Statistical Evidence for the Preference of Frailty Distributions with Regularly-Varying-at-Zero Densities

Trifon I. Missov, Max Planck Institute for Demographic Research
Jonas Schoeley, Max Planck Institute for Demographic Research

Missov and Finkelstein (2011) prove an Abelian and its corresponding Tauberian theorem regarding distributions for modeling unobserved heterogeneity in fixed-frailty mixture models. The main property of such distributions is the regular variation at zero of their densities. According to this criterion admissible distributions are, for example, the gamma, the beta, the truncated normal, the log-logistic and the Weibull, while distributions like the log-normal and the inverse Gaussian do not satisfy this condition. In this article we show that models with admissible frailty distributions and a Gompertz baseline provide a better fit to adult human mortality data than the corresponding models with non-admissible frailty distributions. We implement estimation procedures for mixture models with a Gompertz baseline and frailty that follows a gamma, truncated normal, log-normal, or inverse Gaussian distribution.