Abstract
To the Editor:
Holmes and colleagues raised two points of criticism regarding our PET activation analysis (Halber et al., 1997). First, they pointed out that Standard Parametric Mapping (SPM) and SHERLOCK are not directly comparable. We fully agree. Our profound awareness of the necessity for strong control of Type I error induced the “interest and enthusiasm” for Holmes' approach. The allegedly inappropriate 1.64-threshold for SPM resulted from the defaults of the program, as used in many of the images in the approximately 1,200 plus publications involving this method. As a matter of fact, no validation of Type I error adjustment strategies beyond the per voxel-level has ever been clearly shown, and consequently, we used the default SPM smoothness correction performed by estimating the number of resolution elements in the intracerebral space, as proposed by Friston et al. (1991), for assessing clusters of activation. This has been pointed out by Worsley et al. (1992) to be a prerequisite for evaluation of brain activation modelled by stationary Gaussian random fields. However, no further correction of Type I error was applied to our images, and the SPM images shown represent the voxel-level thresholding results. We considered the complex methodological issues related to SPM Type I error correction (Ford 1995) to be far beyond the scope of our article as defined in its title. Given the proven control for Type I error at the image level inherent in SHERLOCK'S step-down approach, the sensitivity of SPM is—in the absence of valid multiple comparison correction—likely to be incorrectly high. “The principal difficulty in your case“, remarked Holmes in his didactic manner, “lay in the fact of there being too much evidence. What was vital was overlaid and hidden by what was irrelevant” (Doyle, 1894).
Second, Holmes and colleagues claimed “a similar lack of consistency in the visual comparison that the reader is asked to make between Figs. 2 and 3”. In fact, we did not state those two figures to be comparable at all. For visual comparison, we provided Fig. 4 to illustrate the advantage of SHERLOCK over SPM95, as detailed in our text. However, to our disappointment, by applying Holmes' method to a well-known activation paradigm in our study, we showed that its sensitivity can be changed at will by varying the filter kernel size, with more and more of the brain becoming “significantly” activated as the kernel size increases, inevitably paralleled by a decrease in resolution. In general, every claim of validity has to supply a valid model of the activation's spatial distribution—beyond any doubt a still unsolved question, “but it needs a great deal of supplementing before anyone could offer an opinion” (Doyle, 1892).