In the SSO for 5G, which is an economically and geopolitically significant arena in which this policy is implemented, firms compete to have their patented technologies adopted as part of the 5G standards. Using a large language model, I build a new database linking SSO technical documents, 5G policy documents published by the Chinese government, and 5G patents. I show that the policy promotes Chinese technologies in areas where China lags behind foreign competitors. If adopted as standards, these lagging technologies become the basis for subsequent 5G innovation across countries, as measured by 5G patents in close textual alignment with the SSO documents describing these technologies. These follow-on patents account for two-fifths of total 5G patents filed worldwide after standardization. In addition, many of the promoted technologies span both civilian and military applications. China's objectives in 5G extend beyond commercial interests.

Link

10. Patent Protection in Developing Economies: The Role of Market Power and Technology Access

Weili Chen

This paper examines the trade-off between market power and access to advanced technologies in the context of patent protection policy in developing economies. I exploit a policy shock in China that unevenly strengthened patent enforcement across provinces, combining it with a novel dataset that links Chinese firm-level production data to multinational firms’ global patent portfolios. I find that stronger patent protection incentivizes multinationals to adopt their best technologies in Chinese affiliates and encourages domestic inventors to produce higher-quality innovations. However, both groups also raise their markups. To rationalize and quantify these findings, I develop a multi-product model where inventors endogenously reduce markups but withhold higher-quality products due to concerns over local imitation. Enhanced patent enforcement drives out imitators, incentivizing inventors to adopt superior products while raising markups across their portfolios. The calibrated model reveals an inverted U-shaped relationship between patent protection and aggregate welfare, with strong heterogeneity across industries.

Link

11. Early Mentors for Exceptional Students

Ian Calaway

Although we are acquainted anecdotally with extraordinary people like Mozart and Marie Curie, there is little systematic research on how children with exceptional ability develop into truly extraordinary talents. Is the supply of extraordinary talent inelastic, dependent on a rare combination of innate gifts and the availability of mentors who are themselves world-class (Irène Joliot-Curie and her mother Marie)? Or, could the supply be fairly elastic because mentors need only have abilities within the normal range? I analyze these questions in the context of mathematics, where there is a consensus on how exceptional ability presents itself in children. I show that mathematics teachers who organize clubs and competitions can identify and foster exceptional math students, causing them to win honors, attend selective universities, major in STEM fields, and have careers in which they disproportionately spur economic growth. I demonstrate that there are many exceptional math students without mentors who could be reached with modest investments.

Link

12. Minding Your Business or Your Child? Motherhood and the Entrepreneurship Gap

Valentina Rutigliano

Women are less likely than men to start firms and female entrepreneurs are less likely to succeed. This paper studies the effect of childbirth on women’s entrepreneurial activity. Drawing on rich administrative data from Canada and using an event study and instrumental variable design, I show that childbirth has substantial negative effects on women’s founding rates and firm performance, accounting for a large portion of the gender gap in entrepreneurship. The impact spills over onto workers, who experience a decrease in earnings. The effects are permanent: entrepreneurial outcomes never recover to their pre-birth levels. The results are not due to a reduction in risk-taking and cannot be fully explained by household specialization based on labor market advantage. Childcare availability, progressive gender norms, and access to credit reduce the adverse effect of childbirth on the entrepreneurship gap.

Link

13. The Effect of Transport Infrastructure on Innovation: The Role of Market Access in the English Railway Boom

Giorgio Ravalli

What is the impact of transport infrastructure on innovation? I study the historically unprecedented boom of the railway network in 19th Century England, which reduced average travel time between London and Birmingham by 80% between 1830 and 1911. Using "accidentally connected" locations I show that building a rail station caused patents per capita to increase by over 75% in a district. I find no evidence that this effect was due to the relocation of existing inventors. The positive railroad effect is driven by an increase in market access (a demand-induced innovation effect). Across 1823-1861 the median Donaldson-Hornbeck measure of market access approximately doubles; a doubling of market access increases innovation by over 50%. This mechanism is economically and statistically significant. I also find that the railroad induced greater knowledge flows through improved communication between inventors; this effect appears distinct from the market access channel.

Link

14. Knowledge is (Market) Power

Jinglun Yao

US corporate concentration has been persistently rising over the past century with flattening Pareto tails, and productivity growth has been concurrently declining. This paper builds a continuous-time Schumpeterian growth model that interprets higher concentration as a result of lower growth, which complements the existing Schumpeterian literature that focuses on the opposite direction. In the model, laggards benefit from dynamic growth advantage, while leaders possess the static advantage inherent to their leading position. A uniform decline in research productivity hurts endogenous growth of all firms but in particular that of laggards, increasing the relative growth of leaders and fattening the Pareto tail of productivity distribution. With a demand system featuring variable demand elasticities, the proposed mechanism stemming from declining research productivity explains a majority of the changes in productivity growth, corporate concentration, markup, labor share, R&D cost, entry and exit rates, and job creation and destruction rates in the US since 1980s. The model can accommodate increasing concentration with stable markup and labor share in the pre-1980 period by introducing economic integration in addition to declining research productivity.

Link

15. Amenity-Biased Technical Change

Gerard Maideu-Morera

I argue that technical change has raised living standards not only by increasing wages but also by making work more pleasant and safer. Yet, traditional growth, distributional, or welfare accounting abstract from non-pecuniary job characteristics. By estimating shadow prices for job amenities, I first document an amenity-biased shift in labor demand in the US from 1980 to 2015, which reallocated workers from low- to high-amenity occupations. This reallocation significantly alters our understanding of several major macroeconomic changes. First, I theoretically show that the shadow value of amenities should be included in output to measure productivity growth. Otherwise, conventional measures---that only account for the costs of amenities---can underestimate it. Quantitatively, I find that total compensation (wage plus the value of amenities) grew 40% more than wages from 1980 to 2015. Compared to conventional estimates, this implies 25% higher productivity growth but a larger slowdown since 2004. Second, I find no labor market polarization along the distribution of total compensation; employment and relative wages declined the most at the bottom of the distribution instead of in the middle.

Link

16. Asymmetric Information and Digital Technology Adoption: Evidence from Senegal

Deivy Houeix

Digital technologies have the potential to increase firm productivity. However, they often come bundled with data observability, which can be a double-edged sword. Observability reduces information frictions and can increase efficiency, but some agents may lose their informational rent and thus resist adoption. I explore this trade-off between observability and adoption through two field experiments conducted over nearly two years. These experiments, guided by contract theory, introduce digital payments to the Senegalese taxi industry in partnership with the country's largest payment company. In the first experiment, I randomize access to digital payments for drivers (employees) and transaction observability to taxi owners (employers). I find that digital payments reduce drivers' cash-related costs by about half but also serve as effective monitoring tools for taxi owners. Transaction observability substantially increases driver effort, contract efficiency, and the duration of owner-driver relationships. However, 50% of drivers—primarily the worst-performing and poorest—decline to adopt digital payments when transactions are observable. The second experiment shows that the adoption rate doubles when drivers are assured that owners will not be able to observe their transactions. I develop a theoretical framework and use the experimental variations to estimate the welfare impacts of policy counterfactuals. I show that removing transaction observability would maintain moral hazard problems but broaden adoption and thus increase overall welfare—an approach ultimately implemented by the payment company. These findings highlight the crucial role of information embedded in digital technologies, as it magnifies gains for adopting firms but can deter initial adoption.

Link

17. Innovation and Technological Mismatch: Experimental Evidence from Improved Seeds

Sergio Puerto

Biases in research and development create a mismatch between the attributes of new agricultural technology and the preferences of low-income farmers. In this paper, I estimate the impact of this mismatch on farmers’ adoption of new drought-resistant seeds. Using a randomized controlled trial in Costa Rica, I recreated counterfactual scenarios for innovators’ seed development decisions by offering some farmers seed matching their preferences and others a seed variety chosen by crop scientists as a blanket recommendation. Results show that mismatch has a significant impact on adoption, with 41% lower uptake among farmers who were offered the recommended new seed. This gap was larger for farms located farther from the research lab where the new seeds were developed and persisted even in areas with drought exposure. Moreover, the new seeds were 31% more productive among farmers who adopted their preferred variety. To explain these findings, I propose a model where research constraints limit innovators’ ability to account for farmer heterogeneity. Matching new seeds to farmer preferences relaxes those constraints and increases productivity by enabling better adaptation to specific farm-level conditions, which are usually private information unknown to innovators

Link

18. Human Capital and Growth: The Role of High-Skill Labor Concentration

Julio Brandao-Roll

This paper raises and tests the hypothesis that the effects of human capital on economic growth depend crucially on the concentration of high-skill labor across firms. Importantly and surprisingly, an increase in human capital supply can actually lower growth if skill concentration across firms is high enough. Intuitively, large firms have limited financial incentives to innovate because they dominate the market and incur the risk of self-cannibalization when innovating; therefore, when increased skill supply primarily benefits these firms, the equilibrium growth impacts can be negative. I investigate this hypothesis in Brazil, establishing three results. First, in a difference-in-differences design across municipalities, I estimate that new colleges had a positive impact on local economic growth in municipalities with lower concentration of high-skill labor, but a negative effect in municipalities with higher skill concentration. Second, I isolate the causal effect of changes in local high-skill labor concentration on local growth using a shift-share design, leveraging loan shocks to firms. Third, I develop and estimate an endogenous growth model, which quantitatively matches the preceding results and which I use to assess policy counterfactuals. These results help explain why several middle-income countries, including Brazil, have experienced a slowdown in growth despite a fast increase in high-skill supply over the past decades.

