Mapping voxel-based statistical power on parametric images.

Using a classic technique based on the noncentral F-distribution method for computing statistical power, we developed a general approach to the estimation of voxel-based power in functional brain image data analysis. We applied this method to PET data from a large sample (N = 40) of subjects performing the Wisconsin Card Sorting (WCST) paradigm analyzed with SPM95, produced statistical power maps for a range of samples sizes and smoothing filter widths, and examined the effects of sample size and image smoothing on the expected reliability of activation findings. At an uncorrected alpha of 0.01, a fixed filter size of 10 mm3, and a range of power thresholds, maps revealed that the power to reject the null hypothesis in brain regions implicated in the task at Ns of 5 and 10 may not be sufficient to ensure reliable replication of significant findings and so should be interpreted with caution. At sample sizes approaching 20 subjects, sufficient power was found in the right dorsolateral prefrontal cortex (BA 46/9), right and left inferior parietal lobule (BA 40), and left inferior temporal lobe (BA 37), comprising the cortical network typically observed during the WCST. Filter size needed to maximize power varied widely, but systematically, across the brain, tending to follow known neuroanatomical landmarks. Statistical power considerations in brain imaging studies are critical for controlling the rate of false negatives and assuring reliable detection of cognitive activation. The variation of filter size for maximizing power across the brain suggests that the underlying neuroanatomy of functional units is an important consideration in the a priori selection of filter size.