Social Group Appeals and Affective Polarization in Western Europe

Will Horne

Clemson University

Alona O. Dolinsky

University of Birmingham

Lena Maria Huber

MZES, University of Mannheim

James Adams

University of California, Davis

2026-09-04

Two accounts of partisan dislike

Parties don’t just take positions: they stand for and against social groups.

  • Policy account: partisans dislike parties that are ideologically distant (Gidron, Adams & Horne 2023)
  • Identity account: dislike is rooted in who partisans think the parties are for (Mason 2018; Huddy et al. 2018)
  • Comparative work has tested the policy account, because positional data exist. The identity account has lacked data.

Our question. When two parties talk about social groups differently, do their supporters dislike each other’s party more, beyond what ideological distance explains?

This paper

Data

PSoGA manifesto appeals + CSES party ratings: 186,194 respondent–party evaluations across 33 elections, eight West European countries, 1996–2021.

Measure

Rhetorical distance: each party-election is a vector of positive and negative appeal shares over 43 groups; distance = how differently two parties talk about people.

Design

Within-voter. Respondent fixed effects: one voter comparing the parties on her own ballot

Preview: rhetorical distance predicts out-party dislike; strongest among appeals directed at cultural groups, particularly negative appeals.

Measurement: Group Appeals Models

Per manifesto sentence:

  1. Groups mentioned  (token classifier)
  2. Stance  positive / negative / neutral
  3. Policy directed at the group?
  4. Category  (43 + Other, multi-label)

Four fine-tuned mDeBERTa-v3 models, trained on 7,589 hand-coded sentences, validated in 8 languages (F1 0.80–0.92).

The PSoGA Database

791

manifestos

139

parties

8

countries

1970–2025

55 years

  • ~304K group mentions — every one with stance, policy flag, and category
  • 71.5% positive · 8.4% negative · 20.1% neutral
  • Manifesto-level shares merge directly with MARPOR, CHES, CSES

Countries: Austria, Denmark, Germany, Ireland, Netherlands, Spain, Switzerland, UK

Dolinsky, Huber & Horne (2026), BJPS. Open on Harvard Dataverse.

Measuring rhetorical distance

  • Party = vector of positive and negative appeal shares per group (% of coded content)
  • Euclidean distance, within election

Robust to alternative measurements like netting rhetoric, cosine similarity, specific groups, length filter.

Design: one voter’s ballot

186,194

out-party ratings

47,764

voters w/ in-party

867

party pairs

33

elections

  • Outcome: CSES 0–10 like–dislike rating of each out-party; in-party = vote choice (robust to party ID)
  • Respondent fixed effects: identification from one voter comparing the parties in her choice set: absorbs demographics, ideology, baseline warmth, country-year specific context
  • Controls: expert-coded left–right and GAL–TAN distance (CHES)
  • SEs two-way clustered (party-election, respondent); manifestos ≥ 50 coded sentences

Distance predicts dislike

Across the 867 estimation pairs, rhetorical distance correlates just r = .18 with expert left–right and .06 with GAL–TAN

Cultural, not economic

Not for lack of variation: parties diverge as much on economic groups. Economic appeal distance is largely absorbed by left–right; cultural distance survives GAL–TAN too: it is not entirely the expert cultural dimension.

Hostility matters most

The Target Matters

Migrants, religious communities, politicians, criminals: out-partisans punish, the base approves: class polarization. Attacks on class adversaries don’t predict polarization.

Who responds? The attentive

Steeper among the educated and politically informed: consistent with a mechanism that runs through voters registering what parties say.

Take-aways

A distinct dimension

Group-appeal distance predicts partisan affect beyond left–right and GAL–TAN: elite conflict that ideological/issue position scales don’t measure.

Boundary Policing

Hostile talk about identity-relevant communities is associated with polarization.

Infrastructure

PSoGA + models avaialable for others to use. PSoGA = Dataverse, Models = HuggingFace or Python Package GroupAppeals

Elites polarize publics not only by moving apart on issues, but by speaking differently about the people who make up the polity.

Thank you.   rwhorne@clemson.edu  ·  PSoGA: Harvard Dataverse  ·  Paper draft available