A1064
Title: A fast and flexible space-time varying coefficient model selection
Authors: Daisuke Murakami - The Institute of Statistical Mathematics (Japan) [presenting]
Shinichiro Shirota - University of Tokyo (Japan)
Mami Kajita - Singular Perturbations Inc (Japan)
Seiji Kajita - Singular Perturbations Inc (Japan)
Abstract: The space-time varying coefficient (STVC) model attracts attention these days as a flexible tool to explore the spatiotemporal patterns in regression coefficients. However, the model tends to suffer from difficulty in balancing computational efficiency and model flexibility. A fast and flexible STVC modeling method has been developed to break the bottleneck. For flexible modeling, multiple processes are assumed in each varying coefficient, including purely spatial, purely temporal, and space-time interaction processes with/without time cyclicity. While consideration of multiple processes can be time-consuming, a pre-conditioning method is combined, and a model selection procedure inspired by reluctant interaction modelling to select/specify the latent space-time structure computationally efficiently. Monte Carlo experiments show that the proposed method outperforms alternatives in terms of coefficient estimation accuracy and computational efficiency. Finally, the proposed method is applied to a crime analysis with a sample size of 279,360 and confirmed that the proposed method provides reasonably varying coefficient estimates.
