from 3dttest++ to correlations

AFNI Message Board
,
1

Greetings,

I've run 3dttest++ with some covariates and was thinking of converting some of the covariate related Z-maps to correlations, to illustrate brain-behavior correlations?

I tried the code below and I get no error messages, but when I open the newly created corr file there is nothing inside :(

3dcalc -a ~path/tmp_ttest_results_n-26+tlrc[1] -expr 'tanh(a)' -prefix corr

AFNI version info (afni -ver):
Precompiled binary linux_openmp_64: Sep 27 2021 (Version AFNI_21.2.09 'Nerva')

2

That is odd. I just ran that command on a dset I had, and it worked fine.

To check one thing---in some shells, you need to have quotes around the square brackets so they don't get interpreted. For example, trying this:

3dcalc -a ~path/tmp_ttest_results_n-26+tlrc"[1]" -expr 'tanh(a)' -prefix corr

Besides that... Hmm. can you verify that:

3dcalc -a ~path/tmp_ttest_results_n-26+tlrc"[1]" -expr 'a' -prefix test_copy

... copies your dataset?

Separately, I am not sure that it is necessary to convert Z to r for interpretation---many readers would be happy enough with Z? But regardless, it would be good to resolve what is happening on the commmand line here.

--pt

3

Apologies, it does indeed work - I think I just didn't load the corr file in correctly, because I just did "afni corr" instead of "afni corr*".

Separately, I am not sure that it is necessary to convert Z to r for interpretation---many readers would be happy enough with Z? But regardless, it would be good to resolve what is happening on the command line here.

Thank you for raising this! It's my first neuroimaging project.. I do find the Z values to be more informative, but was worried that I couldn't refer to the results as bb correlations..

4

OK, cool. I'm not sure what 'bb corrrelations' are, but I think it is find to just call them Z-scores.

Actually, it is probably more accurate to leave them as Z-scores? If they are coming from model coefficients, wouldn't they be the Z-score value, like the limiting case of an infinite degree of freedom t-test? That is a different thing than a Fisher Z transform of a Pearson correlation r. For example, Pearson correlation r is itself an estimator, and has an uncertainty (which can be expressed, under some assumptions, as a function of the degrees of freedom of the time series that went into it). The Z-score is itself simply a statistic, that's it.

I am not sure which kind of Z you are starting with, but I think there are distinct cases to consider.

--pt

1 Like
5

bb corrrelations ~ brain behavior correlations

Z scores I'm thinking of specifically are the ones coming from the Z maps that 3dttest++ produces by default? I was particularly interested in the ones that survive cluster correction, but now that I think about it the ones that do have such a high Z-value usually that Fisher transforming them doesn't provide a particularly nuanced picture..