Title: Dependent prior processes for panel count data
Authors: Beatrice Franzolini - Bocconi University (Italy) [presenting]
Antonio Lijoi - Bocconi University (Italy)
Igor Pruenster - Bocconi University (Italy)
Abstract: Panel count data occur in observational studies and clinical trials that concern recurrent events, where for each subject cumulative counts are recorded at discrete time points. Both the times points and the cumulative counts are realizations of point processes, namely the observation process and the event process. Even though assuming independence between the two simplifies the inference procedure, the assumption is not realistic in many applications. Prior information on the relation between counts and observation times is often available. We propose a class of Bayesian nonparametric priors over the observation and the event processes that allows for dependence between them, incorporating prior information on the positive association between frequency of observation and counts. The priors are defined modeling the intensities of the two processes through mixtures with respect to GM-dependent completely random measures. We investigate prior and posterior distributional properties of the model and develop a Markov Chain Monte Carlo algorithm to perform posterior inference. The merits of the proposal are further discussed through illustrative examples.