Bibliometric analysis with VOSviewer and Biblioshiny

What is bibliometric analysis, and how is it different from a systematic review?

A bibliometric analysis treats papers as data points. Performance analysis counts things: publications per year, the most productive authors and journals, the most cited papers. Science mapping looks at relationships: who collaborates, which references are cited together, which keywords co-occur, and how those clusters change over time.

A systematic review starts from a focused question, screens studies against criteria, then extracts, appraises and synthesises what they found. Our systematic review guide covers that method. Put simply, a systematic review tells you what the evidence says; a bibliometric analysis tells you how the field is organised. It cannot tell you whether any finding is right.

So the strongest bibliometric papers add a reading of the content: once the maps show four keyword clusters, you read the central papers in each and explain what that strand argues. Our guide to writing a review article compares this with other review types. Bibliometrics suits topics with hundreds or thousands of papers. If yours has sixty, read them all and write a structured review instead.

How do you write and record a search string for Scopus or Web of Science?

A reviewer who doubts the search will doubt every map, so build it carefully.

1. Split the topic into concepts, for example “fintech” and “small firms”.
2. List synonyms for each, with abbreviations, both spellings, and the wildcard * for word endings. Join synonyms with OR and concepts with AND.
3. Choose the fields. In Scopus, TITLE-ABS-KEY searches titles, abstracts and keywords, and DOCTYPE, LANGUAGE and PUBYEAR restrict type, language and years. In Web of Science, the TS= tag searches the topic fields.
4. Test it against papers you know. List ten central papers. If the search misses one, find the term it uses and add it.
5. Record the exact string, database, date and number of results on the day. Databases change daily, so a search is only reproducible with its date.

A Scopus string for the example might be:

TITLE-ABS-KEY ( ( fintech OR "financial technolog*" OR "digital payment*" ) AND ( sme* OR msme* OR "small business*" OR "small firm*" ) ) AND DOCTYPE ( ar OR re ) AND LANGUAGE ( english )

In Web of Science, use TS= for the same terms and set document type and language in the refine panel. The Scopus Support Center’s advanced search page lists the field codes, proximity operators such as W/3, and the order in which operators apply. For choosing between the two databases, see our comparison of Scopus and Web of Science.

How do you export, merge and clean Scopus and Web of Science data?

Export every field, including cited references; without them you cannot run citation, coupling or co-citation analysis.

Limits change. The VOSviewer manual of October 2023 quotes 500 records for Web of Science and 2,000 for Scopus, lower than the vendors’ 2026 help pages, so check the vendor page on the day rather than a tutorial. Above the limit, export in batches (1–1,000, 1,001–2,000 and so on) and name each file with its date.

Merging and removing duplicates

VOSviewer reads several files from one database and removes duplicates itself, but it cannot combine different databases. To merge Scopus and Web of Science, use bibliometrix in R: import each set with convert2df() (for example dbsource = "scopus", format = "csv" and dbsource = "wos", format = "plaintext"), then combine them with mergeDbSources(), which merges the collections and removes duplicate records. Note the counts from each database and the duplicates removed.

The two databases format cited references differently. Before running co-citation on merged data, check that a few well-known references appear once, not twice in slightly different forms. If they don’t, run the reference-based analyses on one database and say why.

Screening and cleaning

Read titles, and abstracts where unclear, and remove off-topic records. Then clean names: “Sharma, R.” and “Sharma, Rakesh” may be one author; “SME”, “SMEs” and “small and medium enterprises” one idea. VOSviewer uses a thesaurus file for this: a tab-separated text file with two columns headed label and replace by. Each line replaces one label with another; leave “replace by” empty and the term is ignored, which is how you drop words like “study”. Keep the file for your supplementary material.

How do you run co-authorship, co-citation, coupling and keyword maps in VOSviewer?

Choose Create, then “Create a map based on bibliographic data”, then “Read data from bibliographic database files”, and select your files. The wizard asks for the type of analysis and the unit.

