Predicting mechanical properties of timber elements by regression analysis considering multicollinearity of non-destructive test results

Abstract

Information retrieved from different sources, which are necessarily uncertain to some extent, must be compiled to have a reliable assessment of a timber element. Means to obtain information are often limited to non or semi destructive tests. Therefore, the available information is an indirect measurement of the element’s mechanical properties, and are incomplete predictors if not complemented with more data. In that scope, data is commonly added using different predictors in a regression analysis. If the predictors themselves are uncorrelated, the determination coefficient, R2, is directly related to the measure being explained. However, predictors are usually correlated and thus multicollinearity exists. In extreme cases, multicollinearity results in imprecise and unstable R2, thus the relative importance among predictors is not accurately measured.