Level of Significance Selection in Engineering Analysis

Abstract

In many statistical analyses for the development of engineering design models, the use of a 5% level of significance is based on tradition. Stepwise regression analysis, using the 5% level of significance, is applied to measured data from engineering systems. The results demonstrate that the commonly used 5% criterion can lead to an incorrect model under certain conditions. Specifically, models with either irrational or inaccurate regression coefficients or too many predictor variables may result. The application of the 5% criterion in stepwise regression analysis is critically assessed, and the conditions under which problems occur are identifled. An alternative decision approach, based on an analysis of the partial F statistic, is developed. This approach enables decisions in stepwise regression analysis to be made using engineering criteria rather than the arbitrarily selected 5% Using the method developed herein with measured data from engineering systems results in models judged more rational than those developed using the traditional approach.