Model parameters.

The means and standard deviations of the ACQ model’s fitted parameters—including from the hysteresis model the (arbitrarily rightward) constant choice bias βR and initial magnitude β0 coupled with inverse decay rate λβ for exponential decay of the perseveration bias—are listed separately for each group, revealing a tendency for Good learners to have lower temperature than Poor learners (M = 0.987, t33 = 2.88, p = 0.004). The logarithm of the ratio between the eligibility-adjusted learning rate and the temperature provides a more precise metric for the sensitivity dictated by the model’s fitted parameters than the temperature alone—especially given the correlation between the eligibility-adjusted learning rate and the temperature [57] exhibited within the Poor-learner group (r = 0.547, t13 = 2.36, p = 0.035) and the lack of such a correlation among Good learners (r = 0.121, t18 = 0.52, p = 0.611). Model sensitivity, which was significantly positive across the Good-learner group (M = 0.440, t19 = 5.59, p < 10-4) but not the Poor-learner group (M = 0.020, t14 = 0.18, p = 0.428), was not only greater for Good learners than for Poor learners (M = 0.420, t33 = 3.23, p = 10-3) but also significantly correlated with the objective metric for choice accuracy (r = 0.409, t33 = 2.57, p = 0.007). The residual deviance D (with degrees of freedom in the subscript) corresponds to the ACQ model’s improvement in fit relative to either a null intercept model or the hysteresis model.