Parsimonious Data: How a single Facebook like predicts voting behaviour in multiparty systems (1704.01143v2)

Abstract: Recently, two influential PNAS papers have shown how our preferences for 'Hello Kitty' and 'Harley Davidson', obtained through Facebook likes, can accurately predict details about our personality, religiosity, political attitude and sexual orientation (Konsinski et al. 2013; Youyou et al 2015). In this paper, we make the claim that though the wide variety of Facebook likes might predict such personal traits, even more accurate and generalizable results can be reached through applying a contexts-specific, parsimonious data strategy. We built this claim by predicting present day voter intention based solely on likes directed toward posts from political actors. Combining the online and offline, we join a subsample of surveyed respondents to their public Facebook activity and apply machine learning classifiers to explore the link between their political liking behaviour and actual voting intention. Through this work, we show how even a single well-chosen Facebook like, can reveal as much about our political voter intention as hundreds of random likes. Further, by including the entire political like history of the respondents, our model reaches prediction accuracies above previous multiparty studies (60-70%). We conclude the paper by discussing how a parsimonious data strategy applied, with some limitations, allow us to generalize our findings to the 1,4 million Danes with at least one political like and even to other political multiparty systems.