Using the president's tweets to understand political diversion in the age of social media

The paper investigates whether President Trump used Twitter strategically to divert attention from politically harmful news, specifically coverage of the Mueller investigation, and whether such diversion influenced subsequent media agendas. The study situates this question in the broader context of agenda-setting theory, where mass media traditionally shape public discourse. With social media, politicians can bypass traditional gatekeepers and potentially impact the news cycle directly. The authors hypothesize a two-step process: (1) increases in threatening media coverage (Mueller/Russia) are followed by increases in diversionary tweeting by Trump on unrelated preferred topics; and (2) those diversionary tweets are followed by reduced subsequent media coverage of the threatening topic. The importance lies in understanding agenda dynamics in modern democracies where social media may allow political actors to influence mainstream media coverage.

The paper reviews evidence that media can drive public agendas and political outcomes, including studies linking news coverage to subsequent terrorist attacks, partisan voting shifts due to outlets like Fox News, and the electoral effects of newspaper endorsements or boycotts. It also notes intermedia dynamics and the role of social media in agenda building, misinformation dissemination, and the responsiveness of partisan media to online narratives. Prior research on Trump’s Twitter use has focused on content and engagement, with mixed findings on Twitter’s agenda-building power, but suggests Trump garners disproportionate attention. The authors position their work within this literature by focusing on diversionary agenda-setting, distinguishing it from censorship and deflection, and connecting to theories such as the diversionary theory of war and rhetorical strategies like pre-emptive framing.

Design: Observational time-series analysis over 731 days from January 20, 2017, to January 20, 2019. Data sources: (1) All tweets from @realDonaldTrump (excluding tweets containing only links or empty after filtering). (2) Full-text New York Times (NYT) articles via the NYT archive. (3) Headlines from ABC World News Tonight via the Vanderbilt Television News Archive. Threatening media coverage measure: Daily counts of items containing the keywords “Russia” OR “Mueller” in NYT and ABC (treated separately and as an average of standardized counts). Diversionary tweeting measures: - Targeted analysis: Daily counts of Trump tweets containing any of the keywords “China”, “jobs”, or “immigration” (alone or combined), chosen a priori as preferred topics. - Expanded analysis: All unordered pairs of words from Trump’s Twitter vocabulary (words occurring ≥150 times, excluding stopwords, numbers, URLs, and excluding “Russia”, “Mueller”, “collusion”), yielding N=55 words and 1,485 unique pairs. The presence of a pair in a day’s tweets serves as the diversionary signal for that pair. Modeling strategy: - OLS approach with two equations estimated independently: (a) Diversion equation: number of diversionary terms in tweets on day t regressed on threatening media coverage on day t. (b) Suppression equation: threatening media coverage on day t regressed on number of diversionary terms in tweets on day t−1, capturing a news-cycle lag. Both equations include substantial controls. - 3SLS (three-stage least squares) simultaneous equations modeling both paths jointly. Two temporal specifications for suppression: (A) today’s coverage as a function of yesterday’s tweets; (B) tomorrow’s coverage as a function of today’s tweets. Controls and autocorrelation handling: All models include 104 week fixed effects (one intercept per week), linear and quadratic time trends (date and date^2), and appropriate lagged dependent variables to model autocorrelation: tweets k=10; NYT k=28; ABC k=21; average k=28. Robust inference uses heteroskedasticity-consistent and Newey–West standard errors. Sensitivity analysis following Cinelli & Hazlett quantifies robustness to omitted-variable bias for key models. Placebo and robustness checks: (1) Placebo threatening topic: NYT coverage of “Brexit” (ABC too sparse), excluding articles mentioning Russia/Mueller to avoid contamination. (2) Additional placebo NYT topics: “economy”, “football”, “gardening”, “skiing”, “vegetarian” (with analogous ABC checks where counts permitted). (3) Randomization test for the expanded analysis by repeatedly randomizing the tweet timeline to generate a null distribution of t-values (visualized via 2D kernel density contours). (4) Negative binomial models explored for count outcomes (reported in Supplementary) but with convergence issues for some specifications; OLS reported as primary. Software: R and Stata (including Stata’s reg3 for 3SLS, called from R).

