Hi Wil,
We encountered the same issue recently. We wanted to add more than one
parametric regressor to a condition without SPM orthogonalization among
them .
I don't think commenting out spm_orth is a good idea as (I think) you do
want to orthogonalize the regressors with respect to the main event
(without doing so the regressors are highly correlated with the main
event due to sparseness, I believe). Also, if you comment it out, there
will be no orthogonalization of higher order modulations (if you chose
to add these).
Instead, we changed slightly spm_get_ons so spm_orth does not take as an
input the entire matrix of the condition+all regressors at once but
instead only orthgonalizes each individual regressor with respect to the
main condition (a simple loop). This seemed to work for us but I will
appreciate any comments from more experienced users.
asaf
William Cunningham wrote:
> Over the past few months, I have been following the discussion of the use of
> spm_orth.m to orthogonalize basis functions.
>
> If I am reading the thread correctly, spm_orth serially orthogonizes
> regressors such that the first regressor remains the same, and each
> subsequent regressor is orthogonized to the one (or ones before it).
>
> This suggests (and was noted in the listserve) that the significance values
> and Beta weights will differ as a function of the order of entry. This is
> not a problem if one is interested in “cleansing” the main effect, but this
> (if I am reading right) becomes a problem if one wants to interpret the
> value of parametric regressors.
>
> Take for example a hypothetical study in which one wants to example the
> activity due to difficulty and reaction time (which are correlated to some
> extent). Ideally, you would want to example the voxels that significantly
> relate to difficulty (controlling for RT) and then the ones that
> significantly relate to RT (controlling for difficulty).
>
> If you get different results putting RT in first, or difficulty in first,
> this makes drawing conclusions difficult.
>
> One solution on the list has been to comment out spm_orth (or comment out
> all the commands in spm_orth itself). What worries me is that SPM was
> written to have that command explicitly in the sequence, and I was wondering
> if there are unforeseen consequences to do this (for example, does the time
> derivative no longer function properly?).
>
> I guess my question simply comes down to:
> (a) how problematic is it leaving spm_orth as it is as default if one wants
> to interpret the beta weights as independent parametric regressors
> (controlling for one another)?
> and
> (b) are there probems with commenting out all the code in spm_orth so that
> it no longer does anything?
>
> Thanks in advance for any insights.
> Best,
> Wil
>
|