The terms "fixed-effects" vs "random-effects" are a twisted mess in the FMRI field that started from the early days.

A random-effects model requires multiple measurements from each subject so that the deviations of each subject from the group average can be modeled as a cross-subjects variable. In this sense the traditional ANOVA with at least one within-subject (or repeated-measures) factor is a random-effects model, but strictly speaking Student t-test is not.

However, the term "random-effects model" is widely used in the FMRI literature unfortunately. What they really mean is that there is an underlying assumption with such a claim: the within-subject variance is the same across all subjects. So it's actually a pseudo random-effects model, and the assumption does not really hold most of the time. Under that assumption, the Student t-test (e.g., with 3dttest++) is basically a special case of the mixed-effects multilevel (or random-effects multilevel) model adopted in 3dMEMA.

On the other hand, the term "fixed-effects analysis" in FMRI is specifically used to refer to the situation with only a few subjects, or when combining a few runs/sessions. And the analysis is done typically with two alternatives: either 1) combine all subjects/runs/sessions' data, and run one regression analysis, or 2) run regression analysis per run/session/subject, and then combine their effect estimates through weighted average but assuming no across-runs/sessions/subjects variability.

But there is another scenario where "fixed-effects analysis" may prevail. Nowadays not many people run "fixed-effects analysis" only. Instead, the typical "fixed-effects analysis" is performed when the individual subject data are analyzed for each run/session separately. In other words, once you have the effect estimates from each run/session, before the group analysis step, you combine the multiple effect estimates from runs/sessions for each subject through the so-called "fixed-effects analysis". In AFNI, you could analyze the individual subject data with all runs/sessions concatenated, so the "fixed-effects analysis" is not always necessary. But in other software packages the concatenation capability does not exist, therefore the "fixed-effects analysis" step is very popular.

Gang



Edited 5 time(s). Last edit at 04/04/2013 12:02PM by Gang.