Selection of a model in reflectometry: use of the linear statistical inference