Senior Policy Researcher, RAND Corporation
Dissertation Title: "Demographic Forecasting"
A number of important public policy decisions rest heavily on the ability to forecast certain cross-sectional time series. In this thesis we consider the task of forecasting time series of all causes and cause specific mortality for individual countries, genders and age groups. Time series of this type are relevant for the design of pension and health benefits, as well as for the prioritization and planning of health interventions. We argue that standard time series methods, by focusing exclusively on the observed values of the time series, fail to take advantage of a huge amount of available, additional information. This information comes under the form of time series of known determinants of mortality, and of experts' prior knowledge. We set the problem of forecasting mortality in the framework of cross-sectional time series with covariates, and adopt a hierarchical Bayesian point of view. We show how prior knowledge on the expected value of the dependent variable, of the type epidemiological experts have, can easily be translated into a prior density for the set of cross-sectional regression coefficients, allowing cross-sections to borrow strength from each other. We consider in details the case in which experts know, a priori, that the expected value of the dependent variable (or its time trend) varies smoothly across age groups, countries and time. While we focus on mortality, the method we developed can be applied to any cross-sectional time series for which the expected value of the dependent variable is known to vary smoothly across cross-sections. In particular, we offer a solution to the problem of pooling time series with different covariates in different cross-sections. We report results obtained applying our method to worldwide cause specific data obtained from the World Health Organization, and report very encouraging comparisons with the state of the art method of Lee and Carter (1992).