An investigation of tests for linearity and the accuracy of likelihood based inference using random fields

We analyze the random field regression model approach recently suggested by Hamilton (2001, Econometrica, 69, 537--73). We show through extensive simulation studies that although the random field approach is indeed very closely related to the non-parametric spline smoother it seems to offer several advantages over the latter. First, tests for neglected nonlinearity based on Hamilton's random field approach seem to be more powerful than existing test statistics developed within the context of the multivariate spline smoother approach. Second, the convergence properties of the random field approach in limited samples appear to be significantly better than those of the multivariate spline smoother. Finally, when compared to the popular neural network approach the random field approach also performs very well. These results provide strong support for the view of Harvey and Koopman (2000, Econometrics Journal, 3, 84--107) that model-based kernels or splines have a sounder statistical justification than those typically used in non-parametric work. Copyright Royal Economic Society, 2002