Normal tissue complication probability (NTCP) models are used to assess the dose-toxicity relationship in radiotherapy. Radiation exposure by organ volume is a functional covariate, and in principle its effect on dichotomous or ordinal toxicity outcomes can be modeled through functional generalized linear models, incorporating a monotonicity restriction which is biologically plausible for dose-toxicity relationships. In this talk we discuss the causal interpretation of monotonic functional regression and identifiability issues involved in such models. As an alternative to functional regression, we consider relating the toxicity outcomes marginally to bivariable dose-volume combinations. For this, we adapt a Bayesian non-parametric monotonic multivariable regression model which can also accommodate ordinal outcomes. The model can approximate arbitrary monotonic regression function shapes without common parametric modeling assumptions such as additivity, linearity or proportional odds.