I'm others can provide more detail on this issue but on a basic level,
exceedance probability is determined in the context of the model space
used for comparison, it is a relative figure limited by the number of
models. Although it is sort of curious that the direction of the
exceedance probability between two models would chance between
comparisons, by adding other models into the space, it is in some ways a
different comparison that you are making which could allow for this
reversal in evidence.
-Drew
On 2/20/14 11:22 AM, "Carmen" <[log in to unmask]> wrote:
>Dear SPM- and DCM-Experts,
>
>I performed a rfx - Bayesian model selection analysis with DCMs modelling
>the influence of a drug on the connectivity of different pain-processing
>regions (direct input: pain, modulatory input: drug). This resulted in
>three models among 64 models that best explain the influence of the drug.
>In a further approach, I estimated the best model of the first study for
>another data set investigating the same drug in the same study design but
>in another experimental pain condition and compared for these data the
>best model (model 1) with a model without drug influence (model 0, no
>modulatory input). Here, the model lacking drug influence turned out to
>be superior.
>However, when I test not only the best, but the 3 best models of the
>first analysis against a model without drug influence (model 0), the
>exceedance probability of model 0 is relatively worse than model 1. Why
>does the presence or absence of other models reverses the ratio of these
>two models?
>
>Thank you for your help!!
|