I.  Introduction

Basic science is a critical input to innovation, but it may be underprovided in competitive markets because discoveries are not directly marketable and property rights are difficult to enforce. Unlike applied research, basic (or “pure”) scientific research advances our fundamental understanding of the world, but typically does not yield immediate opportunities for commercialization (Nelson 1959; Arrow 1962). As a result, credit for ideas, rather than direct profits, is an important potential motivator of innovative activity (Dasgupta and David 1994). Within academia, there is a widespread notion that the first person to publish a new discovery receives the bulk of the credit. Scientists therefore compete fiercely for priority (Merton 1957). Famous examples of priority disputes include Isaac Newton versus Gottfried Leibniz over the invention of calculus, Charles Darwin versus Alfred Wallace over the discovery of natural selection and evolution, and, more recently, Grigori Perelman versus Shing-Tung Yau, Xi-Peng Zhu, and Haui-Dong Cao over the proof of the Poincaré conjecture. This competition for recognition shapes the culture and professional structure of many disciplines, and scientists regularly worry about their work being “scooped” or preempted by a competitor (Hagstrom 1974). However, there is little empirical evidence documenting how credit is allocated in science or how rewards are shared between the “winners” and “losers” of these races. The additional credit given to the winner—what we call the priority premium—is an important parameter because it dictates the intensity of the competition to publish first. A relatively even credit split could lead to less competition than a winner-take-all scenario, which could meaningfully affect the pace, direction, and quality of research.

Therefore, the contribution of this paper is to empirically measure this priority premium. We analyze the impact of getting scooped on the losing project (in terms of probability of publication, journal placement, and citations), as well as on a scooped scientist’s subsequent career. We also investigate whether competition for academic attention contributes to inequality within scientific disciplines.

Conceptually, our goal is to measure the cost of getting scooped by constructing comparisons in which multiple teams of scientists are working independently and concurrently on identical or very similar projects, which we call races. In practice, these races are challenging to identify for three reasons. First, many academic fields use a variety of methods and seek to answer fairly open-ended questions, and so finding near-identical projects is difficult. Second, even if the questions are well-defined, it is difficult—especially without expertise in a given scientific field—to quantify the intellectual distance between two papers in topic space. Third, scooped projects are often abandoned, making them impossible to track in publication data. We tackle these challenges by analyzing project-level data from the field of structural biology. Specifically, we examine projects in the Protein Data Bank (PDB), a repository for structural coordinates of biological macromolecules. The PDB is a centralized, curated, and searchable database of biological details contributed by the worldwide research community and contains over 150,000 macromolecule structures.¹ Several features of the PDB allow us to make headway on the key empirical challenges described above. First, structural biology papers have a well-defined objective, which is to describe the 3-dimensional shape of a known protein. Once the first paper about a protein structure is published, any follow-up publications serve mostly to confirm the result of the first. Second, projects are grouped by the PDB according to molecular similarity, which allows us to identify papers written by separate teams that solve identical or very similar molecular structures. Lastly, the PDB uniquely allows us to observe projects that are scooped shortly after completion but before publication. Scientists are required by journals to upload structures to the PDB prior to publication, so we can see projects that were completed but never appeared in print. Moreover, the rich metadata in the PDB allows us to reconstruct the timelines of projects, and find instances where teams were—unbeknownst to each other—working on the same molecule at the same time. Structural biology is a secretive field: in our data, most teams that lose priority races are scooped unexpectedly near the end of their projects.²

We construct races using two key dates that are recorded for all PDB projects. First, the deposit date marks when the scientist first uploaded her findings to the PDB. Scientists typically deposit their findings shortly after a manuscript has been submitted for publication. The second is the release date, which closely corresponds to the date of publication and is usually 2 to 6 months after deposit. Critically for our design, the data is hidden from the public (and from competing scientists) between deposit and release. To construct races, we find instances where two or more teams have independently deposited a structure discovery for identical macromolecules prior to their competitor’s release date. The order of release then defines the outcome of the race. The first team to release is the winner, and the second team is scooped. We identify 1,611 races in our data. These races consist of 3,279 separate projects out of 67,297 total projects in our sample period from 1999 to 2017, suggesting that 5 percent of all structural biology projects are involved in a late-stage race to publication. These races are composed of a diverse set of scientific teams from different countries, institutional prestige, and experience. In the main analysis of this paper, our definition of scooped projects focuses only on late-stage races where both teams are on the cusp of publication. Focusing primarily on these late-stage scoops is advantageous for the economic interpretation of our results. Since both projects had been completed independently prior to publication, we can infer that the second-place team would have published the priority paper in the counterfactual where they had not been scooped. The estimated difference in observed outcomes therefore isolates the premium for novelty awarded by editors and readers. The downside of focusing on these narrow postdeposit scoops is that the scientists are passive at this point. The research has been largely completed and the timing of release is in many ways out of their hands, so these races offer little insight into the strategic interactions between racing teams, a central topic in the economics of R&D racing. Therefore, as an extension in section V, we study a sample of teams that were scooped after they had begun their experiments but before they had deposited their final project, in order to learn more about these strategic interactions.

While getting scooped is not randomly assigned, we use multiple methods to assess the validity of the causal identification assumptions. We estimate the effect of winning a race using the naturally occurring variation in the priority ordering of races. Therefore, omitted-variables bias is a threat to the causal interpretation of the estimates. If the winners are positively selected on experience, research ability, or university prestige, then our estimates of the scoop penalty will be biased upwards (in terms of magnitudes). However, we find that the outcome of races—even if not perfectly random—is highly unpredictable. We observe cases of high-ranked teams scooping low-ranked teams and low-ranked teams scooping high-ranked teams. Throughout the analysis, we carefully document potential sources of bias and assess treatment balance using the observable team and author characteristics. To further mitigate concerns of omitted-variables bias, we use the post–double-selection Lasso method for control-variable selection (PDS Lasso; Belloni, Chernozhukov, and Hansen 2014; for Lasso, an acronym for “least absolute shrinkage and selection operator,” see Tibshirani 1996).

We find that getting scooped has a moderate impact on the success of the scooped project. Scooped projects are 2.6 percent less likely to be published. Scooped papers appear in journals with impact factors that are lower by 0.19 standard deviations; they are nearly 20 percent less likely to appear in a top-10 journal. Scooped papers receive 21 percent fewer citations, and are 24 percent less likely to be a hit paper, defined as reaching the top 10 percent in citations for their publication year. While these effect sizes are meaningful, they are far from a winner-take-all division of credit. Focusing on citations as an outcome, our estimates imply that the losing paper receives 44 percent of the total citations accrued by both papers, a much higher share than the 0 percent assumed by a winner-take-all model. Much of the citation effect is driven by journal placement, with only a 4 percent difference in citations once we control for journal fixed effects. We provide suggestive evidence that editors and reviewers have a strong taste for novelty. Papers that are scooped prior to submission to a top journal are rarely, if ever, accepted for publication. Some scooped papers do appear in top journals, but only if they were far along in the review process on the date they are scooped.

Does getting scooped have a detrimental impact on the careers of individual authors? We compare the future publications, citations, and academic longevity of scientists on the winning and losing teams. We find that scientists who are scooped are about 6 percent less likely to be actively depositing in the PDB 5 years after this setback, and 2 percent less likely to be publishing in life and medical sciences as a whole. We do not find statistically significant effects on intensive margin publication rates. However, scooped scientists receive 20 percent fewer citations to their future work, an effect that is stronger for novice scientists (34 percent) than their veteran counterparts (16 percent).

