B1200
Title: Bayesian survey design: A new paradigm for social science research
Authors: Seth Flaxman - Imperial College London (United Kingdom) [presenting]
Abstract: Some preliminary work on a Bayesian active learning approach to survey design is described. In many social science fields, the existing gold standard for surveys is a stratified random sampling (i.e. some form of Monte Carlo sampling). Once the data has been collected, regression is used to model a response surface of interest. Inspired by the field of Bayesian experimental design and by recent advances in the machine learning literature on Bayesian Optimization and Bayesian Quadrature, we propose a more efficient alternative to Monte Carlo: ``Bayesian survey design''. Using prior information (e.g. from previous surveys or initial rounds of an ongoing survey), we model the unknown response surface using a Bayesian model (e.g. GLM, Gaussian processes, or Bayesian Additive Regression Trees). Then, conditional on this model, we take an active learning approach to select future respondents so as to optimize some objective function, e.g. we minimize the posterior variance. We present simulation results to demonstrate our approach and preliminary results on a real dataset.