The Empirical Type I Error of Dynamic Statistical Parameter Mapping Estimates

Although functional neuroimaging has provided methods for measuring and localizing change in metabolic or electromagnetic brain activity, the interpretation of the significance of such changes is challenging. This is at least partially due to the lack of a solid statistical foundation for many of these methods as well as the obvious problem associated with simultaneously performing thousands of statistical tests. Dynamic statistical parameter mapping (dSPM) is a well developed anatomically constrained noise-normalized l2 minimum-norm technique used to functionally map magnetoencephalography (MEG) data. dSPM values are theoretically statistically distributed as the square root of an F-distribution with 1 or 3 degrees of freedom in the numerator, depending on how 'loose' the source orientation is fixed. The validity of this statistical assumption for dSPM values is examined in this study. MEG data from six participants were recorded while the participants looked straight ahead at a cross on a screen for 60 seconds during which no task was performed. Triggers were randomly placed throughout the interval and dSPM values were derived as if the participant was performing a task lasting 100ms. When the entire mock task was analyzed, the number of significant dSPM values was close to the expected probability. However, when the dSPM images were examined over time on an individual level, it was clear that the significant dSPM values clustered spatiotemporally. This suggests that a systematic non-stochastic process may have resulted in this pattern of Type I error.