Predictive Machine Learning Models for Exploring Environment and Migration Linkages in Historical Mexico-U.S. Flows

Filiz Garip , Cornell University
Mario Molina, Cornell University

Prior research identifies several linkages through which environmental factors – both gradual changes and sudden-onset events – might shape internal and international migration flows. These linkages point to various explanatory variables, and complex interactions among them. Empirical work tests some of these linkages in isolation, but reports results that vary considerably depending on the particular variables or models used. To address this issue, we propose to use machine learning (ML) tools that allow us to include all potential indicators (and all possible interactions) to predict U.S.-bound migration outcomes among 120,000+ individuals in 1980-2017 in the Mexican Migration Project data. These tools rely on data-driven model selection, optimize predictive performance, but often produce ‘black-box’ results. To overcome this shortcoming, we propose to use the predictions as a starting point, and analyze discrepant communities for which our model offers poor predictions.

See paper

 Presented in Session 6. Climate Change and Migration