On the choice between sample selection and two-part models