The availability of data on social interaction has opened new horizons in the prediction of human behavior. Such data allows viewing the social contacts, or friends, of an individual as data about the individual. We study how connections and interactions on a large online social network predict future electoral turnout. We show that both the aggregate demographics and the past behavior of one's friends are informative per se and improve on models based on individual data. We find that friends' past behavior is less informative than their demographics unless the individual has many friends. Considering all friends or only closer friends in the computation of social network attributes affects prediction accuracy. According to our analyses, the most informative network demographics are computed considering all friends of an individual, whereas the past behavior of the weakest ties is not informative. We find that considering only the strongest friendship ties is usually not preferable due to the low average number of close friends per individual. In addition, close friends appear not more informative than randomly chosen friends.