Dear Phil
What you're describing is best addressed using the PEB approach, rather than RFX BMS:
https://en.wikibooks.org/wiki/SPM/Parametric_Empirical_Bayes_(PEB)
Best
Peter
On 25/02/2021, 22:54, "SPM (Statistical Parametric Mapping) on behalf of Philip Deming" <[log in to unmask] on behalf of [log in to unmask]> wrote:
Hi,
I have a couple questions about RFX-BMS for DCM data.
1. I'm fitting 3 DCMs to fMRI data, and ultimately I'm interested in how the posterior model probabilities relate to psychopathology. It seems like the recommended method would be to treat my sample as two separate populations (healthy vs. patients), run RFX-BMS for each group, and compare (using t-tests) the posterior model probabilities (g_post in the BMS file) between groups. Is there a recommended method for comparing posterior model probabilities to a continuous measure of psychopathology?
My instinct would be to treat the sample as one population, run RFX-BMS for the whole sample, and then regress those posterior model probabilities on my continuous measure. However, I notice that the posterior model probabilities for a given subject seem to depend on the population entered into the RFX-BMS procedure. The model with highest probability for any given subject is the same as the winning model for the population within which I conducted the RFX-BMS. So when I run RFX-BMS on the whole sample, the posterior model probabilities for, for example, healthy subject 1 change so that model 1 instead of model 3 has the highest probability. The posterior model probabilities for a given model are not that variable across the population, so it would be difficult to compare them to a continuous variable.
2. My reading of Stephan et al. (2009) suggests that this dependence on the population might arise from the Dirichlet distribution, which seems to keep a running tally of the number of subjects for which model k generated the observed data. Does this mean that the posterior model probabilities depend on the order of subjects? I reversed the order of my subjects and noticed that, while the model rankings did not change, the exact values of the posterior model probabilities and exceedance probabilities did.
Ultimately, I want to make sure I understand how the population affects the RFX-BMS results so I make the right choices for my data.
Thanks,
Phil Deming
|