Productive Scientists Build Bigger Networks — Not the Other Way Around

For every 1% increase in a scientist’s publication count, their number of distinct coauthors rises by 1.3%. That’s not just correlation—it’s a precise, predictable, superlinear scaling relationship observed across 26,876 Brazilian researchers (Frota et al., 2026). But here’s what’s surprising: the reverse isn’t true. When you look at how many papers a researcher produces based on how many people they’ve collaborated with, the pattern dissolves into noise. There’s no clean functional form, no consistent variance structure—just irregularity.

This asymmetry isn’t a statistical quirk. It’s evidence—strong, structural evidence—that among these scientists, productivity shapes collaboration far more than collaboration shapes productivity. In other words, publishing more doesn’t just go hand-in-hand with working with more people; it appears to drive it. Big networks aren’t the engine of output—they’re the exhaust plume.

The Science

The researchers analyzed career-long publication records from Brazil’s Plataforma Lattes, a national academic CV system where scholars self-report their work. Their cohort? 26,876 recipients of the prestigious Bolsa de Produtividade em Pesquisa—a competitive fellowship awarded to top-tier researchers across disciplines. These aren’t random academics; they’re an elite group selected not just for volume, but for impact, mentorship, leadership, and integration into the global research community.

From each profile, the team extracted two simple metrics: total number of publications ($A$) and total number of unique coauthors ($C$). They then applied a causal discovery framework based on conditional distributional asymmetries—a method that infers directionality not through experiments or interventions, but by asking which variable, when used to condition the other, yields a cleaner, more theoretically plausible statistical structure.

The logic is subtle but powerful. If $A \to C$, then for any given level of productivity, the variation in collaboration should follow a natural stochastic process—say, geometric fluctuations around a mean determined by $A$. Conversely, if $C \to A$, then for any given network size, the variation in output should follow such a process. The framework tests which of these two worlds looks more like reality.

They also examined marginal distributions—how $A$ and $C$ are distributed across all researchers—and tracked how inequality in both measures evolved with career age, using the Gini coefficient as a measure of concentration.

What They Found

First, the correlation: strong and positive. On log scales, the average number of coauthors scales as $\langle C \rangle \sim A^{\beta}$ with $\beta = 1.30 \pm 0.03$

Publications vs. Coauthors: Superlinear Scaling

Label	Value
Bin 1	0 coauthors
Bin 2	0 coauthors
Bin 3	0 coauthors
Bin 4	0 coauthors
Bin 5	0 coauthors
Fit at x=10	0 coauthors
Fit at x=200	0 coauthors

Each point represents the average coauthor count within logarithmically spaced publication bins. Error bars (not shown) are 99% bootstrap confidence intervals. Source: Frota et al. (2026). {{image:0}}

. This means that as scientists publish more, they don’t just add collaborators linearly—they add them disproportionately. A researcher with 100 papers has, on average, not 10 times but nearly 20 times as many coauthors as one with 10 papers.

But correlation is symmetric. Causation isn’t.

When the researchers conditioned on productivity (i.e., looked at collaboration given a certain number of publications), they found something remarkable: the variance in coauthor count followed a near-perfect geometric distribution. Specifically, the standard deviation $\sigma_C$ scaled with the mean $\langle C \rangle$ as $\sigma_C = \sqrt{\langle C \rangle(\langle C \rangle - 1)}$, exactly as predicted by a simple model where new collaborations form probabilistically at a rate proportional to existing output

Mean-Variance Relationship Reveals Causal Direction

Label	Value
Low	0 standard deviation
Medium	0 standard deviation
High	0 standard deviation
Low	0 standard deviation
Medium	0 standard deviation
High	0 standard deviation
Theory Low	0 standard deviation
Theory Mid	0 standard deviation

Left: Variance of coauthor count closely matches geometric model when conditioned on publications. Right: Variance of publication count deviates significantly when conditioned on coauthor count. Alignment with theoretical curve indicates plausibility of generative mechanism. {{image:2}}

.

This kind of regularity suggests a coherent generative mechanism: productivity sets the baseline, and collaboration fluctuates around it in a predictable way.

Now flip it. Condition on collaboration (i.e., look at publications given a certain number of coauthors). The picture breaks down. The mean-variance relationship for publications becomes erratic, deviating sharply from any simple model—including the same geometric expectation. There’s no clean function, no natural stochastic process that explains how output varies once network size is fixed.

This asymmetry alone points toward $A \to C$.

Then there’s the shape of the overall distributions. Publications follow a lognormal distribution—so well, in fact, that a Kolmogorov-Smirnov goodness-of-fit test yields a p-value of 0.46, meaning the data are entirely consistent with being drawn from a lognormal process (Equation 2). This makes mechanistic sense: if productivity arises from multiplicative processes—each paper enabling the next through reputation, funding, or infrastructure—then lognormality emerges naturally.