Link

19. Production Outsourcing and Innovation: Evidence from China’s Pharmaceutical Industry

Shi Gu

This paper examines how removing barriers in production outsourcing affects firms’ innovation activities. I exploit a reform in China’s pharmaceutical industry that permits the outsourcing of drug production. Using a triple-difference identification strategy, I find that drug innovators without production facilities engage in more drug development (clinical trials) after the policy. However, for those firms constrained by limited resources, the increase in drug development is accompanied by a reduction in drug research (patent applications), particularly in low-value patents. This shift occurs because outsourcing increases the marginal value of development more than that of research. Further analysis reveals that the increase in clinical trials is more evident in the development of incrementally innovative drugs, rather than entirely novel ones. These findings suggest that overall, production outsourcing positively contributes to innovation, though its effects are primarily observed in incremental improvements.

Link

20. Learning from Multinationals

Minyoung Song

Multinational firms account for a large share of global trade and production and have a significant impact on productivity growth across countries. We develop a tractable model of international trade and multinational production (MP) to analyze their effects on productivity growth through knowledge diffusion across countries. We connect the theory to evidence by examining the impact of a 2004–2006 Chinese liberalization of outward investment on the productivity growth of its counterparts. Using an instrumental variable approach, we find a 1 percentage point increase in MP share from China resulted in a 1.85% increase in productivity. Mapping the reduced-form estimates to the rate of idea diffusion from foreign firms, we provide evidence on key parameters used in knowledge diffusion models. While long run productivity growth is primarily driven by domestic firms due to its larger knowledge diffusion parameter, MP can have a sizable impact on productivity growth, particularly when productivity differences between multinational and domestic firms are large. Increases in bilateral trade costs shift firm activity away from exporting to producing abroad, leading to productivity increases in the multinational’s production destination at the expense of the exporting nation. We quantify the impacts of Chinese liberalization of trade, as well as inward and outward MP, on the productivity growth of 59 countries.

Link

Innovation Job Market Papers 2024 continues here.

• Last year I put together a post collecting job market papers related to innovation. We’ll do that again this December, and if you are a PhD student going on the job market this year, and you want your paper included, email the title, abstract, and a link to matt@newthingsunderthesun.com.

• Astera is still taking applications for its full-year residency to build public goods related to science. More information here.

• Two new living literature reviews have recently launched, with support from Open Philanthropy. Check them out and subscribe if you are interested in the topic:

  • Scaling in Human Societies (How and Why Size Matters) (substack link) by Michael Goff

  • Bridging Boundaries (on Interdisciplinary research) by Rachel George

Email me to suggest an announcement for the next newsletter. On to the post!

This post was jointly written by me and Caroline Fry, assistant professor at the University of Hawai’i at Manoa! Learn more about my collaboration policy here.

This article will be updated as the state of the academic literature evolves; you can read the latest version here. You can listen to this post above, or via most podcast apps here.

Suppose you wanted to build up the scientific capacity of a country that is far from the scientific frontier. There are good reasons you might want to do that, rather than rely on the scientific efforts of countries on the frontier: where researchers are based affects what they choose to work on, and not all research is relevant everywhere. One part of building capacity is training scientists. In this post, we want to look at the evidence on the effects of training programs for scientists in lower and middle income countries (LMICs).  Training can come in two main flavors: domestic training, whereby activities take place in the LMIC itself, and training that leverages the international community, for example through fellowships to support study abroad or allocation of foreign mentors. The literature on the causal impact of training for LMIC scientists is thin, and predominantly focuses on programs that leverage the international community, as opposed to domestic training programs, so that’s what we’ll focus on today. 

Subscribe now

The STAARS program 

The Structural Transformation of Agriculture and Rural Spaces (STARS) program’s goal is to improve the productivity and impact of early-career economists in low- and lower-income countries. For the first several years of the program, it focused on African nationals who hold a PhD, with an emphasis on those in or planning to return to Africa, and was named the STAARS program (there’s an extra “A” in there; “AA” was for “African Agricultural”). It is predominantly a mentorship program, where fellows propose a research project and are matched to mentors from Cornell University and some other research organizations in high-income countries. Senior and peer mentors guide fellows on their research project, while the program provides accompanying technical and soft skill training and networking opportunities. The program typically lasts 9-15 months.

Schreiber et al (2022) describes and evaluates the impact of participation in the program on the research output of participating fellows. One persistent challenge in estimating the impact of training programs is that you can’t just compare people who get training to those who don’t, because training isn’t randomly assigned. The people who apply might be different than the ones who don’t; maybe more ambitious, maybe better connected (to learn about the opportunity), etc. Or maybe they apply for training precisely because they are struggling to succeed in their field. Meanwhile, among applicants, admission is not randomly assigned. The STAARS program, for example, has two peer reviewers rate the potential of proposals and selects successful fellows among those scored highly by peer reviewers.

So, to assess how well the program works, Schreiber and coauthors focus on finalists to the program, and compare finalists who participated in the program to those who didn’t. Participants in the program were finalists who additionally matched with a mentor, and non-participants are finalists who did not. Schreiber and coauthors argue this process has a lot of randomness baked into it. It might be that two applicants are equally qualified, but one happens to be working in a sub-field where a mentor is available that year and one is not. To ensure that successful fellows and finalists who didn’t enter into the program are comparable, they also employ propensity score matching using pre-fellowship publication output, gender, PhD institution, location and so on. 

The matched sample is small (31 participants and 31 matched non-participants), but results show that on average the participants received an extra 6 citations per year after the program, compared to matched non-participants. That’s a big effect: for comparison, matched non-participating finalists only average 4.5 citations per year. They also find participants publish more papers, but in this case the increase is too small to be confident the difference is not due to random variation. The paper also gathers some qualitative data, which highlights the value participants put on the professional development aspects of the program, including research ethics and peer review lessons. 

The European and Developing Countries Clinical Trial Partnership (EDCTP)

Training programs can also affect the research focus of participants. Let’s turn now to the impact of the European and Developing Countries Clinical Trial Partnership (EDCTP), whose stated mission is to improve drug development for key diseases affecting African countries (HIV, malaria and TB). This program organizes project-based funding and fellowships for African scientists to participate in cutting edge clinical trials with international teams. The teams apply for funding for a project and then work side-by-side for a number of years on clinical trials. Some of these trials ended up being pretty high impact; for example several trials contributed towards the development of the first malaria vaccines.

Fry and Blomfield (2024) look at how participation in an EDCTP funded trial amongst African scientists affected the direction of their subsequent research. To start, they identify 1,190 African scientists involved in an EDCTP trial between 2005 and 2014. They next merge the details of scientists who are involved in the EDCTP trials to their publication record and clinical trial involvement. To evaluate the impact of participation in a trial on the subsequent research trajectory of scientists, they compare the publications of scientists participating in these trials before and after the trial with those of a group of scientists who are also based in African institutions, but not participating in an EDCTP trial. Their goal is to see if scientists start working more on research related to clinical trials after participating in EDCTP, as compared to their research prior to EDCTP and as compared to otherwise similar scientists who never get a chance to participate. 

Of course, they have the same problem as mentioned above, that participation in an EDCTP research project isn’t decided at random. In particular, it may be that scientists interested in transitioning to work on clinical trials seek to join EDCTP and those not interested do not. If that’s so, then it might be the case that even if the EDCTP has no impact on research trajectories itself, we would still see people who joined start working more on clinical trials after it’s done. Fry and Blomfield follow a couple of strategies to deal with this. In their primary analysis, they try to match participants to scientists who don’t participate but are at similar points in their career, work on similar diseases, have similar levels of prior experience working on clinical trials and so on. The following figure compares involvement in clinical trials before and after participation in an EDCTP grant. About five years after a grant starts, careers of participants and matched non-participants begin to diverge. 

