Abstract
A recent paper by Eklund et al. (2012) showed that up to 70% false positive results may occur when analyzing functional magnetic resonance imaging (fMRI) data using the statistical parametric mapping (SPM) software, which may mainly be caused by insufficient compensation for the temporal correlation between successive scans. Here, we show that a blockwise permutation method can be an effective alternative to the standard correction method for the correlated residuals in the general linear model, assuming an AR(1)-model as used in SPM for analyzing fMRI data. The blockwise permutation approach including a random shift developed by our group (Adolf et al., 2011) accounts for the temporal correlation structure of the data without having to provide a specific definition of the underlying autocorrelation model. 1465 publicly accessible resting-state data sets were re-analyzed, and the results were compared with those of Eklund et al. (2012). It was found that with the new permutation method the nominal familywise error rate for the detection of activated voxels could be maintained approximately under even the most critical conditions in which Eklund et al. found the largest deviations from the nominal error level. Thus, the method presented here can serve as a tool to ameliorate the quality and reliability of fMRI data analyses.
Highlights
Even after many years of sophisticated functional magnetic resonance imaging (fMRI) analyses some debate remains about the validity of models based on the general linear model approach (GLM)
Even after many years of sophisticated fMRI analyses some debate remains about the validity of models based on the general linear model approach (GLM)
A main challenge for linear model estimations used in standard software is that the assumption of independent residuals, which is necessary for classical general linear modeling, usually does not hold for fMRI data
Summary
Even after many years of sophisticated fMRI analyses some debate remains about the validity of models based on the general linear model approach (GLM). Due to the experimental setup and the underlying physiological processes, a temporal autocorrelation exists between successive measurements This may mainly result from missing signal components in the model as well as from the extended time course of the so-called hemodynamic response function (HRF): even for very short stimuli the HRF usually rises during the first 6–8 s, declines slowly thereafter, and, after a so-called undershoot, returns to the baseline. An interstimulus interval of typically 2 s will lead to overlapping signals and temporal correlations Other factors causing this correlation may be technical properties of the scanners (changes in the magnetic field, see Smith et al, 1999) or influences of breathing and pulse or other artifacts. Under the assumption that the temporal correlation can be modeled by an AR(1) process, the correlation structure is estimated using a restricted maximum likelihood approach (ReML; Ashburner et al, 2013, Section 8.9) and plugged into the transformation matrix for prewhitening (Friston et al, 2007)
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call