A Frequency Localised Framework for Inference in fMRI
Wednesday, October 11, 2000 - 10:15am - 10:55am
Although fMRI time series have been subjected to a very large number of studies, the noise models in use are generally unreliable and inaccurate, as shown by the general scepticism about P-values. We argue that since the noise sources in question, which are primarily physiological in origin, are well segregated in the frequency domain, most of the problems are obviated by confining the inference to the frequency range relevant to the hemodynamic response. We have developed and tested a methodology for statistical inference as well as for exploratory analysis based on projections of the time series data onto frequency localised basis sets, following the framework of multitaper spectral analysis. The associated statistical tests are both straightforward and provide somewhat more sensible P-values, since Gaussian distributional assumptions are more closely valid for narrowband processes due to the central limit theorem.