A1220
Title: Hierarchical orthogonal subsampling
Authors: Kalliopi Mylona - King's College London (United Kingdom) [presenting]
Songqiao Han - Kings College London (United Kingdom)
Steven Gilmour - KCL (United Kingdom)
Rundong Zhou - King\'s College London (United Kingdom)
Abstract: Structural multicollinearity caused by categorical covariates and high-order polynomial terms is not addressed by existing orthogonality-based Subsampling techniques for linear mixed models, such as Group-Orthogonal Subsampling. Additionally, these techniques lack a strategic method for choosing groups according to invariant covariates. To overcome these limitations, a unified two-stage framework called Hierarchical Orthogonal Subsampling (HOSS) is proposed. HOSS consistently provides superior estimation accuracy while successfully identifying important interactions and high-order effects that are frequently overlooked by traditional Subsampling techniques, as demonstrated by extensive simulations and a real-data application to London Airbnb pricing.
