A0945
Title: Nonparametric kernel and spline cluster weighted models
Authors: John Thompson - The University of British Columbia (Canada) [presenting]
Ling Xue - The University Of British Columbia (Canada)
Abstract: Cluster weighted models (CWMs) are a class of finite mixture of regression models that jointly model random covariates and response variables. A challenge in CWMs is specifying the correct parametric functional form of each cluster's data-generating process (DGP), as well as the distribution of covariates. Nonparametric regression estimators, such as those using kernel and spline functions, can be employed in CWMs to estimate unspecified nonlinear regression functions for each cluster. These estimators do not require specification of the functional form of the DGP, but rather satisfy some smoothness and moment conditions. CWMs with nonparametric regression function estimators can be estimated using an expectation-maximization algorithm. Model smoothness during estimation is controlled using a roughness penalty on spline coefficients and cross-validated bandwidth selection for kernel functions. Kernel and spline CWMs are applied to simulated and real datasets with various linear and nonlinear cluster shapes, and their performance is compared to that of state-of-the-art CWM methods.
