Dynamic and stochastic models of neuroimaging data: A comment on Lohmann et al.

Abstract

DCM is a platform for inferring the architecture of dynamical systems, combining a user-dependent model specification step with a Bayesian model selection scheme. In their critique of the model selection procedure, Lohmann et al confine themselves to models generated from the classic bilinear deterministic DCM. Although brief reference is made to recent modeling advances, such as stochastic DCM and nonlinear DCM, these are negatively cast as guilty of further exploding the combinatorial problem that is proposed to plague model selection. Yet this is only a problem if a naïve approach is adopted towards the model generation process. Where the user draws on prior knowledge of the system being modeled and the statistical properties of the particular data set, these advances can be employed precisely to address the type of concerns Lohmann et al raise in their exemplar analysis (Fig. 6). This note provides a putative generative model for their data by adding stochastic effects, using independent evidence to increase its biological plausibility and challenging the notion that model fit should be assessed using simple linear correlations. Rather than encouraging reliance on future developments in imaging hardware and data-driven multivariate algorithms, informed engagement with causal models of neuronal dynamics allows imaging researchers develop detailed theories of brain function across a broad range of data sets and cognitive phenomena.