First, the full F value for each voxel is a measure of significance on the following hypothesis:

H_0: all regressor coefficients are 0 (the signal was just noise without any activation happening during the time course)

versus

H_1: at least one regressor coefficent is non-zero (there was some activation)

It seems to me that this full F test would only tell you all the activated voxels, but it does not differientiate which predictor/task/condition/stimulus the activation was from. Therefore I don't understand why you would want to do group analysis on the full F values, except that you would like to check how the whole fitting does across all 27 subjects.

Strictly speaking, 3dttest does not work on t values directly. Instead, t test assumes data from original sampling, being normally distributed, and with other assumptions (such as independence). It is usually not a good idea to run t test on t-statistic since it violates the assumption of t test, but from time to time some people do it anyway. In doing so, people usually convert their t values to z scores, and then run t test.

For the full F in 3dDeconvolve, the first degree of freedom is 1 when the difference between minlag and maxlag is 0. Although it is true that t^2 = F in this case, converting from t to F is legitimate while the other way around (i.e., 3dcalc -a "file[1]" -expr "sqrt (a)" -prefix fileFtoT) is unlikely since there is no direct way to figure out what sign (+ or -?) t should be assigned when t is calclulated from F^(1/2).

People usually run group analysis based on the specific regressor coefficients with either 3dttest or 3dANOVA. If you still prefer to do group analysis on the full F, you could use 3dmerge with its option -lzscore to convert the full F to an equivalent Gaussian distribution, and then run t test on it.

Gang