A1948
Title: A deep learning random-effect modelling approach for longitudinal count data
Authors: Jihoon Kim - Pukyong National University (Korea, South) [presenting]
Il Do Ha - Pukyong National University (Korea, South)
Abstract: The deep neural network (DNN) model, a core model of deep learning, provides high predictive power to predict and classify output variables in various fields by modeling the nonlinear functional relationship between input and output variables through hidden layers. However, since the DNN has been mainly developed for independent output variables, applying it directly when output variables are correlated results in greatly reduced predictive performance. Such correlation is typically modeled via random effects, but random effect models have been primarily developed under the assumption of linearity in the functional relationship between input and output variables. A new Poisson DNN random-effect model is proposed herein. The output variables are correlated count outcomes obtained from repeated measures over time. For estimation and learning of the proposed model, an optimization algorithm based on negative marginal likelihood is developed as a loss function. Simulation and real data analysis demonstrate the validity of the proposed method. In particular, simulation results confirm that the proposed DNN model provides higher predictive performance than existing prediction models.
