Asymmetric Affective Polarization in Elite and Public Discourse: Evidence from the Black Lives Matter Uprising

Author: Sylvia Dou Status: Master's Thesis (2026)

Abstract

Does affective polarization respond symmetrically to a major sociopolitical shock? This study uses the death of George Floyd (May 25, 2020) and the subsequent BLM uprising as a natural experiment to examine how partisan sentiment, moral rhetoric, and destructive language shifted among both political elites (U.S. Congress, Governors, Presidents) and the mass public on Twitter/X. Using an original dataset of 139,064 tweets collected within a ±100-day window, I employ zero-shot stance classification (BART-MNLI) for party identification and construct five text-based measures of affective polarization. Interrupted Time Series (ITS) models — complemented by LOWESS smoothing and Gaussian kernel-weighted local linear regression — reveal striking asymmetries: Democrats and Republicans mobilize different affective registers in response to the same event, and elite discourse amplifies these partisan asymmetries relative to the public.

Research Design

Event shock:       George Floyd death, May 25 2020
Observation window: ±100 days (Feb 14 – Sep 2, 2020)
Unit of analysis:  Daily party-source aggregated tweet measures
Identification:    Interrupted Time Series (OLS) + LOWESS + Kernel RD

Measures

Measure	Method	Reference
Ideology Score P(Dem)	Zero-shot BART-MNLI	Lewis et al. (2020)
Sentiment (VADER)	Lexicon-based compound score	Hutto & Gilbert (2014)
Moral Foundations (eMFD)	Dictionary scoring across 5 foundations	Hopp et al. (2021)
Moral Outrage	Dictionary count (19 words)	Brady et al. (2017)
Destructive Toxicity	Toxic-BERT probability	Hanu & Unitary (2020)

Models

Approach	Specification	Purpose
OLS ITS	Y_t = β₀ + β₁·Time + β₂·Intervention + β₃·TimeAfter + ε	Primary: shock effect (β₂) and slope change (β₃)
LOWESS	Local weighted regression, f = 0.18, ±14-day window	Robustness: relaxes global linearity
Kernel RD	Gaussian kernel WLS, h = 30 days, HC1 SEs	Robustness: RD-style local linear estimates

Repository Structure

.
├── 01_data_collection.ipynb           # Stage 1: API collection, cleaning, merging
├── 02_stance_feature_extraction.ipynb # Stage 2: Stance classification, feature extraction
├── 03_ITS_models.ipynb                # Stage 3: ITS, LOWESS, Kernel RD models
│
├── data/
│   └── processed/                     # Aggregated analysis-ready datasets (see note)
│
├── results/
│   ├── tables/                        # CSV + LaTeX regression tables
│   │   ├── its_results_ols.csv        # OLS ITS coefficients (20 models)
│   │   ├── lowess_model_stats.csv     # LOWESS slope estimates
│   │   ├── kernel_model_stats.csv     # Kernel RD intercepts & slopes
│   │   ├── feature_descriptives.csv   # Descriptive stats by group × party × period
│   │   └── feature_prepost_diffs.csv  # Pre/Post mean differences
│   └── plots/                         # Publication-quality figures (PNG)
│
├── build_paper.py                     # Script to generate paper draft (.docx)
├── ANALYSIS_SUMMARY.md                # Analysis documentation
└── README.md

Key Findings

1. Asymmetric affective channels. Democrats responded to BLM with moral intensification (eMFD shock β = 0.384, p < .01 elite; β = 0.125, p < .001 public) and sentiment improvement, while Republicans responded with ideological consolidation (R² = 0.79) and, uniquely, a gradual escalation in toxicity (slope β = 0.0003, p < .05).

2. Elite amplification. Elite affective shifts were 2–6× larger in magnitude than corresponding public shifts across every measure, consistent with elite-cue theory (Zaller, 1992).

3. Temporal dynamics differ. Elite responses are more shock-like (significant β₂, non-significant β₃), while public responses show both shocks and sustained trend changes — suggesting elites mobilize abruptly while the public adjusts gradually.

Technical Stack

• Data collection: Twitter API via twitterapi.io
• Stance classification: Zero-shot NLI with facebook/bart-large-mnli (Hugging Face Transformers)
• Feature extraction: VADER sentiment, Toxic-BERT, eMFD dictionary scoring
• Modeling: statsmodels (OLS, WLS, LOWESS)
• Visualization: matplotlib, seaborn

Reproducibility

The three numbered notebooks (01_, 02_, 03_) form a sequential pipeline:

1. 01_data_collection.ipynb — Retrieves tweets via API, performs bot removal and victim filtering, merges elite + public corpora into tweets_merged_clean.csv.
2. 02_stance_feature_extraction.ipynb — Runs zero-shot stance inference, validates against elite ground truth (accuracy = 83.7%, macro-F1 = 0.755), computes all affective measures, and generates descriptive tables and figures.
3. 03_ITS_models.ipynb — Estimates 20 OLS ITS regressions (5 measures × 2 sources × 2 parties), fits LOWESS and kernel RD alternatives, and produces all time-series visualizations.

Note: Raw tweet data (data/raw/) is not included in this repository due to Twitter/X Terms of Service. Processed aggregate-level results in results/ are fully included to support reproducibility of the analysis stage.

License

For the purpose of academic discussion. Please do not clone or cite.