NSF Award
2428033
Forced displacement is a threat to national, regional, and global security. It impacts economic, political and social systems in host and destination countries, resulting in large-scale resource scarcity and obstacles to development. As conflict and repression rise, environmental risks grow, and economic inequalities increase, more people are being displaced from their homes across the world. Given this situation, effective modeling and timely forecasting of forced displacement is essential for rapidly disseminating humanitarian aid, stabilizing countries that house the displaced, and informing public policy. To build robust forecasting models of forced displacement, this project aims to: 1) identify and measure correlates that may serve as leading indicators of displacement, 2) integrate information from previous displacement contexts to improve the quality of the prediction, and 3) develop a broad statistical modeling framework that adequately incorporates this prior knowledge and applies to spatiotemporal data. To meet these challenges, the investigators will construct statistical models that incorporate a robust set of indicators developed from traditional data sources, digital trace data from Google searches and social media posts, and satellite imagery data, as well as develop methods to transfer knowledge from previous crises to enhance the quality of the data used for forecasting. One important outcome of this project is to improve the practice of forecasting migration by using novel transfer learning methods that encode knowledge about previous migrations in other locations and using innovative organic data sources to obtain timely conflict and climate indicators for emerging forced displacement threats. Ultimately, this approach permits short-term predictions of displaced person flows at the outset of an emerging crisis when the threat of unexpected large-scale displacement is often highest, but predictions are most difficult to make. Moreover, the broader contribution of the project is developing a general approach for modeling and forecasting emerging crises under data sparsity, a typical characteristic of these events, with limited predictor variables, thereby extending the application of this method to forecasting crises that need immediate attention and/or intervention.<br/> <br/>To achieve the stated goals, the investigators will develop a general approach for transfer learning in regression models (Gaussian and non-Gaussian, including complex count distributions) that forecasts threats in spatiotemporal settings, including forecasting forced displacement. In particular, the investigators will develop a novel Bayesian formulation of transfer learning which both encompasses the specific models already present in the literature and applies to a large class of likelihoods. The investigators will also extend transfer learning methods to make them applicable to our non-Gaussian, spatiotemporally correlated datasets and move forward the state-of-the-art in forecasting displacement, among many other applications in emerging threat detection. Critically, the investigators will develop a new prior distribution that encodes and extends the existing transfer strategy for linear Gaussian models. This project will also include a Bayesian technique for deciding which source datasets to consider when transferring to a specific target dataset and allow deployment of transfer learning to a wide spectrum of statistical models, including models with spatiotemporally correlated error. In addition to the specific models that will be studied, the investigators will also lay out a general framework for Bayesian transfer learning. This will lead to statistical research into the theoretical properties of, and efficient computational methods for, these Bayesian transfer techniques. Finally, the predictive models developed for three displacement cases (Ukraine, Bangladesh, and Sudan) will advance migration research by developing new indicators of forced displacement.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