The main analysis focuses only on scoops where the losing team had already deposited and was therefore limited in its opportunity to change its research. When considering cases of predeposit scoops (i.e., scoops that occur before the losing team has deposited their work), we find that scientists are able to strategically respond to being scooped by adjusting the scope and direction of their project and also by integrating insights from the winning publication. We identify this subsample of races using the “collection date” feature of the PDB, which allows us to find teams that had done their initial experiments but had not yet deposited their findings in the PDB. Teams scooped in this intermediate stage take 1.4 years longer from collection to deposit than teams that are scooped after depositing. In that time, they tend to include additional structure deposits in their paper and shift the focus of their writing away from narrowly describing the structure itself by incorporating more analysis of protein function. They are also more likely than our main sample of scoops to build on the priority findings using a technology called molecular replacement. Although some of these strategic responses to getting scooped slow the scientists down, they also help to offset the growing scoop penalty.

We analyze and discuss how the priority-reward system relates to inequality in science. Our sample of races provides unique insight into how reputation affects academic attention, because we see teams of varying reputation and affiliation competing to publish the same discovery first. We find that when a high-reputation lab scoops a relatively unknown lab, they receive 65 percent of the total citations, but when a low-reputation lab scoops a high-reputation lab, they only receive 46 percent of the total citations. We rationalize this asymmetry in priority rewards with a model of academic attention based on the statistical discrimination literature (Phelps 1972; Aigner and Cain 1977). This relationship between priority credit and reputation suggests that compensation in science is not formulaic but may be influenced by the attention constraints and biases of editors and readers.

Finally, we benchmark the size of the scoop penalty by comparing it to the perceptions of active structural biologists. We survey 822 corresponding authors of papers linked to the PDB and pose a hypothetical scenario about getting scooped. The respondents estimate a 27 percent probability of getting scooped between submission and publication, much larger than the 3 percent chance we document in the PDB data. We then ask them to predict the probability of publication and expected citations if they are scooped by a competitor’s paper. They predict that they only have a 67 percent chance of publishing the paper—again, much lower than the 85 percent of scooped projects that we observe being published in the PDB data. Finally, they estimate a 59 percent penalty in citations compared to the hypothetical winner, much higher than the 21 percent penalty we estimate in the PDB data.³ These comparisons suggest that scientists may be overly concerned about the probability and cost of getting scooped and that perhaps better information about the true outcome of races might alleviate concerns about risk and competition in academia.

We choose to focus on structural biology because the unique features of the PDB allow us to estimate an internally valid priority effect in a way that—to the best of our knowledge—would not be possible in other fields of science. However, a narrow focus on a single field naturally raises questions of external validity. Varying norms, institutions, and technology across different academic fields might lead to different distributions of priority and mechanisms for assigning credit. The scoop penalty may be higher in structural biology than in, for example, economics, because structure discoveries are “one right answer” solutions and therefore similar papers are potentially more substitutable. On the other hand, because structural biology is an experimental field, there could be inherent value in replication, which might increase the attention granted to scooped papers as compared to more theoretical fields like pure mathematics. We argue that structural biology is an important area of research per se and is therefore worthy of our attention. However, the research questions and methods structural biologists use are similar to other important fields in the basic life sciences, and so we suspect that our qualitative conclusions may apply to these fields as well.

The size of the priority premium directly relates to the level of competition in science. In a scenario where priority rewards are evenly split between the first- and second-place teams, there is no reason to compete to publish first. At the opposite extreme, if priority rewards are winner-take-all, the competition will be intense. This competition, in turn, has important implications for how science functions. On one hand, sharp priority rewards can encourage intense effort on solving frontier problems. A priority system also has the public benefit of encouraging disclosure, which is critical for fostering follow-on innovation (Williams 2013). On the other hand, some have theorized that R&D racing might induce overinvestment and duplication of effort on particular projects (Loury, 1979; Hopenhayn and Squintani 2021). In a companion paper (Hill and Stein 2024a), we study how high levels of competition generated by unequal priority rewards also impact the quality of scientific work. Our results suggest that the competition to publish first induces scientists to rush and ultimately results in lower-quality research. Some journals—seemingly in response to these rushing concerns—have begun to explicitly offer a grace period in which they will consider scooped papers for publication (Marder 2017; PLoS Biology Staff Editors 2018). This appears to be an effort to directly reduce the priority premium by ensuring more credit for the second-place team. Moreover, competition may affect science along other dimensions. For example, high levels of competition may reduce collaboration and the free sharing of information, ultimately slowing scientific progress. Therefore, measuring the priority premium—which maps directly to the intensity of scientific competition—is a critical first step in this agenda.

The remainder of the paper proceeds as follows. The following paragraphs offer a brief literature review. Section II provides some scientific background and a description of our data. Section III describes the empirical design and identification. Section IV presents results for the short-run impact of scoops on publication, journal placement, and citations, as well as the long-run career results. We also discuss the role of editors and the timing of races for the distribution of priority rewards. Section V studies the strategic response to being scooped in races where the scooped team had not yet completed the project. Section VI describes a model of academic attention and reports results for heterogeneity of the scoop penalty by preexisting reputation. Section VII benchmarks the size of our estimates against the beliefs of surveyed structural biologists about the probability and cost of getting scooped. Section VIII concludes.

Related literature.—This paper contributes to several distinct but connected literatures, both in economics and disciplines interested in the “science of science.” First, and most broadly, it contributes to our understanding of how incentives for basic research are structured. Second, it adds to a more narrow empirical literature about the causes and consequences of innovation races. Finally, it contributes to a literature about career dynamics in scientific labor markets and the role of academic reputation.

Priority races in science are often compared to patent races in industry. However, incentives for basic scientific advances are in many ways distinct from patents. Inventors in a patent race are competing for profits, while researchers in a priority race are competing for journal placement, citations, and recognition from their peers. However, both systems compensate researchers for the production of public goods, incentivize timely disclosure of knowledge, and hasten the pace of discovery. Both systems are usually conceptualized as tournaments for a discrete innovation reward or prize, with the first innovator getting the outsized share of rewards.

Theoretical models of patent races have considered how racing affects the amount of R&D investment (Loury 1979; Lee and Wilde 1980) as well as the pace of research and the amount of risk-taking induced by the structure of races (Dasgupta and Stiglitz 1980). Many of these models presuppose a winner-take-all reward that has implications for the outcome of innovation tournaments and the strategic behavior of the participants. The conventional wisdom in the sciences—and the assumption underlying much of the theoretical economics work on the topic—is that the process of scientific discovery is also a winner-take-all tournament, even if the prize is priority recognition rather than a patent (Merton 1957; Stephan 1996). Dasgupta and David (1994) explain that a discontinuous priority reward might arise in science because of a fundamental verification problem. Because of the public goods nature of new knowledge, a team that tries to publish the second paper cannot credibly prove to the community that they would have successfully completed the project absent the help of the priority paper. Even if it would be socially optimal to share more credit with teams who were working in parallel, these information frictions might make credit-sharing difficult. The discontinuous priority-reward structure has implications for the pace of research and the strategic interaction of teams (Bobtcheff, Bolte, and Mariotti 2017). The literature on innovation systems has yielded influential models but as yet very little empirical evidence about the actual distribution of rewards in these races. Therefore we believe our estimates provide important context for theoretical and policy discussions about the incentives for scientific innovation.

