Co-Authorship Network Metrics: Reading Collaboration Structure

A co-authorship network represents researchers as nodes and shared authorship on a paper as edges connecting them — two authors are linked if they have co-authored at least one publication together, and the link is often weighted by how many papers they share. This is a distinct object from a citation network, which links papers (or authors) through co-citation or bibliographic coupling rather than direct collaboration. A co-authorship network answers a different question: not “which work references which,” but “who actually works with whom.”

Bibliometric software routinely draws these networks as maps of dots and lines. The map alone doesn’t tell a research office anything actionable. What does is the small set of graph-theoretic measures that describe a network’s structure numerically — degree, betweenness, and clustering coefficient are the three most directly useful for reading a collaboration structure, at both the individual-researcher and whole-network level.

Degree centrality: who has the most collaborators

Degree centrality is the simplest measure: it is the count of unique co-authors a researcher has across the papers included in the network. A researcher with a degree of 40 has published with 40 distinct people (not 40 papers — repeat collaborations with the same person don’t add to degree, though they typically do add to edge weight).

High degree identifies prolific collaborators and network hubs — often senior PIs, core-facility staff, or methodologists who get pulled into many projects. It is a genuine signal of collaborative reach, but on its own it conflates two very different roles: someone who collaborates widely within a single large, tightly-connected team, and someone who collaborates across otherwise-unconnected groups. Distinguishing those two requires the next measure.

Betweenness centrality: who bridges separate groups

Betweenness centrality measures how often a node falls on the shortest path between other pairs of nodes in the network. A researcher with high betweenness sits structurally between clusters that would otherwise have no direct connection — removing that person would lengthen, or sever, the path between those groups.

This is the measure that separates a “hub” from a “broker.” A researcher can have only a modest degree — a handful of collaborators — and still carry disproportionately high betweenness, if those few collaborators happen to sit in otherwise-separate labs, departments, or subfields. That person is a single point of structural connection between communities that don’t otherwise talk to each other. High-betweenness, low-degree individuals are the ones a research office should actually worry about losing: their departure doesn’t just remove one collaborator from a team, it can fragment the wider network into disconnected components.

Clustering coefficient: how tight-knit a researcher’s own circle is

The local clustering coefficient for a given researcher asks a narrower question: of that researcher’s own collaborators, what fraction are also collaborators with each other? A value close to 1 means the researcher’s co-authors form a closed, tight-knit team — everyone already knows everyone. A value close to 0 means the researcher’s collaborators are drawn from otherwise-separate circles that don’t overlap.

Read alongside betweenness, clustering coefficient sharpens the hub/broker distinction: a high-degree, high-clustering researcher is embedded in one dense team; a modest-degree, low-clustering, high-betweenness researcher is the connective tissue between teams. Averaged across every node, the same statistic becomes a network-level measure — the global clustering coefficient (sometimes reported as transitivity) — describing how “clique-y” an entire field or department’s collaboration structure is, as distinct from how sparse or dense it is overall.

Network-level measures worth reading alongside the individual ones

• Network density — the proportion of all possible collaborator pairs that are actually connected. Sparse in large fields, denser in small specialist ones.
• Giant component size — whether the network resolves into one large connected mass (typical of mature, well-integrated fields) or splits into many small, disconnected clusters (a sign of fragmentation, or of a young/siloed area).
• Average path length — the typical number of collaboration “steps” separating any two researchers in the connected component. Co-authorship networks in many fields have been documented to show small-world structure: short average path lengths despite low overall density, a pattern associated with the mathematician and physicist collaboration-network studies that established this approach in the early 2000s.

Reading this as collaboration strategy, not decoration

A co-authorship network map is often produced and shown once, as an illustration, and then never used again. The measures above are what make it a working tool instead:

• Succession and key-personnel risk. A high-betweenness researcher’s planned retirement or departure is a structural risk to a department’s cross-group collaboration, not just a loss of one productive collaborator — worth flagging before it happens, not after the network visibly fragments.
• Evidence for interdisciplinary investment. Centers, seed-grant programs, and cluster hires aimed at “building bridges” between departments can point to betweenness centrality as before/after evidence that a specific investment actually changed the connectivity structure, rather than relying on self-reported collaboration counts.
• Distinguishing productive from central. Publication counts and h-index measure individual output; degree, betweenness, and clustering coefficient measure a researcher’s position in the collaboration structure around them. The two rankings frequently diverge, and both are legitimate inputs to different institutional decisions — conflating them is a common misreading.
• Spotting genuine isolation. A subgroup that forms its own small, disconnected component — no path to the rest of the network at all — is a concrete, measurable version of “this group doesn’t collaborate with the rest of the department,” useful for a research office trying to justify pairing or seed-funding interventions.

Tools that compute these measures

None of this requires building a graph algorithm from scratch. VOSviewer constructs and visualizes co-authorship networks directly from a bibliographic export and reports basic network statistics; Bibliometrix, the R package (and its biblioshiny web interface), builds collaboration networks alongside its other bibliometric workflows and can compute degree, betweenness, and clustering coefficient over them. General-purpose network analysis tools — Gephi and Pajek, both widely used outside bibliometrics specifically — accept an exported edge list from either of the above and compute the full range of centrality and cohesion statistics, plus community-detection algorithms for identifying the underlying subgroups these measures describe.

How this differs from citation network analysis

It’s easy to conflate co-authorship networks with the related but distinct family of citation network analysis (co-citation and bibliographic coupling), since both are visualized the same way and often computed in the same software session. The distinction matters: a citation network describes intellectual relationships between pieces of work — which papers draw on similar prior literature, or get cited together. A co-authorship network describes actual working relationships between people. Two researchers can be heavily co-cited without ever having worked together, and two frequent collaborators can cite almost entirely different prior literature. For questions about who collaborates with whom — succession risk, interdisciplinary bridge-building, team structure — the co-authorship network, not the citation network, is the right object to build.

Frequently asked questions

Is a co-authorship network the same as a citation network?

No. A co-authorship network connects researchers who have written a paper together; a citation network connects papers (or their authors) through co-citation or bibliographic coupling — shared references or shared citing papers, not shared authorship. See the citation network analysis guide for that distinct approach.

What’s the difference between a “hub” and a “broker” in a co-authorship network?

A hub has high degree centrality — many direct collaborators, often within one dense team (high clustering coefficient). A broker has high betweenness centrality relative to their degree — comparatively few collaborators, but positioned so that removing them would disconnect otherwise-separate groups. The two roles carry different institutional implications, and a researcher can be one without being the other.

Do I need specialized software to compute these metrics?

Free tools cover this well: VOSviewer and Bibliometrix/biblioshiny both build co-authorship networks directly from a bibliographic database export (Scopus, Web of Science, or OpenAlex) and compute basic network statistics; Gephi and Pajek take an exported edge list and add the full range of centrality measures and community detection.

Does a dense co-authorship network mean a field collaborates well?

Not necessarily. Density measures how connected the network is overall, but a dense network can still be made of several disconnected or loosely-linked dense clusters. Giant component size and betweenness distribution are better indicators of whether a field is genuinely integrated versus siloed into well-connected pockets that don’t talk to each other.

Related on CASRAI: Scientometrics for the wider discipline these measures sit inside, Co-Authorship for the underlying authorship concept, and the Publishing & Research Assessment hub for the full cluster.