Is Bonferroni correction more sensitive than Random Field Theory for most fMRI studies?

Abstract

Random Field Theory has been used in the fMRI literature to address the multiple comparisons problem. The method provides an analytical solution for the computation of precise p-values when its assumptions are met. When its assumptions are not met the thresholds generated by Random Field Theory can be more conservative than Bonferroni corrections, which are arguably too stringent for use in fMRI. As this has been well documented theoretically it is surprising that a majority of current studies (~80%) would not meet the assumptions of Random Field Theory and therefore would have reduced sensitivity. Specifically most data is not smooth enough to meet the good lattice assumption. Current studies smooth data on average by twice the voxel size which is rarely sufficient to meet the good lattice assumption. The amount of smoothing required for Random Field Theory to produce accurate p-values increases with image resolution and decreases with degrees of freedom. There is no rule of thumb that is valid for all study designs but for typical data (3mm resolution, and greater than 20 subjects) residual smoothness with FWHM = 4 times voxel size should produce valid results. However, it should be stressed that for higher spatial resolution and lower degrees of freedom the critical smoothness required will increase sharply. This implies that researchers should carefully choose appropriate smoothing kernels. This can be facilitated by the simulations we provide that identify the critical smoothness at which the application of RFT becomes appropriate. For some applications such as presurgical mapping or, imaging of small structures, probing the laminar/columnar structure of the cortex these smoothness requirements may be too great to preserve spatial structure. As such, this study suggests developments are needed in Random Field Theory to fully exploit the resolution of modern neuroimaging.