Teams, credit, and networks

Five thousand authors

In May 2015 the journal Physical Review Letters published a paper that reported the most precise measurement yet of the mass of the Higgs boson, the particle whose discovery at the Large Hadron Collider was announced in 2012. The paper was 33 pages long. The first nine pages described the measurement and listed the references. The remaining 24 listed the authors and their institutions: 5,154 of them, from two giant experiments called ATLAS and CMS. Guinness World Records recorded it that month as the research paper with the most contributors.

Which of the 5,154 deserves the credit? There is no way to answer, and the Nobel Prize has a rule that makes the question sharper. Its statutes say that a prize may be shared by no more than three people. When the 2013 Prize in Physics was announced for the theory of the particle, it went to two theorists, François Englert and Peter Higgs, who had predicted it in 1964, and not to any of the thousands of experimenters. Science today is done by teams, sometimes very large ones, and prizes, jobs, and reputations are given to individuals. That mismatch is the subject of this lesson, and it takes us through the last four of the ten rules: how teams succeed, how credit is divided among their members, and how the networks around us decide who becomes known.

First, how big is the shift? A 2007 study by Stefan Wuchty, Benjamin Jones, and Brian Uzzi examined 19.9 million research articles in the Web of Science and 2.1 million US patents. In science and engineering, the average number of authors on a paper rose from 1.9 to 3.5 over 45 years, and the work of teams got 1.7 times as many citations as the work of single authors in 1955 and 2.1 times as many by 2000. In the latest period, a team-written paper was 6.3 times as likely as a solo paper to collect 1,000 or more citations. Teams, by this measure, had taken over. In 2019 Lingfei Wu, Dashun Wang, and James Evans added a nuance, from more than 65 million papers, patents, and software projects: large teams tended to develop existing ideas, while small teams tended to disrupt them with new ones. Later critics have found that the size of this effect depends on how disruption is measured, and in one re-analysis it shrank to almost nothing and changed sign for teams of eight or more. Teams matter, but “bigger is better” is too simple.

A smart team

What makes a team good? One famous answer comes from a 2010 study by Anita Woolley and colleagues. They gave 699 people, in 192 randomly assembled groups of two to five, a battery of tasks: puzzles, brainstorming, a moral judgment, negotiating over limited resources, typing, and planning. Groups that did well on one task tended to do well on the others, and a single number summarizing this, which they called the collective intelligence factor, c, accounted for roughly a third to two-fifths of the differences in performance among groups (31 to 44 percent, depending on the method) and predicted how groups did on a later, harder task. It was only weakly connected to the intelligence of the individual members. What went with a high c was the average of the members’ social sensitivity (their score on a test in which you judge emotions from pictures of eyes), more even turn-taking in conversation, and a larger share of women in the group, an effect that the authors said was largely explained by women’s higher scores on the social-sensitivity test.

The lectures turn this into a rule of thumb: a high-performing team needs balance and diversity. It is a good rule when you read it narrowly (groups in which everyone gets a say and reads each other well did better on these tasks), and it needs care when you read it broadly. The correlations are modest (about 0.26 for social sensitivity, 0.23 for the share of women). A 2014 study from the same group found them in online groups too. But outside researchers have challenged the factor itself: a re-analysis of six data sets (2017) concluded that the evidence for a general group factor was insufficient, and a separate study of 312 people (also 2017) found that individual intelligence accounted for around 80 percent of the differences in group intelligence, and that the share of women and the turn-taking did not predict it. A 2021 pooling of 22 studies and 1,356 groups, by the same research group, found the factor again, with 44 percent of the variance. And the wider literature on diversity is mixed: a meta-analysis of team studies found that variety in functional background was slightly positive for team performance, while variety in race and sex had small negative averages and age variety had none. The effects are small and depend on the task and on how the team is run. A popular mathematical argument that “diversity trumps ability” has been criticized as a restatement of its own assumptions. Balance of voices helps; the claim that more women or more diversity causes better teams is not established.

