Title: Inference from non-probability surveys with XGBoost
Authors: Luis Castro Martin - Universidad de Granada (Spain) [presenting]
Ramon Ferri-Garcia - University of Granada (Spain)
Antonio Arcos - Universidad de Granada (Spain)
Abstract: The importance of online methods for data collection has increased the relevance of non-probability surveys, since they depend on self-selection procedures and suffer from coverage problems. These issues result in biased estimations. Some techniques like Statistical Matching and Propensity Score Adjustment have been proposed to compensate this bias using an auxiliary probability sample. However, they usually rely on logistic or linear regression as the chosen machine learning algorithm. We propose using a state-of-the-art algorithm like XGBoost instead. Simulation studies are conducted to demonstrate how much can XGBoost improve the current results.