Harvesting the wisdom of crowds for election predictions using the Bayesian Truth Serum

Social scientists, journalists, politicians, policy makers, businesses, and the general public rely on surveys to understand changing attitudes and behaviors in the society. Prominent among these are election polls. Most election polls ask people for which candidate they will vote (own intention questions). Accuracy of election predictions can be further improved by leveraging the wisdom of crowds: asking people whom they expect will win (election-winner expectations) or whom their friends will vote for (social-circle expectations). Here we investigate a complementary approach that uses the Bayesian Truth Serum (BTS) scoring algorithm and election-winner expectations to improve survey predictions based on own intentions and social-circle expectations. We present a new theorem that gives conditions under which moderate weighting of respondents by their BTS scores should improve survey accuracy even if everyone responds honestly. In a national longitudinal survey (N>4,000) conducted in three waves before the 2018 US House of Representatives election, we compared predictions based on own intentions, state-winner and social-circle expectations, and BTS-scored expectations. Social-circle expectations outperformed own intentions at the national and state levels in predicting the vote-share distribution and the margin between Democrats and Republicans. BTS scores provided additional improvements to predictions of national and state-level vote-share distributions based on social-circle expectations.