Willingness to Say? Optimal Survey Design for Prediction

Abstract

Survey design often approximates a prediction problem: the goal is to select instruments that best predict the value of an unobserved construct or a future outcome. We demonstrate how advances in machine learning techniques can help choose among competing instruments. First, we randomly assign respondents to one of four survey instruments to predict a behavior defined by our validation strategy. Next, we assess the optimal instrument in two stages. A machine learning model first predicts the behavior using individual covariates and survey responses. Then, using doubly robust welfare maximization and prediction error from the first stage, we learn the optimal survey method and examine how it varies across education levels.

Reference

Charlotte Cavaillé, Karine Van Der Straeten, and Daniel L. Chen, “Willingness to Say? Optimal Survey Design for Prediction”, TSE Working Paper, n. 23-1424, March 2023.