Implementation of complex interactions in a Cox regression framework

The standard Cox proportional hazards model has been extended by functionally describable interaction terms. The first of which are related to neural networks by adopting the idea of transforming sums of weighted covariables by means of a logistic function. A class of reasonable weight combinations within the logistic transformation is described. Apart from the standard covariable product interaction, a product of logistically transformed covariables has also been included in the analysis of performance of the new terms. An algorithm combining likelihood ratio tests and AIC criterion has been defined for model choice. The critical values of the likelihood ratio test statistics had to be corrected in order to guarantee a maximum type I error of 5% for each interaction term. The new class of interaction terms allows interpretation of functional relationships between covariables with more flexibility and can easily be implemented in standard software packages.