Examining spread of emotional political content among Democratic and Republican candidates during the 2018 US mid-term elections

The study investigates which specific emotions in political candidates’ tweets influence message diffusion (retweets and favorites) during the 2018 U.S. Senate midterm campaigns and whether effects differ between Democrats and Republicans. With Twitter widely adopted by candidates and integral to modern campaigning, prior work has emphasized valence (positive vs. negative) rather than discrete emotions. The authors aim to move beyond valence to examine how joy, anger, fear, sadness, and confidence expressed by candidates relate to public engagement online and to test for party asymmetries in responses.

Prior research shows negative political content often spreads more and mobilizes more effectively than positive content, particularly on social media, where outrage and moralized language can drive engagement. However, discrete emotions differ in action tendencies: anger tends to motivate approach and action, whereas sadness promotes withdrawal, and fear is linked to avoidance. Evidence on ideological differences is mixed: some motivated social cognition work suggests conservatives are more responsive to negatively valenced stimuli, but other research indicates symmetry or context-dependence. The authors highlight the need to examine specific emotions in naturalistic social media contexts and explore potential party-based asymmetries in responses to anger and fear during campaigns.

Design and data: The authors collected 7,310 original tweets (no retweets) from 65 U.S. Senate candidates (Democrats: 3,711; Republicans: 3,599), including 29 incumbents and 36 significant challengers, across 33 states during the four weeks leading up to Election Day 2018 (09 Oct–06 Nov 2018). Tweets were obtained via Twitter API (Tweepy userTimeline), with direct retweet counts disambiguated from quote retweets using webpage scraping due to API limitations at the time. Metadata included favorite counts, URLs, followers, friends, and timing (Oct–Dec 2018 collection). Race competitiveness was coded from Cook Political Report ratings (0=Safe, 1=Likely, 2=Lean, 3=Tossup).

Pre-processing: Text was cleaned by removing URLs and user mentions, decoding symbols to UTF-8, correcting repeated-character spellings, and retaining hashtag keywords while stripping the “#”. Emoticons/emoji were preserved to capture emotional content.

Emotion detection: IBM Watson Tone Analyzer was applied to each tweet as a whole text block to generate confidence scores for Joy, Anger, Fear, Sadness, and Confidence. Tones were recorded only if the confidence exceeded 0.5. The Tone Analyzer employs an ensemble of machine learning models (e.g., logistic regression, SVM) with n-grams, lexical features, dialog-specific cues, punctuation/emoticon features, and negation handling.

Outcomes: Direct retweet count and favorite (like) count per tweet.

Controls and features: Number of followers, number of friends, incumbency status, race competitiveness; and tweet-level features indicating presence of hashtags, mentions, URLs, and emoji.

Statistical analysis: Linear mixed-effects models (R lme4) predicted log-transformed retweet and favorite counts from the five discrete emotion indicators and controls. Variables were min–max scaled to [0,1]. Random effects accounted for repeated tweets per candidate and temporal structure (week-level factor). Model comparisons using AIC/BIC favored the chosen random-effects structure (e.g., χ² improvements p<0.001). Diagnostics of residuals and random effects were reported satisfactory. Explained variance was estimated via conditional R² and squared correlation between observed and fitted values, with conditional R² ~0.74–0.79.

Descriptive diffusion patterns: Among Democrats’ emotionally charged tweets (n=1,952), approximately 62% of fear-based tweets were retweeted >100 times, followed by anger (58%), sadness (55%), confidence (50%), and joy (33%). For Republicans’ emotionally charged tweets (n=1,765), ~52% of anger-tweets exceeded 100 retweets, with lower shares for confidence (31%), joy (30%), and under 25% for fear or sadness.

Average retweet counts (Table 2): Democrats—Fear: mean 950 (s.e.m. 311.9; 95% CI [316, 1585]); Joy: mean 349 (53.61; [244, 454]); significant differences across emotions χ²(4)=90.54, p<0.001. Republicans—Anger: mean 362 (106.4; [148, 576]); Fear: mean 274 (165.7; [34, 614]); χ²(4)=14.56, p<0.01. Similar patterns for favorites (e.g., Democrats’ fear: mean 3112 vs joy 1389; Republicans’ anger: mean 815 vs joy 698), with significant across-emotion differences for both parties.

