Coupling Bayesian inference and agent based modeling in the context of refugee movement

Abstract:

Information about refugee movement is scarce, but vital for logistical, economical and epidemiological reasons. Predictions of refugee movement are critical for emergency preparedness and response to emerging crises. One way to model the fluxes of refugees is with agent based models (ABMs), governed by a set of rules that dictates how agents - refugees in this case – will move. ABMs are very useful but often are deemed “black boxes”, as many push, pull, and mooring factors are applied in parallel. Here we use the Flee ABM, which simulates refugee movement from conflict zones to refugee camps and attempt to disentangle information generated by the model. Flee’s input parameters include distance, speed, and likelihood to move based on node class (conflict zone, refugee camp, or intermediate location), while agents in the simulation are configured to try and cover as little distance as possible on their search for a camp. First, we use Markov-Chain Monte Carlo sampling to find the appropriate input parameters that match refugee flows from UNHCR data. Based on associated distributions, those parameters provide an uncertainty range for the simulations. Then, we apply a multilevel Bayesian model (MBM) that provides the uncertainty distributions of the contribution from each conflict zone to the arrivals at each camp and infers how distance and attributes of each receiving location determines migration potential in the Flee outputs. We apply the MBM on recent East African conflicts simulated by Flee. We observe that while distance is an important driver to choose a camp, other factors impact the trajectory too. We hypothesize that variables such as camp capacity at the time of conflict, camp condition, host country refugee policies, and road accessibility could reduce uncertainty in the predictions.