Testing for nonlinearity: the role of surrogate data