Applying Bayesian Multilevel Modeling to Single Trial Dynamics: A Demonstration in Aversive Conditioning.

Aversive conditioning changes visuocortical responses to conditioned cues, and the generalization of these changes to perceptually similar cues may provide mechanistic insights into anxiety and fear disorders. Yet, as in many areas of cognitive neuroscience, testing hypotheses about trial-by-trial dynamics in conditioning paradigms is challenged by poor single-trial signal-to-noise ratios (SNR), missing trials, and inter-individual differences. The present technical report demonstrates how a state-of-the-art Bayesian workflow can overcome these issues, using a preliminary sample of simultaneously recorded EEG-fMRI data. A preliminary group of observers (N = 24) viewed circular gratings varying in orientation, with only one orientation paired with an aversive outcome (noxious electric pulse). Gratings were flickered at 15 Hz to evoke steady-state visual evoked potentials (ssVEPs), recorded with 31 channels of EEG in an MRI scanner. First, the benefits of a Bayesian multilevel structure are demonstrated on the fMRI data by improving a standard fMRI first-level multiple regression. Next, the Bayesian modeling approach is demonstrated by applying a theory-driven learning model to the EEG data. The multilevel structure of the Bayesian learning model informs and constrains estimates per participant, providing an interpretable generative model. In the example analysis provided in this report, it showed superior cross-validation accuracy and provided insights into participant-level learning dynamics. It also isolated the generalization effects of conditioning, providing improved statistical certainty. Lastly, missing trials were interpolated and weighted appropriately using the full model's structure. This is a critical aspect for single-trial analyses of simultaneously recorded physiological measures because each added measure will typically increase the number of trials missing a complete set of observations. The present report aims to illustrate the utility of this analytical framework. It shows how models may be iteratively built and compared in a modern Bayesian workflow. Future models may use different conceptualizations of learning, allow integration of clinically relevant factors, and enable the fusion of different simultaneous recordings such as EEG, autonomic, behavioral, and hemodynamic data.