On the use of mixed effects random forests in small area estimation
Conference
64th ISI World Statistics Congress - Ottawa, Canada
Format: IPS Abstract
Keywords: poverty, random forest, smallareaestimation
Session: IPS 212 - Advanced inference on mixed effects models for SAE
Thursday 20 July 2 p.m. - 3:40 p.m. (Canada/Eastern)
Abstract
Estimating the spatial concentration of economic phenomena such as poverty and inequality is imperative for evidence-based policies. The use of national sample surveys to obtain reliable estimates for disaggregated geographical and other domains presents methodological challenges. Small Area Estimation (SAE) methods enable the integration of survey, Census, administrative and other data sources (e.g., remote sensing) using models to estimate domain-specific parameters. In this work we propose the use of mixed effects random forests as a flexible and robust method to produce small area estimates. Random forests excel in terms of predictive performance without making explicit model assumptions. Automated model-selection and detecting complex interactions make the use of such algorithms appealing. We place particular emphasis on critically assessing the need to model the dependence structure in the data, for example via the use of random effects, and the need to use data-driven transformations when employing machine learning algorithms. Comparisons with industry standard small area methodologies using real-world data for poverty mapping are presented, and MSE estimation is discussed. The work aims to inform the discussion on the use of machine learning methods to produce official statistics.