Science as Team Sport

The Large Hadron Collider, accelerating subatomic particles to light speed before crashing them together in spectacular fashion 100 meters beneath the Franco-Swiss border, unites thousands of physicists and engineers from dozens of nations and hundreds of universities in one of the world’s largest scientific collaborations. But calculus, a cornerstone of mathematics that is wielded en masse in the collider’s humming tunnels and glowing control rooms, was developed by just two men, Gottfried Wilhelm Leibniz and Sir Isaac Newton, each of whom worked independently in the latter half of the 17th century.

Recent research by Benjamin Jones (Management and Strategy), Stefan Wuchty (National Institutes of Health), and Brian Uzzi (Management and Organizations) sheds light on how scientific discovery—for ages sprung primarily from the minds of singular giants—is now more likely to arise from large, distributed teams. Published in the journal Science, the study shows that while the reach and influence of teams is growing, the benefits of this evolution are concentrated largely among the nation’s most elite universities.

“There’s the old, classical idea about the lone scientist, like Aristotle or Newton. As recently as the 1950s you had a higher probability of hitting a home run if you came to the plate by yourself,” said Jones, describing how landmark scientific discoveries had been achieved for centuries. “Now, you have a better chance to hit that home run as a team.”

Even among the very best schools—where rock star researchers and founders of fields shared washrooms and water coolers—there was still an advantage to collaborating with other schools. But this advantage was concentrated primarily among the elite.

But as any fan of the Chicago Cubs can tell you, simply assembling a team does not guarantee that you will be able to swing a bat, let alone hit a home run. And with the global, knowledge-based economy depending more and more upon ideas and innovations, the ability to understand and fine-tune research initiatives takes on heightened significance.

“How do you assemble a network and figure out where to plug in your ideas to get the best return? Where do you place your bets? How do you create systems to enhance research?” asked Uzzi.

Following the Research Paper Trail

The beating heart of science can be found in the pages of research journals, where reports of the latest studies are published. Authorship of such papers is a primary indicator of scholars’ productivity. So to take the pulse and measure of the national research apparatus, Jones and colleagues followed the paper trail. This particular trail was weighty and wired, with roots stretching back to the 1960s and the austerely named Institute for Scientific Information (ISI).

The ISI Web of Science (WoS) is an online database that tracks the contents of roughly 8,700 leading academic journals. Jones and colleagues studied thirty years worth of WoS data, a total of 4.2 million research articles published from 1975 to 2005. Researchers from 662 major U.S. universities were represented, spanning 172 fields of science and engineering (SE), from astronomy to zoology. Another 54 fields of social science (SS), such as psychology and economics, were also represented.

Collaboration across Academic Borders

“In the 1950s, worldwide, there were forty thousand publications in science and engineering. Now, there are about a million papers this decade,” said Jones. “To be a renaissance man today, with a million papers published, is tough.”

So tough, perhaps, that savvy scientists are teaming up, not just across campus but increasingly across time zones. In both SE and SS, multi-school collaborations were relatively rare in 1975 when no more than 10 percent of published collaborations involved more than one institution. Over the thirty years that followed, multi-school collaborations grew steadily to account for 30 to 35 percent of publications in 2005. Over that same period, single-author papers became increasingly rare, down from roughly 30 to 10 percent for SE, from 60 to 40 percent for SS. The number of collaborations that were confined to a single school remained fairly constant, roughly 60 percent for SE, 30 percent for SS.

“Some of our earlier work was on this changing nature of the production of knowledge in science, the inversion from sole authorship to team-based publications,” said Uzzi. “This was becoming clear in fields that were capital intensive, requiring big grants and special equipment, like in nuclear physics. But now we’re finding that it’s more universal, even seeing changes in the social sciences.”

While some might attribute this trend to the emergence of the Internet and its telecommunicating kin, the data suggest otherwise, revealing a pattern that largely took hold prior to the communications boom of the 1980s and 1990s. “It’s surprising,” said Jones, “but it doesn’t seem to be a result of increased ease of communication. The trend is mostly smooth back to the 1970s.”

Specialization Drives Team Formation and Collaboration

“Instead,” he continued, “it could be due to increasing specialization. Researchers innovate by becoming more expert along the expanding frontier of knowledge.” Single institutions, let alone isolated researchers, typically cannot amass expertise in more than a slight fraction of a field. To harness increasingly specialized knowledge and skills that are needed to solve ever more complex problems, researchers seek collaborators based on expertise and interests, not location. But, added Jones, “They’re not reaching further, just more often.” While the overall level of multi-school collaboration has increased, the average distance between collaborators has grown only slightly to 800 miles, up from 750 miles for SE, 725 miles for SS.

