Filtering Survey Respondents with Bayesian Truth Serum: Application to the 2018 US House Elections

09.10.2019 12:00 – 13:15


The Bayesian truth serum (BTS) is a 'detail-free' or 'universal' mechanism for incentivizing honest answers about non-verifiable subject matter, such as private behaviors, opinions or beliefs. Respondents are asked to predict a verifiable quantity — the actual survey results — as well as to provide their personal answers to questions. On the basis of these dual inputs, the BTS formula assigns a score to each respondent. BTS scores have two complementary theoretical properties: First, dishonest or careless answers by respondents should reduce their ex-ante expected scores. Second, for questions where a single best answer exists in principle (such as forecasts), a respondent's actual (ex-post) score is diagnostic of expertise. Therefore, weighting respondents by their scores should improve on democratic averaging of opinions in 'crowd wisdom' settings. I will review these two properties of BTS, and then describe an application to election polls. The data come from a national survey by USC’s Dornsife Center for Economic and Social Research, conducted in three waves before and one wave after the November 2018 United States House of Representatives elections. The survey asked for personal voting intentions, estimates of social circle intentions, and respondents' predictions of the state-level election outcome. The results are generally constent with pre-registered hypotheses about the advantages of using BTS-weighted social circle estimates.


Bâtiment: Uni Mail

Boulevard du Pont-d'Arve 40
1205 Geneva

Room: M 3250, 3rd floor

Organisé par

Faculté d'économie et de management
Institute of Management


Drazen PRELEC, Sloan School, Massachusetts Institute of Technology, USA

entrée libre


Catégorie: Séminaire

Plus d'infos

Contact: missing email