|Abstract:|| In this work we propose a novel model for dealing with hierarchical time-to-event data, which is a common structure in healthcare research field (i.e., healthcare providers, seen as groups of patients). The most common statistical model for dealing with this kind of data is the Cox proportional hazard model with shared frailty term, whose distribution has to be specified a priori.
The main objective of this work consists in overcoming this limit by avoiding any a priori hypothesis on the frailty distribution. In order to do it, we introduce a nonparametric discrete frailty, through which we are not just guaranteeing a very good level of flexibility, but we are also building a probabilistic clustering technique, which allows to detect a clustering structure of groups, where each cluster is named latent population.
A tailored Expectation-Maximization algorithm, combined with model selection tech- niques, is proposed for estimating model's parameters.
Beyond the new methodological contribution, we propose a useful tool for exploring big hierarchical time-to-event data, where it is very difficult to explain all the phenomenon variability through explanatory covariates. We show the power of this model by applying it to a clinical administrative database, where several information of patients suffering from Heart Failure is collected, like age, comorbidities, procedures etc. In this way, we are able to detect a latent clustering structure among healthcare providers.|