Variable importance assessment in sliced inverse regression for variable selection

Jerome Saracco (1, 2, 3) , Ines Jlassi (4)

1 CQFD - Quality control and dynamic reliability
2 IMB - Institut de Mathématiques de Bordeaux
3 ENSC - Ecole Nationale Supérieure de Cognitique
4 FSM - Faculté des Sciences de Monastir

The focus is on treating the relationship between a dependent variable $y$ and a $p$-dimensional covariate $x$ in a semiparametric regression model. Since the purpose of most social, biological or environmental science research is the explanation, the determination of the importance of the variables is a major concern. It is a way to determine which variables are the most important when predicting $y$. Sliced inverse regression (SIR) methods allows us to reduce the space of the covariate $x$ by estimating the directions that form an effective dimension reduction (EDR) space. The aim is to propose a computational method based on importance variable measure (only relying on the EDR space) in order to select the most useful variables in SIR model. The numerical behavior of this approach, implemented in R, is studied on a simulation study. An illustration on a real data is also provided.

Domaines: Statistiques [stat]

Jerome Saracco, Ines Jlassi. Variable importance assessment in sliced inverse regression for variable selection. 9th International Conference of the ERCIM WG on Computational and Methodological Statistics (CMStatistics 2016), Dec 2016, Séville, Spain. ⟨hal-01417436⟩