| Popis: |
Political races are often predicted by polls, but such investigations are costly to conduct and the results can be elusive because of their self-report nature. Alternative quantitative modeling approaches are promising to alleviate the prior concerns and therefore, this paper predicted outcomes of the 2020 US Congress elections from computational analyses of language data. Psychological profiles of 918 candidates were created from their social media records (N = 2,836,114 Tweets) and three preregistered dimensions — common words, analytic thinking, and affect — were associated with race outcome. Congressional race winners wrote with common words, communicated in an analytic style, and used more positive affect than congressional race losers. The predictive accuracy of these language dimensions alone was substantial (AUC = 0.844; 79.8% 10-fold cross-validated accuracy, 95% CI [73.2%, 85.4%]), values consistent with or exceeding most polling accuracies. Together, social media language data reveal psychological information about politicians and can indicate race outcomes. |
