CMStatistics 2017: Start Registration
View Submission - CMStatistics
B0363
Title: Time series graphical modelling via partial coherence and lessons from EEG analysis Authors:  Andrew Walden - Imperial College London (United Kingdom) [presenting]
Abstract: In several areas of science, such as pollution monitoring and neuroscience, the ability to create a graphical model to help visualise relationships between time series is very useful. One approach to time series graphical modelling utilises partial coherence between pairs of series: it being zero for all frequencies corresponds to a missing edge in the graph. If partial coherence methodology is to be successfully used on EEG data then careful consideration must be given to practical issues such as mixed spectra and poorly condition spectral matrices. The treatment of spectral lines is discussed. Dealing with poor conditioning via spectral matrix shrinkage reveals very different influences on partial coherence estimates may result, depending on the loss function chosen. The production of a meaningful result in practice is much harder than might be anticipated from the theory.