TL;DR
This paper introduces a novel agent-based modeling (ABM) framework using hazard functions to simulate return migration from conflict zones, specifically focusing on Ukrainian refugees returning from Poland. By incorporating sociopolitical factors and developing computationally efficient surrogate models, they achieve significantly more accurate and nuanced predictions than traditional survey-based methods, offering valuable tools for policymakers.
What Problem Does This Paper Address?
The ongoing conflict in Ukraine has triggered a massive displacement crisis, with millions seeking refuge abroad. A critical, yet understudied, aspect of this crisis is return migration – the process by which displaced individuals decide to go back to their home country. Existing research on return migration heavily relies on survey-based approaches. While these surveys can identify influencing factors, they suffer from significant limitations: demographic bias (as they only capture a subset of the population), lack of validation against real-world data, and an inability to account for the inherent uncertainty in migration decisions. This leaves a crucial gap in our ability to accurately model and predict return migration, hindering effective policy-making for humanitarian aid and reintegration efforts.
Methods
The core of this paper is a novel computational framework for modeling return migration using Agent-Based Models (ABMs). The key innovation lies in grounding these ABMs in hazard functions. Hazard functions, commonly used in survival analysis, model the instantaneous rate of an event (in this case, return migration) occurring at a given time, conditional on it not having occurred yet.
The authors propose three progressively complex ABM models:
- Basic Hazard (BASE): A foundational model where the hazard rate is constant for all agents, representing a baseline propensity to return.
- Conflict Context Influenced Hazard (CL): This model enhances the BASE model by incorporating the conflict context of the origin region (e.g., security situation) into the hazard function. This means the likelihood of returning is dynamically adjusted based on real-time conflict intensity.
- Conflict and Peer Influenced Hazard (CLP): The most sophisticated model, which further adds peer influence to the conflict context. It uses a threshold model where an agent's hazard rate is influenced by the proportion of their neighbors who have already returned.
To address the computational expense of running complex ABMs, the paper also introduces surrogate models. These models operate on aggregated population data rather than individual agents, significantly reducing computation time while aiming to produce comparable results.
Datasets
- Conflict Events Data: ACLED (Armed Conflict Location & Event Data Project) dataset for Ukraine, covering the period of February-August 2022.
- Agent Data: Synthetic household data from [Mortveit et al., 2020], used to create synthetic individuals and households representing the agents in the ABM.
- Ground Truth Data: Border-crossing data from Polish border guards for Ukrainians crossing into Ukraine between February 24, 2022, and August 01, 2022. This data serves as the observed number of returnees for model calibration and evaluation.
Key Insights and Ideas
- Hazard Functions are Powerful for Return Migration: The paper demonstrates that hazard functions provide a natural and effective way to model the "time-to-event" nature of return migration, capturing the probability of an agent returning at any given time.
- Sociopolitical Context Matters: The progression from BASE to CL and CLP models shows a clear improvement in predictive accuracy. This highlights the critical importance of incorporating factors like conflict intensity and social influence (peer effects) for realistic return migration modeling.
- Surrogate Models Offer a Practical Trade-off: While ABMs provide granular insights, their computational cost is high. The surrogate models offer a compelling alternative, achieving significant speedups (seconds vs. minutes per simulation) with only a modest decrease in accuracy, making them ideal for rapid policy analysis.
- Demographic Nuances in Return: The case study on demographic patterns reveals that factors like the mandate for adult males to fight, and the heightened security concerns for families with children and the elderly, significantly influence return propensities. This is crucial for targeted support.
Key Results
- Superior Accuracy: The proposed ABMs significantly outperform baseline linear regression models. The best ABM (C30P) achieved a Normalized Root Mean Squared Error (NRMSE) of 0.089, a 57% improvement over the best baseline (L45 with NRMSE 0.207). Even the simplest ABM (BASE) showed a 42% improvement.
- Surrogate Efficiency: The best surrogate model (surrogate of C30P) achieved an NRMSE of 0.102, a 51% improvement over the best baseline, while running in approximately 3 seconds compared to the ABM's 10 minutes.
- Accurate Trend Capture: All proposed models (ABMs and surrogates) demonstrated high Pearson Correlation Coefficients (PCC) (consistently > 0.9), indicating their ability to accurately capture the trend of observed return migration data.
- Case Study 1 (Return Time): The C30P ABM accurately estimated the length of displacement, aligning well with IOM survey data and showing an increase in displacement duration as the war progressed.
- Case Study 2 (Demographics): The model revealed distinct return hazard rates for different demographic groups, with adult males showing higher propensity to return, and vulnerable groups (elderly, children) showing lower propensity due to security concerns.
Limitations and Future Work
- Generalizability: The authors acknowledge that comprehensive evaluation of generalizability is limited due to the lack of ground truth data for diverse scenarios.
- Surrogate Accuracy vs. ABM Complexity: While surrogates are fast, their accuracy can degrade as the ABM complexity increases, suggesting a trade-off that needs careful consideration. The paper notes that the performance drop in NRMSE for surrogates can be substantial for more complex models.
- Future Work: The authors suggest extending the models to capture the return dynamics of internally displaced people and to model recurring migration patterns. Further research could also explore how limiting the scope of observation for peer influence (e.g., to specific radii of neighboring regions) impacts model outcomes.
Summary updated