and Rao P.

We congratulate the author for this excellent article, which ooers important methods and ideas for inference on functional Magnetic Resonance (fMR) images, with particular application to identifying and testing for theories on working memory utilization. Developing statistical methodology for the analysis of fMR images is an exciting and challenging task, bringing together issues from diverse disciplines such as neurology and psychology and we believe that this article ooers a substantial breakthrough in this regard. Our main comments pertain to some of our recent experiences with fMR experiments which reveal issues of statistical import, some of which may be approached by extending the author's techniques. The blood-oxygenation-level-dependent (BOLD) contrast referred to in the paper uses the diierence in MR signals between oxy-and deoxy-hemoglobin levels that are caused by the excess innow of blood due to neural activation. The diierences in signals are very slight, usually on the order of 2{5 per cent, depending on magnetic eld strength. Unfortunately, there is substantial recent evi-to suggest that such diierences can also arise due to the presence of draining veins in the cortex. Figure 1 illustrates this in the context of a fMR experiment where the goal was to identify cortical regions of activation in a visual task experiment. Note that the areas of signiicant activation (Figure 1a) and the underlying macroscopic vas-cular structure, as depicted through a magnetic resonance angiogram (MRA) (Figure 1b) show substantial overlap, as further seen by the superposition of the activation image on the MRA (Figure 1c). (An angiogram is a MR imaging technique made sensitive to regions with high blood ow rates | of at least 10 cm/s in our experiment.) This causes the concern that the predominance of extra-cerebral signals from these macroscopic vessels may provide inaccurate identiication (in terms of size, shape and spatial location) of regions of neural activation. Further, there is inter-subject variability in the association of activation patterns with vascular structure (Segebarth et al., 1994) which should be accounted for, especially in block-level multi-subject studies. Suppression of vascular signals has been attempted using a variety of approaches (see for example Bandettini et al., 1994; Haacke, et al., 1994; Hlu st ik, et al., 1998) ranging from using a diierent imaging pulse sequence to post-processing, but we wonder whether this can be accounted for in an integrated framework by incorporating this information into the mean baseline signal of the author's Bayesian time-course model. Thus, …