Identifying Timber Performance Classes Using Latent Class Regression*

Latent class regression is a statistical method which is not well known in wood science for predicting distortion of sawn timber from structural wood characteristics. The method identifies unknown subgroups in a dataset and allows more accurate regression models to be derived. We identified two separate classes to describe the relationship between bowdry, springdry, or twistdry and the predictors: initial distortion bowfresh, springfresh, or twistfresh, wood density, ring width, ring orientation, wood type (juvenile or adult), percentage compression wood measured separately on the four faces, and the contagion index which is a measure for the distribution of compression wood. The latent class regression models developed for the separate classes explained the variation in bow, spring, and twist to a higher degree than a single regression model over the entire dataset. For bow, R2 increased from 0.13 to 0.24 and 0.41 for Class 1 and Class 2, for spring from 0.24 to 0.45 and 0.67, and for twist from 0.15 to 0.38 and 0.33. For individual regression models, the predictors showed a varying effect. In classes with significant compression wood on the faces, the effect of wood type seemed weaker, and vice versa. It was concluded that latent class regression analysis allows a more detailed explanation of the effects of wood structure on sawn timber distortion for heterogeneous datasets.