Model-based effects (controlling followers, friends, incumbency, competitiveness; random effects for candidate and week):

• Joy negatively predicted engagement for both parties:
  • Democrats: retweets β≈−0.66 (s.e.m. 0.05; 95% CI [−0.75, −0.56]; F(1,3662)=173.1, p<0.001); favorites β≈−0.35 (s.e.m. 0.04; 95% CI [−0.44, −0.25]; F(1,3661)=52.49, p<0.001).
  • Republicans: retweets β≈−0.40 (s.e.m. 0.04; 95% CI [−0.48, −0.29]; F(1,3364)=67.04, p<0.001); favorites β≈−0.18 (s.e.m. 0.04; 95% CI [−0.26, −0.07]; F(1,3404)=12.93, p<0.001).
• Democrats: Fear significantly increased engagement (retweets B≈0.53, s.e.m. 0.22; 95% CI [0.10, 0.98]; F(1,3648)=5.60, p<0.05; favorites positive and significant). Anger was not a significant predictor of retweets (B≈0.27, s.e.m. 0.23; p=0.254).
• Republicans: Anger significantly increased engagement (retweets β≈0.60, s.e.m. 0.18; 95% CI [0.33, 1.06]; F(1,3343)=9.90, p<0.01; favorites β≈0.43, s.e.m. 0.19; 95% CI [0.16, 0.89]; F(1,3381)=5.22, p<0.05). Fear was not significant (retweets β≈0.23, s.e.m. 0.23; p=0.342).
• Confidence and sadness did not significantly predict retweet or favorite counts for either party.
• Tweet features: Mentions, URLs, and hashtags generally had negative or non-significant associations with engagement (e.g., Democrats: mentions β≈−0.32, URLs β≈−0.44; hashtags β≈−0.23; Republicans: hashtags β≈−0.16), all p<0.001 where noted.

Model fit: Conditional R² and r² ranged ~0.744–0.794, indicating 74–80% variance explained by the models and random effects.

The findings show that discrete emotions, not just valence, shape the spread of political messages on Twitter. Positive joy signals reduced diffusion across parties, while negative emotions mattered asymmetrically: fear boosted engagement for Democrats and anger boosted engagement for Republicans. This addresses the research questions by demonstrating that specific emotions differentially predict online engagement and that patterns vary by party. The results align with literature on negativity advantages online but refine it by distinguishing anger (approach-related) from fear (avoidance-related). The party asymmetry may reflect contextual dynamics such as group identity, uncertainty, and power status during 2018, where Democrats (out of presidential power) may have been more responsive to threat/fear appeals, whereas Republicans responded more to anger/approach appeals. These insights contribute to understanding how candidates can gain exposure and how emotional rhetoric may reinforce partisan echo chambers and polarization.

This study analyzes over 7,000 Senate-candidate tweets from the 2018 U.S. midterms using IBM Tone Analyzer and mixed-effects modeling to disentangle the role of discrete emotions in social media diffusion. Key contributions are: (1) demonstrating that joy reduces engagement for both parties; (2) identifying party-specific effects where fear increases engagement for Democrats and anger for Republicans; and (3) showing no robust effects for confidence or sadness. The work highlights the importance of discrete emotions in political communication strategies online. Future research should replicate across election cycles and contexts, examine other platforms (e.g., Facebook), assess how references to news content and high-profile individuals affect engagement, and test how group power status moderates responses to fear and anger appeals.

• Context specificity: Focused on the 2018 U.S. midterm Senate races during the Trump presidency; generalizability to other periods, offices, or countries is uncertain.
• Class imbalance: Relatively few anger- or fear-labeled tweets, potentially amplifying their observed effects; a more balanced dataset would improve inference.
• Platform scope: Twitter-only analysis; patterns may differ on other platforms (e.g., Facebook).
• Measurement constraints: Reliance on automated tone detection (threshold >0.5) and API limitations for disentangling retweet types (addressed via scraping); emotional labeling may be imperfect despite model validation.
• Data sharing: Original tweets cannot be publicly shared due to platform policy, limiting external replication with the exact dataset (sample data and code provided).