This thesis compares different parametric hazard functions, which are fitted to the survival data of Japan, Norway and the US. It begins by discussing the special characteristics of actuarial mortality modelling. The concepts of left-truncation and right-censoring, which are common features of the data collected at insurance companies, are explained. This is followed by the introduction of mortality ratios, which can be used to compare mortality at different ages. By comparing the initial ratio of mortality, which is based on population size, with the central ratio of mortality, based on the time lived by the observed group (exposure to risk), it will be concluded that analysis based on exposure should be preferred. Chapter 3 presents the survival data from the human mortality database (HMD) used throughout this thesis. For Japan, Norway and the US the number of deaths and exposures for ages 60 to
95 are aggregated over the years 2015 to 2019.
The mathematical model describing human lifetime is introduced. The hazard rate 𝜇𝑥,𝑡 which plays a central role throughout the thesis is defined and it is shown that all survival probabilities can be derived from it. As proposed in Macdonald, Richards, Currie (2018) the number of deaths occurring in the aggregated data sets are then modelled as a Poisson variable. This is done by making the simplifying assumption, that 𝜇𝑥,𝑡 is constant over the course of one year. By using the laws of the Poisson distribution, the likelihoods of the data depending on 𝜇𝑥,𝑡 can be determined. Richards (2008) proposes six parametric functions of the Gompertz-Makeham family, which can be used to model the hazard rate. Using the software R, these functions are fitted to the survival data by maximizing the corresponding likelihoods with respect to the parameters. Several statistics, which are used to compare the different models as well as statistical tests, used to formally describe the goodness of fit, are introduced and applied to the fitted hazard functions. However, all the models share the common defect of having strongly positive autocorrelated residuals, which indicates over respectively underestimation of mortality over specific age ranges.
«
This thesis compares different parametric hazard functions, which are fitted to the survival data of Japan, Norway and the US. It begins by discussing the special characteristics of actuarial mortality modelling. The concepts of left-truncation and right-censoring, which are common features of the data collected at insurance companies, are explained. This is followed by the introduction of mortality ratios, which can be used to compare mortality at different ages. By comparing the initial ratio...
»