Variable Selection in Multivariate Data

A problem of selecting a subset of the original variables based on the likelihood is considered. An appropriate perturbation scheme is attached to the variables and the likelihood displacement is computed that forms a curve in the Euclidean plane. The direction vector associated with the maximum curvature of this curve provides information about a variable that greatly change the likelihood displacement. Variables associated with large absolute components of the direction vector would be included in the model. A numerical example is given for illustration. The simulated data set shows that the variable selection procedure works well. This procedure of choosing variables would be useful for analyzing multivariate data.