Assemble the Data (isxconcat-sess) (Volume)

The first step in the group analysis is to "assemble" the data. This means creating a single 4D file with where the 4th "time" dimension is actual all the subjects concatenated together in a common space. is a different command, depending upon whether the common space is volume- or surface-based.

To run the volume-based concatenation, run the command below. Note that the data from this command already exist in the group-analysis-tut directory.

isxconcat-sess -sf <sessid> -analysis <analysis name> -c <contrast name> -o <output dir>

In the output directory, you will see a series of files that start with "tal":

In the new contrast name folder, you will see:

Group GLM Analysis (Volume)

When you perform a group analysis, you are looking for effects of the task across the population. Now decide if you are running a Random Effects or Weighted Random Effects or Fixed Effects Analysis.

Random Effects (RFx, OLS) (Volume) This is a program for intersubject averaging using a random-effects model, by computing one number for each subject (at each voxel) based on the contrast matrix.

Weighted Random Effects (WRFx, WLS) (Volume) In a real experiment, some subjects are noisier than others, and it is a good idea to take this into account since we have information about how noisy a subject is through the lower-level analysis. In weighted least squares (WLS), this is handled by weighting each subject by the inverse of their noise (i.e., noisier subjects get lower weight). The first-level noise variances (tal.cesvar.nii) are passed with the --wls option.

Fixed Effects (FFx) (Volume) This is a program for intersubject averaging using a fixed effects model, where data from all subjects are treated as if they came from a single subject.

Correction for Multiple Comparisons/Cluster Analysis (Volume)

With so many voxels in fMRI maps, it is very likely that many voxels will appear to be active purely by random chance (ie, a false positive). The is known as the "Problem of Multiple Comparisons". One way around this is to do a cluster analysis in which active voxels are eliminated unless they appear in a cluster, the idea being that false positives will not appear next to each other.