–2–
Conclusions First
¥ Serial correlation does not appreciably impact the activation magnitudes (b s) estimated using 3dDeconvolve (= Ordinary Least Squares solution)
¥ Group activation maps made from combining these b s using 3dANOVA, 3dLME, etc., are essentially the same using 3dDeconvolve or 3dREMLfit (= Generalized Least Squares solution)
Ø In other words, there is no need to re-run old group analyses to see if allowing for serial correlation will change the results
¥ Thresholded individual subject activation maps are potentially affected, depending on the task timing and on the scanner
H The biggest effect of serial (AKA temporal ) correlation—when this correlation is significant—is on the estimates of the variance of the individual subjectsÕ b s
H If the variance is under-estimated using 3dDeconvolve, then the individual subject t- and F-statistics will be over-estimated
H Individual subject variances and statistics are not usually carried forward to the group analysis level
o Since inter-subject variance is much larger than intra-subject variance
H Thus, group results are only marginally affected by serial correlation