Revisiting forced migration: A machine learning perspective

Maja Micevska

European Journal of Political Economy (2021)


Using machine learning techniques, this paper investigates the factors that influence refugee flows and asylum applications from African countries. The author argues that, compared to traditional estimation methods, machine learning techniques are better suited to estimating complex relationships and interaction effects.

The analysis draws on several data sources including: UNHCR data on refugee stocks (from which refugee flows are derived) and asylum applications from 45 African countries from 1997 to 2017; data on various types of conflict from the Armed Conflict Location and Event Dataset (ACLED); GDP data from Penn World Tables; data from the World Bank on population, internet penetration, and net official development assistance and aid received; regime type from the Polity IV index; data on political terror from the Political Terror Scale website; number of disasters from the Centre for Research on the Epidemiology of Disaster; and country-level data on precipitation and temperature from the Climatic Research Unit of the University of East Anglia.

Main findings:

  • Country fixed effects are most important in explaining forced migration flows. Country fixed effects account for factors, such as persistent conflicts, chronic poverty, geographic position, or proximity to refugee routes, which are not observed in the data and do not vary over time. This finding holds after controlling for variables that have been used as determinants of forced migration in previous studies.
  • Riots are the most important type of conflict for explaining asylum applications. Fatalities and violent conflicts are important drivers of refugee flows, but not for asylum applications.
  • Internet penetration rates are important in explaining forced migration flows and asylum applications. The author suggests that Internet access lowers communications costs and facilitates access to information about migration routes and destinations.
  • Population growth is an important driver of asylum applications.
  • GDP per capita does not seem to play a prominent role in predicting forced migration.
  • Riots are the most important driver of asylum applications from Nigeria, but population growth and increasing Internet penetration rate also play a role in explaining asylum applications. Riots are an important driver of asylum applications from Cote d’Ivoire. Internet penetration is important for explaining asylum applications from Gambia, Senegal, Guinea, Cote d’Ivoire, and Mali. Algeria and the Democratic Republic of the Congo are important sources of asylum applications due to factors not included in the model.

The author notes that the salience of country fixed effects points to the importance of within-country differences (e.g. between regions or ethnic groups), which is not adequately captured in country-level data. The author calls for further research on the most important origin countries to scrutinize factors that have been neglected in previous research.