Predictive model in the presence of missing data: the centroid criterion for variable selection