Hi Anthony,

> If the t scores are equivalent between scaled and unscaled, why go to the
> trouble of converting to percent signal change before group analysis? Why not
> just use the t score?

Well, if your only interest is t scores and if you don't care for percent signal change, yes, you don't have to scale the original signal before you run individual subject analysis.

Actually the above statement is true if there is only one run of data. With multiple runs of data, prior scaling is still desirable if all runs are concatenated in the analysis, otherwise variations of baseline across runs could make the analysis go awry.

> As I understand it, the t is a measure of whether the parameter estimate
> (beta) is significantly different from zero. It is given basically as t = beta/standard
> error (beta).

Suppose there is only one run of data. Without scaling let's call the regression coefficient of a condition b_0, and the corresponding t_0 = b_0/se(b_0).

Now if you scale the original signal by a constant C, the new coefficient would b_1 = b_0/C which can be interpreted as percent signal change, but the new t should be the same as before,

t_1 = b_1/se(b_1) = b_0/se(b_0) = t_0.

> This has the effect of scaling because the t value means the same thing
> across subjects, and baseline is taken into account during the deconvolution.
> Higher t's mean better fit to the model. By the way, please correct me if I am
> wrong here.

I'm not so sure what exactly you mean here by scaling: scaling the original signal, or scaling in terms of t score? Well, still assume one run of data for simplification purpose. Keep in mind se(beta) is different across subjects. So whether beta or t score goes into group analysis will have different result in group analysis.

> If fit to the model is what you are interested in, why not calculate the
> average fit to the model and compare across conditions? I am kind of playing
> devil's advocate here, because something about using a test statistic as a DV
> seems wrong, but maybe fMRI analysis is different from a standard behavioral
> study in this respect.

Basically I agree with you here. When percent signal change is not feasible (for example correlation coefficient in functional connectivity), you might have to use z-score. I don't know the context in which z-score from t is used for group analysis: Can you provide the literature if you don't mind?

Gang