This is informative, to the extent that the non-participating matched scientists are actually very similar to the participants. One way to assess that is to identify some types of scientists who are more likely to be interested in clinical trials, and see if the results are different for that sub-population. To take two examples, Fry and Blomfield look to see if participation in an EDCTP grant has a larger impact for researchers who do more applied research, or who already work on HIV, Malaria, or TB. It doesn’t. Finally, interviews with both foreign and African scientists in these project teams were expressive about how important this collaborative project-based learning was in advancing their careers. There could still be the concern that EDCTP is just selecting people with a high propensity to excel in clinical research, rendering the question of whether or not the trial was effective in its training capacity. Recall however, the primary goal of the program is about increasing drug development, rather than training scientists; it’s possible a program primarily focused on training might have bigger effects.

Spillovers

So far we’ve looked at one paper (with a small sample) that showed participants who participated in training saw some increases in quantitative research outcomes, plus another that indicated participation in a program affected research interests. Next, we widen our scope to look for impacts on people who did not receive training themselves, but work with, or are exposed to, people who did. If training programs help individuals build better networks and skillsets, perhaps they can share those with their peers and build research capacity more broadly. Indeed, some other papers, such as those discussed in the post Local learning, suggest working physically together is an important channel for the dissemination of new ideas. On the other hand, as discussed in An example of successful innovation by distributed teams: academia, a number of papers have also looked at whether it helps a researcher’s productivity when they have other great researchers in their academic department. In the era of the internet, it’s not clear this makes much of a difference. 

In LMICs in particular, there are additional reasons why spillovers to other people might be higher or lower than they are in high income countries. On the one hand, it’s possible spillover effects will be unusually large, as the impact of new ideas and networks in a resource poor and constrained environment might just be bigger on the margin. On the other hand, it may be that spillover effects are unusually small, given constraints that make it difficult to leverage the potential benefits from these programs. For example, scientists might be limited in their capacity to take advantage of new connections or knowledge coming from colleagues who participated in advanced training programs. 

That said, two studies of foreign training programs imply that spillovers on peers at home are large. 

The NIH FIC AITRP 

The National Institutes of Health Fogarty International Center’s flagship program was the AIDS International Research Training Program: a training program for scientists from LMICs. Started in 1988, the goal of this program was to control the AIDS epidemic by training the next generation of leaders in research and development, on the ground where the epidemic was most devastating. The training offered to LMIC scientists was varied, but the focal program offered short- and long-term training (in the form of masters and PhDs) from public health departments and medical schools at top US institutions. This provided participating scientists the opportunity to train with some of the leading HIV researchers in the US, with explicit incentives to return home after training (like re-entry funding and visa restrictions). 

Fry and Ganguli (2024) study the impact of this program on HIV-AIDS research capacity in Africa. The basic idea is to look at what happens in academic departments among faculty working on diseases related to the research of the faculty members who participated in the training program. They evaluate the impact of the return of a trainee to an African institution across a variety of metrics: publications, grants, clinical trials and contributions to policy documents. To assess the impact of having a peer who participated in the training, they set up two comparison groups: peers working in the same institution, but in fields unrelated to the trainee’s field, and peers working in other institutions, but in fields closely related to the field of trainees (neglected tropical diseases - of which HIV is one). Comparing peers in related fields, at the same institution, to peers in unrelated fields at the same institution or scientists in related fields at different institutions, they document that the impact of the return of a trainee is large: peers publish more, particularly in HIV (with some peers moving into HIV who were not already publishing in the disease), and more often in journals with higher impact factor ratings. They also show an increase in grants obtained, HIV clinical trials, and contributions to policy documents. 

Most of the increase in publishing involves papers with international collaborators, not papers with the returning trainee. That suggests networking is an important channel through which these programs have spillovers. Indeed, another paper studying the same program, Fry 2023, studies the same program using a similar method, albeit exploring alternative outcomes. This paper documents specifically that returnees from the program impacted their peers through connecting them to the US based researchers in the training institution. In that paper, networking for previously isolated peers in Africa seemed to have the largest impact. If networking is in fact the primary spillover of training programs, that has implications for the optimal design of training programs. 

The Fulbright foreign student program 

Tying together those direct and indirect (spillover) effects of these foreign training programs, a series of papers from Kahn and MacGarvie study the impact of the well known US international training program: the Fulbright Foreign Student Program. This program, established in 1946, provides scholarships to students from other countries to pursue graduate study in the United States but requires them to return to their home country upon completion of their studies.

The first paper in the series, Khan and MacGarvie (2016), documents that Fulbright students were less productive upon return to their home country than comparable peers who were able to remain in the USA after graduation, especially if their home country had a lower GDP per capita (this paper is discussed more in the post Innovators who immigrate). But another 2016 paper by Kahn and MacGarvie explores the impact of return of these trainees on the diffusion of knowledge in their home countries. They find that articles by Fulbright Fellows who return home are cited more frequently in their home countries than articles by similar scientists who train in the USA but do not return home. This effect is especially strong among scientists from countries with a weak science base. This implies that the potential spillover benefits from the fellows who might be producing cutting edge research during and after their studies in the US are stronger if they return home. This is in contrast to the private research productivity benefits of remaining in the US, documented in that Kahn and MacGarvie’s first paper. 

Summing Up

While these latter papers focus on spillovers, rather than the impact of training on the trainee, it seems likely that if people who did not attend a training benefit from being around someone who did, then the training is probably doing something for the participant as well. More broadly, another reason to be a bit less skeptical that training programs really do change participants’ research trajectory comes from evidence on the impacts of training programs in other contexts. The post Teachers and the transmission of excellence surveys a few studies that indicate students of great researchers tend to have more successful research careers than otherwise similar students who do not. Another post, Students get interested in what their mentors are interested in looks at other studies that explore how who you study under affects the topics a researcher chooses to study. And Teaching innovative entrepreneurship looks at the (mixed) evidence that you can train science and engineering students to be good entrepreneurs. Viewed in that light, the hypothesis that training programs for LMIC scientists simply do what they seem to do - help participants improve their skills and networks - seems reasonable!

Still, there’s a lot more we don’t know. We’ve focused on training organized by foreign groups, but “training” is a bucket for a lot of different kinds of programs: mentorship, project-based learning, short-term workshops, and so on. Most of the studies we’ve examined focus on a bundle of programmatic elements, rather than isolating the impact of their specific elements, though we have seen some suggestive evidence that networking might be an especially important channel for impact. Moreover, while the outcomes of many of the empirical exercises in the above studies center on publication output in the years after training, Schreiber et al (2022) (which studied the STAARS program) notes that the goals of training programs can vary widely, and it would be valuable to have more careful evaluations of some of these other goals.

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.

If you want to read more, the following posts were mentioned above:

• Geography and what gets researched

• When research over there isn’t helpful here

• Local learning

• An example of successful innovation by distributed teams: academia

• Innovators who immigrate

• Teachers and the transmission of excellence

• Students get interested in what their mentors are interested in

• Teaching innovative entrepreneurship

What’s a living literature review?

“Living Literature Review” is the best name I’ve come up what I’m writing here. The key characteristics I have in mind are:

• Collection of articles

• Synthesize recent academic research.

• Written for non-specialists

• Rigorous and credible

• Updatable

• One primary author with relevant expertise

See this longer explanation for more discussion and some justification for these characteristics.

In my case, I have a website that meets the above characteristics, but I also send out new articles and updates to existing articles via this substack.

What are some examples of living literature reviews?

As I’ll discuss towards the end of this post, Open Philanthropy has a program to support people writing living literature reviews. Besides New Things Under the Sun, we’re currently supporting three other reviews, presented here in order from newest to most established.

Some Are Useful

Tom Gebhart, a University of Minnesota Research Scientist writes about how AI is used in different domains of science. For example, he is in the midst of publishing a series on the use of AI in weather forecasting. But his most recent piece (not on weather forecasting) is particularly relevant to New Things readers.

Some Are Useful

How Much New Knowledge is Hidden in Scientific Text?

Imagine that the pieces of a puzzle are independently designed and created, and that, when retrieved and assembled, they then reveal a pattern-undesigned, unintended, and never before seen, yet a pattern that commands interest and invites interpretation. So it is, I claim, that independently created pieces of knowledge can harbor an unseen, un-known, an…

Read more

2 years ago · Tom Gebhart

Check out the website or subscribe to the substack version.

Good Questions Review

Paul Kellner, Research Fellow at the Monash Sustainable Development Institute, recently launched a review on designing and implementing more valuable social science research. It’s very meta: academic social science on producing good academic social science! His first two posts are Do research findings need to be timely to influence policymaking? and Choosing policy relevant research questions.

Check out the website here.

Existential Crunch

Florian Jehn, Data Science Lead at ALLFED, writes on what we know about societal collapses, both in the distant past and (potentially, but I hope not) the near future. Here’s a fairly recent post:

Existential Crunch

What factors allow societies to survive a crisis?

