On Marginally Correct Approximations of Dempster-Shafer Belief Functions from Data