Network Centrality Analysis

Thinking Tools3.3 Mapping the System Around the ProblemAlso known as Organizational Network Analysis

Developed by Alex Bavelas (A Mathematical Model for Group Structures, 1948), Linton C. Freeman (Centrality in Social Networks: Conceptual Clarification, 1979), and Phillip Bonacich (1972), set out by Stanley Wasserman and Katherine Faust (Social Network Analysis, 1994) and by Stephen P. Borgatti, Ajay Mehra, Daniel J. Brass, and Giuseppe Labianca (Network Analysis in the Social Sciences, 2009), and brought to managers by Rob Cross and Andrew Parker (The Hidden Power of Social Networks, 2004).

Measures where each person sits in a web of working ties (degree, betweenness, closeness, and eigenvector centrality) to find the connectors, brokers, and bottlenecks an org chart hides.

Twelve people on a product team, measured four ways

Each line is a working tie from a two-question survey of the team. Switch the measure and the squares resize: the engineering manager has the most ties; the coordinator, with only three, sits on the most routes and is nearest to everyone; and on eigenvector the engineers tied to the manager outscore the design lead, who has as many ties. Select a person to read their numbers, and take one out to see the measures recompute. Illustrative.

How many ties the person has. Highest: Ben, UX researcherAna, design leadEli, engineerFay, engineerCam, visual designerCy, program coordinatorDev, engineering managerGil, engineerJo, sales engineerIda, product marketingKai, support leadHal, QA leadHow often the person sits on the shortest route between two others. Highest: Ben, UX researcherAna, design leadEli, engineerFay, engineerCam, visual designerCy, program coordinatorDev, engineering managerGil, engineerJo, sales engineerIda, product marketingKai, support leadHal, QA leadHow few steps the person is from everyone else. Highest: Ben, UX researcherAna, design leadEli, engineerFay, engineerCam, visual designerCy, program coordinatorDev, engineering managerGil, engineerJo, sales engineerIda, product marketingKai, support leadHal, QA leadHow well connected the person's own contacts are, 1 for the best. Highest: Ben, UX researcherAna, design leadEli, engineerFay, engineerCam, visual designerCy, program coordinatorDev, engineering managerGil, engineerJo, sales engineerIda, product marketingKai, support leadHal, QA lead

Select a person to read their numbers.Ben, UX researcher. Works inside design; his findings reach engineering only through Ana and Cy. Degree 2 · Betweenness 0.00 · Closeness 0.32 · Eigenvector 0.10Ana, design lead. Design's one tie outside the group runs to Cy. Degree 3 · Betweenness 0.33 · Closeness 0.44 · Eigenvector 0.21Eli, engineer. A solid engineer with two ties, both inside engineering. Degree 2 · Betweenness 0.00 · Closeness 0.38 · Eigenvector 0.55Fay, engineer. Pairs with Eli and Gil; never hears from design directly. Degree 3 · Betweenness 0.01 · Closeness 0.39 · Eigenvector 0.73Cam, visual designer. Two ties, both inside design. Degree 2 · Betweenness 0.00 · Closeness 0.32 · Eigenvector 0.10Cy, program coordinator. Three ties, low on the org chart, and design's only link to the rest of the team; marketing's shortest route to engineering also runs through him. In the month he was on leave, the design work stalled. Degree 3 · Betweenness 0.56 · Closeness 0.58 · Eigenvector 0.47Dev, engineering manager. The most ties on the team, nearly all inside engineering. Degree 5 · Betweenness 0.45 · Closeness 0.55 · Eigenvector 1.00Gil, engineer. Three ties, the same count as Ana, but all to well-connected engineers, so he scores far higher on eigenvector. Degree 3 · Betweenness 0.02 · Closeness 0.41 · Eigenvector 0.75Jo, sales engineer. One tie, to Ida: customer requests reach the team through two hops. Degree 1 · Betweenness 0.00 · Closeness 0.32 · Eigenvector 0.09Ida, product marketing. Marketing's link to Cy, and through him to everyone else. Degree 3 · Betweenness 0.25 · Closeness 0.46 · Eigenvector 0.27Kai, support lead. Talks to Ida and to Hal about bugs customers report. Degree 2 · Betweenness 0.05 · Closeness 0.41 · Eigenvector 0.29Hal, QA lead. Engineering's only direct tie to support, by way of Kai. Degree 3 · Betweenness 0.09 · Closeness 0.44 · Eigenvector 0.65

PersonDegreeBetweennessClosenessEigenvector
Ben, UX researcher20.000.320.10
Ana, design lead30.330.440.21
Eli, engineer20.000.380.55
Fay, engineer30.010.390.73
Cam, visual designer20.000.320.10
Cy, program coordinator30.560.580.47
Dev, engineering manager50.450.551.00
Gil, engineer30.020.410.75
Jo, sales engineer10.000.320.09
Ida, product marketing30.250.460.27
Kai, support lead20.050.410.29
Hal, QA lead30.090.440.65

An example

A software company's twelve-person product team had missed two launch dates, and each group blamed the handoffs. The vice president of product asked everyone two questions: whose help do you need to get your work done, and who do you share project news with? She turned the answers into a list of ties and computed three centrality scores for each person.

