Coalition Structure and Polarization in Parliamentary Voting Networks: Evidence from the Italian Parliament

Ensuring legislative accountability in multi-party systems requires quantitative tools that reveal actual voting behavior beyond formal party affiliations. We present a network-based framework for analyzing parliamentary dynamics at multiple scales, capturing coalition structure, group coherence, and individual influence. Applied to over 4 million vote expressions from the Italian Parliament across three government formations (2018-2021), the methodology combines network modularity, voting distance metrics, and betweenness centrality to map the structure of collective decision-making. Using this framework, we show that system-level polarization, as captured by network modularity, varies systematically with coalition structure rather than coalition size. Technical governments display paradoxically lower global polarization despite broader formal support, reflecting structurally mixed voting patterns rather than unified blocs. On polarizing issues such as immigration, network polarization depends strongly on the fragmentation or cohesion of the opposition, even when the governing coalition votes cohesively. Betweenness analysis reveals that mediator roles are highly concentrated, with only about 2% of parliamentarians acting as structural bridges between communities. Community detection further uncovers implicit coalitions that are not apparent from declared alliances. The framework relies exclusively on public roll-call data, enabling reproducible analysis and direct applicability to any legislature with transparent voting records. By linking individual voting behavior to emergent system-level structure, the methodology provides quantitative infrastructure for comparative analysis of legislative voting networks and coalition monitoring, enabling systematic assessment of legislative behaviour.

💡 Research Summary

The paper introduces a comprehensive, reproducible network‑based framework for analysing parliamentary voting behaviour at three hierarchical levels – coalition structure, party cohesion, and individual influence – using only publicly available roll‑call data. The authors apply the framework to the Italian Chamber of Deputies and Senate during the 18th legislature (March 2018 – September 2021), a period that saw three distinct government formations: the right‑populist Conte I coalition, the left‑populist Conte II coalition, and the technocratic “grand coalition” led by Mario Draghi.

Data collection was performed through automated SPARQL queries to the Italian Parliament’s open‑data endpoints, yielding more than four million individual vote records (yes, no, abstain, absent) together with MP identifiers, party affiliations, and timestamps. For each government, the authors isolate the first 50 days after cabinet formation – a window that captures coalition stabilisation while keeping sample sizes comparable – resulting in roughly 160 roll‑call votes per chamber per government.

From these records the authors construct weighted, undirected co‑voting networks: nodes represent MPs, and the weight of an edge equals the number of roll‑calls on which the two MPs cast the same vote. The weight matrix is normalised to lie between 0 and 1, and basic network statistics (density, weighted degree) are reported.

Three complementary analytical tools are then applied:

1. Modularity‑based community detection – Using the greedy optimisation algorithm implemented in NetworkX, the authors maximise Newman‑Girvan modularity (Q) to uncover latent voting blocs. Modularity is interpreted as a system‑wide measure of polarization, not merely a proxy for government‑opposition split. Results show that the two partisan coalitions (Conte I, Q≈0.42; Conte II, Q≈0.38) exhibit substantially higher polarization than the technocratic Draghi government (Q≈0.27). This “coalition‑structure paradox” indicates that broader formal support does not automatically translate into a more unified voting bloc; instead, parties retain distinct identities and engage in selective defection.

2. Voting‑distance metrics – Votes are encoded numerically (yes = 1, abstain/absent = 0, no = ‑1). The absolute difference between two MPs’ encoded votes defines a pairwise distance; averaging across all roll‑calls yields a normalized distance matrix ranging from 0 (identical voting patterns) to 1 (maximally opposite). From this matrix the authors compute intra‑group cohesion (average within‑party distance) and inter‑group separation (average between‑party distance). In the Conte I period, the average inter‑group distance between government and opposition is 0.68, whereas under Draghi it drops to 0.51, confirming that the technocratic cabinet’s policy stance is comparatively moderate.

3. Betweenness centrality – For each MP, the standard Freeman betweenness (proportion of shortest paths that pass through the node) is calculated on the weighted network. Only about 2 % of MPs (≈ 30 individuals) display statistically significant betweenness values. These high‑betweenness MPs are disproportionately “party‑switchers” and long‑serving legislators, acting as structural bridges between otherwise separate voting communities. During the Draghi government, the bridge‑players frequently vote with both coalition and opposition, highlighting their role as mediators in a fragmented legislative environment.

The authors also perform a focused case study on immigration‑related bills, demonstrating that issue‑specific polarization is driven more by the fragmentation of the opposition than by the internal cohesion of the governing coalition.

Methodologically, the study emphasizes reproducibility: all code (Python, SQLAlchemy, NetworkX) and a snapshot of the relational database are deposited in an open repository, and the modularity optimisation is repeated with multiple random seeds to confirm stability of community assignments. The authors argue that the framework is readily transferable to any parliament that publishes roll‑call data, requiring only minor adjustments to the SPARQL queries and party‑affiliation tables.

Limitations acknowledged include the treatment of abstentions and absences as a single “neutral” category, and the focus on the first 50 days after cabinet formation, which may not capture longer‑term dynamics. The paper suggests future extensions that combine textual analysis of bill content with the voting‑distance approach to enrich the semantic interpretation of the identified clusters.

In sum, the paper delivers a robust, multi‑scale analytical pipeline that links individual voting actions to emergent system‑level structures, revealing how coalition composition, opposition fragmentation, and a small elite of bridge‑building legislators shape legislative polarization. The work contributes both a methodological toolkit for comparative legislative studies and substantive insights into the functioning of multi‑party parliamentary systems.