Disaster Analytics for Society
The extent of damage to buildings caused by earthquakes is affected by different structural properties as well as the location of the sites themselves. Nearby measurements tend to be more related than those further away. While factors such as ground motion intensity can explain part of this, there is residual spatial correlation due to unknown factors and/or factors we do not have covariate data for. Since ignoring this effect result in the underestimation of damage extremes for regional portfolios, I’m currently working with Sabine Loos and David Lallemant on a project to account for spatial correlation in probabilistic damage models.
Classical spatial statistical models assume that observations are randomly distributed in space. However, hazard measurements are often spatially clustered and/or preferentially taken in areas where higher or lower than average values are expected. As misleading inferences can be made when observations do not give the same degree of information, I am researching the effects of such sampling schemes on statistical inferences. This work is in collaboration with David Lallemant and Maricar Rabonza.
Malaria Atlas Project
Malaria is an age-old disease which is caused by Plasmodium parasites transmitted by mosquitoes. According to the 2017 World Malaria Report, there was an estimated 216 million cases of malaria in 91 countries in 2016. While I was at MAP, I was part of the Global Malaria Epidemiology team and worked with researchers from fields including Public Health, Ecology and Geography. Together, we developed spatio-temporal modelling infrastructure to predict transmission and burden in malaria endemic countries. Our work continues to contribute to the Global Burden of Disease project at the Institute of Health Metrics and Evaluation as well as World Malaria Report produced by the World Health Organisation. Two Lancet papers outline our modelling strategies and results for Plasmodium falciparum and Plasmodium vivax.
In line with our efforts to support control agencies in their decisions, I had also been leading a project to model and map malaria seasonality patterns. It is hoped that by understanding when malaria cases peak in a year, we can time interventions to have optimal impact as well as prepare ahead to reduce the number of people affected. This research is in collaboration with colleagues at Imperial College London, the National Malaria Control Programme in Madagascar and Institut Pasteur de Madagascar.
PhD in Statistics @ Imperial
In my PhD, I studied different spatial and spatio-temporal models which are all centred on the relatively new modelling framework of ambit fields. Without equations, one may describe an ambit field as a random field (that is, random variables indexed by time and space) where the major stochastic part is a stochastic integral. By choosing the kernel function, the shape of the ambit set (i.e. the region over which we integrate), the stochastic volatility term (if required) and the Lévy basis (our integrator), we can vary the characteristics of the model. In many cases, ambit fields can be viewed as solutions to stochastic partial differential equations (SPDEs). Studying ambit fields directly is an alternative approach to constructing SPDEs and finding their solutions.
Recent interest in ambit fields started from the modelling of energy dissipation characteristics in turbulence. It has since been used in for example, financial modelling and tumour growth modelling. There is ongoing research on both the probabilistic and statistical aspects of ambit fields and its sub-models.
My PhD consists of three projects. Each seeks to develop the theory, simulation and inference methods for particular subclasses of ambit fields. The latter in turn highlight different aspects of the more general ambit field model:
- Project 1: Spatio-temporal Ornstein-Uhlenbeck processes
- Project 2: Volatility modulated moving averages
- Project 3: Mixed spatio-temporal Ornstein-Uhlenbeck processes