A team and its leader

The other half of the rule says that a team that is successful, which means popular and widely noticed, needs a leader. The evidence is from open-source software. In 2016 Max Klug and James Bagrow studied about 150,000 teams on the code-sharing website GitHub between 2013 and 2014. A team was the set of people who could change a project, and success was the number of “stars” a project had been given by other users, a kind of popularity vote. The work was concentrated in every team: the top contributor often did two or three times as much as the second. And the 10 percent of teams with the most stars were more likely than average to be “dominated” by one person, meaning that the top contributor did more than all the others put together. The strongest single predictor of a team’s success, though, was not a lone leader but the number of its members who were the leader of some other project. The authors’ own summary: a successful team is “a diverse support system for a specialist core.” The diversity here is not about gender or background: it is whether the members have worked on different sets of other projects.

Two things are worth noticing. The lectures say that the leader’s share of the work grows with team size. The paper says it falls slowly. And the lectures set “balance” and “leader” against each other, as though they were the same quantity measured twice. They are not. One is a laboratory score for strangers working together for an afternoon; the other is the popularity of a volunteer project over months. Stars, the authors warn, are not a perfect measure of quality, and success may also recruit members rather than the other way round. The data suggest that a successful project often has one or two people who do most of the work, supported by others who lead their own projects. They do not show that a team needs a boss.

Who gets the credit?

Return to the paper with the five thousand authors, and shrink it. Suppose a discovery has two authors. If the paper is the only evidence, you cannot tell who did what. But there is another source of evidence: how the scientific community treats the paper. In 2014 Hua-Wei Shen and a colleague proposed a way to read that evidence. For a given paper, find all the papers that cite it, then find the other papers those papers cite alongside it. These are the papers the community thinks of together with it. Each co-cited paper gives credit to its authors, split equally among them, and the credit of each author of the original paper is the total, weighted by how often the papers are cited together. The result is a share for each author. The idea is that credit is assigned not by the authors, or by the order of their names, but by the community’s perception of whose work this is.

The lectures use two Nobel Prizes of 2010. The Chemistry prize went to Richard Heck, Ei-ichi Negishi, and Akira Suzuki for chemical reactions that join carbon atoms using palladium. One of the key papers was written in 1976 by two authors, Baba and Negishi. According to the lectures, the papers most often cited together with it were Negishi’s, and none was Baba’s. The algorithm gave Negishi nearly three quarters of the credit (72 percent, according to the first author’s talks) and Baba a little more than a quarter. The Physics prize went to Andre Geim and Konstantin Novoselov for graphene. In their 2004 paper, which has eight authors, the algorithm gave Geim and Novoselov almost the same share, about a quarter each, well above the others. Across 63 multi-author Nobel-winning papers in physics, chemistry, and medicine, the author or authors with the most credit were the actual laureates in 51 cases, 81 percent.

Look at what the method says and does not say. It does not say who had the idea or did the experiments. In the authors’ words it captures the community’s perception, not the precise role of an individual. It is validated only on prize-winning papers, where the “right answer” is itself a committee’s decision, so it measures agreement with the committees, not deserved credit. And it can disagree with them: for the particle of 2013, the algorithm ranked Higgs first, then Tom Kibble, then Englert, and the Nobel committee skipped Kibble. The authors noted that the committee had “clearly deviated” from the community’s perception, and they cautioned against using the algorithm alone to decide credit. Credit is also not fixed: a prize pushes the laureates’ shares up afterwards.

Credit is perceived

The ninth rule, then, is that credit follows the community’s perception of an individual’s work. It is not enough to be on a winning team. You also have to be seen as the one who belongs to it. In 1968 Robert Merton had already observed that eminent scientists receive more than their share of the credit for joint work. A 2021 study by Heather Sarsons and her colleagues examined this for gender. Using the CVs of 644 economists who came up for tenure at one of the 35 leading US economics departments between 1985 and 2014, they found that men were tenured at a similar rate whether they coauthored or published alone, while for women an extra coauthored paper added less: after controlling for the quality of the work, an extra coauthored paper was linked to an 8.2 percent increase in the chance of tenure for men and 5.6 percent for women, and the gap was smaller for women who coauthored with women. In online experiments, observers asked to predict someone’s future performance from a joint score gave women less credit for a math quiz, though not for a grammar quiz. The authors read this as evidence that stereotypes shape credit. They also noted limits: a similar sample of sociologists showed no gap in tenure rates, the CV data cannot rule out differences in actual contributions, and the economists studied are at elite US departments.

