The influence of EM estimation of missing nodes in DCM on model ranking

In Dynamic Causal Modeling (DCM) group analyses alternative model(s) are commonly specified and compared against each other. A model comparison problem is generally encountered by any kind of modeling approach where model selection is required given some observed data and several alternative models. The goal would be to select the optimal model by deciding between competing hypotheses represented by different DCMs. These hypotheses can involve any part of the structure of the modeled system, i.e. the pattern of intrinsic or extrinsic connections to the system. However, the underlying assumption is that the comparison is only valid if the data is the same in all models. In DCM for fMRI, where the data results from concatenation of all the time series of all areas in the model, the comparison requires that only models containing the same areas are included. In previous work, we have shown that Expectation Maximization can address this limitation by estimating time series in subjects with missing areas. This opposed the traditional approach of using a less conservative p-value which creates noisy time series to enforce topological comparability across models. Alternative methods for inference include the Fixed Effects (FFX) Analysis and the Random Effects (RFX) Analysis which can be used to rank prospective models in a model comparison/selection problem. Furthermore, Bayesian Model Averaging (BMA) can also be applied enabling subject specific mean parameters to represent summary statistics for a standard group analysis. Significant differences in a given parameter between a control group and a patient group, for instance, could then be calculated by computing the two-sample t-test on the average data from the two groups.