Dear Esa,
> We are applying subject specific subtraction analysis (PET/SPECT
> models: Multi-subj: cond x subj interaction & covariates) as a fixed
> effects analysis in a PET study of emotion (2 conditions (neutral and
> specific emotion), 4 replications/condition, 11 subjects, 88 scans). We
> are wondering which is the better: subject specific subtraction or
> replication by condition model.
>
> In our data, the residual variance of subject specific model is greater
> than of replication by condition model and thus subject specific model
> yielded less extensive activations. In addition, emotions are subject
> specific experiences which might be handled by subject specific
> models.
>
> Which of the two models would you recommend? Any comments are greatly
> appreciated.
This is an issue of model selection. The principled approach would be
to test for the contribution of the extra effects included in
progressively more comprehensive models to justify their inclusion (in
a step up procedure). There is no right answer here; it is a balance
between modelling various interactions among condition, replication and
subject (reducing error variance) and omitting these interactions
(increasing the degrees of freedom). It sounds as if the condition x
replication interaction (i.e. the differences in adaptation among
conditions) is greater than the condition x subject interaction (the
differences among subjects).
Note that if you modelled all the interactions you would have no d.f.
to estimate the error variance.
I hope this helps - Karl
|