Median-based regression methods in wood science applications

Abstract

Regression analyses using the traditional, mean-based, least squares (LS) and the median-based, least absolute deviations method (LAD) were performed on twenty-five sets of bending strength and stiffness data. Comparisons of regression models were made to determine the effect of mathematical technique. In all cases, model coefficients demonstrated some degree of sensitivity to regression method. In cases where the residual data appeared to approximate a normal distribution, the models produced fairly similar results. However, the LS technique was found to be highly influenced by even a single extreme residual value. LAD regression proved to be far more stable in the presence of extreme residual values due to its more rational distance function. © 1993 Springer-Verlag.