Fig 1.
How we formed unsupervised and analytical clusters.
Fig 2.
Number of days each visual frame dominated left-and right-leaning media outlets coverage.
This plot reflects the number of days when each of the visual frames dominated the rhetoric of left- and right-wing media outlets correspondingly. The plot only includes the days of the publication in the range from Aug to Dec 2018 (media coverage of the second large migrant caravan) and only days that had more than 2 publications in total. The ideology of the media outlets is based on the AllSides, combining “left” and “lean left” outlets in the group of left-leaning, and “right” and “lean right” outlets in the group of right-leaning ones. “Center” media outlets are excluded from the sample.
Fig 3.
Media outlets with different ideologies amplify and downplay different visual frames.
This plot illustrates the relationship between two nominal variables: (1) the ideology of media outlets (from very liberal—“left” according to AllSides—to very conservative—“right” according to AllSides) and (2) visual frame. It is based on images attached to tweets (extracted via the Twitter API), excluding those labeled as “Other.” Colors indicate both the direction and strength of the association between outlet ideology and frame, as measured by Pearson standardized residuals: blue shades denote positive associations (more cases than expected under independence), red shades denote negative associations (fewer cases than expected), and gray indicates no meaningful association. The magnitude of each residual reflects how strongly each cell deviates from the values expected under independence. Numerical values displayed in each cell indicate the Pearson standardized residual for that cell. The p-value displayed corresponds to a Chi-square test of independence, which rejects the null hypothesis of no association between these two variables.
Fig 4.
How people react to visual frames within partisan groups.
Each whisker represents the estimated regression coefficient for a visual-frame predictor (factor) from linear models fit separately on Democratic and Republican subsamples, using the ‘violations’ frame as the reference category. Partisanship was self-reported (1 = Democrat; 0 = Republican). Accuracy and attitude were measured on 7-point scales ranging from 1 (image gives a faulty representation/evokes a negative attitude) to 7 (image gives an accurate representation/evokes a positive attitude). Media outlet ideology guess was coded as a binary variable: responses of 1–3 on the question “Do you think this image is from a liberal or conservative outlet?” were classified as left-wing (1), and responses of 5–7 as right-wing (0); moderate responses (4) were excluded. All models report 95% confidence intervals, include random effects (random intercepts for images and respondents), and control for gender, age, ethnicity, education, income, and interest in politics.