Dear Rajan,
The idea behind model selection is to identify the model (hypothesis) with the highest model evidence among a set of alternative models. If there isn't one particular model that best explains your data, you should use Bayesian model averaging (BMA). This allows you to do inference on the connection strengths using an estimate from the entire model space.
Best wishes
Martin
> On 02 Sep 2015, at 10:44, Rajan Kashyap <[log in to unmask]> wrote:
>
> Dear SPM people
>
> I have ERP data of two conditions (Suppose A and B). Based on the data i created two models F and FB.
> I fitted the data on individual subjects.
> I found on the group level :(1) F model was preferred over FB model in condition A
> (2) FB model over F for condition B.
>
> However, then i found that few participants who show a correlation to behavioural data (in my case faster reaction time)in condition A preferred FB model though F model was winning over there at group level.
>
> This motivated me to take the connection parameters (posterior means) of the non winning models too. Then i compared for significant connections between the F models connection posterior means for 2 conditions and FB models posterior means for the two conditions, to see what variations in connections can cause such outcomes
>
> In general i always found in DCM papers that only winning models are preferred.
> In this regards i would like take the view of DCM experts, if this could be done.
>
> Regards
> Rajan
> PhD student
|