Did you visualize some of your power spectra, both before and after taking the log (or log-log, in your case) of it? Are you sure you aren't including any super-low freqs that are essentially baseline, that might skew your results? A super high exponent for a power spectrum suggests that baseline might be left in your data.
I would try visualizing your data (here, power spectra, etc.) as a first means for QCing it.
I am also not sure what a statistically appropriate way to average PLEs across voxels would be.
--pt