A popular approach in Bayesian modelling of partially exchangeable data consists in imposing hierarchical nonparametric priors, which induce dependence across groups of observations. In survival analysis, hierarchies of completely random measures have been successfully exploited as mixing measures to model multivariate dependent mixture hazard rates, leading to a posterior characterization which may also accommodate censored observations. Such framework can be easily adapted to a competing risks scenario, in which groups correspond to different diseases affecting each individual: in this case, the multivariate construction acts at a latent level, as only the minimum time-to-event and the corresponding cause of death are actually observed. The posterior hierarchy of random measures, as well as the posterior estimates of both survival function and cause-specific incidence functions are explicitly described, conditionally on a suitable latent partition structure which fits the Chinese restaurant franchise metaphor. Marginal and conditional sampling algorithms are also devised and tested on synthetic datasets. The performances of this proposal are finally compared with those of its non-hierarchical counterpart, which models the hazard rate of each disease independently: leveraging the information borrowed from other groups, the hierarchical construction is empirically shown to recover the shape of the incidence functions more efficiently, in presence of proportional hazards.