Search

Life expectancy has been on the rise in most countries due to the continuous development in healthcare over the past century. This is positive. However, with the rising average age of humans, current plans for pensions and healthcare may need to be revised to remain affordable for a country. Therefore to inform appropriate changes to these plans, mortality forecasts need to be reliable with various sources of uncertainty incorporated. The goal of this project is to develop a model ensemble approach, through Bayesian model averaging (BMA; Hoeting et al. 1999) to forecast human mortality. This approach forecasts mortality by probabilistically combining a suite of forecasting methods, a feature that aims to incorporate model uncertainty. This source of uncertainty is ignored in a conventional single-model forecasting approach. Applications of this method are done in two settings. The first involves the use of registry based, area-level data from the Human Mortality Database. Here, a number of unique poisson based models can be combined through a model selection weighting technique. Results will be shown comparing the single model approaches, a two-stage model averaging approach (Kontis et al, 2017) and this model averaging technique in a cross validation setting. The second application involves the use of individual level, under-five mortality, data from the Demographic and Health Surveys. DHS data are collected from developing countries through surveys and are available for a limited numbers of years. The limited survey data pose a number of challenges and allow for different interpretations to modelling. Recently, this has been done in the context of using Poisson or Bernoulli based models (Mejia-Guevara et al., 2019; Wakefield et al., 2018). Here, a different view of the data is shown through using a novel survival analysis approach. The model selection approach to estimate model averaging weights is then extended to the survival setting, with a view to incorporating model uncertainty in forecasts.