This post is part of a living literature review series of societal collapse. You can find the most recent version here…

Read more

2 years ago · 8 likes · 4 comments · Florian U. Jehn

Check out the website or subscribe to the substack.

Hopefully this is just the start: we’ve made some additional grants to support more living literature reviews which will be coming soon. Meanwhile other academics write newsletters that I think are quite similar in spirit: Emily Oster’s ParentData comes to mind, or Alice Evans’ The Great Gender Divergence.

Time for more living literature reviews

The genre may be tiny today, but someday I hope there will be hundreds of living literature reviews, maybe thousands. For anything you want to know about, I would love for you to able to read an accessible, short, current, credible synthesis of frontier research.1

I think such a world would be a healthier innovation ecosystem. It would be easier for policymakers, voters, and other decision-makers to understand what research says about a particular topic. It would be easier for aspiring researchers to survey the landscape of possibility and race to the academic frontier of their choice. It would also be easier for active researchers to quickly pivot to new topics.2 And maybe it would even be easier to justify public support for academic research, when academia can more easily point to accessible explanations of what it is doing.

But for today, I think it would be great if we could just get some more examples of the genre, to test out if they’re really as valuable as I think they might be.

Should you write a living literature review?

If you’re an academic who believes in the value of the work done in your specialty, you should consider writing a living literature review. We have a program at Open Philanthropy to provide financial support to help you work on it. If your employer is supportive, these grants can be used to reduce teaching loads, provide summer salary support, or allow you to reduce other duties to allow you to work quarter-time on writing and maintaining a living literature review.

I think of writing a living literature review as sort of analogous to teaching a college course on a topic. Indeed, when I was at Iowa State University I did create and teach a course on the Economics of Innovation. A key difference, of course, is that the reach of writing online can be dramatically higher. I’m proud of the course I created, but it had 12 students and New Things Under the Sun has more than 16,000 subscribers. And nobody is reading New Things Under the Sun just to meet some kind of credit requirement for graduating.

Like a college course whose syllabus you might update once a semester to reflect research, the updatable format of a living literature review lets you keep what you write reflective of frontier research. But my efforts on New Things Under the Sun accumulate in a way that teaching efforts didn’t. The majority of my time working on New Things Under the Sun goes into writing new posts, rather than updating old ones. In contrast, after an iteration or two, the majority of my teaching sessions were a variation on something I had taught before. Over time that adds up. Today there are 85 posts up on the website, delving3 much deeper into niches than I would ever be able to justify during a 15-week course.

More selfishly, writing a living literature review is also valuable to me on a personal level. It’s a commitment device to keeping up with the literature. And figuring out how to explain something in plain language forces me to understand what I read at a higher level than if I was reading purely for my own interest. New Things Under the Sun has also become a very useful personal database / set of notes - when I need to find or refresh my memory about an article, it’s the first place I turn.

More broadly, writing online can be good for your career, especially when that work is closely related to your other professional duties, as writing a living literature review is for an academic. I think this is especially the case now. Every field has a lot of good researchers. Most fields do not have someone writing a good living literature review. Personally, I owe most of the professional opportunities I have had to New Things Under the Sun.

So I hope that if you are an academic yourself, you’ll consider it. If you’re interested, please drop me an email (matt.clancy@openphilanthropy.org) to chat more about it.

Announcements

• Nature has a nice write up of the UK Department of Metascience, whose latest grant program Open Philanthropy is co-funding. We want to support science agencies and organizations interested in using science to improve their policies and processes. If that’s you, email me!

• NeuroLibre Day: On September 27, in Montreal, there will be a symposium on reproducible publishing and the beta-releast of NeuroLibre - an open-source server for hosting reproducible research objects.

Email me to suggest an announcement for the next newsletter. On to the post!

Subscribe now

Here’s a chart from FT columnist John Burn-Murdoch, showing how language about progress in English, French, and German books has changed over the last few centuries. The share of words associated with progress rose during the era of industrialization, but is down since the 1950s. Meanwhile, words associated with worry and risk are up.

Who cares? Well, there’s a school of thought that cultural attitudes towards progress are an important driver of innovation. The general idea is that societies which valorize innovation and progress get more of it: they inspire more people to become innovators, their governments place a higher priority on supporting innovation in regulation and education policy, and their affluent class are more willing to invest in the future. Indeed, Burn-Murdoch shows that the prevalence of progress-oriented language rises in England well before Spain, and also that GDP per capita began to rise in England before Spain as well.

Burn-Murdoch’s chart is inspired by Almelhem et al. (2023), which looks at how written language in England changed between 1500 and 1900. Their goal is to find some quantitative support for an influential theory of the industrial revolution from economic historian Joel Mokyr. Mokyr (most notably in his book A Culture of Growth) argued that one important cause of the British industrial revolution was a belief in the possibility of progress and the virtue of finding tangible improvements in things like industry. When this belief collided with the fertile ground of Britain’s artisanal class and a growing base of useful knowledge (derived in part from the sciences), then sustained productivity growth began. Almelhem and coauthors want to look for evidence consistent with this story by seeing if there is an uptick in language about progress in the runup to the industrial revolution.

To do that, they use word frequency data for more than 170,000 works from the Hathitrust digital library, a collaboration of research libraries that has digitized their holdings. Their analysis focuses on all the English-language works published in England over 1500-1900, in this collection. They pull three different kinds of information out of these texts:

• Topics discussed: there is a standard set of algorithms in natural language processing which seeks to identify “topics” as sets of words that tend to co-occur in documents. They use these algorithms to construct 60 distinct topics that are discussed in their corpus.

• Progress sentiment: how frequently does the work use the words progress, advance, improvement, rise, stride, amelioration, or betterment (all synonyms for progress that do not have double-meanings and were in use prior to 1643).

• Industrialization focus: how much does the work use words found in the index of Appleby’s Illustrated Handbook of Machinery, vol. 1-5.

As an additional step, they do some complicated work to measure how closely related each of the 60 topics is to one of three major categories: science, religion, and what they call political economy.1

Here’s a chart that is packed with information from this exercise. Let’s walk through what it shows.

First, note that we have four triangles, corresponding to four different 50-year periods: 1700-1750, 1750-1800, 1800-1850, and 1850-1900. Within each triangle, we have a set of circles, each of which corresponds to a text published in that time period. The position of these texts tells us how closely related they are to the three main categories, science, religion, or political economy. We can see that over the time period covered, there was significant growth in the number of texts about science and political economy.

One thing that’s notable is that whereas many texts are about religion and political economy (they appear roughly halfway between the two vertices, on the left edge of a triangle), and many texts are about science and political economy (stretched out along the base of the triangle), we don’t really see any texts that are about both religion and science.

Now let’s turn to the colors in these diagrams. The yellow colors correspond to more progress-oriented language. Overall the figures get much more yellow as time goes on, matching the Burn-Murdoch chart we opened with. But we can also see that the political economy axis, which is associated with words of human institutions (law, govern, trade, etc.) is the center of gravity for progress sentiment. So the growth in progress oriented language was associated with growth in a new kind of literature, which discussed human institutions explicitly.

Finally, the size of the circles is a measure of how focused a text is on industrialization. In general, there does seem to be a correlation between how focused a text is on the topic of industry, and how progress oriented the language is. Since these trends began in the 1700s (other data in their paper shows very little change in progress or industrialization language in the 1600s), before industrialization began in earnest, Almelhem and coauthors take this to be consistent with Mokyr’s theory: in England, there began to be a belief in the possibility of progress associated with texts on industry.

Back to Today

OK, so changes in text patterns in England, circa 1500-1900 may have anticipated changes in the rate of technological progress. That certainly doesn’t mean changes in language always have that effect - reverse causality seems also to be possible, where faster technological progress leads people to write more favorably about progress. But let’s return to the present and more carefully examine the evidence that there is a genuine change in our language about progress.

Burn-Murdoch uses the google ngram dataset to generate his figure of changing word frequency. This dataset is a collection of about 8 million books that have been scanned by google. Using the dataset to track broad cultural changes is controversial, so in this section I’m going to basically kick the tires of the above chart in several different ways. As a spoiler, I will conclude that the claim that our interest in progress has declined can’t be easily dismissed.

The big problem with using Google Ngram data is that its text does not represent a random sample of text, and any time you are working with non-random samples all sorts of biases associated with how a sample was created can creep into your analysis. Younes et al. (2019) provide some guidelines for using google ngrams for studies. First, they advise checking that trends hold across different languages - a trend that is consistent across multiple languages is less likely to be driven entirely by compositional changes, if the corpuses google assembles for different languages have different biases. Younes et al. (2019) also advise looking at whether synonyms display similar trends. We have a bunch of synonyms for progress, but adding them all together might mask variation in the constituent words.