This paper joins a small literature that aims to study innovation races empirically (Lerner 1997). Most related to our work, Thompson and Kuhn (2020) document that winners of patent races do more innovation in the future, and that this innovation is more likely to be related to the original patent. The authors identify patent races by looking for patents that were rejected for lack of novelty. Bikard (2020) studies the phenomenon of simultaneous discovery in science, and documents many cases of papers that are similar in content, are published around the same time, and are frequently cited together. However, our method of using biological details to link competing papers allows us to find simultaneous discoveries where one paper goes unpublished or is cited infrequently in the future.

Our heterogeneity by reputation estimates contribute to work in sociology and economics about path-dependent advantage in academic prestige, commonly called the Matthew effect (Merton 1968). Our results build on recent empirical work that has documented evidence of the Matthew effect in life sciences (Azoulay, Stuart, and Wang 2013), astronomy (Hill 2019), and grant funding (Jacob and Lefgren 2011; Bol, de Vaan, and van de Rijt 2018; Wang, Jones, and Wang 2019).

II.  Background and Data Construction

D.  Defining Priority Races

Broadly speaking, we define a priority race as an instance where two or more teams are working on the same protein independently and concurrently and are likely uncertain about the identity or progress of their competitors. Following Brown and Ramaswamy (2007), we define the same protein as meaning two proteins within the same 50 percent or higher sequence-similarity group (called a cluster in the PDB). This is a conservative cutoff, as 30 percent has been suggested as sufficient similarity for building homology models (Moult 2005; Dessailly et al. 2009). In other words, the first publication within these 50 percent similarity clusters is often highly cited because it provides a novel structure model that other crystallographers can build on to solve very similar proteins.⁷ The PDB assigns identification numbers to clusters of similar proteins, and we say that the first structure released in that cluster is the priority structure deposit. There are often many subsequent deposits that report similar structure coordinates as the priority deposit, only some of which we define as being scooped. These follow-on deposits appear for a variety of reasons. Some are concurrent projects by authors that were racing to be first but were scooped or are replication projects of the same protein by future teams, while others are new projects that solve the structure for closely related proteins that either derive from different organisms or else are bonded with different macromolecules in a novel way.⁸

We use project timelines reported in the PDB to determine whether a follow-on deposit qualifies as scooped by the priority deposit. The PDB provides two key dates at the structure level that help us determine whether two teams were working concurrently: the deposit date and release date.⁹ The deposit date corresponds to the date that the scientist uploaded her solved structure to the PDB. Importantly, the structure is not yet visible to the public. Nearly all scientific journals require that authors upload their structures to the PDB prior to publication, so deposit typically occurs slightly before or after the date that the scientist first submits her paper. The release date is the date that the PDB deposit is made public. This typically corresponds to the publication date. In cases where the structure is never published, the PDB releases the deposit by default 1 year after the deposit date. Figure 1 provides a visual timeline of these dates, as well as some summary statistics. Throughout this analysis we will always use the release date as the relevant marker of priority. An alternative approach would be to use paper publication dates to determine priority ordering. But these dates are often unavailable, especially for older publications, or are ambiguous in recent data because online publication may come before print edition publication. Further, we treat publication as an outcome variable; we would risk potential bias if we conditioned on publication as a requirement for treatment assignment. Lastly, PDB releases are publicly salient events that the community pays attention to, so the release dates are therefore good markers of priority order. Appendix A4 discusses implications and presents evidence about the concordance between release dates and publication dates in greater detail.

Figure 2 illustrates how we define a scoop event. Consider two projects, A and B, authored by two distinct teams working on the same protein. Suppose project A is a priority project in one of the similarity clusters. We say that project A scoops project B if (i) A is released before B is released but (ii) after B has deposited to the PDB. Condition i guarantees that A finishes first, while condition ii guarantees that B did not know about A until after the structure was deposited in the PDB. Since B had already deposited a completed structure, they likely would have been the priority deposit had they not been scooped by A. Requiring that B has deposited before A is released ensures that we observe abandoned projects, since all deposited structures appear in our data even if they are scooped and fail to publish. We allow the priority project to scoop more than one team, and 5.6 percent of the races we identify have three or more competitors. Appendix B provides a more detailed description of the data work necessary to construct these races in practice. In our main analysis, we exclusively focus on these clean, but narrowly defined scoops that occur after B has already deposited. However, in section V we expand our analysis to include earlier-stage scoops, that occur before B deposits.

1.  An Example

To help understand our procedure, consider the example outlined in table 1. The table shows two structures: 4JWS and 3W9C. Both are structures of the Cytochrome P450cam protein complexed with its redox partner, putidaredoxin (Pdx-P450cam complex). This enzyme is involved in metabolism and clearing toxins, such as in the human liver. Figure 3 shows the nearly identical biological assembly models that each team deposited independently and confidentially to the PDB. The scientists at Leiden University (3W9C) collected their data a few months before the scientists at the University of California, Irvine (4JWS) (February 3, 2012 versus September 14, 2012). However, by the time of deposit, the UC Irvine team had pulled ahead, depositing one week before the Leiden team (March 27, 2013 versus April 3, 2013). Ultimately, UC Irvine won the priority race, with their structure being released two months before Leiden (June 19, 2013 versus August 21, 2013). Importantly, when Leiden deposited their structure on April 3, 2013, UC Irvine had not yet released their structure. This means that Leiden was likely unaware of their competitor’s progress or results when they were preparing their publication and depositing the structure. Comparing the outcomes of the winner (4JWS) and the loser (3W9C), we observe that the winning paper was more successful. It was published in a better journal (Science, with an impact factor of 31.5, vs. Journal of Molecular Biology, with an impact factor of 4.0) and received about 30 percent more citations over the next 5 years (Hiruma et al. 2013; Tripathi, Li, and Poulos 2013). In this case, the Leiden authors became aware that they were scooped during the manuscript review. In the conclusion of their paper, they write, “While this manuscript was under review, Tripathi et al. published the crystal structure of the Pdx–P450cam complex that was obtained via cross-linking of the two proteins. It is interesting to compare our complex with those reported in that study. Tripathi et al. found a position and orientation of Pdx relative to P450cam that is essentially identical with ours” (Hiruma et al. 2013).¹⁰

	Winning Project	Scooped Project
PDB structure ID	4JWS	3W9C
Protein name	Pdx-P450cam complex	Pdx-P450cam complex
Paper title	“Structural Basis for Effector Control and Redox Partner Recognition in Cytochrome P450”	“The Structure of the Cytochrome P450cam-Putidaredoxin Complex Determined by Paramagnetic NMR Spectroscopy and Crystallography”
Key dates:
Collection	September 14, 2012	February 3, 2012
Deposit	March 27, 2013	April 3, 2013
Release	June 19, 2013	August 21, 2013
First author affiliation	University of California, Irvine	Leiden University
Journal	Science	Journal of Molecular Biology
JIF	31.5	4
5-year citations	52	39

Note.  This table presents an example of a racing pair identified in the PDB using the scoop rules outlined in section II.D See fig. 3 for the image of the structure models deposited by each team.

