Hello all,
I'm interested in getting local estimates of temporal autocorrelation in SPM, and for that purpose used Bayesian 1st-level estimation.
The fMRI data I used to test that have 3 sessions of 128 scans at a TR of 2 s and 64x64x23 voxels of size 4x4x4 mm, unsmoothed, of which approximately 13,000 are within brain.
I then extracted the AR coefficients (order 6) for the first session (Sess1_AR_0001.nii to Sess1_AR_0006.nii) and used the Yule–Walker equations iteratively to obtain the corresponding autocorrelation function across lags 0 to 128.
The results are strange (see attached plot):
– In 16 voxels the AR coefficients describe a non-stationary process. After excluding them:
– At lag 127, 7906 voxels have an autocorrelation > 1e-6, 1651 voxels > 1e-3, and 113 voxels > 0.1.
– The largest negative autocorrelation at lag 127 is -2.02e-9. If I take that as an indicator of numerical/estimation precision, there are 8185 voxels where the autocorrelation at lag 127 is different from 0 (> +2.02e-9).
This makes me suspect that the AR estimation in "Bayesian 1st-level" is not very reliable. Is there something I might have done wrong?
Is there a recommended postprocessing for the AR coefficients or autocorrelation functions?
I thought about tapering à la FSL, or clustering as a crude form of spatial regularization.
Or should I simply exclude voxels with unbelievably long-range autocorrelation?
Thank!
Best,
Carsten
|