Pole-Augmented Multidimensional Scaling Framework for Survival-Analysis Visualization

We introduce a multidimensional scaling framework that represents survivaloutcomes and covariates in a single two-dimensional map. Two reference points called polesrepresent the event direction and survival or non-event direction at a fixed follow-up time.The poles and covariates are represented in a multidimensional scaling space, such that eachcovariate is closer to one pole. This representation is intended to combine two readings in aunified visualization, namely, the relational structure between covariates and each covariatelocation relative to the two survival-outcome poles. The event pole is built from the martingaleresidual of a null Cox model, while the non-event pole is built from a Kaplan–Meierpseudo-survival quantity. To produce a clearly interpretable unified visualization, Pearsoncorrelations between items are converted into distances using three category-specific exponentsfor the pairs of covariate–covariate, covariate–pole, and pole–pole. The proposedframework is evaluated on the Rossi recidivism dataset (N = 432). Financial aid and maritalstatus are closer to the non-recidivism pole, while prior convictions and race are closer to therecidivism pole. A simulation under Cox proportional hazards provides a known coefficientstructure, with positive- and negative-coefficient covariates near the event and non-eventpoles, respectively.