Appearance
To interpret centrality measures responsibly, pick the measure that matches your historical question, normalise before comparing, and always read scores against the biases of your sources. A node with high betweenness is structurally a broker in your data — whether that reflects a real historical broker depends entirely on what survived and what you recorded. Centrality describes the network you built, not the past directly.
This is a working best-practice checklist for keeping centrality results consistent, documented and defensible across a whole collection.
Which centrality measure answers which question?
Each measure encodes a different idea of "important". Choose deliberately:
| Measure | Intuition | Good historical question |
|---|---|---|
| Degree | How many direct ties | Who had the most contacts? |
| Betweenness | How often on shortest paths | Who brokered between groups? |
| Closeness | How near to everyone else | Who could reach the network fastest? |
| Eigenvector | Connected to well-connected | Who was embedded in the elite? |
| PageRank | Directed influence flow | Who received the most attention? |
Reporting one measure you fully understand is worth more than four you cannot defend in a seminar.
Why can't I compare raw scores across networks?
Centrality values scale with network size and density. A betweenness score of 0.4 in a 50-node graph and 0.4 in a 5,000-node graph mean entirely different things. Two rules keep you safe:
- Use normalised versions (most tools offer a normalised flag).
- Compare ranks or quantiles, not raw values, across networks.
In NetworkX:
python
import networkx as nx
bc = nx.betweenness_centrality(G, normalized=True)
ranked = sorted(bc, key=bc.get, reverse=True)[:10]State findings as "Clarke ranks third for brokerage" rather than "Clarke scores 0.31".
How does missing data distort the picture?
This is the deepest pitfall in historical network analysis. Centrality — betweenness above all — is acutely sensitive to absent edges. If one pivotal letter was lost, the broker who connected two circles may vanish from your scores entirely, or a marginal figure may appear pivotal because the true bridge is missing.
Before interpreting, ask:
- What fraction of original ties plausibly survived?
- Is survival correlated with status, so important figures are over-represented?
- Would adding one plausible missing edge reorder my top ten?
If the answer to the last question is "yes", flag every centrality claim as provisional.
Should I trust eigenvector centrality on directed data?
Often no. Eigenvector centrality can fail to converge on directed graphs with dangling nodes (recipients who never write back), producing zeros or errors. For directed historical networks — correspondence, citation, patronage — prefer PageRank, which is designed for exactly that structure and handles dead ends with a damping factor.
python
pr = nx.pagerank(G, alpha=0.85)How do I keep results consistent across a collection?
Standardise a small procedure and apply it identically to every sub-network:
- Fix the graph model (directed/undirected, weighted/unweighted) once.
- Compute the same normalised measures for every slice.
- Record the software version and parameters (e.g. PageRank
alpha). - Save scores to CSV alongside the graph file.
- Report ranks in prose; keep precise scores in the data.
Documenting parameters matters: weighted versus unweighted betweenness can completely reorder your brokers.
What does a defensible centrality claim look like?
A strong claim names the measure, the normalisation, the data caveat and the rank — for example: "In the surviving 1690s letters, Locke ranks first for normalised degree and second for PageRank; given the Bodleian's collection bias toward his own papers, his apparent centrality is partly an artefact of provenance." That sentence is honest, reproducible and seminar-proof.
Key Takeaways
- Match the measure to the question; degree for contacts, betweenness for brokerage, PageRank for directed influence.
- Never compare raw scores across networks; normalise and compare ranks.
- Treat betweenness as the measure most sensitive to missing edges.
- Prefer PageRank over eigenvector centrality on directed graphs with dead ends.
- Record graph model, software version and parameters for every analysis.
- Report ranks or quantiles in prose; retain full-precision scores in the data.
- Always read centrality against your sources' survival and collection biases.
Frequently Asked Questions
Which centrality measure should I report first?
Start with degree centrality because it is the easiest to interpret and audit, then add betweenness or eigenvector only if your question is about brokerage or influence. Reporting one well-understood measure beats four you cannot explain.
Does high betweenness mean a person was historically important?
Not necessarily. Betweenness measures structural position in your sampled data, so a high score can simply reflect a survival bias in which letters were kept. Always interpret it against what your sources do and do not record.
Can I compare centrality scores between two different networks?
Raw scores are not comparable across networks of different sizes. Use normalised centrality, or better, compare ranks rather than values, and never state that node A in network 1 is 'more central' than node B in network 2 from raw numbers.
How does missing data affect centrality?
Centrality is highly sensitive to missing edges, and betweenness especially can swing dramatically when a single bridging tie is absent. Document your data's completeness and treat scores as provisional whenever survival is patchy.
What is the difference between eigenvector centrality and PageRank?
Both reward being connected to well-connected nodes, but PageRank handles directed graphs and dangling nodes gracefully, while eigenvector centrality can fail to converge on them. For directed historical networks, PageRank is usually the safer choice.
Should I round or rank centrality scores when reporting?
Report ranks or quantiles for narrative claims and keep full-precision scores in your dataset. Tiny score differences are rarely meaningful given historical data uncertainty, so ranking communicates findings more honestly.