Overview
I'm an applied statistician with experience in applying statistical models to solve problems in a variety of areas, including environmental sciences, hydroinformatics, epidemiology and public health. My research broadly focuses around applied Bayesian modelling. More specifically research interests include:
- Space-time modelling of environmental processes
- Space-time modelling of point processes
- Semi-parametric regression models
- Stochastic weather generators
- Post-processing of weather and climate forecasts
- Hidden Markov and semi-Markov models
- Spatio-temporal modelling of extreme European storms. European windstorms are the second largest source of damage worldwide after hurricanes, so quantifying the associated risk is of great importance. The work involves the implementation of Bayesian spatio-temporal extreme value models for quantifying the intensity of storm peaks over Europe and the potential influence from large-scale atmospheric variability such as the North Atlantic Oscillation. Results indicate that windstorms of hurricane strength are not unlikely over North Europe.
- Decision theoretic severe weather warning system. Early warning systems play a major role in reducing monetary, structural and human loss from natural hazards. Optimally issuing warnings requires input from both involved entities: the best prediction of the forecaster and the potential losses from the end-user. In this project, a system based on Bayesian decision theory was proposed to tackle this challenge, applied to the current UK Met Office warning system. Ensemble forecasts were used to quantify the probability of severe weather and an interactive tool was developed to elicit loss functions from potential end-users.
- Statistical correction of under-reporting in natural hazard data. Observations of certain natural hazards, such as hail and tornadoes, rely on actual observations/sightings or proxy measurements, compared to rainfall and wind, which can be measured using instruments. As such data tend to be under-reported and information is only partial. This work, illustrated on UK tornado counts, involves use of semi-parametric regression models (GAMs) and information on population density to correct data for under-reporting.
Medical/epidemiological applications
- Use of Bayesian hierarchical models, to estimate parameters in disease transmission models. A widely used model to describe infectious disease dynamics in populations at risk is the so called Susceptible-Infected-Resistant (SIR) model: a system of ordinary differential equations. The work involves fitting dynamic hierarchical models using data on swine counts infected with Salmonella, to estimate the transmission parameters of the SIR model.
- Estimating the spatio-temporal variability of risk from hip fracture in Portugal. A series of models, including a Bayesian Poisson geostatistical model and a Negative Binomial Generalised Additive Model (GAM), were implemented to quantify how risk from hip fractures varies across Portugal, while allowing for factors such as gender, socio-economic status and weather.
- Assessing the influence of factors affecting the development and progression of parastomal hernias, using Bayesian multilevel models.
- Predictive models for underground water pipe failures. The prediction of pipe failures in water distribution systems is an essential planning component for water companies. This work involves the use of Bayesian non-homogeneous Poisson process models with zero-inflation, to predict failure occurrence while allowing for pipes that are more resilient to failure than expected. The models are further extended to a hidden semi-Markov formulation, to allow for unobserved temporal processes (e.g. weather, maintenance, soil movement).
