Multi-subject analyses with dynamic causal modeling.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Statistical Modeling

Background:

• Dynamic Causal Modeling (DCM) commonly uses random-effects (RFX) analysis for group inferences.
• Fixed-effects (FFX) analysis presents an alternative for specific group inference scenarios in DCM.
• Existing FFX methods include Bayesian parameter averaging (BPA), posterior variance weighted averaging (PVWA), and temporal averaging (TA).

Purpose of the Study:

• To systematically compare the statistical properties of different FFX approaches in DCM.
• To investigate the impact of signal-to-noise ratio (SNR) and population heterogeneity on group-level parameter estimates.
• To provide guidance on the appropriate application and interpretation of FFX methods in DCM.

Main Methods:

• Simulated data from a two-region network were generated to assess FFX methods.
• Analyses included Bayesian parameter averaging (BPA), posterior variance weighted averaging (PVWA), and temporal averaging (TA).
• Simulations varied signal-to-noise ratio (SNR) and population heterogeneity (homogeneous vs. heterogeneous parameters).

Main Results:

• Temporal averaging (TA) showed advantages at lower SNR but has limited applicability.
• BPA and PVWA can produce non-intuitive results with high SNR, parameter interdependencies, or violated FFX assumptions (inhomogeneous groups).
• The issues with BPA and PVWA diminish with decreasing SNR and are absent for models with independent parameters or when FFX assumptions hold.

Conclusions:

• FFX approaches in DCM, specifically BPA and PVWA, necessitate careful interpretation of group results.
• Consideration of parameter dependencies is crucial, particularly under conditions of high SNR and population heterogeneity.
• The choice of FFX method and interpretation should account for data characteristics and model assumptions.