 |
 |
Dear SPM list, I have a quick question about what assumptions need to hold in order for an FDR-correction to be valid. I have a bunch of fMRI and behavioural measures. I then calculated the matrix of correlations, and applied FDR to correct for multiple comparisons. A significant correlation that emerged, which I was particularly interested in, is between an fMRI measure and a behavioural measure. The following objection was raised: because the brain-with-brain correlations and behaviour-with-behaviour correlations will tend to be strong, and will therefore contain a lot of small p-values, this artificially makes it easier for brain-behaviour correlations to pass the FDR test, as their not-so-low p-values will get put into the same sorted p-value list as the brain-brain and behav-behav ones. Therefore, according to this objection, the FDR test of the brain-behav correlation is not valid, because it is not "conceptually homogeneous". I can see a certain logic to that argument, but consider the following different example: suppose I want to apply brain-wide FDR to, say, a visual working memory task, where the contrast is simply task-minus-fixation. Because of the visual stimulation, there will be large areas of strong activation in the visual cortex. There will probably also be a smaller area of weaker activation in prefrontal cortex. If I apply brain-wide FDR and find that the small prefrontal region passes the FDR test, is this rendered invalid by the fact that the sorted p-value list also contains lots of low p-values from the visual cortex activation, which is "conceptually inhomogeneous" because it isn't memory-related? This strikes me as quite an analogous situation, but in this case all of the data are from the same fMRI measurement. Although the idea of "conceptual homogeneity" does make some sense, it certainly seems to be a consideration that is outside of the domain of statistics. It therefore seems to me that it isn't the kind of thing that can determine whether or not a statistical test is valid or not. E.g. it's not like homogeneity of variance, which is purely statistical and well-defined, and which of course does enter into deciding the validity of tests. So, my question is: is my mixed-bag brain-behaviour correlation FDR-correlation valid or not? Or, put another way, is "conceptual homogeneity" an implicit extra-statistical criterion that any valid FDR test must satisfy? Any comments greatly appreciated. Many thanks, Raj ------------------------------------------- Rajeev Raizada, Ph.D. Postdoctoral Research Fellow Institute for Learning and Brain Sciences University of Washington Box 357988, Seattle WA 98195 E.mail: [log in to unmask] Tel: 206 221 6415 Fax: 206 221 6472 WWW: http://faculty.washington.edu/raizada