Data Quality in Bibliometric Network Analysis
A network can be mathematically correct and still be analytically misleading if the underlying entities are wrong. In bibliometric collaboration analysis, small metadata problems propagate quickly: one duplicated institution becomes two nodes, one missed alias splits its edges, and one duplicated publication inflates several relationships at once.
The highest-impact quality risks
| Risk | Network effect | Control |
|---|---|---|
| Affiliation variants | One institution appears as multiple nodes | Canonical affiliation dictionary |
| Incorrect merging | Distinct organizations collapse into one node | Manual review of ambiguous matches |
| Duplicate publications | Edge weights are inflated | Stable document identifiers and deduplication |
| Missing affiliations | Edges disappear | Completeness checks and exclusion reporting |
| Consortium papers | Dense bursts of edges | Full vs fractional counting decision |
| Changing organization names | Artificial temporal breaks | Historical identity rules |
Affiliation normalization is not simple string matching
Institutional names can vary by language, abbreviation, punctuation and organizational hierarchy. Exact matching misses legitimate aliases, while aggressive fuzzy matching can merge different institutions.
A robust workflow combines automated candidate generation with explicit review rules. Each normalized node should have a canonical identifier, preferred display name and list of accepted source variants. Ambiguous cases should be logged rather than silently guessed.
Decide what an institution means
Research organizations are hierarchical. A paper may list a university, faculty, department and institute in the same affiliation. Another paper from the same research group may list only the institute.
Before normalization, decide the analytical level:
- Parent organization level for broad institutional comparisons;
- Institute level for a research-organization network;
- Department or laboratory level for fine-grained internal analysis.
Mixing levels creates nodes that are not comparable.
Watch the consortium effect
A publication involving 50 institutions generates 1,225 undirected institution pairs under full counting. One consortium paper can therefore create a dense block of edges that visually dominates smaller-team collaboration.
There is no universal rule for handling this. Options include:
- full counting, where every pair receives one unit;
- fractional counting, where the contribution of a paper is distributed across its institutional pairs;
- separate analysis of large-consortium documents;
- a sensitivity analysis comparing networks with and without them.
The chosen method should be reported because it affects both edge weights and centrality.
Time creates identity problems
Institutions merge, split, move and change names. A ten-year analysis must decide whether historically different names represent one continuing node or distinct organizational periods. The correct answer depends on the question.
| Question | Preferred treatment |
|---|---|
| Long-term organizational lineage | Map historical names to a continuing identity where justified |
| Structure at a specific historical moment | Preserve the name and organization active in that period |
| Before/after restructuring analysis | Model old and new entities explicitly |
Build quality checks into the pipeline
Do not wait until the final visualization to notice suspicious patterns. Useful automated checks include:
- top affiliation variants by frequency;
- nodes with near-identical labels;
- edges with extreme weights;
- documents with unusually large institution counts;
- records without a mapped institution or country;
- sudden year-to-year jumps in node volume;
- changes in the number of connected components after cleaning.
Validate with domain knowledge
Network statistics can identify anomalies, but domain knowledge is often needed to interpret them. A very strong edge may be a data error, or it may reflect a legitimate joint laboratory. A cluster of nearly identical institute names may indicate failed normalization, or several genuinely separate units in the same research campus.
Quality control is analytical work
Data cleaning is sometimes treated as preparation for the “real” analysis. In bibliometric networks, it is part of the analysis because entity resolution directly determines the topology of the graph.
A transparent network project should make normalization decisions, counting rules and exclusions inspectable. That is what allows readers to trust not only the visualization, but the structure it claims to reveal.