View Table Image

2.  Additional Sample Restrictions

We make three further restrictions to minimize cases of ambiguity in the race-construction procedure. First, we drop some proteins that are exceedingly complex. Some very large proteins are composed of many entities that are sometimes solved piece by piece over many years instead of all at once. This introduces the possibility that a scientist could be scooped on only a fraction of their project.¹¹ Second, we drop projects that are published in a paper that is linked to 15 or more other structures. Among the set of papers included in our final analysis sample, 46 percent are linked to more than one structure, and the average number of structures per paper is 1.9. Multistructure papers are at risk of being scooped on a fraction of the full project. This restriction allows for some fractional scoops to enter our data, but ignores papers where each protein becomes a very small fraction of the full contribution of the paper. Finally, we drop races that end in a near or exact tie. Occasionally, two racing papers will be submitted to the same journal and the editor will publish them as companion pieces in the same issue, and we drop these cases. We also drop races where the two papers were released closer than 2 weeks apart from each other. We make this restriction to help ensure that the first project has a clear claim of priority and that the order of release is more likely to correspond to the order of publication.¹²

III.  Empirical Design

The analysis is designed to identify the causal effect of getting scooped on the short-term success of the project (publication, journal placement, and citations), as well as on subsequent academic success of the scooped authors. We estimate the difference in outcomes between the winners and losers of the priority races in the PDB. In an ideal setting for causal inference, the winners and losers would be randomly assigned. In reality, the outcome of these late-stage races is not exactly random but is highly unpredictable. We present evidence that although some characteristics of the teams are correlated with winning a race, these observables can only explain very small differences in outcomes. In this section, we present the main estimating equations of our analysis, describe and test for potential sources of bias, and explain the control selection strategy we use to deal with potential selection bias.

B.  Identification and Balance

Comparing outcomes of winners and losers of the PDB races identifies the causal effect of getting scooped if the race ordering is as good as randomly assigned. There are many reasons a team might win or lose a priority race, and it is plausible that the order of completion is somewhat idiosyncratic. The randomness of the scientific process, day-to-day operation of scientific labs, and the vagaries of the journal review process leave ample opportunity for random chance to dictate the timing of these races. Anecdotal accounts of ill-timed personnel issues, lab accidents, or unlucky experiment failures suggest that the timing of project completion is oftentimes out of the hands of even the most diligent and skilled scientist (Ramakrishnan 2018; Yong 2018). Furthermore, after the deposit date and submission of a manuscript, the team has very little discretion over the timing of the review process, which may be delayed by editor preference, reviewer inattention, or publisher congestion. Moreover, scientists typically have little information about the identities or progress of their competitors.

On the other hand, skill, experience, or resources could provide an advantage to certain teams that would allow them to systematically start earlier or work faster and therefore win priority races. This is a threat to identification because these characteristics may simultaneously increase the probability of winning and improve project outcomes. For example, suppose a technological breakthrough marks the starting point of a race that many diverse teams enter. If one team from Harvard has exceptional resources to adopt the technology and complete the project first, we will observe them win the race and receive many citations. But since Harvard is a high-reputation university and has a track record of success, they would likely have received many citations even in the counterfactual where their competitor won the race. Therefore, we rely on the assumption that well-resourced or otherwise high-reputation teams are not able to systematically win priority races, and we test this using observable characteristics of each team.

If winning a priority race is random, then winning and losing teams should look balanced based on observables. We assess this observed balance between winners and losers in table 3. Using the information disclosed by the teams in the PDB, we inspect a variety of observable characteristics that might reasonably be correlated with the probability of treatment or with outcomes. These include the number of authors, the location of the lab, the rank of the university affiliation, and the experience in years of the first and last authors. We also calculate measures of the authors’ productivity in PDB-related publications in the 5 years prior to the racing deposits. These include the number of PDB deposits, publications, and publications in top-ranked journals.¹⁶

	Nonracing	Losing Projects (LP)	Winning Projects (WP)	Difference (LP − WP)	Difference (SE)
Variable	(1)	(2)	(3)	(4)	(5)
	A. Team Characteristics
Number of authors	7.454	7.183	7.056	.127	(.205)
Affiliation:
North America	.351	.262	.320	−.057	(.022)***
Europe	.158	.134	.170	−.036	(.018)**
Asia	.134	.224	.156	.067	(.018)***
University (rank 1–50)	.240	.223	.278	−.055	(.021)***
University (rank 51–200)	.261	.248	.228	.020	(.020)
Other	.499	.529	.494	.035	(.023)
Industry or nonprofit	.171	.153	.152	.001	(.018)
First author experience (years)	5.983	5.744	5.134	.611	(.279)**
Last author experience (years)	7.806	7.521	7.311	.210	(.313)
	B. First Author Productivity (prior 5 Years)
Deposits	12.361	3.791	5.504	−1.714	(.687)**
Publications:
Total	2.893	2.591	3.139	−.548	(.464)
In top-10 journals	.656	.709	.670	.038	(.065)
In top-5 journals	.222	.262	.239	.023	(.032)
	C. Last Author Productivity (prior 5 Years)
Deposits	44.269	30.937	28.991	1.946	(4.327)
Publications:
Total	9.905	12.513	13.398	−.884	(2.241)
In top-10 journals	4.027	4.669	4.622	.047	(.511)
In top-5 journals	1.421	1.653	1.799	−.146	(.190)
	D. Project Quality Metrics (Lower Is Better)
Resolution (Å)	2.244	2.328	2.315	.013	(.062)
R-free goodness-of-fit	.236	.245	.243	.002	(.002)
Observations	64,018	1,668	1,611	F-statistic: 4.019***

Note.  This table compares characteristics of winning and losing projects in order to check for treatment balance. Observations are at the structure level. Column 1 shows the means of the nonracing sample, col. 2 shows the means of the losing projects in the racing sample, and col. 3 shows the means of the winning projects in the racing sample. Column 4 shows the difference between the losing and winning projects, and col. 5 shows the heteroskedasticity-robust standard error of the difference. The F-statistic and associated p-value is calculated in a regression in which all of the variable values are stacked into a single left-hand side outcome variable and the treatment indicator is interacted with variable fixed effects on the right-hand side.

** p < 0.05.

*** p < 0.01.

View Table Image

Table 3 shows the mean values of each covariate for the winning and losing teams, as well as for the teams in the non-racing sample, for reference. We report test statistics for the difference in means between the winning and losing teams, as well as an F-statistic for a test of joint significance of all covariates. We find that many of the covariates are balanced between the winning and losing teams. But winning and losing teams are statistically different in a few notable dimensions. North American and European teams are more likely to win than lose, while Asian teams are more likely to lose than win. Scientists from top-50 ranked universities are more likely to win, as well as first and last authors with slightly less experience. The prior productivity of these labs is more balanced, with most measures of productivity being statistically insignificant for both first and last authors (though winning first authors appear to have deposited more). We also test whether the scientific results that are being deposited by both teams are similar. Refinement resolution and R-free are two variables reported by the PDB that describe the objective quality of the experimental data and model in each deposit. Resolution describes the degree of precision in the diffraction data produced during crystallography experiments, and R-free measures the goodness-of-fit between the experimental data and the proposed structure model. For both of these measures, smaller values imply better quality. These two measures are very close to balanced between winners and losers, suggesting that the quality of the science or the skill of the scientists is likely not driving our results. Taking the table as a whole, we reject the null hypothesis of balance on the full battery of covariates based on an F-statistic of 4.02.

