14
Outline of SPM and FSL Approaches
¥ SPM5 and SPM2
H Estimate fixed ARMA(1,1) (more precisely, AR(1)+white noise) model for all Òvoxels of interestÓ (pass an OLSQ F-test)
oBy averaging estimated auto-covariance matrix from OLSQ residuals over these voxels
oSPM assumes AR parameter a È 0.2, and approximates ARMA(1,1) correlations via linear Taylor series, to make correlation parameter estimation easier to program
H Use GLSQ (same for each voxel) to solve for b s
oSPM99: Use OLSQ and adjusts DOF downwards to allow for serial correlation
¥ FSL and FMRIstat (similar, but differ in important details at several points)
HUse OLSQ to get first-pass residuals; use these to estimate each voxelÕs auto-correlation matrix; smooth these matrices spatially (FSL & FMRIstat vary here)
HEstimate AR(1) parameter for each voxel separately from smoothed matrices
HUse GLSQ (different for each voxel) to solve for b s
¥ All these programs use a non-REML method to estimate serial correlation parameter(s) from the OLSQ residual auto-correlation matrix, and then adjust these estimates to reduce the bias thus introduced