Geographical Ratings with Spatial Random Effects in a Two-Part Model

Rating areas are commonly used to capture unexplained geographical variability of claims in insurance pricing. A new method for defining rating areas is proposed using a two-part generalized geoadditive model that models spatial effects smoothly using Gaussian Markov random fields. The first part handles zero/nonzero expenses in a logistic model; the second handles nonzero expenses (on log-scale) in a linear model. Both models are fit with R package INLA for Bayesian inferences. The resulting spatial effects are used to construct more representative ratings. The methodology is illustrated with simulated data based on zipcode areas, but modeled on zipcode- or county-level. Address for Correspondence: jun.yan@uconn.edu