No, this is not a good idea to do t-tests on posterior probabilities.
From BMA you should have average parameter estimates (averaged over models and weighted by model evidences).
It should be sufficient to do two-sample t-tests (between groups) on the parameter estimates.
Or, if you want an inference about groups at the model level it should be sufficient to just report the posterior probability values - eg in group one model 1 was favoured (post prob=0.91) whereas in group 2 model 3 was favoured (post prob=0.93). If you don't
have peaky posterior probs for the models you are best off doing parameter level inference (see above).
Hi SPMers and experts in DCM,
I have a question about DCM, BMA. I've recently performed a random-effects analysis with 8 DCM models in two groups, healthy controls, respectively a clinical group. Two models showed differences between the two groups.The "model level results" comprise the
posterior model probabilities (
.g_post). Is it OK to perform some statistical tests (i.e. two-sample t-test) on these posterior probabilities?
Would it be necessary to further transform these values?
Thanks in advance for your advice,
Eugenia
Dr. Eugenia Radulescu, MD, PhD
Postdoctoral Research Fellow
Sackler Centre for Consciousness Science,
Brighton and Sussex Medical School,
Brighton, UK