Print

---

Dear Pierre, Barry and Others,

I share some of the views raised by Pierre.  I also have some
problems understanding why we look for condition-specific
differences in an ill-fitting model.  What is, after all, the
point of interpreting parameters that are meaningless (in
modelling terms)?

There are consistency problems that are prior to this one.
What if the same model is statistically significant (p>0.05)
for one subject and not for another subject?  Should we just
report group data even though none of the subjects really behave
like the group data?  Should we conclude that anatomy is
different for different subjects?

I think we must be careful with maverick interpretations on
ill/good fitting models or condition-specific differences
encountered in group/single subject analysis.  We need to
stop drawing conclusions about stimulus-induced connectivity
that are more appropriately attributed to subject-specific
effects.  I think a far greater body of neuroimaging data
is required for a fair critical assessment of models and
condition-specific differences.

Barry raises a very important point when he says that "a
reasonable fit of model to data is desired, but having
some arbitrary threshold that one always uses may be
unwarranted."  This is no excuse for the goodness of fit
of a specific model not being published.  Independently of
the criteria one uses to accept/reject a model one must
still inform the reader of the goodness of fit indices.
Without these the reader may have the impression that
the authors didn't look at them or the fit was so awful
that they didn't have the courage to print it.

Part of the goodness of fit problem is that in fMRI we are
modelling minute changes in signal using the covariance
matrix.  The main component of the correlation among
different areas is not the signal, rather it is the
noise/smoothing/etc that is driving the covariance matrix.
If noise component changes between two sessions or two
days you may have one subject fitting your model and the
other not. The question here is how reliable will be the
differences that you may find between different conditions?

Quite frankly I don't know the answer to most of the issues
that I've raised.  I just wanted to alert people to think
about them when they carry out their modelling.

Miguel