Abstract
We introduce a new variable selection method, called SgenoLasso, that handles extreme data. Our method relies on the construction of a specific statistical test, a transformation of the data and by the knowledge of the correlation between regressors. It is appropriate in genomics since once the genetic map has been built, the correlation is perfectly known. This new technique is inspired by stochastic processes arising from statistical genetics. We prove that the signal to noise ratio is largely increased by considering the extremes. Our approach and existing methods are compared on simulated and real data, and the results point to the validity of our approach.