Estimating sample size in functional MRI (fMRI) neuroimaging studies: Statistical power analyses

Estimation of statistical power in functional MRI (fMRI) requires knowledge of the expected percent signal change between two conditions as well as estimates of the variability in percent signal change. Variability can be divided into intra-subject variability, reflecting noise within the time series, and inter-subject variability, reflecting subject-to-subject differences in activation. The purpose of this study was to obtain estimates of percent signal change and the two sources of variability from fMRI data, and then use these parameter estimates in simulation experiments in order to generate power curves. Of interest from these simulations were conclusions concerning how many subjects are needed and how many time points within a scan are optimal in an fMRI study of cognitive function. Intra-subject variability was estimated from resting conditions, and inter-subject variability and percent signal change were estimated from verbal working memory data. Simulations derived from these parameters illustrate how percent signal change, intra- and inter-subject variability, and number of time points affect power. An empirical test experiment, using fMRI data acquired during somatosensory stimulation, showed good correspondence between the simulation-based power predictions and the power observed within somatosensory regions of interest. Our analyses suggested that for a liberal threshold of 0.05, about 12 subjects were required to achieve 80% power at the single voxel level for typical activations. At more realistic thresholds, that approach those used after correcting for multiple comparisons, the number of subjects doubled to maintain this level of power.