Probabilistic Models for Human Migration Forecasting and Residency Imputation

In this talk, I discuss new probabilistic models for forecasting international migration and imputing residency.

The first aspect of my work develops a model to generate probabilistic forecasts of global bilateral migration flows among the 200 most populous countries. Using a Bayesian hierarchical approach, we produce the first probabilistic forecasts of international migration flows by age and sex through 2045. This approach improves prediction interval calibration compared to leading alternative methods, reduces error rates compared to existing approaches, and integrates seamlessly into a full population projection framework.

Forecasting global bilateral flows is computationally demanding and sometimes met with skepticism over longer forecast horizons due to the reliance on just six periods of flow data. Net migration estimates are available over a much longer period (1950–2020). The second aspect of my work leverages the strengths of both data sets to provide a more computationally tractable approach for long-term international migration forecasting. Specifically, we account for the influence of population age structure on past and projected net migration rates, a factor often overlooked in existing models. To address this, we introduce the Migration Age Structure Index (MASI), which rescales historic and forecasted migration rates relative to a reference population. Using this approach, we generate joint probabilistic forecasts of net migration rates for the 200 most populous countries through 2100. Incorporating age structure reduces uncertainty in migration projections and predicts less severe population declines in rapidly aging populations compared to models that ignore age distribution effects.

Finally, we introduce the Bayesian Person-Place Model (Bayes PPM) for residency imputation in countries without population registers. This Bayesian hierarchical discrete choice model addresses longstanding limitations of existing Person-Place Model specifications. By eliminating approximations that constrain statistical inference and subsequent imputation, our method provides a rigorous framework for census operations and intercensal population estimation from administrative records data. We demonstrate that our approach produces accurate uncertainty intervals and effectively leverages information across sometimes sparsely populated subgroups.