Computational modeling of Pavlovian response bias and instrumental learning bias.

(a) Model evidence, relative to the simplest model, M1, favors M3c (marked by darkest color). The simplest model, M1, contains a feedback sensitivity (ρ) and learning rate (ε) parameter. Stepwise addition of the Go bias (b), Pavlovian bias (π), and instrumental learning bias (κ) parameter improves model fit. (b) Absolute post hoc model fit. The model predictions of winning model M3c (black lines; shaded areas indicate standard error of the mean) capture the key features of the data (colored lines); responses are learned (more Go responding for Go cues versus NoGo cues) and a motivational bias is present (more Go responding for “win” versus “avoid” cues). (c) Posterior densities of the winning model M3c. The group-level estimates are positive for the Go bias (100% samples > 0), the Pavlovian bias (95.5% samples > 0), and the instrumental learning bias (100% samples > 0). (d) Subject-level parameter estimates of the winning model M3c. The feedback sensitivity and learning rate were strongly anticorrelated over subjects (R = −0.77, p 0 + κ) and decreased learning rates for punished NoGo responses (ε0 − κ), which we refer to as the instrumental learning bias. Note that κ is added to ε0 prior to {0 1}-constraining; see S2 Text. Underlying data can be found at http://hdl.handle.net/11633/di.dccn.DSC_3017033.03_624. WAIC, Watanabe–Akaike Information Criterion.