B0297
Title: Robust and flexible model selection for multivariate local linear regression
Authors: Montserrat Guillen - University of Barcelona (Spain)
Jens Perch Nielsen - City, University of London (United Kingdom)
Dimitrios Bagkavos - University of Ioannina (Greece) [presenting]
Abstract: The focus is on the local linear nonparametric regression of a response variable, against an arbitrary number of independent explanatory variables (covariates). To the best knowledge, an introduction is developed to a consistent model selection procedure using an estimate of the mean integrated square error (MISE) of the estimated regression function. The basis of the development is the fact that the inclusion of an irrelevant covariate in the model entails a substantial increase in the model MISE. On the other hand, the inclusion of a relevant covariate results in a reduced model MISE, thus acting as an argument for including the variable in the model. This approach turns out to have an important extra feature that is new to modern model selection as shown that it can pick a different preferred model for each value of the set of independent variables and hence can cherry-pick the best model for each different combination of explanatory factor levels.
