Statistics Postgraduate Talks I
Thursday 4th July, 2013 14:00-16:00 Maths 203
Kathakali Ghosh - Characterisation and mixed effects models for EEG signals (2pm)
Brain signals from cognitive functional neuroimaging experiments (EEG,MEG, etc.) may be characterised into functional forms of phase/frequency and amplitude to reduce the dimension of the data. In this presentation, a characterisation method based on Euler's formula is proposed and applied to artefact corrected and smoothed EEG signals. A linear mixed effects model is fitted to repeated measurements from such characterised frequency curves. Results from this model is presented to assess the changes in mean and standard deviation over various experimental conditions as evidence of entrainment of brain signals.
Katie Alsion - A Spatio-Temporal Air Pollution Indicator for Glasgow (2.20pm)
The main aim of my thesis is to develop an environmental indicator for Glasgow based on air pollution. Within my thesis I discuss and analyse two main datasets - PM10 monitoring site data and annual mean PM10 model output which leads onto building a model which combines both the temporal and spatial aspect of PM10 in Glasgow.
Craig Anderson - Identifying discontinuities in spatial data (2.40pm)
The aim of disease mapping is to identify clusters of areal units which have a high-risk of a particular disease, so that the appropriate remedial action can be taken. Conditional autoregressive (CAR) models are generally used to model the spatial correlation present in areal disease data, but they have shortcomings when attempting to identify the spatial extent of high-risk clusters. Here, we propose a two-stage method for identifying such high-risk clusters, and are motivated by a study of respiratory disease in the Greater Glasgow and Clyde health board. The first stage is a spatially-adjusted clustering technique, which is applied to data prior to the study period and produces multiple potential cluster structures for the disease data. The second stage uses the real data to choose between the set of potential cluster structures using Bayesian hierarchical models.
Coffee break (3pm)
Amira Elayouty - Making Sense of the Environment (3.20pm)
In the past, monitoring programs involved monthly, weekly, sometimes daily data but not sub-daily or sub-hourly data. Nowadays, sensors technology advances and facilitates the ability of monitoring programs to involve sub-hourly data. Therefore, there is a need for statistical methods to analyze and model such high frequency data. My PhD research focuses on modeling high frequency water quality data in space and time. The research is based on a data set measuring the excess in partial pressure of carbon dioxide (a measure for river water quality) on a small river system every 15 minutes throughout four hydrological years (2003-2007). Modeling such high frequency data is very complex due to the high correlation between successive observations, the huge amount of recorded data, the presence of singularities and the non-stationarity of the series. Also, the series vary across different time scales and witness many changes attributed to changes in other covariates such as flow and temperature which are also affected by preceding events in the river system. Accordingly, the research aims to build a model for the partial pressure of carbon dioxide that is able to handle this huge amount of data and that accounts for all the time components; hour, day, month, year and its relation to other covariates by including more environmental specific information into the statistical model.
Daniel Molinari - Posterior approximations for Gaussian models with "non-detect" data (3.40pm)
In many applications one has to deal with censored data, i.e. data for which it is only known that their value is below or above a certain threshold. We propose a posterior approximation similar to the Laplace approximation for partly censored data with Gaussian response. The approximation can be computed very efficiently and is typically close to the full Bayesian solution, which requires the use of sampling strategies such as MCMC. The proposed method is illustrated using an example from environmental modelling.