And indeed, we do see some evidence that adding together the frequency of synonyms for progress does obscure some interesting variation. In the figure below, I separately examine trends in English, French, and German.2 For all three of the languages above, the word “future” in English and its equivalents in French and German tends to enjoy a long upward rise, while the other words associated with progress are much more consistently down at the end of the period. In the figure below, I’ve broken out “future” and its French/German equivalents, but left the other progress word synonyms as a bundle to keep the charts tidy. But I did check and “advance”, “rise”, “improvement” and “progress” do display broadly similar trends.

We see a fairly consistent story across these languages. Notably, if we exclude the equivalent words for “future” from our list of progress synonyms, there is a significant decline in progress words across English and French around 1960, and a similar decline in Germany in the 1970s (the blue line). Words for “future”, on the other hand, tend to be up over the whole period in each language.

Compared to “progress”, “rise”, “advance”, and “improvement”, it’s notable that “future” is a more ambiguous marker of sentiment about progress. The word can be used just as easily to convey concern and worry over the future, as it can be used to convey an expectation of improvement.

As another check, we can abandon Google and turn to the same Hathitrust digital library that Almelhem et al. (2023) used in their analysis of words in the run-up to the industrial revolution. The Hathitrust bookworm project makes it possible to study changing word frequencies in the digitized collection of participating academic libraries. To the extent the curation of these library corpuses is distinct from the curation of the google ngram dataset, then if they display similar trends that’s another reason to believe composition effects are not the main driver of changing word frequencies.

Below I look at how frequencies for the words “progress” (left) and “future” (right), and their foreign equivalents, change across the English, French, German, but also Japanese and Italian collections of participating libraries.3

A few broad trends match the google ngram data. Across the five languages, the frequency of the word “progress” and foreign equivalents rises and then falls (left figure), with a particularly marked decline beginning in the 1960s and 1970s. Meanwhile, words for “future” (right figure) generally rise, level off, and then rise again, with French and possibly Italian serving as possible exceptions.

Is the accelerated drop in the frequency of the word “progress” across English, French, German, Italian, and Japanese texts, all in roughly the same decade, evidence of a global vibe shift? It seems at least possible - if I was going to think of an event that might lead to a global reassessment of technological progress, the 1962 Cuban Missile Crisis would probably make the short list.

But more benign explanations are also possible. For example, after WWII, the modern scientific ecosystem was born, and with it a massive rise in the number of new scientific publications. In Almelhem et al. (2023), scientific publications were associated with neutral language, not progress oriented language. Could it be that the decline in the frequency of progress words in text can be attributed to the entry of large-scale, value-neutral scientific publishing into the global text corpus? I investigated this a bit (see the appendix on the website version), and while it’s true that the word “progress” is used less often in scientific text, so is the word “future.” If the rise of academic publishing explains the decline in the use of the word “progress”, it’s surprising that it would not also have a similar effect on usage of the word “future.” Moreover, even looking within scientific texts, use of the word “progress” is falling over the twentieth century.

Still, there could be lots of other compositional things going on. For that reason, it’s useful to restrict our attention to a single type of text, to minimize composition change issues. Ideally, this would be a set of text that we have reason to believe broadly reflects broader cultural attitudes. Fiction seems to be a good candidate for that. Below, I plot the frequency of progress synonyms and the word “future” in Google’s corpus of English fiction. English fiction exhibits significantly stronger and longer-run declines in the frequency of progress synonyms then the overall English corpus. Notably, it also shows declines in mentions of the word “future.”

We can also return once more to the Hathitrust digital library. In the figure below, for example, we have the frequency of the word “progress” in their digital fiction collection, and in their digital non-fiction collections. We can see, in proportional terms, that the decline of “progress” in the English language fiction corpus is larger and longer here than for the non-fiction corpus, and that the decline in frequency of the word “progress” accelerated in English generally around 1960 seems to be mostly a story about non-fiction.

Restricting our attention to fiction removes a lot of concerns that changing word frequencies just reflect the composition of a corpus. However, another issue with tracking cultural change via changes in word frequency is that not all words are created equal. The above approaches treat every text the same, whether the books are highly influential or obscure tracts that are never read. It’s not clear a one-word one-vote system is the best way of tracking cultural change.

Varieties of Influence

As an alternative way of assessing changing attitudes towards progress as reflected in fiction, we can look at some work that tries to analyze more carefully selected sets of texts. For example, there are at least two different ways a novel can be considered influential. First, we could equate influence with overall readership. Second, we could equate it with critical acclaim. A number of papers have assembled datasets based on best-selling novels, or novels short-listed for major novel-of-the-year prizes, as a way to identify and study these different kinds of influence. They don’t focus on attitudes around progress, per se, but they do provide interesting insights into how interested novels are in new technologies, and their interest in the present (or future), relative to the past.

To start, let’s look at two studies that document a rift opening up between contemporary best-sellers and critically acclaimed literature. English (2016) looks at when novels are set, while Manshel (2020) looks at how often novels mention new technologies. Both find that critically acclaimed novels are increasingly disinterested in the present.

The following figures, from English (2016) track when novels are set - in the past, present, or future - among US/UK best-sellers and novels short-listed for major awards. Among bestsellers, the share of books set primarily within twenty years of publication (the higher bars on the left) is up slightly from where it was in the 1960s. Meanwhile, the share of best-selling books set in the more distant past (the white bars) has fallen pretty sharply, from a high of 50% in the late 70s to around 10% in the early 2010s. The share of best-selling books set in the future (in dark gray) remains small overall, but has increased by a lot in proportional terms.

But the story is quite different among critically acclaimed novels. There, the share of critically acclaimed books set in the contemporary moment has steadily dropped, from over 70% in the 1960s to below 50% in the 2000s. Meanwhile, the share of critically acclaimed books set in the past has risen steadily, from 20% to over 50% over the same period. Critically acclaimed books about the future remain rare.

Manshell (2020) looks at the share of novels that mention new technologies, comparing it to data on the diffusion of these technologies across the US economy. For radio, television, computers, and the internet, he finds diffusion of these technologies into novels generally follows diffusion through the economy with a lag. But when he breaks his sample into best-sellers and critically acclaimed fiction, he also finds a comparatively recent divergence in novels that mention new technologies. The figure below, for example, compares the share of novels that mention television or computers to the share of households owning a TV or the total number of computers purchased. Prize winners and bestsellers begin to diverge in the 1990s, with best-sellers continuing to track broader adoption of new technologies, and prizewinners falling behind.

A final paper is Piper and Portelance (2016), which examines variation in the language used in different subsets of books, mostly from the 2000s. In one analysis, they try to identify sets of words that best differentiate books in different categories: what kinds of words are most frequently used in, say, novels shortlisted for major awards but least often used in best-selling novels or science fiction award winners?

One such set of words is the following: afterward, age, born, boy, child, childhood, children, daughter, dream, family, father, girl, life, little, live, memory, mother, old, older, once, recall, remember, school, sometimes, son, winter, young, youth (they actually use the word stems, rather than words, so that boys, families, remembered, etc. would all be counted). Piper and Portelance label this set of words “nostalgic” since they seem to be associated with looking backward. In the figure below, they plot the frequency of these nostalgic words across seven different categories of novel.

Working from left to right, these “nostalgic” words are least common in books in the science-fiction, romance, and mystery genres. Nostalgic words are only slightly more common in best-sellers. But they are increasingly common as we look at nineteenth century realist fiction (C19), books reviewed in the New York Times (NYT), and most common in books short-listed for literary awards.

In another analysis, they manually identify fifty nostalgic passages from each of the above categories (excluding nineteenth century realist fiction). They then use these passages to train an algorithm to classify a 1,000-word passage as nostalgic or not (note, this is a paper from 2016, so they aren’t using contemporary LLMs to do this). They then turn this classifier loose on their text corpus and have it assess what share of 1,000-word passages in a novel are nostalgic. This is a noisier approach, but it also finds nostalgia is much more common among books with critical influence (those reviewed in the New York Times or short-listed for awards) than those with broad readership (genre novels and bestsellers).

Taken together, I think this small literature suggests critically influential fiction is increasingly interested in the past, and as a corollary, probably increasingly less likely to reflect a worldview with progress and optimism about the future of society as important constituents. At the same time, we don’t see much evidence of this kind of shift among more widely read books.

Synthesis

What to make of all this?

I think the academic work discussed here is pretty compelling that critically influential literature is increasingly fixated on the past, at least through roughly 2015. Meanwhile, the word frequency data for fiction seems reasonably clear that words associated with progress and the future have been on the decline for more than a century.

I have more doubts about the broader word-frequency data that documents a sharp decline in progress-related language in the post-war era. The Hathitrust data suggests this recent decline is driven mostly by changes in non-fiction (the google data is a bit more ambiguous). But the non-fiction corpus is a collection of a huge variety of different kinds of texts, so there is a lot of scope for composition changes to drive changes in word frequency. And yet, you see similar trends across many different languages and different collections, and I don’t think it’s easy to explain this away with the rise of academic writing. In the end, I can’t dismiss this line of evidence.

