A flexible approach to testing the hypothesis of no regression interaction is to test the hypothesis that a generalized additive model provides a good fit to the data, where the components are some type of robust smoother. A practical concern, however, is that there are no published results on how well this approach controls the probability of a Type I error. Simulation results, reported here, indicate that an appropriate choice for the span of the smoother is required so that the actual probability of a Type I error is reasonably close to the nominal level. The technique is illustrated with data dealing with cannabis problems where the usual regression model for interactions provides a poor fit to the data.
Wilcox, Rand R. and Earleywine, Mitchell
"Inferences About Regression Interactions Via A Robust Smoother With An Application To Cannabis Problems,"
Journal of Modern Applied Statistical Methods: Vol. 4
, Article 6.
Available at: http://digitalcommons.wayne.edu/jmasm/vol4/iss1/6