Robus groupwise variable selection.

Variable selection methods for statistical models aim to find a small number of explanatory variables with higher prediction performance and better interpretability than the full model. One example for such a procedure is least angle regression (LARS), which produces a sequence of candidate predictors in the order of their predictive content. Since LARS can be expressed only in terms of correlations, its robust counterpart replaces them by high-breakdown robust correlation estimates. However, LARS and its robust counterpart were designed for cross-sectional numerical variables. For the classical case, extensions of LARS to groupwise variables have recently been proposed in the literature. In the case of time series data, these groups are formed by the variables themselves and their respective lags. For cross-sectional data, the groups can be blocks of binary variables representing factors. In any case, groupwise LARS can be expressed only in terms of R-squared measures from short regressions and correlations. The aim of this project therefore is to robustify groupwise LARS based on robust regression, R-squared and correlation estimates. Various configurations of robust estimates will thereby be investigated and evaluated with respect to the performance of the resulting robust groupwise LARS procedure.

Keywords:Econometrics, Data analysis, Robust statistics

Disciplines:Applied mathematics in specific fields