, 2006 and Preuschoff et al., 2008). Consequently, we modeled subjects’ trial-by-trial estimates of the correlation coefficient and regressed those model-predicted time series against simultaneously acquired fMRI data. We found BOLD activity UMI-77 research buy in right
midinsula varied with the correlation strength between the outputs of the solar and wind power plants (xyz = 48, 5, −5; Z = 4.12; p < 0.001 familywise error (FWE) corrected; Figure 3A). Right insula was the only region to survive cluster level whole brain correction and we provide a comprehensive list of all activated areas at a lower threshold (p < 0.001 uncorrected) in Table 2. We next determined whether the correlation strength is represented either as covariance, a raw measure of how much the two variables fluctuate together, or as the correlation coefficient, a scale invariant metric of the covariance normalized by the standard deviation of each resource. We estimated two additional models using Bayesian estimation, with either the covariance or the correlation coefficient as parametric modulator, and compared the ensuing log-evidence maps selleck products in a random effects analysis. Activity in right midinsula was better described by the correlation coefficient than by covariance (exceedance probability of
p > 0.999). The linear relationship between correlation coefficient and BOLD is visualized in a binned effect size plot (Figure 3B). We then verified whether this signal was more strongly represented at the time of outcome, when new evidence is available to update estimates, or at choice when subjects actively readjust their allocated weights for the two resources (Figure 3C). In addition to plotting the effect time course we tested these neural hypotheses by estimating a design where the correlation coefficient acted as an unorthogonalized parametric modulator of activity at both the time of outcome and time of choice. In this analysis
we observed significant effects of correlation strength solely at the outcome time (Z = and 3.60, p = 0.01 FWE corrected) but not at the time of choice (Z = 2.40, p = 0.02 uncorrected). If our behavioral model explains subject’s choices and subjects’ brain activity represents crucial decision variables in this process then we would expect that brain activity should be particularly well explained in those subjects in whom our model also provides a good choice prediction. This would be expressed in a relationship between the behavioral model fit and the model fit in the general linear model (GLM) against BOLD data. Consistent with our conjecture, we found a significant positive correlation between R2 in the behavioral model and R2 in the MRI analysis (r = 0.50, p < 0.03; Figure 3D).