Catching Fire via "Likes": Inferring Topic Preferences of Trump Followers on Twitter

Yu Wang, Jiebo Luo, Richard Niemi, Yuncheng Li, Tianran HuarXiv 2016 · arXiv

TL;DR

This paper presents a computational framework to quantify what makes a political tweet "viral" by analyzing the relationship between tweet content and user engagement. By combining Latent Dirichlet Allocation (LDA) with negative binomial regression, the researchers found that for Donald Trump’s followers in late 2015, attacks on Democratic opponents were the single most effective driver of "likes."

What Problem Does This Paper Address?

While much of the existing social media research focuses on retweeting behavior (which measures information diffusion), this study shifts the focus to "likes" as a proxy for follower preference and attractiveness.

The core challenge is that a single tweet is rarely about just one thing; it is a mixture of various themes, mentions, and rhetorical styles. The authors sought to disentangle these mixtures to determine which specific topics actually resonate with a candidate's base. Understanding these preferences provides a window into the "echo chamber" dynamics of political social media and offers a generalizable framework for analyzing other public figures.

Methods

The researchers employed a two-stage pipeline to analyze a corpus of tweets:

  1. Topic Extraction: They used Latent Dirichlet Allocation (LDA) to treat each tweet as a probability distribution over a set of latent topics. To determine the optimal number of topics, they used the "elbow method" on the Mean Absolute Error (MAE) of their subsequent regression, settling on four distinct topics: News Media, Republicans, Democrats, and Trump Campaign.
  2. Engagement Modeling: Because the "like" counts exhibited significant over-dispersion (where the variance far exceeds the mean), standard Poisson regression was unsuitable. Instead, they used Negative Binomial Regression.

The model treated the topic weights as explanatory variables while controlling for external factors that influence engagement, such as:

  • The number of followers at the time of posting (which grew significantly during the study).
  • Whether the tweet was posted on a weekend.
  • Proximity to Democratic or Republican primary debates.
  • The hour of the day the tweet was sent.
Figure 4: Using MAE to find the optimal number of topics for the LDA model.

Datasets

The study utilized a custom-built dataset titled US2016, which tracked Twitter activity for major 2016 presidential candidates.

  • Scope: September 18th to December 27th, 2015.
  • Volume: 2,120 tweets posted by Donald Trump.
  • Features: The dataset includes a time-series tracking of Trump’s follower count (growing from ~4.2M to ~5.5M) and the specific "like" counts for every tweet.
Figure 2: The steady growth of Trump's follower count during the study period.

Key Insights and Ideas

  • The Power of the "Out-Group" Attack: The most striking finding is that attacking the "out-group" (Democrats like Hillary Clinton and Barack Obama) earned significantly more likes than discussing the "in-group" (Republican rivals) or general campaign updates.
  • Debate Dynamics: Tweets posted during or immediately after Democratic debates saw a significant spike in likes. Conversely, tweets during Republican debates actually received fewer likes on average. This suggests that Trump’s followers were most engaged when he acted as a critic of the opposing party's platform.
  • Follower Count as a Baseline: Unsurprisingly, as the follower count grew, the baseline number of likes increased. The model successfully isolated this growth to ensure the "topic preference" results weren't just a byproduct of a larger audience.
  • Weekend Lull: Despite the 24/7 nature of political news, tweets posted on weekends tended to receive fewer likes, suggesting a cyclical nature to political social media engagement.

Key Results

The regression analysis (Table 2 in the paper) provided quantitative backing for these insights:

  • Topic Preference: The "Democrats" topic had the highest positive coefficient ($0.545^{***}$), followed by "Trump Campaign" ($0.084^*$). The "Republicans" topic was not statistically different from the "News Media" baseline.
  • Engagement Spikes: The coefficient for Democratic Debates was $0.286^{***}$, indicating a strong positive correlation with engagement, while Republican Debates had a negative coefficient of $-0.165^{***}$.
  • Visualizing Preference: Figure 5 clearly illustrates that the 95% confidence interval for the "Democrats" topic sits significantly higher than all other categories, marking it as the clear favorite among the follower base.
Figure 5: Estimated topic coefficients showing the clear preference for tweets about Democrats.

Limitations and Future Work

The authors acknowledge several areas for expansion:

  • Demographics and Sentiment: The current model treats "likes" as a monolithic signal. Future work aims to incorporate follower demographics and sentiment analysis to understand why certain topics attract specific subsets of followers.
  • Generalizability: While the framework is designed to be general, this specific paper only applies it to Donald Trump. Testing the model on other candidates (e.g., Hillary Clinton or Bernie Sanders) would reveal if "out-group" attacking is a universal engagement driver or specific to certain political brands.
  • Retweet vs. Like: The study does not deeply compare the drivers of "likes" versus "retweets," which might reveal different psychological motivations (e.g., private agreement vs. public endorsement).

Summary updated

Abstract

In this paper, we propose a framework to infer the topic preferences of Donald Trump's followers on Twitter. We first use latent Dirichlet allocation (LDA) to derive the weighted mixture of topics for each Trump tweet. Then we use negative binomial regression to model the "likes," with the weights of each topic serving as explanatory variables. Our study shows that attacking Democrats such as President Obama and former Secretary of State Hillary Clinton earns Trump the most "likes." Our framework of inference is generalizable to the study of other politicians.
BibTeX
@misc{wang2016catching,
  author = {Yu Wang and Jiebo Luo and Richard Niemi and Yuncheng Li and Tianran Hu},
  title = {{Catching Fire via "Likes": Inferring Topic Preferences of Trump Followers on Twitter}},
  year = {2016},
  eprint = {1603.03099},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1603.03099},
}
Figure