Coupling and co-citation are often confused. Bibliographic coupling links two papers that cite the same work, grouping recent papers by shared foundations. Co-citation links two works cited together by later papers, showing the classics the field rests on.

Counting method and thresholds

Full counting is the default. With fractional counting, a paper with many authors contributes less to each link, so a few large-team papers don’t dominate a co-authorship map. Say which you used.

The wizard then asks for a minimum, such as keyword occurrences or documents per author. That threshold decides what appears on the map, so treat it as a method decision. Report how many items met it and whether the main clusters held at a lower and a higher value: “clusters were stable at thresholds of 3, 5 and 10; we report 5, which retained 84 of 2,316 keywords.”

Reading the maps

Circle size shows weight, colour shows the cluster, and distance roughly shows relatedness. The overlay visualisation colours items by a score such as average publication year, which shows which topics are newer. VOSviewer only numbers the clusters; naming them from their items is your job.

How do you use Biblioshiny for Lotka’s law, Bradford’s law and thematic maps?

Install R, run install.packages("bibliometrix") once, then library(bibliometrix) and biblioshiny() each session. The app opens in your browser. Load your export or merged file and apply the filters you reported for the search.

Lotka’s and Bradford’s laws

Lotka’s law describes author productivity: in Lotka’s 1926 formulation, the number of authors with n papers is roughly 1/n² of the number with one, so most authors contribute a single paper. Bradford’s law describes how papers concentrate in journals: sources fall into zones that each hold about a third of the papers, and the first, core zone has only a few journals. Biblioshiny plots both. Report them as descriptions of your dataset; a table of top authors is context, not a contribution.

Thematic map and three-field plot

The thematic map places keyword clusters on two axes: centrality (links to other themes) and density (internal development). The quadrants are read as motor themes (high on both), niche themes (dense but peripheral), emerging or declining themes (low on both) and basic themes (central, less developed). Only a reading of recent papers tells you whether a lower-left theme is emerging or fading. The three-field plot links three fields, such as authors, keywords and sources; keep each to ten or fifteen items or it becomes unreadable.

Cite the software: Aria and Cuccurullo (2017), Journal of Informetrics 11(4), 959–975, for bibliometrix, and van Eck and Waltman (2010), Scientometrics 84(2), for VOSviewer.

Most scholars use VOSviewer for network figures and Biblioshiny for statistics and thematic maps. CiteSpace earns its learning curve when your question is how a field changed over time.

How should you report a bibliometric analysis in a paper or thesis?

Write the method so that someone could rerun it and get the same dataset.

A PRISMA-style flow diagram. PRISMA 2020 was written for systematic reviews, but its flow diagram is the clearest way to show your records. Adapt the “databases and registers” template: records from each database, duplicates removed, records screened, records excluded with reasons, and the final dataset. Say in the text that the diagram is adapted.

A table of search parameters, like this one:

State your research questions before the results, and organise the results by question rather than by software menu.

Why do reviewers reject bibliometric analysis papers?

Editors see a steady stream of bibliometric submissions, and a paper that reads like unedited software output is easy to return. These objections come up again and again:

• Maps with no interpretation. Each figure needs a paragraph on what the clusters mean and which papers sit at their centre.
• Thresholds with no rationale. “Minimum 5 occurrences” needs a reason and a check that the clusters hold at other values.
• One database, no justification. If you used only Scopus, say why, and what you might have missed.
• Descriptive only. Papers per year, top countries and top journals, with no research questions and no conclusion about the field.
• No screening. A keyword search always pulls in off-topic records; unscreened data gives noisy maps.

The remedy is mostly the same. Set two to four research questions the maps can answer, interpret every figure against them, read the central papers in each cluster critically (our guide to critiquing an article helps), and end with a research agenda that follows from what you found. If you can’t say what a map shows in one plain sentence, leave it out.