English
Español
I will describe a new “Bayesian bootstrap” method for election forecasts that combines traditional polling questions with expectations about their social contacts will vote as well as their predictions of the election outcome. The bootstrap forecast treats each survey respondent’s election winner prediction as an optimal forecast given all evidence available to that respondent. Using linear regression, it infers the amount of independent and shared evidence, and then, in a second step, aggregates the evidence across respondents. In large national samples before the 2018 and 2020 U.S. elections, we find that the (preregistered) bootstrap outperforms forecasts based on any single type of input. The forecast tends to put about 65% weight on social circle predictions, 25% on election winner predictions, and only 10% on own voting intentions. The method may be applied to other categories of social science surveys. Based on joint work with Henrik Olsson and Mirta Galesic (Santa Fe Institute), and Wandi Bruine de Bruin (University of Southern California).