The engineering manager had the most ties, as everyone expected, and almost all of them ran inside engineering. The surprise was the program coordinator. He had only three ties, sat near the bottom of the org chart, and scored highest on both betweenness and closeness: design, engineering, and marketing each reached the others mostly through him. The launch slips matched the month he had spent on leave. She made the coordinator role a formal one, gave him a backup, and put the design lead into the engineering standup, so the team no longer depended on a single path.

Why it works

An org chart records who reports to whom; work runs through who asks whom. Centrality turns the second into numbers, and each measure answers a different question. Degree counts ties, so it finds the busy, visible people. Betweenness counts how often a person sits on the shortest route between two others, so it finds the brokers whose absence would cut groups apart. Closeness measures how few steps a person is from everyone, so it finds whoever hears things first. Eigenvector centrality, the fourth common measure, scores people by how well connected their contacts are.

Computing several at once matters because they disagree, and the disagreement is the finding. A person low on degree and high on betweenness is a bottleneck the chart does not show.

The thinking behind it

Alex Bavelas Alex Bavelas ran the small-group communication experiments at MIT in the late 1940s. His paper A Mathematical Model for Group Structures (Applied Anthropology, 1948) is credited by Linton Freeman with bringing centrality into the study of human communication: the idea that where a person sits in a network of ties shapes how much influence they have, whatever their formal role.

Linton C. Freeman Linton C. Freeman, a founder of social network analysis, published a paper in the journal Social Networks in 1979, Centrality in Social Networks: Conceptual Clarification, that sorted the field's many centrality measures into three conceptions: degree, betweenness, and closeness. He gave each a standard formula, and the paper is still the field's reference for them. Three of the four measures in the figure on this page are his three conceptions.

Phillip Bonacich Phillip Bonacich, a mathematical sociologist at UCLA, proposed in Factoring and Weighting Approaches to Status Scores and Clique Identification (Journal of Mathematical Sociology, 1972) that each member of a network be scored by the scores of the people they are tied to. That is the eigenvector measure, which network centrality analysis still uses to find people whose few contacts are themselves central; in the figure it lifts the engineers tied to their manager above a design lead with as many ties.

How to run it

  1. Define the people and the tie. Name who is in the network and what counts as a tie, such as asking for help, sharing news, or trusting someone's advice.
  2. Collect the ties. A two-question survey works for a team; message metadata or observation works for a larger group, with consent.
  3. Build the list of ties. One row per pair, cleaned of duplicates and names spelled two ways.
  4. Compute several measures. Run degree, betweenness, closeness, and eigenvector before deciding which one answers the question.
  5. Compare the leaders with the org chart. Look hardest at people high on betweenness and low on rank; they are the hidden bottlenecks.
  6. Remove a person and recompute. If one departure splits the network, add a second path before that person leaves.
  7. Share findings as structure. Report patterns and roles to the team; hand out individual scores sparingly.

Where it goes wrong

The first failure is picking one measure and reading it as importance. A high degree score can belong to someone who is merely busy. Run several measures and explain where they differ.

The second is bad tie data: a survey half the team skipped, or message counts that capture volume and miss trust. Spot it when the map shows people as isolated who obviously are not, and fix the collection before interpreting anything. The third is turning scores into a ranking of people. Staff who learn their number game the next survey, and brokers who are thanked publicly for holding the network together can become targets for poaching. Present the results as roles and paths.

When to use something else

  • Influence Sociogram. When the network is small, the question is one decision, and a hand-drawn map from observation will do.
  • Power-Interest Grid. When the question is how much attention each stakeholder needs on a project, and the ties between them matter less.
  • Value Stream Mapping. When the delays live in the steps of a process more than in the relationships between people.

Where it sits in SOLVE eX

Go to the source

  • A Mathematical Model for Group Structures, Applied Anthropology, 1948
  • Centrality in Social Networks: Conceptual Clarification, Social Networks, 1979
  • Factoring and Weighting Approaches to Status Scores and Clique Identification, Journal of Mathematical Sociology, 1972
  • Social Network Analysis: Methods and Applications, Cambridge University Press, 1994
  • The Hidden Power of Social Networks: Understanding How Work Really Gets Done in Organizations, Harvard Business School Press, 2004
  • Network Analysis in the Social Sciences, Science, 2009

Back to the catalog