 |
 |
Hi, I was wondering if anyone can help me with a strange observation I made. I was analyzing structural MRI data with the general linear model in SPM8 (27 subjects, images were smoothed with FWHM 8mm). First, I did estimate the effects of interest with the ML estimation and everything seemed fine. However, when applying the Bayesian estimation I observed some strange effects for the conditional paramters: The centers of the regions with the largest effects (positive and negative) were deleted. This affects only the centers, the outlines are estimated well. To analyze this issue in more detail I produced a coefficient image with a gaussian blob (in analogy to Penny et al., 2005) and computed simulated images which were smoothed prior to analysis (FWHM 8mm). The results (estimated coefficient maps) for 30 subjects of my fake data set can be found in the attached PDF. As it can be seen the ML estimation detects the whole blob whereas the Bayesian estimation seems to overwrite part of the blob. What I found out so far is that this strange effect decreases with increasing FWHM, but only for the simulated data set. Increasing sample size also helps to decrease the effect, but not in the same amount as increasing FWHM. I am very grateful if someone could help me with this problem. Best. Paul