Selection of ordinally scaled independent variables with applications to international classification of functioning core sets

TLDR: In this article, the ordinal structure is taken into account by use of a difference penalty on adjacent dummy coefficients, and an alternative blockwise boosting procedure is proposed to select ordinally scaled independent variables in the classical linear model.

Abstract: Summary. Ordinal categorial variables arise commonly in regression modelling. Although the analysis of ordinal response variables has been well investigated, less work has been done concerning ordinal predictors. We consider so-called international classfication of functioning core sets for chronic widespread pain, in which many ordinal covariates are collected. The effect of specific international classification of functioning variables on a subjective measure of physical health is investigated, which requires strategies for variable selection. In this context, we propose methods for the selection of ordinally scaled independent variables in the classical linear model. The ordinal structure is taken into account by use of a difference penalty on adjacent dummy coefficients. It is shown how the group lasso can be used for the selection of ordinal predictors, and an alternative blockwise boosting procedure is proposed. Both methods are discussed in general, and applied to international classification of functioning core sets for chronic widespread pain.