Thank you, Alfonso, for your post -- yes, that helped a bit. You are right the authors did use the Fisher z-transform.
In the meantime I have gathered a bit more info which was not reported in the papers that clarify things for me a bit, which I'll summarize for others who might be interested. (Thank you also to Christian Büchel and Mikaël Naveau for their help!)
The key point I was missing (which probably was obvious to everyone else) was that they did not simplify z-ify the t-maps but transformed them to a correlation coefficient using a formula that is also used in Christian Gaser's cg_spmT2s() function which apparently is now part of the VBM toolbox:
https://www.jiscmail.ac.uk/cgi-bin/wa-jisc.exe?A2=ind0805&L=SPM&D=0&P=22138
the relevant parts being:
% --------------------------------
% correlation coefficient:
% --------------------------------
% sign(t)
% r = ------------------
% df
% sqrt(------ + 1)
% t*t
df = [xCon(Ic).eidf SPM.xX.erdf];
t2x = sign(Z).*(1./((df(2)./((Z.*Z)+eps))+1)).^0.5;
where the 'sign(Z)' is so we don't lose the sign when we take the square-root of the square, and 'eps' is to add a tiny number so we don't have any division by zero.
Then the output of this calculation was put through the ln form of the Fisher z-transformation
0.5*log((1+i1)./(1-i1))
which, to just expand on Alfonso's explanation, is to make the correlation coefficient, which is bounded between [-1,1] and only behaves normally when r = 0, to behave more normally so you can perform the usual parametric tests and confidence interval calculations when r is not 0 (also has effects of making the variance less dependent on the mean value/stabilizing variance). A *wonderful* discussion of this which also beautifully situates explanations in historical context is in Nicholas Cox's correlation article in a 2008 issue of the Stata Journal (https://doi.org/10.1177/1536867X0800800307).
Now I still have the remaining question of how this t-stat-map to r formula is derived??
I'm stuck at the t-statistic of a regression coefficient estimate b being t = b/SE_b = (b * n^0.5) / SD_b and that b can be shown to be a standardized correlation coefficient, or equivalently r = b_yx * (SS_xx)/SS_yy)^0.5 ....
Gina
|