The map of the art world

The tenth rule completes the circle. When performance cannot be measured at all, networks decide. The best test is the art world, where, as we saw in Lesson 1, nothing about the object tells you its worth. In 2018 a team led by Samuel Fraiberger put together the exhibition history of 496,354 artists between 1980 and 2016: 497,796 shows in 16,002 galleries, 289,677 shows in 7,568 museums, and 127,208 auctions in 1,239 auction houses, in 143 countries. The data came from a company, Magnus, whose founder is one of the authors.

From the records they drew a map. The nodes are galleries and museums, and a link joins two institutions whenever an artist has shown at one and then at the other. An institution is prestigious if it is well connected to other well-connected institutions, a calculation closely related to the PageRank score behind Google search. The score agreed with experts’ grades of 9,392 institutions and with a ranking based on prices. The core of the map is a dense cluster of major European and North American institutions; regional communities in Europe, Asia, South America, and Australia were relatively isolated from it.

The question was what an artist’s early exhibitions predict. The researchers looked at the first five. Among artists with at least ten recorded exhibitions, of those who began in the top fifth of institutions, 58.6 percent were still showing at that level at the end of their recorded career, and only 0.2 percent had ended in the bottom 40 percent. Of those who began in the bottom 40 percent, 10.2 percent eventually reached the top fifth. A model with memory, which looks at the prestige of an artist’s past exhibitions, predicted where an artist would show next: an artist whose past shows had been in the bottom tenth was 17 times as likely to show next at a low-prestige institution as at a high-prestige one, and an artist in the top tenth was 42 times as likely to show at a high-prestige one as at a low-prestige one. It predicts the level of the next venue, not which gallery. The 240 or so artists who began low and broke through all did so within their first ten years, and they tended to shop around more among different institutions at the start.

Where do the two writers of Lesson 1 fit? Not in this study, which does not name Basquiat or Diaz, and not in any proof. Their story has a shape the study makes plausible: a person who gets into the high-prestige part of the network early is likely to stay there, and one who does not, may not. For one person it is still a story. For half a million artists it is a pattern, and the authors’ own warning applies: art not shown in galleries and museums, such as performance art, is underrepresented, and a map of access does not measure talent.

Ten rules, weighed

The lectures set out ten rules. After the checking we have done, here is how much weight each one can bear.

• 1. Performance is about you, success about us. A useful distinction, with a blurry boundary.
• 2. Where performance can be measured, it drives success. Supported by one study of men’s professional tennis (about 60 percent of the variation in attention); not shown for other fields.
• 3. Performance is bounded. Strong, where performance is physically measured: the top eight in an Olympic final are within a few percent.
• 4. Success is unbounded. Strong: rewards are heavily skewed, and the reasoning is clear.
• 5. Success breeds success. Supported by randomized experiments, with real but modest effects and diminishing returns.
• 6. Success can come at any time. Supported in large samples of careers that lasted, and replicated in other fields; hot streaks add structure; says nothing about those who stopped.
• 7. The Q-factor decides. A useful description of citation records; its meaning and causes are disputed.
• 8. Teams need balance for performance and a leader for success. The weakest: the evidence is mixed, the measures differ, and the contrast is the lectures’ own.
• 9. Credit follows the community’s perception. Supported: an algorithm matched Nobel laureates in 81 percent of cases, and studies find gaps in credit by gender; it measures perception, not merit.
• 10. Networks drive success where performance cannot be measured. Strong correlational evidence from one large data set of the art world; it does not measure talent.

The common thread is the first rule. A great deal of what we call success is what a community does with a performance: it rewards it, forgets it, or credits someone else. That is not a reason for despair, and it is not a reason to stop doing good work, because in most fields some part of success does follow performance. It is a reason to be modest about the winners, and to pay attention to the networks we are in and to the ones we are not.