Unbalanced covariates lead to biased estimates only if they are systematically correlated with the outcome variable. Therefore, to further assess potential selection bias, we visually inspect the difference in expected citations between winners and losers. We estimate a paper-level model of citations using a Lasso regression of 3-year citation counts on the battery of team covariates. This model is estimated only in the sample of nonracing deposits. We then take the selected variables and estimated coefficients to predict citations in the racing sample in a post-Lasso ordinary least squares (OLS) procedure. The covariates we include are counts of publications, citations, and journal placements in the 5 years prior to the deposit for the first and last author, as well as the squares of these variables. We also use the career age of the first and last authors, the rank of the first author’s institution in 10-school bins, and the country and university of the first author. The Lasso model selects many of the variables one would expect to be important, including dummies for being in the US and dummies for university rank. The full Lasso results are reported in table A1.

Figure 4 plots a histogram of the difference in predicted citations between each pair of winning and losing teams (races with three or more teams are omitted here). A perfectly balanced sample would be centered around zero and symmetric. If winners were systematically better resourced, had a better reputation, or had more experienced, then the histogram would be skewed to the right. As a benchmark for perfect balance, we compare this distribution to a simulated distribution where we randomly assign one of the paired teams as the winner. We simulate this coin flip 100 times per pair. The true distribution is shifted slightly to the right of the randomly simulated distribution, suggesting that winners are slightly more likely to be high-reputation than would be predicted by chance. But the differences in the distribution are small. The difference in means between the two distributions is 0.68 predicted citations with a p-value of 0.065 (for reference, the sample average is about 12 citations, so this represents a 6 percent difference). This slight lack of balance motivates our control strategy discussed in the next section.

IV.  Results

C.  Mechanisms: Role of Scoop Timing in the Publication Process

Scooped projects receive about 21 percent fewer citations than their winning counterparts, suggesting that academic researchers pay less attention to the projects that are scooped. In this section, we investigate how the editorial process affects the scoop penalty, and we argue that journal placement is a primary driver of the citation penalty. Further, the size of the penalty is highly correlated with the timing of races. Teams that are scooped early (very shortly after they deposit their findings) receive a much larger penalty than teams that are scooped late (shortly before publication). We provide evidence that top journal editors are unlikely to accept scooped papers; therefore, scooped papers consistently fall to lower-ranked journals, excepting those already deep into the review process at the time they were scooped. These results suggest that editors and reviewers are key policymakers in determining the distribution of academic credit for novel research.

1.  Decomposing the Citation Effect by Journal

First we show that the citation penalty is largely driven by journal placement. We decompose the citation effect into an editor/reviewer effect and a reader effect by controlling for journal placement. Column 1 of table 6 replicates the citation penalty effect from column 4 of table 4, but uses a subsample of races in which both papers were published in ranked journals. When both papers are published, the citation penalty is 16 percent for scooped papers. In columns 2 and 3, we add controls for JIF, first as a linear term and then as a cubic polynomial. The citation effect falls to 10 percent, but remains statistically significant. Finally, in column 4 we include journal fixed effects to control completely for any direct effect of the publication outlet on citations. The effect falls to 4 percent. These results suggest that nearly three-fourths of the citation penalty comes through the channel of the publishing journal. Any remaining effect on citation attention comes through readers differentially citing winning and losing papers in similar journals.

	5-Year Citations
Dependent Variable	(1)	(2)	(3)	(4)
Scooped	−.155***	−.111***	−.102***	−.044*
	(.032)	(.029)	(.028)	(.027)
Journal controls	None	Linear JIF	Cubic JIF	Journal FE
Winner Y mean	34.7	34.7	34.7	34.7
Observations	1,891	1,891	1,891	1,891

Note.  This table reports the scooped coefficients in regressions with 5-year citations as the outcome where we control for JIF. The citation counts are transformed with the inverse hyperbolic sine function in the regression, but the winner Y mean is reported in levels for ease of interpretation. The regression sample is restricted to races where both papers were published in a ranked publication. Column 1 reestimates the table 4, col. 4 regression in this subsample. Columns 2 and 3 add linear and then cubic controls for JIF. Column 4 includes fixed effects for journal. All regressions also include PDS-Lasso–selected controls and protein (i.e., race) fixed effects.

* p < 0.1.

*** p < 0.01.

View Table Image

2.  Editors’ Role in Priority Credit

We further explore the role of editors in adjudicating priority credit by focusing on the submission, review, and publication timelines of scooped projects submitted to leading science journals. Academic journals compete fiercely to publish the highest quality and most novel scientific articles. Many of these journals have explicit policies for accepting only highly original and novel research. For example, Science provides the following guidelines to peer reviewers: “Recommend in your review whether the paper should be published in Science and provide a more detailed critique based on the following…. Novelty: Indicate in your review if the conclusions are novel or are too similar to work already published” (AAAS 2019). Editors and reviewers therefore likely drive much of the scoop penalty if they choose to reject scooped papers when they come across their desk. In this section we look at how the scoop penalty is affected by the timing of journal submissions. Many of the papers in our sample had already been submitted to a journal when they were scooped, and a few papers had already been accepted. Even if an editor would prefer to reject a scooped paper, they may be unable to do so if the paper had already been accepted or was far along in the review process. We use the supplementary data collected from journal websites to examine how the scoop penalty is affected by the timing of the review process. Ideally, we would compare the scoop date to rejection dates at leading journals, but data on rejected papers is not publicly available. Therefore, we use instead the timing of submission and acceptance to present suggestive evidence that editors at top journals are reticent to publish scooped papers.

In our data, scooped papers occasionally appear in top journals like Science, Nature, and Cell, but 90 percent of those papers were already under review on the date that they were scooped. Furthermore, about 60 percent of those papers were scooped after they had already been accepted. Figure 5 further shows that this pattern varies greatly by the impact factor of the journal that eventually publishes the scooped paper. For lower-ranked journals, such as PLoS One, only 60 percent of scooped papers had been received by the journal on the date they were scooped, and just over 20 percent had been accepted. Among the 11 large journals for which we have information about received and accepted dates, there is a positive and statistically significant relationship between the share accepted before the scoop date and the impact factor: a scooped paper published in a journal whose rank is 1 standard deviation higher is 8 percentage points more likely to have already been accepted on the scoop date. Although we cannot directly observe scooped papers being rejected from these journals, we can infer from this pattern that top journals are less willing to accept papers that were scooped before submission or early in the review process. Many of these scooped papers fall to lower-ranked general interest journals or highly specialized structural biology journals.²² Some of these lower-ranked journals, such as PLoS Biology, have explicit policies of accepting scooped papers. PLoS Biology editors write, “Just as summiting Everest second is still an incredible achievement, so too, we believe, is the scientific research resulting from a group who have (perhaps inadvertently) replicated the important findings of another group. To recognize this, we are formalizing a policy whereby manuscripts that confirm or extend a recently published study (‘scooped’ manuscripts, also referred to as complementary) are eligible for consideration at PLoS Biology” (PLoS Biology Staff Editors 2018). But even some lower-ranked journals are concerned about the fierce competition for novel research. When we approached one publisher about sharing their data on received and accepted dates, they only offered to provide the data anonymously, stating their concern about presenting public evidence that they publish scooped papers.

