Elite Cleavages: Concept, Measurement, and a Verifiable AI Search Agent Methodology

John Gerring | Connor Jerzak | Erzen Öncel

[PDF (Preliminary & Incomplete)]

Setting the Stage

How do social distinctions map onto political parties? The field can lack a replicable, continuous, and cross‑national way to track whether elites—legislative party delegations—are integrated across ethnic lines or segregated into parallel partisan pillars.

Our paper reframes elite ethnic cleavages as an alignment problem: To what extent do parties sort MPs by ascriptive identity (e.g., ethnicity, religion, language, region)? Instead of binary labels (“ethnic party” vs “not”), we deliver interpretable scales at three levels:

  • System: overall ethnic cleavage of a legislature
  • Party: degree to which a given party is an ethnic party
  • Group: cohesion of an ethnic group across parties

Together, these let researchers see not only whether a system is segmented, but also where segmentation resides—within particular parties or groups.

What We Did

1) A transport metric of elite cleavage

We first cast cleavages as a discrete optimal‑transport problem with a crisp interpretation:

Our cleavage index equals the minimal share of seats that must switch parties—holding each MP’s identity fixed—to make every party’s delegation mirror the chamber’s overall ethnic mix.

  • The index is bounded in [0,1] and comparable across time, chambers, and countries with different party and group configurations.
  • The same logic yields party‑level and group‑level diagnostics (ethnic party and group cohesion indices), so scholars can zoom from system patterns down to specific delegations.

2) A global, auditable data pipeline for elites

We pair the metric with a global dataset of parliamentary delegations built from expert coding and a verifiable retrieval‑augmented AI search agent:

  • Country‑specific closed codebooks constrain labels; the agent cites sources, archives queries, and abstains when evidence is weak or conflicting.
  • Disagreements and abstentions route to human coders, preserving cross‑national comparability while scaling coverage.
  • In the current snapshot (figures may update as we refresh the corpus):
    • Countries: 153
    • Chamber‑elections: 368 (1990–2022)
    • Coverage gain from the agent on missing party labels: ≈21% new leader‑records (>15,000 leaders)
    • High‑confidence agreement with expert codes: ≈86% on expert matched cases

3) Validation and first findings

Our indices recover well‑known cases from the literature and reveal within‑country and within‑party variation that binary typologies and survey gaps miss. The cross‑national distribution is right‑skewed: most legislatures sit near low segregation, with a long tail of highly segmented systems. Illustratively, Lebanon, Bosnia and Herzegovina, and Malaysia appear near the high‑segregation end; many Western European and East Asian cases cluster near the low end, and some hegemonic systems display low measured cleavage because elite sorting is muted (rather than cross‑ethnic integration being achieved).

Substantively, we find robust support for a simple demand–supply account:

  • Demand: cleavage rises with underlying ethnic heterogeneity.
  • Supply: party fragmentation amplifies (rather than creates) ethnic segmentation.
  • Institutions matter in intuitively plausible ways, but their independent effects are less stable than the demand fundamentals in reduced‑form models.

Why the Results Matter

  • For theory: We provide a replicable baseline for testing classic claims about ethnic politics, without relying on ecological inference. The continuous scales enable comparisons across countries and over time, and the party/group diagnostics let scholars locate the mechanics of segmentation.
  • For policy and institutional design: Because the index is literally the seat‑reassignment work needed to achieve integration, it doubles as a diagnostic for reform.
  • For future research and reproducibility: The provenance‑preserving search‑agent pipeline opens over 700,000 elite records with embedded citations at low marginal cost, lowering barriers for teams outside large, well‑funded centers. Every agent assignment is auditable; we make all of the half a million agent records available open sources.
  • For extensibility: The same transport‑based design and agent workflow carry over to other ascriptive cleavages (religion, language, region) and other elite arenas (cabinets, local councils). The method is flexible to different transport “costs” (e.g., adding geographic frictions) and to multi‑membership identities.

Related Work

John Gerring, Alan Hicken, Connor T. Jerzak, Robert Moser, Erzen Öncel. Where Minorities are the Majority: Electoral Rules and Ethnic Representation. OSF Preprint, 2024.
@article{gerring2024minorities,
  title={Where Minorities are the Majority: Electoral Rules and Ethnic Representation},
  author={Gerring, John and Alan Hicken and Connor T. Jerzak and Robert Moser and Erzen Öncel},
  journal={OSF Preprint},
  year={2024},
  volume={},
  number={},
  pages={}
}

John Gerring, Connor T. Jerzak, Erzen Öncel. The Composition of Descriptive Representation. American Political Science Review, 118(2): 784-801, 2024.
@article{gerring2024composition,
  title={The Composition of Descriptive Representation},
  author={Gerring, John and Connor T. Jerzak and Erzen Öncel},
  journal={American Political Science Review},
  year={2024},
  volume={118},
  number={2},
  pages={784-801}
}
[Overview][Data][Code]