Finally, I have even greater doubts about what this all means for progress. How much do the values and attitudes of critically acclaimed fiction writers matter? How well does word frequency data track broader attitudes towards progress, especially in a world where more of culture is arguably disseminated by other kinds of media? How much do our broad attitudes towards progress matter for overall rates of progress? But I came away from writing this post more concerned than when I started.

Prior to the 2000s, many European countries practiced something called “the professor’s privilege” wherein university professors retained patent rights to inventions they made while employed at the university. This was a “privilege” because the norm is for patent ownership to be assigned to the organization that employs an inventor; professors were an exception to this norm. American universities, in contrast, had long followed a different approach, where patent rights were typically assigned to the university, who managed commercialization efforts. Professors then split the proceeds of commercializing their inventions with the university.

There had long been a sense that commercialization of university research worked better in America, and in the 2000s a number of European countries reformed their laws to move them closer in spirit to the American system. Professors lost their privilege and universities got more into the commercialization game. 

If the goal of this reform was to encourage more professors to invent things that could be commercialized, several papers indicate this policy was a mistake.

Announcements

• 8th Annual Research Policy Online Conference for Early Career Researchers: Sponsored by the Research Policy journal. Apply to present by August 25.

• Good Science Project request for R&D reform policy ideas: If selected, they will help you write a policy brief.

• 7th Annual Research on Innovation, Science and Entrepreneurship (RISE) Workshop: Organized by the Max Planck Institute for Innovation and Competition in Munich. Application deadline is July 26 (today!)

Email me to suggest an announcement for the next newsletter.

Subscribe now

Declining Commercialization

Each of the following papers takes a conceptually straightforward strategy to evaluate the impact of dropping the professor’s privilege: look at how the number of university inventions per researcher changes before and after the policy change, as compared to some other group that was not affected by the policy change.

We’ll start with Hvide and Jones (2018), which looks at the impact of this policy in Norway. Hvide and Jones identify all the Norwegian patents granted to inventors with addresses in Norway - this includes university and non-university inventors. They then match the names of inventors to a list of the names of all Norwegian University researchers. In the figure below, they plot the average number of patents per researcher (red-dashed line, left-hand axis) and the average number of patents per non-university worker (solid blue line, right-hand axis). Following the 2003 cessation of the professor’s privilege, there was a sharp drop in the patenting rate among university researchers, but only a modest decline among non-university workers. All in all, Hvide and Jones estimate stripping university researchers of full patent rights led to a 50% reduction in their patenting rate.

Let’s next press on to Ejermo and Toivanen (2018), who study the same kind of policy in neighboring Finland. Finland adopted a similar policy in 2007, though it paired that policy with a substantial increase in funding to support commercialization efforts. Ejermo and Toivanen have a dataset listing the home address and employment of seemingly everyone in Finland, which lets them be quite confident in how they match patents to university or private sector businesses (or non-university public research institutes), since they can check that the listed address on the patent matches an individual’s listed home address. In the figure below, they compare how the number of patents from firms, universities, and public research institutes (which were not affected by the policy change) changed over 1995-2010. Patenting at the three institutions grew at a similar rate from 1995 to 2004, but then patenting at universities began to diverge to a permanently lower growth rate. Ejermo and Toivanen note the decline began in 2004, rather than 2007 when the policy was enacted, and suspect this is because discussion of the policy change started in the earlier year. Forward-looking university researchers realized that any patent applications they filed might not be granted before the policy became official, when their ownership rights would be given to the university. All in all, Ejermo and Toivanen estimate the policy reduced university researcher patenting by 29-46% relative to trends in the private sector, depending on whether they consider the policy as starting to have effect in 2004 or 2007.

Czarnitzki et al. (2017) examine the impact of a similar policy in 2002, down in Germany. In their case, they focus on comparing university researchers to researchers in non-university public research organizations, such as the Max Planck, Fraunhofer, and Helmholtz institutes. Inventors at these institutions never enjoyed default patent rights to their inventions, but were in other ways similar to university researchers (for example, both relied heavily on state funding). In the figure below, Czarnitzki and coauthors plot the average deviation of an individual researcher’s patents per year, over time. When the line is above zero, this figure is saying that, on average, a researcher tended to have more patents than they usually do in that year. Conversely, when the line is below zero, it means that is a year in which researchers tended to have fewer patents than their long-run average. The chart depicts a steady decline in patenting among university researchers, relative to their peers at non-university public research institutes. The years prior to 1998, before the policy had begun to be discussed, were relatively good ones for patenting among university researchers but not so abnormally good for researchers at public research organizations. The two kinds of institutions converged over 1998-2002, when the policy was debated. And after 2002, when the policy was enacted, university researcher patenting was abnormally low, relative to their long-run average, while patenting at public research organizations was not so unusually low. All in all, Czarnitzki and coauthors estimate the policy reduced university researcher patenting by about 17%.

Next, over to Denmark. Valentin and Lund Jensen (2007) obtains similar results in a study comparing university commercial research in Denmark (which dropped the professor’s privilege) to Sweden (which did not). Their context is a bit more narrow than the preceding papers; specifically, they look at collaborations between industry and university biotech researchers that result in patents. In this industry, it was relatively common for university researchers to partner with industry, with university researchers typically ceding their patent rights to industry partners. This negotiation got more complicated in the year 2000, when Denmark (but not Sweden) ended the professor’s privilege. Now, three parties - the university researcher, the industry partner, but now also the university itself - negotiated over patent rights. Valentin and Lund Jensen find that after Danish professor’s lost full rights to their patents in the year 2000, university-industry biotech collaborations in Demark declined, relative to Sweden: there were fewer patents owned by industry with a university inventor, and the increase in university patents was too small to offset this decline.

Before moving on, let’s circle back to the first paper we discussed, Hvide and Jones (2018). One issue with the results discussed so far is that they rely entirely on patent data.1 What these policy reforms ultimately cared about was not patenting but the commercialization of university research. Perhaps the drop in patenting doesn’t reflect a drop in actual commercialization; maybe all that these papers document is that professors rationally shift their commercialization efforts away from patents (which universities will own) to alternatives. Hvide and Jones deal with this by also looking at what happens to university startups after the policy change. After matching data in a national startup registry to employment data for the Norwegian population they look at the rate at which university researchers launch startups, relative to the general Norwegian population. As displayed in the figure below, this exercise tells us basically the same thing as the patent analysis: university startups, like university patenting, declined sharply after the end of the professor’s privilege, relative to trends outside the university sector.

With the startup data, Hvide and Jones have data on the educational background and sector of employment for most of the Norwegian population, which lets them tighten up their comparison group to PhD holders in the general population. But comparing startup rates between university professors and other PhDs in Norway generally obtains similar results.

Incentives Matter?

OK, so across these studies we find that when you move from a regime where university professor’s retain 100% of the rights of their patented inventions to a regime where the university owns the patents and pays professors a portion of royalties, then university professors begin to patent a lot less. While the magnitudes of the effect are perhaps surprisingly high (17-50%), in other ways this seems like a supremely non-surprising result: when there is less reward for doing something, people do less of that thing. Incentives matter.

But it doesn’t actually seem to be that simple.

First, if this is primarily a story about researchers responding to the size of the rewards to invention, then we should expect to see a positive correlation between the extent of patenting and the size of rewards in other contexts. If patenting goes down because you go from earning 100% of patent licensing fees to 33% when the professor’s privilege is abolished, then it should also be true that universities that gives researchers 25% of the royalties from their patents should see less patenting than those that give researchers 50%, all else equal.

That’s not Ouellette and Tutt (2020) find when they study the relationship between the patent policy of 152 American universities and patenting by faculty. Taking the average royalty share at a university over 1991-1999 and comparing across universities, there is no relationship between this measure and patenting or invention disclosures by faculty. Neither is there a relationship over 2000-2013. Universities may use complicated formulas for splitting patent profits which makes it tricky to estimate a single “royalty share”, but there is no relationship when Ouellette and Tutt try alternatives, or when they restrict their sample to universities that have simple splitting formulas. Perhaps the issue is that the size of royalties is negatively correlated with how much support there is for commercialization, so that the positive effect of a higher royalty share at a given university is perfectly offset by lower support for commercialization there? But Ouellette and Tutt also look to see if changes in a given university’s royalty share are followed by changes in patenting per researcher at that university. Again, no. Finally, Ouellette and Tutt look to see if faculty who patent a lot and change university tend to disproportionately move to universities that offer a higher share of patenting revenues to their faculty: they don’t.

Maybe the issue is just that the royalty share at American universities doesn’t move around very much? That seems unlikely too: by Ouellette and Tutt’s estimates, they span the 20%-80% interval.