3.  Time Lag and the Scoop Penalty

The severity of the scoop penalty is correlated with the time lag between the release of the winning and losing projects. In figure 6, we plot the difference in outcomes separately for 3 terciles of races divided by the time between the release dates of the winning and losing projects. The points are placed on the x-axis at the average delay time within the subset of races. The first panel shows the JIF penalty and the second panel shows the citation penalty. Both plots have a strong decreasing trend in the penalty—in other words, the longer the lag between the priority paper and the scooped paper, the less credit the scooped paper receives. The JIF penalty is 0.1 standard deviations in the first 3–4 months, then drops to 0.3 standard deviations by 8 months. Similarly, projects released within 1 month of each other have no difference in citations. The scoop penalty grows to 50 percent for scooped projects with an 8 month delay. In fact, much of the negative effect that we present in table 4 is driven by the tercile of races with the longest delays. An important caveat to these results is that the delay to release after being scooped is potentially endogenous. While much of release lag may be due to idiosyncrasies of the publication process that are out of the researchers’ hands, teams may also make strategic decisions about whether to rush to publish, revise and delay, or give up publication altogether, so the delay times should be viewed as potentially selected on team or project characteristics. We explore some of these forces in more detail in the next section. These results suggest that the delay time between projects is relevant for editors and readers, perhaps because the community can more clearly attribute priority credit with more time separating similar projects.

V.  Strategic Responses to Getting Scooped before Project Completion

Thus far, we have focused exclusively on races where two teams had completed the project before the knowledge of the scoop is revealed. We chose this restriction because it minimizes the scope for researchers to endogenously respond to the scoop event. In cases where scientists are scooped after depositing, they are usually preparing a manuscript or have submitted to a journal already. The PDB also mandates that the project is released to the public 1 year after deposition at the latest, and this forced disclosure likely puts pressure on the team to publish quickly if they have already deposited. Therefore, they have less flexibility to respond to the scoop event by repositioning their research, changing direction, using insights from the winning paper, or abandoning the project altogether. This allows us to estimate the impact of being scooped, all else equal. However, the endogenous response itself is interesting. How do scientists use the knowledge that they have been scooped to reoptimize? In this section, we compare projects that were scooped before and after deposit to show how scientists respond when they learn that they have been scooped before completing the project.

The classic patent-race literature has focused on the strategic decisions of a follower in a race for a discontinuous reward, typically the profits from a patent (Loury 1979; Dasgupta and Stiglitz 1980; Lee and Wilde 1980; Gilbert and Newbery 1982; Reinganum 1983). Depending on the modeling assumptions, these models predict a range of outcomes: for example, the follower will persist at a steady R&D pace, the follower will increase effort in an attempt to leapfrog the leader, or the follower will choose to drop out of the race altogether. The optimal strategy is dependent on the R&D technology, the information structure of the game, and other features such as whether the race has a single or multiple stages (Fudenberg et al. 1983). Our setting differs from those models for important reasons, but insights from this literature are relevant for interpreting scientist behavior in our setting, especially for those scooped before they had deposited.

Like these classic innovation-race models, researchers in our setting can choose to accelerate a research project or abandon it altogether. However, there are other important choice margins in our setting. First, unlike the models described above, the game does not automatically end when the first team releases their structure. Instead, the second-place team still has an opportunity to adjust the pace, direction, and scope of their project. This is more akin to recent patent-race models where races are multistaged or endless (Judd 1985; Aoki 1991; Doraszelski 2003; Horner 2004). Second, early models rarely grappled with the public goods nature of innovation, where a loser can benefit from the winner’s discovery through imitation or improvement of the winner’s disclosed discovery (Arrow 1962; Dasgupta and David 1994).

In the remainder of this section, we study the strategic decisions of a scientist who is scooped early enough in the project’s life that she still has an opportunity to reoptimize the path of the project. The key idea is that once a scientist learns that she has been scooped, she faces a tradeoff. She knows that on one hand, she will get more credit if she publishes quickly because the scoop penalty grows with time (as shown in fig. 6). On the other hand, she can expand the project in new directions (e.g., by adding additional structures or experiments). This will take time—leading to a larger penalty—but will also make the project more valuable overall. Moreover, because she can now take advantage of informational spillovers from the first paper, it might be easier to expand the project than before that paper was released. We formalize this trade-off in appendix C. Broadly speaking, there are three possible cases. In one case, the scientist speeds up when she learns that she has been scooped to minimize the penalty. In a second case, she slows down and improves or broadens her project to maximize its value. Finally, it is possible that the cost of completing a project is no longer offset by the reward, leading to a third case where she abandons the project upon learning it has been scooped.

In our data, we can observe the behavior of some scientists that were scooped before they had a chance to complete their projects, and thus have the flexibility to reoptimize. Although not required by the PDB, many deposits (81 percent) report a collection date, which is the date that the scientist collected the x-ray diffraction data at a synchrotron. Using these dates, we can identify races where scientists had successfully crystallized their protein and collected diffraction data but then learned they were scooped by another team prior to depositing their completed structure model (see fig. 2).

Overall, the empirical evidence is consistent with the second case: researchers spend longer to expand the scope of their projects when they know they have been scooped. Figure 7 compares the timeframe of projects between postdeposit scoops (our original sample) and predeposit scoops. On the left, we show the number of years that pass between the original collection of the data and the time of being scooped. Not surprisingly, predeposit scoops tend to be slightly earlier in the life of the project (mean of 1.6 years for predeposit scoops and 1.7 for postdeposit scoops). There are very few projects that have a short (less than 4-month) lag between collection and scoop in the postdeposit sample of races because the scientists would not have had time to analyze the experimental data and deposit their structure. However, there is considerable overlap of the distributions, suggesting that these two types of scoops occur in similar timeframes on average.

However, the right panel shows the number of years between the scoop and final release of the scooped paper. This release gap is much longer on average for predeposit scoops (mean of 0.36 years for postdeposit scoops, 2.13 for predeposit scoops), suggesting that, for scientists who know they have been scooped but decide to continue, the preferred strategy is to invest more time into the project rather than abandon it. One important point of context is that postdeposit scooped projects are mandated to release the findings after 1 year, so even if postdeposit scooped teams wanted to change their research, add experiments, or rewrite their paper, they have much less flexibility after they have already deposited.

This delay in release appears to be consistent with scientists electing to add additional experiments and differentiate their project from the race winner. Table 7 presents regression results using the full sample of races associated with 1,778 predeposit scoops and 979 postdeposit scoops combined for which we have available data.²³ We regress a series of project characteristics that relate to the margins of adjustment discussed above on a scooped indicator and an interaction between a predeposit indicator and the scooped indicator.²⁴ In column 1, we can see that there is a very large increase in maturation time (time between collection and release) for the predeposit scooped teams relative to the postdeposit scooped teams, with the predeposit teams spending 1.4 more years on average. Next, we consider how trailing scientists may adjust the scale or scope of their research to offset the scoop penalty. We find in columns 2 and 3 that predeposit scooped teams are much more likely to include multiple protein structures in their paper relative to the postdeposit scooped teams, suggesting that predeposit scooped teams expand the scope of their papers. Next, we look at how scooped teams may have adjusted the content of their paper by analyzing keywords from the paper titles. Column 4 suggests that predeposit scooped teams are much less likely to use the word “structure” or “structural” than postdeposit scooped teams. However, as seen in column 5, they are more likely to use words like “function,” “mechanism,” or “analysis” in the title. It appears that if teams have the flexibility to adjust the direction of their research after being scooped, they choose to shift the focus away from the structure determination itself and toward describing the function or biological mechanisms that the structure implies.

