Citation Context

...l models (Stephan et al., 2009, 2010) with inference over families of models (Penny et al., 2010) to identify the most likely model structure from the model space described above. This was done in two stages. First, for each subject the model evidence was computed for each model and each run using the negative free-energy approximation to the log-model evidence. The free-energymetric formodel evidence balances model fit and complexity taking into account interdependencies amongst parameters and has been found to outperform other more conventional methods of model scoring for model comparison (Penny, 2012). The subject-specific sums of log evidences across runs (equivalent to a fixed effects analysis across runs) were entered into group random effects (RFX) BMS to identify the most likely model across subjects (Stephan et al., 2009). This procedure requires that all subjects have the same number of runs (c.f. SPM DCMmanual), so only the first four runs were used for DCM for all subjects (as mentioned previously, three subjects had only four usable runs due to motion artifacts). The RFX approach to groupmodel selectionwas preferred over fixed effects because it does not assume that the optimal m...

Citation Context

...) that are returned by the lme4 package for each model (see Table 4). Although derived within different frameworks, both the BIC and AIC can be thought of as approximations to the true model evidence =-=[36]-=-, both containing a term reflecting the likelihood of the model given the data (the ‘accuracy’ term) and a penalization term reflecting the number of parameters in the model (the ‘complexity’ term). A...

Citation Context

...he ratio of the maximized probabilities of the data under the alternative and null hypotheses. By contrast, the Bayes factor (Kass and Raftery, 1995), which has been widely used in neuroimaging (e.g. =-=Penny, 2012-=-), is the ratio of integrals of the data probability, with respect to the prior distributions for two models. The high-dimensional case Modern predictive analyses, in neuroimaging as in other fields, ...

Citation Context

... For the linear Gaussian models underlying MEG source reconstruction, the model evidence is well approximated by the negative variational free energy (henceforth “Free Energy”) (Friston et al., 2007; =-=Penny, 2012-=-; Wipf & Nagarajan, 2009). The free energy allows one to determine the most adequate model for a given dataset. For the model associated with a given head location hk it can be expressed as a trade of...

Citation Context

...model p(Y|M), known as the model evidence or marginal likelihood. The log evidence is approximated by a negative (variational) free energy that is used for Bayesian model comparison or scoring (e.g. (=-=Penny, 2012-=-)). Further details about DCM for fMRI responses can be found elsewhere (Friston, 2011b; Seghier et al., 2010b). Priors on the parameters of DCM for fMRI In this work, we used the latest release of DC...

Citation Context

...oximated using the variational free energy which consists of an accuracy and complexity term, thus enabling the comparison and selection of competing models (Penny et al., 2004; Friston et al., 2007; =-=Penny, 2012-=-). We employed the same three-level hierarchical general linear model (GLM) as Mars et al. (2008). For model fitting and calculation of the model evidences we used parametric empirical Bayes (PEB) fro...

Citation Context

...ree energy value represents a balance between a model's goodness of fit to the data and the model's complexity which additionally takes into account interdependencies or covariances among parameters (=-=Penny, 2012-=-). Given a particular network, a model space is generally a subset of all theoretically testable models built by the investigator to probe specific questions in a computationally pragmatic manner. The...