 |
 |
Dear Rezvan You asked about summarising parameters across four first level models (DCMs) from multiple subjects. Just to be clear on some terminology: - Bayesian Parameter Averaging (BPA) is typically used when you have one model fitted to different data – e.g. to multiple subjects. - Bayesian Model Averaging (BMA) is typically used when you have multiple models fitted to one or more subjects. BPA accumulates the parameter estimates from the same model across all subjects (this is also called Bayesian Belief Updating). Unlike an arithmetic average, it takes the variance (and optionally the covariance) of the parameters into account. Note that if you take the covariance into account, you can get some counter-intuitive results. See discussion at https://en.wikibooks.org/wiki/SPM/Bayesian_Parameter_Averaging_(BPA)#Frequently_Asked_Questions . BMA is a weighted average of the parameters over models (weighted by the evidence or posterior probability of each model). So in your example, you would expect model 3 to make the largest contribution to the average, and then there would be smaller contributions from models 4, 2 and 1. The arithmetic average of the parameters ignores the variance and covariance – which is a waste of useful information, given that these are estimated by the DCM. I suggest you use BMA – because otherwise you’re ignoring the contribution of models 4, 2 and 1. Best Peter From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Rezvan Farahi Sent: 14 December 2018 12:50 To: [log in to unmask] Subject: [SPM] DCM for ERP- BMA vs. BPA Dear SPM experts, I'm using DCM for ERP and I'm comparing 4 models (5 ROIs per model). We have model 3 as a clear winner (mean F values are -609.40, -608.12, -526.34, -596.70). In this case, I was wondering if it is correct to expect (a) straight parameter averaging of model 3 to yield similar results as (b) BMA and (c) BPA? For our data, while (a) and (b) are similar, BPA estimations are different (table below). I had a brief look at spm_dcm_bpa code and bpa.Ep seems to include parameter averaging but also weighting by covariance etc. I'm wondering if it's normal to get different results from (a), (b), (c) in the presence of a clear winning model, or under which circumstances they should agree? Thanks a lot! Rezvan 0-250ms Model 3 parameters BPA 0 0.155 -0.160 -0.480 0.208 nocd=0 -0.141 -0.045 0 0 0 -0.078 0 0.220 0 0 -0.364 0 0 0.064 0 0.402 0 0 0 -0.034 BPA 0 -0.015 -0.103 0.019 -0.053 nocd=1 -0.055 0.011 0 0 0 -0.038 0 0.039 0 0 0.035 0 0 0.016 0 0.004 0 0 0 0.007 BMA 0 -0.037 -0.151 -0.010 0.047 -0.033 -0.02 0 0 0 0.0096 0 -0.006 0 0 -0.048 0 0 0.051 0 0.0009 0 0 0 -0.012 meanpar 0 -0.033 -0.152 -0.008 0.047 -0.032 -0.022 0 0 0 0.0081 0 -0.006 0 0 -0.047 0 0 0.051 0 -0.0013 0 0 0 -0.012