Dear Takayuki,
I'm not sure what SPM is exactly doing when setting variance to unequal for factor subject, but this has a rather drastic effect on the covariance structure in any case (see e.g. the 19th subject in your model "Variance: subject=two within-subject factors = Unequal"). Maybe you should run some additional models with variance set to equal for within-subject factors and unequal for subject to allow for some comparisons.
In general the idea seems to be to use equal variance for factor subject (as this is assumed for SPM's one-way ANOVA within-subject). However, I'm not sure whether setting variance to unequal is theoretically plausible (I guess variance is likely going to be different, often some subjects show a rather large effect, others don't) or whether it would actually produce incorrect/invalid results.
Looking forward to some more information on that issue,
Helmut
|