ИСТИНА

In our talk we consider model in which binary response variable $Y$ depends on predictors $X_1, \dots, X_n$ where $X_j$ takes values in some discrete set for all $j$. The aim is to find all the predictors that essentially influence on response. There are different methods for this task solution. One of them is MDR (multifactor dimensionality reduction) method which was proposed in the paper by M.Ritchie et al. \cite{Ritchie}. A vast set of papers were devoted to generalizations and modifications of MDR method. We extend approach developed in \cite{Bulin} to the case of stratified samples. We modify method introduced in \cite{Dehman} by another hierarchical clustering which allows overlaps between groups of the predictors. Besides, we compare performance of our procedure clustering procedure in combination with different variable selection methods for overlapped groups. This work is partially supported by RFBR grant 13-01-00612.