A0236
Title: Random forests and mixed effects random forests for small area estimation of general parameters
Authors: Nora Wuerz - Otto-Friedrich-Universitaet Bamberg (Germany) [presenting]
Patrick Krennmair - Freie Universitaet Berlin (Germany)
Timo Schmid - Otto-Friedrich-Universitaet Bamberg (Germany)
Nikos Tzavidis - University of Southampton (United Kingdom)
Abstract: Random forests are highly effective for prediction due to minimal tuning parameters, automated model selection, and the ability to capture complex relationships. There is notable research on tree-based methods for survey data. More recently, theoretical properties of random forests have been explored for complex survey data. The focus is on random forests and extensions for estimating small area parameters, proposing mixed effects random forests to incorporate random effects crucial for small area estimation. This method extends prior work and uses a non-parametric bootstrap to correct the bias in the estimated residual variance before estimating the variance of the random effects. Estimators of general small area parameters are derived using area-specific smearing. For MSE estimation, a non-parametric block bootstrap with appropriate scaling of the residuals is used. Evaluation includes simulations and real data from poverty assessment in Mozambique, comparing forest-based estimators to industry standard methods like the Empirical Best Predictor. Findings highlight the impact of including random effects, the importance of data transformations, and the performance of estimators.
