Robust Model Selection with LARS Based on S-estimators

摘要: We consider the problem of selecting a parsimonious subset explanatory variables from potentially large collection covariates. are concerned with case when data quality may be unreliable (e.g. there might outliers among observations). When number available covariates is moderately large, fitting all possible subsets not feasible option. Sequential methods like forward or backward selection generally “greedy” and fail to include important predictors these correlated. To avoid this Efron et al. (2004) proposed Least Angle Regression algorithm produce an ordered list (sequencing) according their relevance. introduce outlier robust versions LARS based on S-estimators for regression (Rousseeuw Yohai (1984)). This computationally efficient suitable even exceeds sample size. Simulation studies show that it also presence in compares favourably previous proposals literature.