Test-retest reliability of dynamic causal modeling for fMRI.

Dynamic causal modeling (DCM) is a Bayesian framework for inferring effective connectivity among brain regions from neuroimaging data. While the validity of DCM has been investigated in various previous studies, the reliability of DCM parameter estimates across sessions has been examined less systematically. Here, we report results of a software comparison with regard to test-retest reliability of DCM for fMRI, using a challenging scenario where complex models with many parameters were applied to relatively few data points. Specifically, we examined the reliability of different DCM implementations (in terms of the intra-class correlation coefficient, ICC) based on fMRI data from 35 human subjects performing a simple motor task in two separate sessions, one month apart. We constructed DCMs of motor regions with fair to excellent reliability of conventional activation measures. Using classical DCM (cDCM) in SPM5, we found that the test-retest reliability of DCM results was high, both concerning the model evidence (ICC=0.94) and the model parameter estimates (median ICC=0.47). However, when using a more recent DCM version (DCM10 in SPM8), test-retest reliability was reduced notably. Analyses indicated that, in our particular case, the prior distributions played a crucial role in this change in reliability across software versions. Specifically, when using cDCM priors for model inversion in DCM10, this not only restored reliability but yielded even better results than in cDCM. Analyzing each component of the objective function in DCM, we found a selective change in the reliability of posterior mean estimates. This suggests that tighter regularization afforded by cDCM priors reduces the possibility of local extrema in the objective function. We conclude this paper with an outlook to ongoing developments for overcoming the software-dependency of reliability observed in this study, including global optimization and empirical Bayesian procedures.