In this paper, we investigated rates of admission to hospitals (or other health facilities) due to respiratory diseases in a United States working population and their dependence on a number of demographic and health insurance-related factors. We employed neural network (NN) modelling methodology, including a combined actuarial neural network (CANN) approach, and model admission numbers by embedding Poisson and negative binomial count regression models. The aim is to explore the gains in predictive power obtained with the use of NN-based models, when compared to commonly used count regression models, in the context of a large real data set in the area of healthcare insurance. We used nagging predictors, averaging over random calibrations of the NN-based models, to provide more accurate predictions based on a single run, and also employed a k-fold validation process to obtain reliable comparisons between different models. Bias regularisation methods were also developed, aiming at addressing bias issues that are common when fitting NN models. The results demonstrate that NN-based models, with a negative binomial distributional assumption, provide improved predictive performance. This can be important in real data applications, where accurate prediction can drive both personalised and policy-level interventions.
- statistical models for insurance
- machine learning and data science in insurance
- predictive modelling
- neural network
- k-fold validation
- nagging predictor