Hi,
I'm working on implementing Bayesian Model Selection for an analysis with 603 subjects and 16 models. I have .mat files with the DCM results for each subject and for each model (so a total of 603*16 mat files). I want to use spm_BMS() to decide which model is the best, and I want to check on a few things:
(1) Is the log model evidence the F value or the negative of the F value inside the DCM structure? Reading this email on JISC (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;27167107.1306) I can't tell if I should be feeding in the F value or the -F value when preparing the lme matrix for input into spm_BMS(). The F value for a given subject and model with DCM mat file 'sub001_model1.mat' I am accessing by typing
a = load('subj001_model1.mat')
a.DCM.F
My understanding is that F is the free energy which is the *negative* of the log model evidence, and so I should be formulating the lme matrix for input to spm_BMS() by using -1*a.DCM.F, but I get some surprising results when I use -1*a.DCM.F, so wanted to check what's going on.
(2) The alphas returned in either case (a.DCM.F or -1*a.DCM.F for lme) are all greater than one, while they are supposed to represent a vector of model probabilities. Should I assume that we normalize these values by e.g.
model_probabilities = alpha/sum(alpha);
best,
Spencer
PhD student, UCLA Statistics
|