Dear SPM- and DCM experts,
I am currently working on a way to describe functional hemispheric
dominance by means of connectivity. Following the example of Frässle
et al. (2016, “mechanisms of hemispheric lateralization: …”) I want to
compare intrinsic and modulatory (fMRI-)DCM parameters from both
hemispheres. More precisely, I aim to assign a value to each of my
subjects in a fashion close to the lateralization index often used for
activation magnitude or cluster size, i.e. (left - right)/(left +
right). It turns out, however, that there are some obstacles inherent
to my approach which I am going to describe in the following:
The main issue preventing me from simply applying the classic LI to
DCM parameters here is that this only works for two positive values.
In the case of DCM parameters, the LI should be suitable for two
negative parameters as well (by inversion of the result’s sign), but
it is certainly not suited for one positive and one negative parameter
and yields results outside the limits of -1 and +1. While computing
the difference between a positive and a negative parameter would work
perfectly fine (numerator), the problem seems to be the normalization
(denominator), meaning I am now left with the evaluation of how big
this difference is in relative terms.
I have tried transforming sets of parameters of interest across all
subjects into a positive range to prevent negative values and then
computing the classic LI. However, it seems inappropriate to just
treat the most negative/positive parameter or basically any arbitrary
value as the limits of this set. Additionally, this results in smaller
normalization terms for parameters that were negative before
transformation and larger ones for parameters that were positive
before transformation. Therefore, I abandoned this approach in favor
of a more promising one.
This time I transformed each subject’s entire set of parameters (A, B
and C) into a common size. I did that by shrinking or stretching each
set (preserving proportionality within the set) so that the average
absolute parameter size is equal across all subjects. This essentially
serves as a sample-dependent normalization so that the simple
difference of one subject’s two parameters of interest should now
serve as a comparable measure of lateralization in the context of this
sample.
However, for this procedure I am assuming that the absolute size of a
DCM parameter is not clearly interpretable per se or between subjects
but only when comparing it with other parameters from the same
subject/DCM. To my knowledge there is no rule of thumb on what is a
small, medium, or big parameter of effective connectivity (is there?),
which leads me to believe that it differs from one DCM to another. To
me, this also corresponds with the fact that in some DCMs I find the
parameters in general to be pretty large and in others rather small,
while showing related connectivity patterns.
A brief example:
Let’s say I have inverted a DCM with 3 regions in each hemisphere and
many connections between and within both hemispheres and now I am
interested in the lateralization of the intrinsic connectivity from
left to right area M1 (“l2r”) and from right to left area M1 (“r2l”).
Both subject X and subject Y have l2r = 1.0 and r2l = 0.5 implying a
stronger intrinsic information transfer towards the right M1. However,
subject X’s parameters in general are large while subject Y’s are
small. Let’s say subject X’s average absolute parameter size is 0.9
and subject Y’s is only 0.3.
Am I now correct in assuming that the two parameters of interest
should be considered more powerful within subject Y’s set of
parameters? And if yes, is it adequate to assume that they are in fact
about 3 times more powerful (0.9/0.3 = 3) and that this also applies
to the difference ((1.0 – 0.5)*3 = 1.5)? Or am I better off just
interpreting the absolute difference (1.0 – 0.5 = 0.5 in both cases)
regardless of each subject’s average parameter size?
I am well aware that this form of normalization is quite different
from the one applied in the classic LI, but I could not find any
simple solutions in the literature or by myself. I have applied my
normalization approach to a large sample of simulated DCM parameters
(10,000 subjects x 38 parameters) and correlated the resulting
differences with the classic LI for cases where both parameters of
interest share the same sign (allowing the classic LI to yield valid
results), Pearson’s r was ~ .88. When using the unnormalized
differences instead, r drops to ~ .76, suggesting that I should go
with the normalized differences instead of the “raw” differences to
better approximate the classic LI. However, my knowledge regarding the
interpretation of DCM-parameters is still somewhat limited, so I hope
you can provide me with some insights. Any help would be greatly
appreciated! Thanks in advance!
Best regards,
Rafael Maleki
|