B1047
Title: Bayesian nonparametric covariate-driven clustering: An application to blood donors data
Authors: Raffaele Argiento - Università degli Studi di Bergamo (Italy) [presenting]
Abstract: Blood is an important resource in global healthcare and therefore an efficient blood supply chain is required. Predicting arrivals of blood donors is fundamental since it allows for better planning of donations sessions. With the goal of characterizing behaviors of donors, we analyze gap times between consecutive blood donations. In order to take into account population heterogeneity we adopt a Bayesian model for clustering. In such a context, defining the model boils down to assign the prior for the random partition itself and to flexibly assign the cluster-specific distribution, since, conditionally on the partition, data are assumed iid within each cluster and independent between clusters. In particular, we drive the prior knowledge on the random partition by increasing the probability that two donors with similar covariates belong to the same cluster. The resulting model is a covariate-dependent nonparametric prior, thus departing from the standard exchangeable assumption. Specifically, we modify the prior on the partition prescribed by the class of normalized completely random measures by including in the prior a term that takes into account the distance between covariates. First, briefly discuss the theoretical implications of this mathematical operation, finally, we fit our model to a large dataset provided by the Milan department of AVIS (Italian Volunteer Blood-donors Association) the largest provider of blood donations in Italy.
