correlation of time series with 3D+time

AFNI Message Board
1

I need to find correlation of 3D+time with a time series I found function 3dTcorr1D which seems to do just that

2 questions:

  1. I need to get the corrcoef along with the significances (or some stat that I can threshold for significance) in the description of function there is no mention of whether p values can be accesses ( e.g. looks like only 1 volume is returned per column of 1D input)

  2. -pearson correlation option is mentioned to be a dot product. Is this calculated after subtracting the averages from to be correlated signals or should we demean the inputs before ourselves?

thanks
Ali

2

Ali, if you just have one time series to correlation the 3D+time dataset, check out step (4) on the following page: https://afni.nimh.nih.gov/SimAna

3

Also from the help of 3dTcorr1D:

  • No detrending, blurring, or other pre-processing options are available;
    if you want these things, see 3dDetrend or 3dBandpass or 3dcalc.
    [In other words, this program presumes you know what you are doing!]
4

Hi Ali,

Jumping in for more fun…

  1. Gang’s suggestion of 3dDeconvolve or 3dfim+ can get you p-values,
    or you could run cdf yourself. For example, if a time series of length 100
    yields a correlation of 0.3, then the p-value might be 0.002426, from:

cdf -t2p fico .3 100 1 1

That assumes only demeaning the data (as is done with 3dTcorr1D).
If quadratic detrending were done, the ort-dof would be 3, leading to
a p-value of 0.00269, from:

cdf -t2p fico .3 100 1 3

See the output of “cdf -help” for details.

  1. 3dTcorr1D removes the mean, but does no trend removal.
  • rick
5

Thank you my friends for suggestions and clarifications. I want to used 3dTcorr1D because I want to do multiple correlations at the same time ( i prefer to avoid for loops). signals to be correlated are concatenated 3D beta values from individual runs and a behavioral measure relevant for each run. so I do not need detrending.

In this case say I use z transform to calculate p values for correlations (probably using 3dcalc), I can then use this pvalue brick to threshold the correlations overlay (unless there is a better solution in afni?). Also may I suggest to add stats to 3dcorr1D in future? no one can report values without stats these days.

6

I ran 3dTcorr1D and realized it outputs p values that are shown in afni. That is what I was looking for so I do not have to calculate myself.