Cluster-based lack of fit tests for nonlinear regression models

Abstract

Checking the adequacy of a proposed parametric nonlinear regression model is important in order to obtain useful predictions and reliable parameter inferences. Lack of fit is said to exist when the regression function does not adequately describe the mean of the response vector. This dissertation considers asymptotics, implementation and a comparative performance for the likelihood ratio tests suggested by Neill and Miller (2003). These tests use constructed alternative models determined by decomposing the lack of fit space according to clusterings of the observations. Clusterings are selected by a maximum power strategy and a sequence of statistical experiments is developed in the sense of Le Cam. L2 differentiability of the parametric array of probability measures associated with the sequence of experiments is established in this dissertation, leading to local asymptotic normality. Utilizing contiguity, the limit noncentral chi-square distribution under local parameter alternatives is then derived. For implementation purposes, standard linear model projection algorithms are used to approximate the likelihood ratio tests, after using the convexity of a class of fuzzy clusterings to form a smooth alternative model which is necessarily used to approximate the corresponding maximum optimal statistical experiment. It is demonstrated empirically that good power can result by allowing cluster selection to vary according to different points along the expectation surface of the proposed nonlinear regression model. However, in some cases, a single maximum clustering suffices, leading to the development of a Bonferroni adjusted multiple testing procedure. In addition, the maximin clustering based likelihood ratio tests were observed to possess markedly better simulated power than the generalized likelihood ratio test with semiparametric alternative model presented by Ciprian and Ruppert (2004).