Responding to rating questions with inaccuracy: a multivariate mixture model