Dear Lisa,
from my point of view this issue is not "DCM-specific" because
statistical significance is not solely reflected by a large difference
from zero, but depends on this difference related to its variance. As
you wrote all your parameters are positive (ranging from .00002 to
.0044), which means that 100 % of your observed data lie beyond the
value under the null-hypothesis.
In other words it seems as if the rate of change you report might be
(quite) small but it is measured with high precision yielding the
magnitude of the effect significant.
From my point of view different DCMs should primarily be evaluated in
terms of their posterior probabilities relative to the rest of the model
space (especially in multi-subject designs). The parameters of the
winning model might of course be important for interpreting the model
superiority. One might think, however (at least hypothetically), of
crucial connections whose mean across subjects is zero, because one
subpopulation in the sample increases the respective rate of change
whereas another decreases it (and both do it in a rather systematic than
a random fashion).
So if your winning model is interpretable and is in line with your
hypotheses about the "truth", you shouldn't worry about small effect
sizes in the coupling parameters as long as the winning model's
superiority is evident. There is even less reason to worry in your case
because the parameters *are* significant although the effect size in
terms of rate of change is small.
Go ahead and publish your data,
Thilo
On 07/11/2013 11:04 AM, Lisa Bulganin wrote:
> Dear DCM experts,
>
> few days ago I sought your assistance in the matter of coupling parameter in nonlinear DCM. Please see original message below. Is any information missing to solve the problem?
>
> Kind regards, Lisa
>
>
>
> Dear DCM experts,
>
> I am a PhD student using DCM for the first time. In the current project, I am using nonlinear DCM to answer the question whether the connections between different areas in network A (consisting of 3 areas) are modulated by two experimental conditions and by the activity of network B (consisting of 2 areas).
>
> I constructed 44 models, of which one was superior. I tested whether the coupling parameters for the B- and D- matrices are different from zero and found significant effects that are consistent with our hypotheses. But I am surprised by the small Hz values of the coupling parameters (ranging from .00002 to .0044). My first question is: Can I trust the significant effects despite the small values? Does anyone have an idea why these values are so small?
>
> Also, our GLM analysis showed different activation peaks in network B for condition 1 vs. control than for condition 2 vs. control. One of the two areas of network B was only activated by condition 1, but not by condition 2. I used only the peaks from condition 1 in the DCM analysis but tested for modulatory effects of both conditions. Now my second question is: should I run two different DCM analyses to consider the varying peaks, i.e. run a separate analysis for each condition?
>
> Many thanks in advance!
> Best, Lisa
>
|