AFNI Message Board

History of AFNI updates  

|
May 16, 2022 09:27AM
Hi, Philipp-

With this kind of resting state-like processing, the regression model is basically full of regressors of no interest. The "output" signal for further use is then the leftover residuals. From your applied case, it seems like the question is: is the quality of fitting in the first 400 time points different than that of the remaining ones, such that the residuals would have very different properties? While sleep and rest do have different frequency content and time series characteristics, I would suspect that the quality of time series fit wouldn't be appreciably different (as long as there is no big change in motion profiles, say). I suspect that it should be reasonable to take your full, processed output/residual/errts time series and now, afterwards, chop out the first 400 time points and estimate a power spectrum without too much difference from having chopped ahead of time and processed the "sleep" segment separately. Your errts for time points with indices 400..3000 will likely be quite similar whether you process the full 3000 time points and chop, or chop and then process---with the important caveat that there is nothing "weird" happening or a major difference between the first 400 and later ones (e.g., lots of motion differences, etc.).

I suspect your choice of polynomial regressor and bandpassing was due to this AP note:
-regress_polort DEGREE  : specify the polynomial degree of baseline

        e.g. -regress_polort 2
        default: 1 + floor(run_length / 150.0)

    3dDeconvolve models the baseline for each run separately, using
    Legendre polynomials (by default).  This option specifies the
    degree of polynomial.  Note that this will create DEGREE * NRUNS
    regressors.

    The default is computed from the length of a run, in seconds, as
    shown above.  For example, if each run were 320 seconds, then the
    default polort would be 3 (cubic).

  * It is also possible to use a high-pass filter to model baseline
    drift (using sinusoids).  Since sinusoids do not model quadratic
    drift well, one could consider using both, as in:

        -regress_polort 2         \
        -regress_bandpass 0.01 1

    Here, the sinusoids allow every frequency from 0.01 on up to pass
    (assuming the Nyquist frequency is <= 1), modeling the lower
    frequencies as regressors of no interest, along with 3 terms for
    polort 2.

    Please see '3dDeconvolve -help' for more information.
so that sounds good, as well. That should mean that a pretty reasonable baseline should be fit for the full time series, hopefully.

You could re-process one subject as a "quick" check/verification, I suppose, but I don't see that as necessary, providing other time series checks.

--pt
Subject Author Posted

Cutting a time-series after vs. before preprocessing

Philipp May 15, 2022 01:54PM

Re: Cutting a time-series after vs. before preprocessing

ptaylor May 16, 2022 06:47AM

Re: Cutting a time-series after vs. before preprocessing

Philipp May 16, 2022 07:43AM

Re: Cutting a time-series after vs. before preprocessing

ptaylor May 16, 2022 08:38AM

Re: Cutting a time-series after vs. before preprocessing

Philipp May 16, 2022 08:54AM

Re: Cutting a time-series after vs. before preprocessing

ptaylor May 16, 2022 09:27AM

Re: Cutting a time-series after vs. before preprocessing

Philipp May 16, 2022 09:46AM

Re: Cutting a time-series after vs. before preprocessing

rick reynolds May 16, 2022 10:42AM



Sorry, only registered users may post in this forum.

Click here to login