Modul:   MAT959  Seminar in Data Science and Mathematical Modelling

Disaster & conflict modelling: I've done it all wrong.

Talk by Dr. Hamish Patten

Date: 17.10.24  Time: 12.15 - 13.45  Room: Y27H12

Inherent biases in the methodologies of data collection of humanitarian and government organisations of conflict and disaster impacts render risk modelling extremely difficult. Mortality is one of the most robust types of impact due to government regulated death certification (but not for all countries). However, deaths are not very informative for disaster response: population displacement is the key metric. We now have the inverse problem: a key metric, but one that is inaccurately estimated and that changes in time, as displaced persons return home or are relocated. An alternative is to look at the number of events, but this too, is far from a clean metric to predict. There is no point in talking about predictive performance or feature importance if the data the models are trained on is not appropriate for the problem at hand. After an introduction on data and reporting bias, this talk will move on to two applications of risk modelling: conflict prediction and earthquake impact prediction. For the conflict setting, we propose Hawkes processes, a self-exciting stochastic process used to describe phenomena whereby past events increase the probability of future events occurring. The aim being to monitor the risk of political violence and conflict events in practice and characterise their temporal and spatial patterns. For the earthquake setting, we provide a novel approach whereby both spatially aggregated estimates of impact and geotagged building damage are used to train a model to predict mortality, population displacement, and building damage on a 1x1km gridded spatial level. Both the conflict and disaster models are developed in a Bayesian framework, parameterised by Markov Chain Monte Carlo (MCMC) and Approximate Bayesian Computation - Sequential Monte Carlo (ABC-SMC) algorithms, respectively.