Nonparametric regression as a general statistical modeling methodology: A Monte Carlo investigation of factors influencing statistical power and robust performance in the presence of moderator variables

摘要

This paper is an introduction to nonparametric regression for research applications in the social sciences. Nonparametric regression is a regression technique, which does not require specification of an a priori functional form. The paper focuses on kernel regression, which is a particular nonparametric regression estimator, for predictive purposes. It is also suggested that the kernel estimator can be used as a test for unspecified nonlinearity in a parametric regression function. A Monte-Carlo study is used to study the performance of kernel regression in prediction and hypothesis testing. Results suggest that kernel regression yielded better out-of-sample prediction of moderated regression models than did a misspecified ordinary least squares regression equation. Results also suggest that a kernel F-test for lack of fit yielded approximately correct Type I error rates with large samples, and strong statistical power, for detecting the existence of an unspecified moderator variable in simulated regression data.