 |
 |
Dear Clark, For the record, the cluster size random field test for Gaussian images was proposed in Friston, Worsley, et al, 1994, HBM 1:210-220, and extended to t & F fields by Cao, 1999, Adv Appl Probability 31:579-595 (for a concise listing of RF results, see Cao & Worsley's chapter in the book "Spatial Statistics: Methodological Aspects and Applications", Springer Lecture Notes in Statistics, Vol 159; also, Poline, Worsley, Evans & Friston, 1997, NI 5:83-97, develops a cluster test for Gaussian images using a joint size-height statistic.) As for the reviewer's comment, I would have to understand that the reviewer was specifically referring to Gaussianization when they said that the "use corrected P-values based on the cluster-size statistic" is not state of the art, and not cluster size inference in general. But I think you have an easy reply that will silence the reviewer: Both FSL *and* SPM(5 & older) use Gaussianization for cluster size inference; only Keith Worsley's fmristat offers an implementation of all the current results. SPM's voxel-wise results use the current t/Chi^2/F results, but the cluster size results for t just use the alpha-level of the cluster-defining threshold and then use the Gaussian results. Karl has said this method doesn't Gaussianize, as much as it assumes a Gaussian ACF. Whatever the details, neither FSL nor SPM are using the "state of the art" t or F cluster size results. It is currently on the plan for the next version of SPM to incorporate Keith Worsley's statthreshold code into SPM to use the most up-to-date results. FSL? TBD :) Finally, to echo Steve's response, FDR is a different beast altogether, and just another voxel-wise inference method. -Tom ____________________________________________ Thomas Nichols, PhD Director, Modelling & Genetics GlaxoSmithKline Clinical Imaging Centre Senior Research Fellow Oxford University FMRIB Centre