Psephological investigations: tweets, votes, and unknown unknowns in the republican nomination process

This paper analyzes the utility of using information contained within Twitter posts in predicting electoral outcomes. Particularly, we are interested in patterns in Twitter communications that can help explain differences between published opinion polls and the actual vote. We consider three categories of models. The first is a mentions model that examines the correspondence between the prevalence of communications about a candidate and electoral outcomes. The second series of models treat Twitter similar to a prediction market, aggregating not candidate preferences but predictions of the electoral result. Last, we consider whether the rediffusion of tweets about a candidate is a reliable predictor of the candidate's performance. The results find inconsistent support for the predictive value of Twitter mentions as an estimate of the overall vote, but these communications provide some evidence of otherwise undetected shifts in momentum with respect to the aggregated predictions of candidate performance and message rediffusion via retweets containing information about a particular candidate. Given the nature of the information extractable, these data are most sensitive to detecting changes in momentum.