NCRM International Visitor Exchange Scheme (IVES)
Bayesian Multiregional Population Forecasting (2016 - 2017)
Arkadiusz Wiśniowski (University of Manchester) visited James Raymer (Australian National University).
The main objectives of the visit were:
- Preparation and submission of the journal article on forecasting international migration by age and sex using bilinear (Lee-Carter) type models and Bayesian inference to produce measures of uncertainty. The article has been submitted to Population Studies journal, which is a high-impact outlet in demography.
- Preparation of the journal article on Bayesian Multiregional Population Forecasts and adapting it to the Australian data; targeted journal Demography is a top methodological journal in the field.
- Extending the model for international migration by using a Bayesian hierarchical approach developed in Wiśniowski and Raymer (2016) Bayesian multiregional population forecasting: England. Working paper presented at the Joint Eurostat/UNECE Work Session on Demographic Projections, Geneva, 18-20 April 2016. This includes preparation of Australian data obtained from the Australian Bureau of Statistics as well as testing various modelling assumptions regarding spatial and temporal patterns of population components.
Statistical offices produce population forecasts which aid planning for housing, education, health, pensions and social integration expenditures. Realistic planning should incorporate not only the inherent uncertainty about the future, but also uncertainty related to the imperfect data collection mechanisms, natural fluctuations in the data, as well as uncertainty about the method used to forecast future. Current methods used by vast majority of statistical offices rely on deterministic projections that reflect a set of assumptions about the future developments of population components. Probabilistic forecasting methods allow quantification of uncertainty and provide evidence for robust decision making under risks mentioned above. International migration component currently contributes most of the uncertainty of the population change, especially in developed countries.
The second important piece of information are transitions that population group may experience throughout their life course. They may include those between states of residences, employment, marriage or health. However, despite the many theoretical and analytical advantages, these models have been relatively unexplored because of the large amount of input data required (which are not always available over time, particularly between censuses) and complex calculations required to perform the estimations and analyses. In fact, most national statistical agencies choose to rely on relatively simple accounting models to produce estimates of population by age, sex and region, which do not include uncertainty or population covariates, such as state of employment, internal interregional movements, health or marriage, that are of particular interest to policy makers and local planning agencies.
In the project we test the Lee-Carter approach to various types of international migration data from Sweden, Australia and South Korea. We assess the forecasting performance of the model by evaluating how well it can forecast age and sex profiles for the withheld part of the data. We also demonstrate how multivariate models can lead to more realistic assessment of uncertainty in the migration forecasts.
We also extend the method of forecasting population at sub-national level which was developed by Andrei Rogers and colleagues in 1970’s and 1980’s to be fully probabilistic. Multiregional, or more broadly multistate models provide a general and flexible platform for modelling and analysing population change over time and with various transitions. We hope to change the reliance on simple models by (1) creating frameworks for dealing with various types of the available data and (2) estimating the components of population change, that is, births, deaths, international immigration and emigration and internal migration, by age and sex with various transitions.
To provide measures of uncertainty, we utilise a statistical Bayesian methodology that offers a natural framework for incorporating various sources of uncertainty and expressing population forecasts in terms of probabilities. The results demonstrate the differences that arise from different modelling choices and the promise of the general approach.