B1203
Title: Robust switching models and fuzzy clustering: A robust approach based on trimming
Authors: Francesco Dotto - Sapienza - University of Rome (Italy) [presenting]
Alessio Farcomeni - Sapienza - University of Rome (Italy)
Luis Angel Garcia-Escudero - Universidad de Valladolid (Spain)
Agustin Mayo-Iscar - Universidad de Valladolid (Spain)
Abstract: We propose a robust switching regression model based on fuzzy clustering. We efficiently estimate the underlying linear relationships between a dependent variable and a set of predictors in each cluster. Estimation is based on an iterative trimming procedure. During each iteration we trim the observations with largest current regression error, and the remaining are used to estimate the parameters of the linear model within each cluster. Fuzzy weights computation is based on the current cluster-specific residuals. A constraint on the relative variability of the residuals protects against the occurrence of spurious maximizers of the objective function. The procedure is robust to Tukey-Huber contamination and the use of fuzzy weights allows to recover the underlying linear structures even with moderate cluster overlap.
