Section IX: Across-Subject Group FEAT
The group analysis shows activation (or deactivation) trends across a population of subjects.
If it is done within subjects it is called ‘fixed effects’ because the brain is constant across subjects. Here it is done across subjects it is called a ‘mixed effects’ because the brain is not a constant across subjects. These comparisons have less power than within subject comparisons. It is a historical anomaly within functional neuroimaging that SPM users in the 1990s and early 2000s regularly applied ‘fixed effects’ models to populations (when they should have been using a ‘mixed effects’ model.). This produced positive results where there should have been none.
Something special happens, conceptually, at the group level: you can tell the computer that individual subjects differ from one another along a single variable, and the computer can find out whether this variation predicts brain activity. This kind of thing would not have made any sense at the individual level.
It is at the group level that you can classify people by how depressed they are, whether they are male or female, whether they suffered a distant or recent loss, or whether they had a cat or dog.
As a result, two new pieces of information will become available to you:
1. Whether two populations are different
2. Whether a score on a continuous variable (eg height) correlates with brain activity (eg amygdala activation)
All of this will happen through
Please note that as described above in the section on preparing behavioral data as regressors, it sometimes can happen that one subject will lack a regressor that another subject has. For example, one subject may have claimed to feel no anxiety throughout a study, while another states that he did. The first subject cannot have an anxiety regressor. THIS SCENARIO MAKES COMPARISONS OF .FEAT DIRECTORIES ACROSS SUBJECTS IMPOSSIBLE. In this case you are going to have to compare individual 3D cope images (cope.hdr files within cope directories), selecting ONLY those that are common to all subjects. This will be explained below.
Step 1: Misc Tab Differences
In group level analyses you have the option to cleanup first level standard space images; this wasn’t possible at the individual level. This is a bad thing and you should say no to it.
Cleanup first-level standard-space images
= NO yellow check box. This will screw you up big time if you leave it checked.
Under the DATA tab of FEAT we will start from the bottom.
Saving the .fsfSaving the .fsf file, so you can run it again on a different subject, or re-run it on this subject. Press save and in the box that pops up, determine what directory you want to save it in (for programs that can be used on multiple subjectsthink of a good name
you will specify that inputs are lower-level FEAT directories – what this means is you are doing a group analysis. In this case, the group is within a given subject. You tell it to expect two analyses – meaning two FEAT directories. You then select these feat directories, by clicking on the select FEAT directories tab, as shown below.
This box pops up, and you navigate two the two directories you want. REMEMBER THE ORDER you put these in, it is crucial for determining your subtractions.
Group Analysis Regressors – Loading them Into FEAT
Once you have run all the individual analyses, you are ready for a group analysis, as described above. However, what if individuals in a group vary along some variable – say their gender, their HAMD score, their reaction time, or some other variable that you think may account for some of the variance in the data? Easy: you create a regressor.
First demean these values. In SPSS this is as simple as running a descriptive statistic on the variable, yielding the mean score (eg HAMD avg = 6). Subtract this from each individual subject by computing a new variable, DM (for demeaned) variable name – eg, DMHAMD.
SPM does not require the following step, but FSL DOES require that you normalize these demeaned values by dividing by the maximum, ensuring a range from –1 to 1. For some reason FSL can’t deal with numbers greater than 1 or less than –1.
The details of entering this information are kind of annoying because FSL is fussy. On the FEAT template, click stats -> full model setup -> Evs -> add an EV under ‘number of Evs’ -> in the resultant box type in the regressor values IN THE FOLLOWING WAY:
1. Type the digits from the keyboard not the numeric keypad
2. Type in the digits without any decimal points
3. Once the digits are correct, arrow key to the place where the decimal is, and enter it
4. IF it’s a negative number, now put that in
5. Don’t ask why this crazy system is necessary
6. Don’t try to paste values in
7. Don’t think you can avoid normalizing to a –1 to 1 range.