What’s going on here? Why did losing 100% of the rights to a patent lead to large drops in inventive activity across universities in Europe, but faculty at American universities seem completely indifferent to their share of patent profits?

A second line of evidence that this isn’t just down to “incentives matter” comes from Czarnitzki et al. (2017), discussed above. As noted previously, they find the overall effect of eliminating the professor’s privilege in Germany was to reduce patenting per university researcher. However, this overall effect conceals some interesting variation. In one analysis, Czarnitzki and coauthors identify university researchers who have been listed as an inventor on a patent owned by a private firm in industry; that’s an indication that the professor had some connections with industry. When they separate researchers with and without industry connections, they find that patenting among researchers with an industry connection fell about 23%. But among those who didn’t have a connection, patenting after the reform increased by 32%. The reason there was a net decline is because among patenting researchers, many more had industry connections than did not, at the time of the reform. Why would a decline in the royalties from patenting lead researchers without prior industry experience to seek more patents?

After thinking about all these different papers, here’s one working hypothesis (but email me if you have a better one!).

Maybe royalty rates don’t actually matter very much, because the returns to commercialization are distributed so unevenly. Most patents are not even worth paying the fees to keep them active for their full 20-year life, but a few are extremely valuable. So whether you earn 20% or 100% of the profits from commercialization, what really matters is if you have a big success or not. If you do, even a 20% share of that is a lot of money and fully justifies the commercialization effort.

In that case, what matters most for the decision to commercialize is whether you think you’ll succeed, not the share of royalties you get conditional on success. And maybe successful commercialization is more about things like “is there a network of entrepreneurial scientists who can advise you?” Or “is there a local VC scene?” Or “Do I know some people in industry who will pay for this research if I give them the patent?” If the environment is supportive of commercialization, you’ll go for it, with any influence of the royalty share too small for statistical tests to identify.

In Europe, during the era of the professor’s privilege, one of the most important questions was likely “how good am I at commercializing research?” The ones who chose to commercialize must have decided they were good enough. The ones who did not commercialize must have decided they lacked the skill (or interest). But following the reforms, commercialization leadership went to the university, and the relevant question became “how good are my university’s commercialization people?” Maybe some of the former commercializers answered this question as “not as good as me, and not good enough for it to be worth the effort.”2 Accordingly, patenting from these folks dropped. And maybe some of the researchers who formerly did not commercialize gave the opposite answer. These folks began to commercialize more, but since there weren’t many of them the net effect of transferring commercialization from the individual to the university was negative. Meanwhile, in the USA, commercialization has always been university-led, and so we don’t see any kind of similar story.

A Metascience Parable

Whatever the real story, I think this episode can also serve as a useful parable about improving the productivity of our research and innovation ecosystem. First off, it demonstrates that policy questions can be quantitatively important: the magnitudes of the effect of this policy change are strikingly large. University commercialization is a small part of overall technological progress, but it’s disproportionately likely to draw on frontier science, and so policies that affect it matter.

Second, it demonstrates that the right policy is not always obvious. In this case, it seems that several European countries decided they wanted better commercialization of university research, and so they looked to the perceived leader in commercialization (the USA) and copied some of its features. This seemingly straight-forward approach to designing policy didn’t work as anticipated - possibly the copied features only work in conjunction with other things the US has going for it; or maybe they don’t matter at all (or are actively bad), and US commercialization outcomes are down to something else entirely. When getting innovation policy right matters and is non-obvious, careful academic research can have a big social impact.

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.

Before we get to this week’s post…

Announcement: There is still time to apply to work with me on Innovation Policy at Open Philanthropy! Details here.

Meta-announcement: if you have an announcement that you would like to disseminate to readers of New Things Under the Sun (17,000+ followers), please email me the details and if appropriate I will include it as an announcement ahead of the next newsletter (free of charge). If I get a lot of announcements, I’ll bundle them into a dedicated announcements post, like the one I put out earlier this month.

Announcements that would be a good fit for this are grant opportunities, conferences, and job opportunities. But I’m open to other ideas too. Just make sure it’s relevant to readers of New Things Under the Sun.

On to the post!

Subscribe now

A classic topic in the study of innovation is the link between physical proximity and the exchange of ideas. I’ve covered this pretty extensively on New Things Under the Sun, and one theme has been that scientists/inventors have various ways of keeping up with relevant new discoveries: they read journals, attend conferences, talk with people in their professional circle, etc. Especially with the advent of digital communication technology, none of that is particularly constrained by geography anymore - you don’t need to be close to where a discovery happens, in order to learn about it. While these methods are pretty good for keeping you up-to-date on things happening in your particular niche, they aren’t great for serendipity. If you don’t know that a field from outside your niche is relevant to your work, how do you know to read the journals, attend the conferences, or talk to people in that field?

One thing that does provide serendipitous encounters with new ideas is physical proximity: being around people who work on lots of different things means you can bump into people and learn what they’re working on, even if you wouldn’t normally expect their work to be relevant - but sometimes it is.1 But I’ve long been interested in a relatively new kind of serendipity engine, which isn’t constrained by physical proximity: twitter (later renamed x, but forever twitter to me). Lots of academics use twitter to talk about new discoveries and research. Today I want to look at whether twitter serves as a novel kind of knowledge diffusion platform.

We’ll mostly look at citations, an imperfect but widely available way of assessing the impact of a scientific paper.2 We could also look at article page views or downloads, but citations arguably provide a stronger signal both that an academic author learned about an idea and thought it was related to their work. Many papers have started by gathering data on how often academic papers get tweeted about and compared that to the number of citations they receive. These are just observational papers, documenting correlations in the wild, but they generally find that papers that get more tweets also get more citations, as indicated in the set of figures from four different papers below.

Let’s quickly walk through this figure. In the upper left, we have a figure from de Winter (2015), which looks at articles published in PLOS ONE in 2012 and 2013. The figure compares the number of tweets (horizontal axis) to various average outcomes. The bottom line in this figure shows citations, which move (slightly) up and to the right. Meanwhile, the upper right figure from Jeong et al. (2019) compares the number of citations for roughly 400 articles published in coloproctology journals. Those that got a tweet tended to have more citations. On the bottom left, we’ve got a figure from Peoples et al. (2016) which looked at tweets and citations in ecology papers. The figure shows how their statistical model predicts citations as a function of tweets (also as a function of time); again, (very) slightly up and to the right. Lastly, in the bottom right, we’ve got a figure from Lamb, Gilbert, and Ford (2018) looking at ecology and conservation papers. The figure shows the predicted number of citations as the number of tweets increases, for a paper that is average along other dimensions. A number of other papers document similar relationships, for example in fields like urology, ornithology, orthopedics, AI research, political science, and communication.

If twitter is acting as a way to spread ideas, we would expect to see this kind of correlation. But other stories would also give us the same correlation. Papers that people think are good are more likely to be tweeted about, but also more likely to be talked about offline, more likely to publish in good journals, and more likely to be cited. The fact that there is a correlation between tweeting and citation doesn’t tell us whether twitter is mediating awareness of the paper or if it’s merely that good papers get cited and also talked about on twitter. To estimate whether twitter is having an impact on knowledge diffusion, we need to look at scenarios where tweeting about a paper is unrelated to the paper’s quality.

Chan et al. (2023) argues they have found one such scenario in economics. There is a popular online publication series, called VoxEU, where economists write accessible short columns, often about their own research. These columns often get tweeted about, which also drive twitter attention towards the underlying academic research. While much of this tweeting is driven by how interesting people find the research, Chan and coauthors also find it’s related to when the VoxEU columns are published. Specifically, columns that are published on weekdays and in the summer are more likely to be tweeted about. They argue the timing of when columns are published is not related to how interesting the underlying research is, and so this provides a natural experiment to assess the impact of tweeting on citations. The basic idea is, if you know the typical relationship between tweets and subsequent citations, and you also know the typical relationship between publication date and number of tweets, then you can back out the impact of a tweet that you received purely due to differences in the timing of publication.undefined That’s arguably the impact we’re interested in: tweets that are unrelated to a paper’s underlying quality.

When they do this exercise on a sample of about 1,000 articles published on Friday through Monday, they find receiving any tweets about your research is associated with about 16% more citations. Note though that the error bars on this are pretty wide - a 95% confidence interval would include “no effect” (though a slightly tighter confidence interval of 90% would not). That’s a theme we’ll be seeing a lot in this literature.

There is also a growing literature that directly attempts to measure the impact of tweeting (unrelated to paper quality) through a different method: twitter experiments.

RCT Evidence (Randomized Control Tweeting)

There have been several studies that randomly select some papers to tweet about, some papers not to tweet about, and then compare the citations received by papers that were tweeted about to those that were not. Since papers get randomly allocated to the tweeting or “no tweet” side of the experiment, any systematic difference in citation rates can be attributed to the impact of tweeting. When we look at these randomized control tweet experiments, rather than observational data, we don’t see much evidence that tweeting affects paper citations.