To examine how collaborations influenced the quality of research, Jones and colleagues first determined how often average, run-of-the-mill research articles were cited by subsequent research papers, focusing on the decade spanning from 1995 to 2005. To find the crème de la crème—the game-changing, high impact publications that arose from particularly fertile collaborations—they then identified papers that were cited significantly more than the average paper. For example, the average SE paper published in 2001 was subsequently cited about 17 times, while a high impact paper was cited at least 38 times.

Collaborating beyond a single school improved the odds of publishing high impact research. In SE, multi-school collaborations had a 35.6 percent chance of having high impact, 2.9 percent higher than for single-school collaborations. A multi-school boost of 5.8 percent was seen for SS as well, increasing the probability of publishing high impact research from 34.1 to 39.9 percent.

Jones and colleagues next considered ways to rank and categorize the more than 600 institutions, recognizing that not all schools are created equal, and that those inequalities could influence—or be influenced by—the structure and strength of collaborations. Schools were ranked according to the total number of times that their researchers were cited by others, considering only single-authored papers and within-school collaborations. The top 5 percent of schools were categorized as Tier I. Tier II encompassed schools from 6 to 10 percent, Tier III from 11 to 20 percent, and Tier IV the remaining 80 percent.

From 2001 to 2005, researchers from Tier III and Tier IV schools, which account for 90 percent of the institutions in the study, participated in only 18 percent of multi-school collaborations in both SE and SS. At the other end of this spectrum, Tier I schools, a mere 5 percent of the institutions, participated in 55 to 60 percent of multi-school collaborations. Even among these very best schools, where rock star researchers and founders of fields shared washrooms and water coolers, there was still an advantage to collaborating with other schools. Within Tier I, multi-school collaborations were 6.19 percent (SE) to 11.7 percent (SS) more likely than single-school collaborations to achieve high impact status. “A Harvard-Harvard connection doesn’t do as well as a Harvard-Stanford connection,” said Jones.

But this advantage was concentrated primarily among the elite. Multi-school collaboration offered diminished advantages within Tiers II and III, and was ineffective—even slightly detrimental—within Tier IV. In the 68 percent of collaborations that spanned multiple tiers, the higher tier schools blunted the impact of their research and realized less benefit than if they had collaborated within their own tier. However, the higher tier school pulled the overall impact up more than the lower tier school pulled it down.

With so much at stake, researchers probably did not choose their collaborators like slumber-partying teens making prank calls, blindly flipping through the phonebook and dialing numbers at random. But what if they did? Assuming that the researchers were not self-destructive or intent on career suicide, such a selection process would suggest that all potential collaborators, the whole phone book, were deemed equally likely to share valuable skills, efforts, and insights, regardless of institution or tier. Jones and colleagues modeled this random scenario to see if it reflected real collaborative processes.

Contrary to the egalitarian predictions of the random matching model, researchers collaborated in a skewed, stratified manner, staying largely within their own tiers. Compared to the random model, same-tier collaborations among Tier I schools were 14 percent more common in SE, 27 percent more common in SS. Similarly, same-tier collaborations among lower tiers were more frequent by as much as 89 percent in the case of SE among Tier IV schools. But collaborations across tiers were unexpectedly rare. Compared to the random model, Tier I-IV collaborations were 19 percent less common in SE, 32 percent less common in SS.

“We see increasing stratification more than diversification,” said Jones, who noted also that these between-tier gaps have been growing over time. Since the late 1970s, within-tier collaborations among Tier I schools have increased, while between-tier collaborations of Tier I schools with partners in Tier IV have become more rare. And once the decisions were made to look beyond one’s own school for collaborators, said Jones, “If you’re going to reach a few kilometers, you might as well reach a few thousand kilometers.” He pointed to analyses of specific research hotspots, such as Boston, the Bay Area, Chicago, and North Carolina’s Research Triangle. These showed, again, that social proximity and the relationships with one’s peers are becoming far more important than spatial proximity. Top tier schools in those cities increased their collaboration with same tier peers in distant locales, while partnerships with local schools declined.

These findings, far from an esoteric exercise in library science, could advance policy and practice and have significant impact on economic development. Said Jones, “Economists are concerned with production functions, inputs and outputs, efficiency. We looked at knowledge production functions. How do you put individuals together? How does that influence decisions, career paths? And how do those individual functions aggregate up to the level of the economy?”

Further reading:

Whitfield, John (2008). “Collaboration: Group Theory,” Nature, October 8, 455(7214): 720-723 (Accessed October 10, 2008; subscription required)