In medical research, ordered categorical outcomes (such as seriousness, side effects, and grade of treatment) are sometimes used as response variables. Typically, the influencing factors are explored by ordered logistic regression. Recently, the tree-structured method has been extended to ordered categorical outcomes (Piccarreta, 2008; Archer, 2010), but the predictive outcomes of this approach are poor. In this paper, we newly develop a nonlinear ordered categorical regression method, named PO-MARS, which extends multivariate adaptive regression splines (Friedman, 1991). The PO-MARS method is developed on a proportional odds model framework, and model selection is based on the modified Akaike's information criteria (AIC) proposed by LeBlanc and Crowley (1999). The effectiveness of the PO-MARS method was illustrated through a practical example. In small-scale simulations, this method demonstrated higher predictive performance than existing methods.
View full abstract