Case study

Network Analysis of Political Corruption in Bosnia

United States Institute of Peace · Transparency International · 2021–2025

Mapping the structural relationships between political parties, state-owned enterprises, and corporate actors in Bosnia and Herzegovina — so anti-corruption work can target structure, not just individuals.

Domain

Governance & anti-corruption

Client

USIP / Transparency International

Timeframe

2021–2025

Methods

Graph data science (NetworkX, SciKit-Network), centrality

Output

Report to the European Council on Foreign Relations

Context

The problem

Corruption in Bosnia and Herzegovina is usually described one scandal at a time — a director here, a contract there. But the durable part of the problem is structural: political parties, state-owned enterprises, and corporate actors are connected through shared directors, financial flows, and political appointments, and those connections persist even when individuals change. The question this project set out to answer: where in that political-economic network does influence actually concentrate?

Force-directed network graph of political and corporate actors — dense clusters of nodes joined by shared-director and financial ties — A force-directed network layout. In graphs like this, the analytically interesting nodes are often not the largest — they are the ones bridging otherwise separate clusters.

The method

I built a custom data pipeline in Python, using NetworkX and SciKit-Network, to clean, link, and analyze records on three kinds of relationships: shared directors, financial flows, and political appointments. Record linkage carries the most analytic risk in work like this — deciding that two records refer to the same person or company is where errors enter — so the pipeline kept cleaning and linking as explicit, inspectable steps rather than burying them.

On the assembled graph I applied standard centrality measures — betweenness, PageRank, and degree — each capturing a different sense of "influence": sitting on paths between others, being pointed to by well-connected actors, and raw connection count. Where the measures agree, confidence in a hub is higher; where they disagree, that disagreement is itself informative.

Illustrative structure — not real entities or data.

hover, tap, or Tab to a node

An illustrative node-link sketch of the three relationship types this method links and scores — shared directorship, financial flow, and political appointment — across three classes of actor. Hover, tap, or keyboard-focus any node to trace only its ties. Edge type is shown by line pattern and node class by shape, so the picture reads without relying on color. It contains no real names and no real financial data: it shows the shape of the problem, not any finding from the engagement.

The result

The analysis surfaced the hubs of influence within Bosnia's political-economic network. The findings were SourceCo-authored report presented to the European Council on Foreign Relations. Public link pending partner sign-off; see Provenance. and are used by in-country anti-corruption partners. The engagement ran through March 2025. Given the subject matter, I do not name individuals or entities from the analysis here.

Provenance

What traces to what:

Data. Records on shared directors, financial flows, and political appointments, cleaned and linked in the pipeline. Every edge in the graph traces back to a source record; linkage decisions are where judgment enters, and they were kept inspectable for exactly that reason.
Code. A Python pipeline built on NetworkX and SciKit-Network. It is not public — the sensitivity of the subject matter argues against publishing the linkage layer.
Publication. A co-authored report presented to the European Council on Foreign Relations.

More selected work Discuss a similar engagement