Finally, we use another unique feature of the data to test whether trailing teams benefited from seeing the priority deposit if they were scooped before completing their own work. The PDB contains a flag for a technology called molecular replacement, which is a crystallography technique that improves model prediction. Importantly, it relies on using another similar structure model as a pattern to refine the new model from diffraction data (see Kim 2023 for a detailed explanation of the technology). In other words, trailing teams can use molecular replacement—which makes completing their projects easier—if they can observe the winning structure before they finish their own structure. Column 6 suggests that scooped teams are more likely to use this technology, but only if they were scooped before they deposited. If the winning structure is released after the trailing team deposits, as is the case in our postdeposit sample, the trailing team is unable to take advantage of insights from the winning structure model in their own process. However, if the winning structure is released before the trailing team deposits, as is the case in our predeposit sample, they are able to benefit from this information. This suggests there are meaningful knowledge spillovers that benefit the losing team. In addition to being consistent with our model, we think this provides strong empirical evidence that the release of the project represents a meaningful information shock. Overall, it appears that scientists who are scooped before they have a chance to deposit their findings are more likely to delay the release of their structure, increase the scope or change the direction of their research, and integrate the knowledge from the first discovery into their project.

Finally, we compare the cost of being scooped before and after deposit. We interpret the results of this exercise cautiously because of the endogenous selection into the predeposit sample and the additional flexibility that predeposit scooped teams have to strategically respond to the scoop. Table A10 reproduces table 4 in the predeposit sample. We find that the difference in most outcomes between the winners and losers is about 15 to 40 percent larger in predeposit scoops compared to our primary postdeposit sample. The citation gap is 28 percent in the predeposit scoops compared to 21 percent in the postdeposit (main sample) scoops. The relative reduction in the probability of publication is comparable between the two groups. Scientists who persist despite being (knowingly) scooped are likely a selected set who are determined to publish.

VI.  Reputation and the Scoop Penalty

Scientific races provide a unique setting to study how academic recognition is affected not only by priority, but also by the preexisting reputations of winners and losers. We find that when a high-reputation team scoops a low-reputation team, they receive 65 percent of the total citations, but when a low-reputation team scoops a high-reputation team in a comparable race, they only receive 46 percent of the total citations. This asymmetry in attention suggests that the distribution of priority rewards is not formulaic and may be affected by the institutions, norms, or biases of the academic community. In appendix C, we present a model of academic attention based on a standard statistical discrimination model (Aigner and Cain 1977). Here we present empirical results that support the predictions of the model.

Priority rewards are allocated by a decentralized set of actors, including journal editors and readers, in a market for academic attention. Because scientists have limited time for reading and reviewing new papers, it may be difficult to determine the quality of new research. Therefore, editors and readers may rely on signals of ability based on the reputation of the researchers or their institution to supplement their judgement of a paper’s quality. The model considers cases where two types of teams, high- and low-reputation, publish identical papers. Readers decide who to cite based on priority and reputation. In cases where teams are of the same type, the priority effect is isolated, and the first team to publish receives more than 50 percent of the total citations. However, in cases where teams are of different types, the priority and reputation effects will either work in the same or opposite direction, depending on which team finishes first. If the high-reputation team wins the race, the two effects reinforce each other, meaning the high-ranked team will have an equal or greater share of citations compared to the low-ranked team than they would competing against another high-reputation team. If the low-reputation team scoops the high-reputation team, the net effect is ambiguous. If the reputation effect is stronger than the priority effect, the low-reputation team may receive less than 50 percent of the total citations, despite publishing first.

To test our model, we measure the share of total citations received by winning and losing labs, and compare these shares in races where the reputation varies between the two racing teams. More specifically, if lab A and lab B race to write a paper about the same protein, we compute CitationShareA=CitationsA/(CitationsA+CitationsB). This citation share maps to the probability of citation outlined in the model above.²⁵

We proxy for the preexisting reputation of each lab using the Lasso-estimated predicted citations from the nonracing data sample as described in section III.B. Labs with above-median predicted citations are categorized as high-reputation labs, while teams below median are called low-reputation labs. In figure 8 we plot the predicted citations of the losers on the x-axis and the predicted citations of the corresponding winners on the y-axis. Each point on this scatterplot represents the observed match between two racing labs. If all labs were equally matched in preexisting reputation, all points would lie on the dashed 45-degree line. Of course, labs are rarely perfectly matched in the data, providing variation in the difference of reputation between the winners and losers.

The median lines in figure 8 conveniently partition the sample into four subsamples that line up with the four types of “matchups” we discuss in our model. The top right and bottom left corners represent subsamples of closely matched races where both labs were either high-reputation or both low-reputation. The top-left and bottom-right subsamples represent mismatched races where an above-median team scooped a below-median team and vice versa.

In mismatched races, we interpret the difference between citations as being caused by an additive effect of priority and reputation. One potential confounder in that interpretation is that high- and low-reputation teams might produce different quality of scientific outputs for the same structure discovery. If high-reputation teams produce higher-quality or more convincing results, then the additional citations they receive may not be caused by their high-profile reputation alone. Although it is difficult to quantify all aspects of paper quality, we examine two important measures of quality reported by the PDB: resolution and R-free (goodness-of-fit), described in more detail in section III.C. Table A11 compares the average resolution and R-free of the winning and losing structures in each of the four subsets of races. We find very little evidence of statistical difference in quality metrics between high- and low-reputation teams engaged in a race. This suggests that any difference in citations is not driven by the quality of science that each team is producing.

Figure 9 shows the average citation counts by matchup type, as well as the citation shares. Panel A shows the evenly matched races, which isolates the priority effect. As predicted by the model, the winning labs receive more citations. Moreover, if we look at the share received by the winning team, we see that it is identical in the high-versus-high–reputation matchups and the low-versus-low–reputation matchups (in each case, the winning team receives about 55 percent of the total citations), which is consistent with the model.²⁶

Panel B shows the unevenly matched races. When a high-reputation lab scoops a low-reputation lab, the priority effect and the reputation effect work in the same direction. Here we see that, consistent with proposition 2, the winning team receives an even larger share of the total citations (65 percent). Conversely, when a low-reputation lab scoops a high-reputation lab, the priority effect and the reputation effect move in opposing directions. In this case, it appears that the reputation effect is the stronger of the two, with the winning team receiving less than half (46 percent) of the total citations. Again, this matches the prediction outlined by proposition 2 of the model.

Collectively, we interpret this as evidence that statistical discrimination based on prior lab reputation can rationalize our heterogeneity results. The lack of symmetry exhibited in panel B suggests that being first is not the sole determinant of credit in science. In science, there is no central arbiter that gives legally binding credit or property rights to the first-place team. Here the teams vie for attention, and although the low-reputation teams may benefit by winning a race, there appears to be built-in inequality in attention that prevents them from capturing as much of the credit as their high-reputation competitors.

