Stephen Smith1, Achim Gass2, Andreas U. Monsch3, Anil Rao4, Brandon Whitcher4, Paul M. Matthews4, Christian Beckmann1,5
1Oxford University FMRIB Centre, Oxford,
Oxon, UK; 2Depts. Neurology and Neuroradiology, University Hospital
Basel, Switzerland; 3Memory Clinic Basel, Switzerland; 4GlaxoSmithKline,
Clinical Imaging Centre, London, UK; 5Clinical Neuroscience
Department, Imperial College London, UK, UK
Model-based analysis methods are generally reliable in identifying expected responses, aiding interpretability of results. Conversely, model-free methods are able to find surprising effects in the data, or separate out structured confound processes from signals of interest. Little work has been carried out to combine the best aspects of the different approaches. We present initial results from inserting model information into a multivariate model-free approach and compare this approach to a GLM analysis as well as to PLS and CVA. We find that approaches containing both model-based and data-driven aspects are almost always more interpretable than rigidly enforcing model structure.