This essay reviews some of the undesirable things that can occur when parametric models are used without proper concern for checking that they are at least approximately correctly specified. It shows that nonparametric methods can, to a certain extent, avoid these problems.