Hi ,
My current purpose is to analyze univariate datasets of MxN dimensions from M trials and N subjects using a mixed effect model. My dataset is from functional near infrared spectroscopy (fNIRS), so I cannot use FSL. But I am interested in a mixed effect inference, as explained in Beckmann et. al. paper [ref below], in a frequentist framework To begin with, I want to apply the method as explained in example IV-A (Average group activation) in the same paper as follows:
At the first level, I obtain contrast means, variances, and dfs for all
subjects. Then, following example IV-A in the above paper (I refer to it as MFX here), I compute the weighed contrasts for all the subjects, and perform a t-test; the weights are inversely proportional to subject-specific within-subject variances. In effect, it sounds like the SPM type of two-level GLM (lets refer to it as RFX), except that in MFX model, the lower level variances are considered explicitly at the higher level.
As I understand it, the above method (MFX) is a simpler version of the multilevel bayesian FLAME analysis, and should account for any possible heterogeneity among subjects more optimally than RFX method. But there may be some potential caveats in following this simplistic MFX method of estimation. So, I would like to know if my reasoning is correct.
Thanks in advance for your response.
Archana
Ref:
General multilevel linear modeling for group analysis in FMRI
NeuroImage, Volume 20, Issue 2, October 2003, Pages 1052-1063
Christian F. Beckmann, Mark Jenkinson, Stephen M. Smith
___________________________________________
Archana K. Singh
Sensory & Cognitive Food Science Lab
National Food Research Institute
2-1-12, Kannondai, Tsukuba,
Ibaraki 305-8642, Japan
|