Comparison of Some Bayesian Analyses of Heterogeneity in Group Life Insurance (Q2703232)

A possible approach to describing heterogeneity between policy holders in general is as follows: Each policy holder is equipped with a heterogeneity parameter. The heterogeneity parameters are assumed independent, identically distributed drawn from an unknown heterogeneity distribution. A tractable parametric assumption is imposed on the heterogeneity distribution. The parameters in the heterogeneity distribution are estimated from data, and based on these estimates the posterior distribution of the individual heterogeneity parameters is found. This approach is called parametric empirical Bayesian. This paper focuses on two possible extensions. The uncertainty with respect to the estimated heterogeneity distribution is disregarded. That can be taken care of by extending the model to a parametric fully Bayesian model, which is done by specifying a prior distribution of the parameters in the heterogeneity distribution. The problem of parameterization of the heterogeneity distribution is considered. The heterogeneity parameters are, contrary to the individual policy records, not observable. Therefore it might be argued that in their choice of model the company should express relatively high uncertainty with respect to the (parametric) shape of the heterogeneity distribution, compared to the uncertainty with respect to the conditional distribution of the individual policy records. There are different ways of expressing high uncertainty with respect to the shape of the heterogeneity distribution. A possibility is to work with a very broad class of parametric distributions. The author uses a nonparametric approach. \textit{R. Norberg} [Scand. Actuarial J. 1989, No. 4, 194-224, (1989; Zbl 0714.62099)] analysed the heterogeneity in a portfolio of group life insurances using a parametric empirical Bayesian approach. In the present paper, the Norberg model is compared with a parametric fully Bayesian model and with a nonparametric fully Bayesian model.