Survival data analysis becomes complex when the proportional hazards assumption is violated at population level, or when crude hazard rates are no longer estimators of marginal ones. We develop a Bayesian survival analysis method to deal with these situations, based on assuming that the complexities are induced by latent cohort or disease heterogeneity that is not captured by covariates, and that proportional hazards hold at the level of individuals. This leads to a description from which risk-specific marginal hazard rates and survival functions are fully accessible, 'decontaminated' of the effects of informative censoring, and which includes Cox, random effects and latent class models as special cases. Simulated data confirm that our approach can map a cohort's substructure, and remove heterogeneity-induced informative censoring effects. Application to data from the ULSAM cohort leads to plausible alternative explanations for previous counter-intuitive inferences on prostate cancer. The importance of managing cardiovascular disease as a comorbidity in women diagnosed with breast cancer is suggested on application to data from the AMORIS study.