Political memes and influence in the Global South: Calibrated prediction of misinformation and opinion change

Le résumé fourni par la source

Memes have become a routine infrastructure of political talk across platforms, blending humor, affect, and compressed commentary. Yet scholarship often alternates between strong causal claims and descriptive accounts, leaving open what can be predicted—reliably and with well-calibrated probabilities—from lightweight survey signals. We compare an interpretable penalized logistic regression (ridge) and a flexible random forest to model three self-reported outcomes among Salvadoran university students (cross-sectional survey; n ≈ 162): perceived misinformation potential of memes, perceived manipulation potential, and self-reported opinion change after viewing memes. Predictors were derived from closed and multiple-selection items capturing (a) most-used platforms (TikTok, Instagram, Facebook, and X), (b) common emotions when seeing memes (laughter, reflection, annoyance, and indifference), (c) verification practice (yes/no), and (d) confidence in meme messages (ordinal 0-2). Two operational indicators captured multiple selection (multi_platform and multi_emotion). Models were evaluated with stratified 5-fold cross-validation; classification thresholds were selected on training folds only. We report discrimination (AUC-ROC), threshold-dependent metrics (balanced accuracy, F1, sensitivity, specificity), probabilistic accuracy (Brier score), and decile-based calibration. Predictive performance was highest for perceived misinformation (AUC-ROC ≈ 0.85-0.86; Brier ≈ 0.10), moderate for perceived manipulation (AUC-ROC ≈ 0.74-0.75; Brier ≈ 0.17), and lowest for opinion change (AUC-ROC ≈ 0.70-0.74; Brier ≈ 0.21-0.23). Across models, multi_platform selection and confidence emerged as recurrent signals. Findings are presented as predictive rather than causal and motivate richer measures of exposure and context to model opinion change.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.