Tonia et al. (2020) report on an experiment on 130 original articles published in the International Journal of Public Health between December 2012 and December 2014. Half these articles received a brief social media promotion on Twitter, Facebook and a blog post, and the other half did not. They then followed up about 2-4 years later (depending on when the paper was published), but found the promoted papers did not receive a statistically significant number of additional citations.

One shortcoming of Tonia et al. (2020) is that the social media intervention came from the journal’s twitter account, which wasn’t highly followed (403 followers at the time the study started). Maybe that’s just not enough people to test the knowledge diffusion hypothesis.

But another study, Branch et al. (2024), doesn’t have this issue and gets similar results. Branch et al. (2024) is a collaboration between 11 active researchers who are also influential twitter users, most with more than 5,000 followers - for comparison, this would put them in the top 13% by follower count, in a sample of academic economists on twitter. Each of the authors picked a journal in their field (for example, Journal of Animal Ecology, Polar Biology, Ecological Entomology, etc.) and then each month five research articles were randomly selected. From this set of five, one was randomly selected to be tweeted about by the participating twitter account, and the other four were not tweeted about. Branch and coauthors tweeted brief summaries of the paper in their usual style of tweeting about papers. Three years after each tweet, the authors collected data on the papers that had been tweeted and the other four from the same journal that had been randomly selected not to receive a tweet. Across the 110 papers that were tweeted about and the 440 that were not, the ones tweeted about did not receive a statistically significant number of additional citations.

So that’s two studies that find tweeting does not influence citations over a multi-year period. On the other hand, there are two more studies that do. But both of these studies have come under scrutiny since publication. Luc et al. (2020) looks at an experiment to tweet or not tweet 112 representative articles published in Thoracic and Cardiovascular surgery journals in 2017-2018. A year later, they found tweeted articles got 3.1 citations, compared to 0.7 among the non-tweeted group, which was statistically significant. However, a later attempt to reconstruct this dataset and replicate this result, by Phil Davis over at the Scholarly Kitchen website, was not able to find a difference between the two groups. In another case, Ladeiras-Lopes et al. (2022) report on an experiment to randomly promote 347 articles published between 2018 and 2019 in European Society of Cardiology journals, and randomly not promote another 347. Roughly 2.7 years later, they found the tweeted articles had received a statistically significant 12% more citations than the non-tweeted articles (though Ladeiras-Lopes and coauthors point out that tweeted articles also benefited from 24-hour open access). A reanalysis of their data, again by Phil Davis, successfully replicated this result, but found the statistical significance of the result was sensitive to choices in the method of statistical analysis. Arguably preferable methods found the difference between the two was not statistically significant.

Interlude

So it seems that twitter isn’t the vehicle for serendipitously discovering unexpected papers, that I personally hoped it would be.

Well, not so fast. There are two wrinkles that warrant further study I think.

First, it is notable that every study listed here actually does find tweeted papers receive more citations than ones that are not tweeted. It’s just that the differences are not statistically significant, or not replicable, or not robust. For example, Branch et al. (2024) find that papers they tweeted about are 7-12% more highly cited than papers they did not tweet about. Ladeirras-Lopes et al. (2022) and Tonia et al. (2020) also find similar effect sizes. These levels seem plausible to me - as a point of comparison, Azoulay, Wahlen, and Zuckerman Sivan (2019) find that elite life scientists who unexpectedly pass away and are subsequently memorialized in journal articles see a similarly sized bump to citations to their work.

The trouble is citations are so noisy that given the sample sizes in these twitter studies there is a reasonable chance the tweeted papers would have gotten this many additional citations simply by chance, and so we cannot be confident the increase in citations (in the neighborhood of +10%) is statistically significant. For example, Branch et al. (2024) estimate they would need a study that is 3-7 times larger than theirs to reliably detect effect sizes of this magnitude. That suggests to me that either a larger study, or perhaps a meta-analysis pooling results across all these studies might find that, indeed, tweeting can increase citations on the order of 10%.

A second wrinkle: I don’t think we should assume the impacts of tweeting will be uniform across articles. Recall, researchers already have a lot of ways to learn about relevant work - they read the journals, attend the conferences, talk to their network, etc. I would anticipate that the impact of twitter would be much attenuated for papers that are likely to diffuse through these traditional channels; whether anyone tweets about them or not, you’ll learn about these papers. But for papers that might not normally diffuse via these channels, the impact of twitter might be a lot stronger, because if you see a tweet you learn about it, and if not, you don’t. I described some similar dynamics in the post Steering Science with Prizes. In that post, we found some evidence that scientific prizes have a bigger impact on citations to work that is less well known at the time of the prize.

Finch, O’Hanlon, and Dudley (2017) provides some observational evidence consistent with these effects for twitter too. Their study looks at more than 2,500 articles published in 10 ornithology journals between 2012 and 2016 and look at correlations between social media attention and subsequent citations. Unlike any other study (as far as I can identify), they look at this effect separately for journals at different impact tiers. They find the correlation between social media attention (mostly twitter) and citations is substantially stronger for articles published in journals with low impact factors, compared to high ones. In other words, for articles in journals that don’t normally attract a lot of citations, getting mentioned on twitter a lot is associated with more citations to a much stronger degree than is the case for articles published in highly cited journals.

Related, if twitter is most useful for serendipity, it might be that it has the largest impact on citations from outside the field. These citations might be more valuable than average (since connecting ideas from across disciplines is associated with high impact), but they’re probably not common and you would need a large study to detect them.

What’s a Citation Really?

OK, so while we have probably ruled out that twitter has a very large impact on citations, I’m not sure we’ve ruled out the possibility that it has a small average effect, and possibly a larger effect for articles that would otherwise be less well known.

But a critique of this entire exercise is that it’s focused on citations. I’m ultimately interested in whether twitter is a platform for disseminating important knowledge for innovators. While I think citations are a useful measure of knowledge flows at sufficiently large scales, it’s also true that the majority of citations made don’t seem to be particularly important.3 I think it’s plausible that citations formed via twitter might be disproportionately of the unimportant type. Maybe the really important stuff you learn through the usual channels, and twitter just helps you round out your citation list before submission to a journal (and not even very much).

For that reason, Qiu et al. (2024) is quite helpful. Their study is also a randomized control trial, where some papers get tweeted about more, and others do not. Unlike the other papers discussed so far, Qiu et al. (2024) is not focused on citations (though they do say they plan to look at that in the future), but instead on hiring decisions. They set up their experiment around the 2023 economics job market for new PhDs. In economics, it is the norm for students to prepare a “job market paper” which showcases their research to prospective employers. In this experiment, tweets about their job market papers were submitted by 519 new PhDs. All of these tweets were tweeted by a specially created twitter account designed to promote job market candidates. Qiu and coauthors then arranged for 81 academic economists with more than 4,000 twitter followers to tweet about a random subset of these submitted tweets.

They then look at how the number of job interviews, the number of invitations to fly out and present their research to the potential employer, and the number of job offers received by students who receive a tweet about their work from a large twitter account and those that don’t. Echoing the findings of our earlier papers on citation, they find that applicants who (randomly) get a tweet about their work get more interviews, more flyouts, and more job offers, but also that these results are often not statistically significant. They can never reject zero impact of tweeting on tenure-track jobs outcomes, though depending on what factors they adjust for they can sometimes find the effects of tweeting is statistically distinguishable from zero for all job outcomes of PhD economists.

What’s useful about this study is that, unlike the decision to cite, hiring decisions have real stakes. If twitter actually does disseminate valuable knowledge that would not be circulated via conventional channels, one place we could see that is in hiring. It’s quite common for hiring committees to receive hundreds of applications, and so they don’t have the time to read every job market paper they receive. So having heard about the research previously - via conferences, peers networks, or twitter - could be very valuable in allowing a paper to get past the first cut and examined more closely. If people subsequently go on to hire those people, that suggests twitter is playing an important role in disseminating information people find valuable, that they wouldn’t otherwise see. The paper provides some weak evidence this does happen, though not especially for tenure track jobs.4

So in the end, I think this experiment is broadly consistent with the other findings from randomized controlled trials on the effects of twitter. Does twitter matter? Maybe!5

As a final point, the fact that randomizing papers to be tweeted about doesn’t yield unambiguously large effects on academic citations is one more (small) reason to think academic citations contain some useful information; they’re not whipped around wildly based on something as seemingly inconsequential as a tweet.

Thanks for reading! As always, if you want to chat about this post or innovation in generally, let’s grab a virtual coffee. Send me an email at matt@newthingsunderthesun.com and we’ll put something in the calendar.