Gauging Public Opinion in the Age of Social Media

Can social media data be used to make reasonably accurate estimates of electoral outcomes and public opinion? Given that social media users—particularly more active ones—are not representative of the general population, and that the data they generate is both unstructured and unsolicited, how could such analyses yield reasonably accurate estimates of public opinion? In this meta-review of published research, we examine the three main approaches to social media-based predictions of elections and public opinion: (1) volume-based analysis; (2) sentiment analysis, based on lexicons and machine learning; and (3) network analysis. In comparing the predictive power of these three approaches, we find that network analysis outperforms both volume-based and sentiment analysis, while volume-based analysis outperforms sentiment analysis. Finally, we find that methods which combine network analysis with either volume- or sentiment-based analysis yield the most accurate predictions when benchmarked against voting results or public opinion surveys.