A0980
Title: Small area estimation with quantile regression forests
Authors: Nicolas Frink - Otto-Friedrich-University Bamberg (Germany) [presenting]
Abstract: A small area estimation method that employs quantile regression forests to enhance the estimation of disaggregated means in small domains is proposed. The use of a machine learning technique allows predictive, non-linear relationships to be captured directly from the data. The approach utilizes quantile-like predictions of the conditional distribution of the outcome variable given the explanatory variables. Thus, analogous to the M-quantile modelling procedure used to estimate means in small areas, the characterization of domain-specific distinctions is facilitated through the generation of domain-specific predictions, thereby overcoming difficulties associated with the identification of random effects. The performance of the quantile regression forests is compared with that of other small area estimation techniques, including both parametric and semi-parametric approaches, using model-based simulations.