VII.  Benchmarking Magnitudes: Survey Results

We estimate that getting scooped causes a decrease in the probability of publication, leads to publication in lower-impact journals, and reduces citations. However, priority races are not winner-take-all. Our citation estimate suggests that winners get 56 percent of the total citations, a far cry from the 100 percent often assumed in the theoretical literature. But how does this estimated share of credit compare to scientists’ beliefs? In an email survey of structural biologists, we pose a hypothetical situation about a late-stage race to publication. The full text of the questions can be found in appendix D. First we ask, “Suppose you have just completed a very promising research project … what do you think is the probability that your project will be scooped between now and when it is published?” We next state that their hypothetical project has indeed been scooped by a paper in the journal Science. In this scenario, we ask them the following questions: “Would you choose to abandon your manuscript? Assuming you submit, what is the probability the article will eventually be published? What is the best journal that would accept your paper? If your competitor receives 100 citations, how many citations do you expect your publication to receive?”

Table 8 reports the average responses of the biologists in columns 3–6 and compares them to the magnitudes estimated in the PDB data in columns 1 and 2. The hypothetical scenario in the survey was designed to match the instances of racing that we have in our data. However, because we tried to pose the survey questions as concretely as possible for clarity, the racing situation does not exactly match the average situation in the PDB. In particular, in the survey the losing team is scooped early in the submission process, and the project is very high quality, with an expected journal placement in Science. Therefore we report estimates in column 2 from a subset of the PDB data where (1) the losing team is scooped soon after they deposit their data,²⁷ and (2) one of the teams published in one of the three highest-impact journals (Science, Nature, or Cell). These restrictions make some of the PDB estimates smaller or larger, but we still consistently find evidence of pessimism among respondents. Surveyed scientists report a 27 percent chance of being scooped between submission and publication, more than three times the 8 percent scoop probability in the comparable PDB sample.²⁸ Six percent of respondents report that they would abandon the project, but only 68 percent think they would succeed at publishing conditional on submitting, implying a 67 percent unconditional probability of publishing as shown in column 3. This is much lower than the 85 percent of scooped papers that are actually published in the PDB data, and the 98 percent that are published in the comparable subsample. Scientists are very pessimistic about the potential journal placement of scooped papers, expecting that the best journal they could publish in would be almost 3 standard deviations below Science, which has a standardized impact factor of about three in most years. Finally, we ask about expected citation effects. When asked to guess the number of citations they would receive compared to the hypothetical winner’s 100 citations, the average guess was only 41 citations, which translates to a 59 percent penalty or a share of 29 percent of the total citations. The corresponding estimate in the PDB is no more than a 21 percent penalty or a 44 percent share. Ultimately, PDB scientists expect much worse consequences from being scooped than can be found in the data.

Table 8 also reports survey responses separately for high- and low-reputation scientists. We split the survey sample using the same Lasso-predicted citation measures used in section VI. Column 4 reports the average responses for below-median–reputation scientists, column 5 reports the average responses for above-median–reputation scientists, and the difference with standard errors is reported in column 6. High- and low-reputation respondents predict equal probabilities of being scooped. Low-reputation respondents are more pessimistic however about the probability of publishing conditional on being scooped, with 7 percentage points lower probability that they will be able to publish their scooped paper. Perhaps surprisingly, both types of respondents had similar expectations for the types of journals that they would publish in, all expecting that the scooped papers would fall to field journals or middling general interest journals with average impact factor. But they again depart on their expected citations, with high-reputation scientists expecting a citation penalty that is about 4 percentage points smaller than low-reputation scientists (57.5 percent penalty vs. 61.4 percent penalty). This difference in expectations is consistent with our results about the role of reputation in determining priority rewards. Since both types of authors suggest they would submit to similar journals, it may be that the difference in citations is driven by statistical discrimination of editors, reviewers, and readers as explained in the model in section VI. It appears that although all scientists are pessimistic about the cost of getting scooped, less-prominent authors are particularly concerned. Our estimates of significant inequality in citation patterns suggest that these beliefs may be justified.

VIII.  Conclusion

Priority races are a common feature of academic science, and credit for priority is considered an important motivator for the generation of new knowledge. Yet, we have little empirical evidence on how these priority rewards are structured. Racing is hard to analyze empirically because proximate research projects are difficult to link in data and many scooped projects are abandoned before entering the scientific record. This paper makes progress on these empirical challenges by focusing on project-level data in a setting that captures the near universe of completed projects in structural biology. By taking advantage of the unique data collected by the PDB, we are able to construct credible estimates of the priority premium in the field of structural biology. We find that rewards are far from winner-take-all; rather, our preferred estimates suggest a 56-44 split in citations between the winning and losing paper.

This paper contributes to our understanding of the role of priority and the structure of incentives in basic research. Academic science is an atypical marketplace of productive activity. New ideas are valuable for the world but are not immediately marketable, and are therefore unlikely to be produced by private firms or individuals seeking profits. A patent system is therefore a less-effective instrument for encouraging investment, risk-taking, effort, or disclosure of scientific studies. Instead, a system of priority rewards has developed to encourage research investment, which is reinforced through norms in the scientific community. Individuals who produce new knowledge are given credit by the community that can accumulate into a reputation that likely has both intrinsic and monetary value to the scientist. Although R&D races have been posed as winner-take-all tournaments in past literature, we find that priority rewards are not winner-take-all, but are potentially still an important motivator of both effort and novelty in science. Even if the result of one race has a small impact on careers, the accumulation of credit may still be important.

In this paper, we establish that priority is a relevant incentive in science, but we do not analyze the overall welfare implications of the priority system and size of the priority premium, nor do we consider alternative systems or policies. How would a larger or smaller priority premium affect the efficiency of science? There are many margins to consider, including how changes would affect effort, project selection, collaboration, and even participation in science. A particularly interesting concern raised in popular and academic writing is that priority may be pursued at the expense of quality. Racing to complete projects may stimulate effort and hasten the pace of discovery, but it may lead scientists to cut corners on the quality of the results that they disclose. If the incentives for replication are low and the costs of replication are high, science as a whole may suffer as quick and sloppy research becomes the norm. In Hill and Stein (2024a), we analyze objective measures of the quality of crystal diffraction data and corresponding structure models to study how racing in science affects quality outcomes. We find that proteins with high ex ante potential have more competitors racing to complete the structure, are deposited faster, and are completed with lower quality. This evidence suggests that racing in science does indeed hasten disclosure but has negative effects on quality. Concerns about the cutthroat nature of racing have led to suggestions of policies that might dampen the strong incentives for novelty. These include allowing a grace period for journal acceptance a few months after being scooped, providing opportunities to establish priority for early-stage work through preprints, or directly incentivizing replication efforts through directed grant funding.

Finally, the results of our survey suggest that scientists are very pessimistic about the cost and probability of being scooped. If the perceived threat of being scooped has a negative influence on the pace, direction, quality, and openness of science, we believe that this paper should help assuage concerns about competition for priority and foster a more productive research environment.

Data Availability

Data and code for replicating the tables and figures in this article can be found in Hill and Stein (2024b) in the Harvard Dataverse, https://doi.org/10.7910/DVN/TJ5VCW.