An agent-based framework to study forced migration: A case study of Ukraine

Abstract

The ongoing Russian aggression against Ukraine has forced over eight million people to migrate out of Ukraine. Understanding the dynamics of forced migration is essential for policy-making and for delivering humanitarian assistance. Existing work is hindered by a reliance on observational data which is only available well after the fact. In this work, we study the efficacy of a data-driven agent-based framework motivated by social and behavioral theory in predicting outflow of migrants as a result of conflict events during the initial phase of the Ukraine war. We discuss policy use cases for the proposed framework by demonstrating how it can leverage refugee demographic details to answer pressing policy questions. We also show how to incorporate conflict forecast scenarios to predict future conflict-induced migration flows. Detailed future migration estimates across various conflict scenarios can both help to reduce policymaker uncertainty and improve allocation and staging of limited humanitarian resources in crisis settings.

Keywords: Ukraine; agent-based modeling; digital twin; forced migration; policy analysis; social theories.

Fig. 1.

Architecture of the model. After extracting the agent data (from synthetic population) and conflict data (from ACLED) for the current timestep, (1) Each person agent interacts with conflict events in their vicinity and develops an initial perceived migration intent (Attitude and PBC), (2) Agents in the same household combine their decisions to form a unified initial migration decision at the household level (Subjective Norm), (3) A network of household is constructed (see Appendix SA for details of the construction mechanism), (4) Each household agent communicates with household agents in their neighborhood through the threshold function to ultimately decide whether they migrate or not (Subjective Norm), and (5) Migrating agents are removed from consideration for next timestep.

Fig. 2.

Spatial, temporal, and demographic dimensions of ABSCIM estimation capability. a) shows the daily estimation of total individuals crossing borders, compared against reported border crossings for validation. Both data are displayed with a 7-day moving average applied to reduce noise associated with the reported data (unsmoothed observed counts appear in Appendix SA). b) presents stack plots disaggregating refugee totals into four different demographic groups. Validation is not possible due to the absence of observed data. However, we present validation against some early qualitative reports indicating aggregated statistics, presented in Appendix SB. c) shows the cumulative total median outflow estimation at Oblast level. A high density of estimated migrants can be observed across the Eastern regions, where early conflict events took place. Dnipropetrovska, Donetska, Kharkivska, and Kyivska oblasts were reported to be among the top five oblasts of origin for the refugees during the early period of the war (UNHCR), which are among the top five oblasts to have the highest refugee estimates by the ABM. d) disaggregates the outflow at Raion level. See Appendix SA for visualizing the ratio of refugee agents with respect to the total population. (a) Total. (b) Demographic. (c) Oblast and (d) Raion.

Fig. 3.

Visualization of spatial origin of fleeing Ukrainians in response to three possible conflict scenarios. To develop a better sense of the origin locations for refugees in these scenarios, we subtract the estimated Raion aggregates from the Raion aggregates estimated using the observed conflict data. (See Appendix SB for visualizing outflow difference with status quo.) The figures highlight the Raions most likely to produce greater refugee flows (darker red) under the given scenario relative to the observed estimates. In the status quo scenario in a), most Raions exhibit similar outflows as the outflow generated using observed conflict data. However, the few Raions with higher estimates primarily lie on the periphery of the conflict space and therefore these Raions likely did not have any observed conflict events but, owing to forecast uncertainty, did have events in the status quo scenario resulting in larger outflows. Under the Belarus Offensive scenario in b), the bulk of new refugees originate from Raions along or near the Ukrainian border with Belarus. In fact, Raions like Sarnenskyi and Zviahelskyi observe greater than 10 K refugee estimate differences under this scenario compared to the observed scenario. c) shows similar differences in outflow across Raions under the Kherson counteroffensive scenario where the Raions with large estimate differences are observed around the center of the offensive. a) Status quo. b) Belarus offensive, and c) Kherson counter-offensive.