Toward an objective and reproducible model choice via variable selection deviation.

Various model selection methods can be applied to seek sparse subsets of the covariates to explain the response of interest in bioinformatics. While such methods often offer very helpful predictive performances, their selections of the covariates may be much less trustworthy. Indeed, when the number of covariates is large, the selections can be highly unstable, even under a slight change of the data. This casts a serious doubt on reproducibility of the identified variables. For a sound scientific understanding of the regression relationship, methods need to be developed to find the most important covariates that have higher chance to be confirmed in future studies. Such a method based on variable selection deviation is proposed and evaluated in this work.