Coauthor counts, meanwhile, do not follow a simple lognormal or power law. Instead, the best fit comes from a mixture model: a geometric distribution whose mean is itself lognormally distributed. In other words, heterogeneity in productivity generates the broad distribution of collaboration. You first draw a productivity level from a lognormal distribution, then generate a coauthor count from a geometric distribution centered on a function of that productivity. This hierarchical model reproduces the empirical coauthor distribution far better than any single-component alternative.

Finally, the dynamics over time. The researchers grouped scientists by career age and computed Gini coefficients for both publications and coauthors within each cohort. The results were striking: inequality in publication output increased steadily with career age, rising from about 0.42 at 5 years post-first-publication to 0.57 at 20 years

Inequality Grows in Output, Not Network Size

Label	Value
Year 5	0 Gini coefficient
Year 10	0 Gini coefficient
Year 15	0 Gini coefficient
Year 20	0 Gini coefficient
Year 5	0 Gini coefficient
Year 10	0 Gini coefficient
Year 15	0 Gini coefficient
Year 20	0 Gini coefficient

Gini coefficients computed within cohorts of similar career age. Publication inequality increases over time; coauthor inequality remains flat despite high initial levels. {{chart:2}}

. This is the signature of cumulative advantage—success breeds more success.

But coauthor inequality? It started high (Gini ≈ 0.58) but remained flat. No compounding. No snowballing. The rich got richer in output, but not necessarily in network breadth.

Why This Changes Things

We’ve long assumed that collaboration fuels productivity. Team science is celebrated. Funding agencies prioritize collaborative grants. Universities build interdisciplinary institutes. The implicit theory is that bringing smart people together generates more knowledge than they could produce alone.

And sometimes it does. No one denies that collaboration can enable breakthroughs, especially in complex, resource-intensive fields. But this study suggests that at the level of individual careers, the dominant causal arrow may point the other way.

Productivity isn’t just correlated with collaboration—it appears to cause it. How? Through visibility. Reputation. Opportunity. A scientist who publishes consistently becomes known. Others seek them out. Invitations to join projects, write reviews, lead consortia—all flow more readily to those already producing.

This reframes big networks as a consequence of success, not its cause. It suggests that policies aimed at boosting collaboration—mandating team composition, rewarding coauthorship, funding network formation—may have limited returns if they don’t also address the underlying drivers of productivity: time, resources, training, autonomy.

It also deepens our understanding of inequality in science. The “Matthew Effect”—where the rich get richer—is often invoked to explain why a few scientists dominate citation counts or funding. This study shows it operates most powerfully on output, not connectivity. Once you’re publishing a lot, everything else—collaborators, invitations, influence—follows.

Interestingly, the effect was strongest in fields organized around labs and research groups: Exact and Earth Sciences, Engineering, Biological Sciences. In these domains, productivity inequality grew most sharply with career age, and the causal signal from $A \to C$ was clearest. In contrast, in Humanities and Social Sciences, the asymmetry weakened, though never reversed. This makes intuitive sense: in lab-based sciences, productivity is tied to infrastructure, teams, and grant funding—all of which scale with past success. In solo-authored fields, the link may be looser.

Still, even in those areas, the data never favored $C \to A$. Where the signal was weak, it was silent—not contradictory.

What’s Next

This study doesn’t prove that collaboration never boosts productivity. It shows only that, when looking at total career outputs, the dominant statistical signature points from productivity to collaboration. There may be important nonlinearities: perhaps collaboration does give a boost, but only after a threshold of productivity is reached. Or perhaps it accelerates output in specific contexts—say, cross-disciplinary work or international teams—but not on average.

Future work could test these ideas by moving beyond totals to longitudinal models. Does gaining a new collaborator predict a jump in future output? Do team compositions matter? Are there tipping points?

There are also limitations. The data come from self-reported CVs, though Lattes is considered reliable. Coauthor count doesn’t capture depth of collaboration—some ties are fleeting, others transformative. And the cohort is elite: findings may not generalize to early-career researchers or less productive scientists.

Most importantly, this is observational causal inference. The method detects asymmetries in joint distributions, not effects of intervention. We can’t say what would happen if we randomly assigned scientists to large or small networks. But in the absence of randomized trials—which are impossible at this scale—the structural regularities uncovered here offer compelling evidence.

The takeaway isn’t that collaboration is unimportant. It’s that we may have the story backwards. Rather than seeing collaboration as the engine of scientific progress, we might better view it as the signature of it. The most connected aren’t necessarily the most productive because they collaborate more—they collaborate more because they’re already productive.

That doesn’t diminish the value of teamwork. It refines our understanding of how credit, attention, and opportunity circulate in science. And it suggests that if we want more collaboration, we might start by supporting more productivity—through stable funding, protected time, and equitable access to the tools of research.

Because in science, as in so many systems, momentum begets connection.