Targeted analysis (keywords China/Jobs/Immigration): - Diversion path (media → tweets, same day): Increased Mueller/Russia coverage is associated with more diversionary tweets. • NYT: coefficient ≈ 0.012 (SE 0.004), p=0.005 (Newey–West p=0.007). • ABC: coefficient ≈ 0.213 (SE 0.084), p=0.012 (NW p=0.006). • Average (standardized): coefficient ≈ 0.226 (SE 0.067), p=0.001 (NW p<0.001). Interpretation: Each additional ABC headline mentioning Russia/Mueller is associated with about +0.2 additional mentions of diversionary keywords in tweets that day. - Suppression path (tweets → next-day media): More diversionary tweets are followed by less Mueller/Russia coverage next day. • NYT: coefficient ≈ −0.403 (SE 0.207), p≈0.052 (NW p≈0.066; borderline). • ABC: coefficient ≈ −0.038 (SE 0.020), p≈0.052 (NW p≈0.061; borderline). • Average: coefficient ≈ −0.050 (SE 0.020), p=0.016 (NW p=0.024). Interpretation: Each additional diversionary keyword mention is associated with roughly 0.5 fewer instances of “Russia” or “Mueller” in the next day’s NYT. - Sensitivity analysis indicates omitted variables would need substantial associations to overturn diversion results; suppression is weaker but consistent in sign. 3SLS analysis (simultaneous estimation): - Diversion coefficients are consistently positive and often significant across specifications. • Panel A (yesterday’s tweets → today’s coverage): NYT→tweets 0.018 (p=0.007); ABC→tweets 0.225 (p=0.236, n.s.); Average→tweets 0.293 (p=0.009). • Panel B (today’s tweets → tomorrow’s coverage): NYT→tweets 0.012 (p=0.009); ABC→tweets 0.220 (p=0.001); Average→tweets 0.228 (p<0.001). - Suppression coefficients are uniformly negative but vary in significance. • Panel A: NYT −0.399 (p=0.054); ABC −0.038 (p=0.037); Average −0.050 (p=0.010). • Panel B: NYT −0.732 (p=0.238, n.s.); ABC −0.090 (p=0.095); Average −0.094 (p=0.099). Expanded analysis (all word pairs): - Across OLS and both 3SLS variants, a notable share of word pairs fall in the bottom-right quadrant (significant diversion with significant next-day suppression), exceeding patterns expected under randomized tweet timelines. - The pattern is similar for NYT and ABC and for their average, indicating synchronicity across outlets. Placebo analyses: - Brexit (NYT): No significant diversion (0.028, p=0.174) and no suppression (0.010, p=0.858). SUR comparisons confirm Russia-Mueller coefficient differs from zero while Brexit does not. - Other placebo topics (economy, football, gardening, skiing, vegetarian): Virtually no word pairs show the diversion-plus-suppression signature (≈0.03% across 7,425 OLS-tested pairs).

Findings are consistent with a strategic diversion mechanism: when faced with increased Mueller/Russia coverage, President Trump tweets more about unrelated preferred topics, and this behavior is followed by reduced subsequent coverage of the threatening topic in leading U.S. media (NYT, ABC). The results hold under extensive controls for time trends, weekly fixed effects, and autocorrelation, and they generalize beyond preselected keywords to broader elements of Trump’s Twitter vocabulary. Placebo topics that are not politically threatening (e.g., Brexit, hobbies) do not exhibit the diversion-plus-suppression pattern, supporting specificity to threatening coverage. While the evidence is correlational and cannot definitively establish causality or intentionality, the joint pattern across OLS and 3SLS, together with randomization and placebo analyses, suggests a meaningful interaction between presidential tweeting and mainstream media agendas. The authors discuss implications for editorial processes, noting that suppression effects appear even in outlets publicly committed to resisting presidential framing, implying unintentional contextual influences on coverage. The results are linked to broader theories of diversion in politics and to rhetorical strategies such as deflection and pre-emptive framing.

The study provides the first empirical test indicating that President Trump’s tweets are associated with diversion from and subsequent suppression of media coverage on politically threatening topics (Mueller/Russia) in leading U.S. outlets. The effect appears robust across modeling approaches, keyword choices, and robustness checks, and is absent for placebo topics. The findings highlight the power of social media to influence mainstream media agendas and raise challenges for journalistic practice in the face of rapid, direct presidential communication. Future research should examine causality more directly, extend analyses to additional media outlets and platforms, assess generalizability to other leaders and contexts, and explore mechanisms (e.g., newsroom dynamics, audience attention) that mediate diversion and suppression.

- Observational design limits causal inference; endogeneity from omitted variables cannot be ruled out despite sensitivity analyses and extensive controls. - Results are based on two outlets (NYT and ABC); patterns may differ in other media ecosystems. - Suppression effects, while consistently negative, are sometimes only marginally statistically significant and less robust than diversion effects. - Multiple testing across many word pairs was not adjusted in the main analyses. - Negative binomial models faced convergence issues; OLS results are emphasized. - Intentionality cannot be inferred—whether diversion is strategic or intuitive remains unknown. - Keyword-based measures may miss nuances in topic framing beyond selected terms.