[FieldTrip] ft_regressconfound

[Stolk, A. (Arjen)]

Dear Raghavan,

> I have another question. I understand ft_regressconfound must be the
> last
> step. However, can time frequency calculation be done on data after
> employing regressconfound? Or would you suggest doing trial by trial
> time
> frequency calculation, then compensate using regressconfound and then
> compute timelock? Is there a reason not to do the former way?

I would recommend doing the latter, provided that you meant to say 'compute statistics' instead of 'compute timelock'. When ft_regressconfound is applied to raw/ERF data, trial-by-trial adjustments are made to the data, according to explained variance by head movements (and thus differences in distances to the sensors). Subsequently performing time frequency analysis on this data may give a distorted view of signal frequency powers. 

Note that if a source reconstruction analysis is also one of your follow-up steps, it is recommended to do this on the 'original' data, and then use ft_regressconfound at the source level. See 'Practical issues' on this page for the rationale behind this:

http://fieldtrip.fcdonders.nl/example/how_to_incorporate_head_movements_in_meg_analysis

In sum, ft_regressconfound is best used as a last step prior ft_xxxstatistics.

Yours,
Arjen


----- Oorspronkelijk bericht -----
> Van: "Raghavan Gopalakrishnan" <gopalar.ccf at gmail.com>
> Aan: fieldtrip at science.ru.nl
> Verzonden: Maandag 10 februari 2014 17:02:34
> Onderwerp: Re: [FieldTrip] ft_regressconfound
> Dear Arjen
> Thanks for your prompt response. I understand your point regarding
> using
> ft_regressconfound with all 4 blocks together. I think that makes more
> sense.
> I have another question. I understand ft_regressconfound must be the
> last
> step. However, can time frequency calculation be done on data after
> employing regressconfound? Or would you suggest doing trial by trial
> time
> frequency calculation, then compensate using regressconfound and then
> compute timelock? Is there a reason not to do the former way?
> 
> Thanks,
> Raghavan
> 
> Dear Raghavan, It is indeed recommended to use ft_regressconfound as a
> last
> step prior to statistical assessment. It will remove trial-by-trial
> variance
> in the neural data that can be attributed to trial-by-trial variance
> in head
> position. The latter is approximated with regressors containing
> trial-by-trial information on head positions deviating from the
> session/experiment mean. Because the head position timeseries is
> mean-subtracted, the mean neural activity over trials is not affected
> by
> ft_regressconfound; only the variance over trials is, which should
> result in
> a cleaner representation of the data. In order to estimate the
> contribution
> of different head positions to the neural data, ft_regressconfound
> relies on
> general linear modeling. Applying ft_regressconfound to the four
> blocks
> separately (your option 1) will involve four different model
> estimations and
> their associated errors. Because these errors may differ per
> estimation, the
> quality of treatment of the neural data may also differ per block.
> This will
> not affect the mean neural activity in each block, but it may affect
> the
> grand mean over all four blocks as for trials in one block more
> contribution
> from head position may be regressed out than for another. Applying
> ft_regressconfound on the data of the four blocks together (your
> option 2),
> will not affect the grand mean over the trials from all four blocks.
> It will
> reduce the influence of head movement on trial-by-trial variance in
> neural
> activity. This can be for better, or for worse: namely, if there are
> consistent differences in head positions between two conditions
> (captured in
> those four blocks), it may bring the means of neural activity evoked
> in
> these two conditions closer to each other, reducing effect sizes. In
> fact,
> the employment of ft_regressconfound allows one to make a good case
> that an
> observed effect (i.e. differential neural activity between the two
> conditions) cannot be attributed to differences in head positions when
> recording those conditions (note that the same analysis could also be
> performed with eye-movement related activity, or any other measure of
> a
> potential confound). Hope this helps, Arjen ----- Oorspronkelijk
> bericht
> -----
> > Van: "Raghavan Gopalakrishnan" <gopalar.ccf at gmail.com>
> > Aan: fieldtrip at science.ru.nl
> > Verzonden: Vrijdag 7 februari 2014 17:03:40
> > Onderwerp: [FieldTrip] ft_regressconfound
> > Dear all,
> > I am using regressconfound on Neuromag data. In CTF data, data for
> > coil1, coil2 and coil3 are generated and then circumcenter function
> > is
> > called to compute 3 translational and 3 rotational dof.
> > Unlike, CTF, Neuromag maxfilter allows to compute CHPI or QUAT
> > channels that provide quarternion parameters q1 through q6, where
> > q4,q5 and q6 are translations in x, y and z directions. I am using
> > these q4,q5 and q6 to directly compute the rotational orientations
> > (using the last part of the circumcenter.m script).
> > It is said regressconfound must be used as a last step prior to
> > stats.
> > I have 4 blocks of data for each subject. Which option below should
> > I
> > follow?
> > 1. Apply regress confound separately to four blocks? But, then I
> > have
> > to average these four blocks once again using ft_timelockanalysis,
> > then grand average using ft_timelockgrandaverage before computing
> > stats.
> > 2. Or should I append the four blocks first, then perform
> > regressconfound? In this case, I directly go to grandaverage and
> > stats.
> > Any suggestion is appreciated.
> > Thanks,
> > Raghavan
> 
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip

-- 
Donders Institute for Brain, Cognition and Behaviour
Centre for Cognitive Neuroimaging
Radboud University Nijmegen

Email:  a.stolk at donders.ru.nl
Phone:  +31(0)243 68294
Web:    www.arjenstolk.nl