Figure 5. Model evidence and parameter inference of extended MPH models.

(A) Model evidence (WAIC) relative to winning base model M4. We tested whether MPH alters the strength of the Pavlovian response bias (π_MPH; M5a), the instrumental learning bias (κ_{MPH-selective}; M5b), or has a diffuse effect on the learning bias (κ_MPH-diffuse; M5c; Figure 1C). Model selection favoured the composite model M6, including the π_MPH and κ_MPH-diffuse parameters. (B) Posterior densities of the top-level parameters of M6. (C) Subject-level estimates of MPH-induced Pavlovian bias parameter (upper) and the MPH-induced diffuse learning bias parameter (lower; logistic scale) correlated significantly with Listening Span. (D) One-step-ahead model predictions and posterior predictive model simulations of M6 using subject-level parameter estimates. The model predictions and simulations echo the observed data, i.e. that the motivational bias correlates positively with working memory span (Figure 4B), confirming the winning model M6 captures the MPH-induced increase in Go responses to Win vs. Avoid cues.

DOI: http://dx.doi.org/10.7554/eLife.22169.015

Figure 5—figure supplement 1. Illustration of the behavioural effects of MPH related to the Pavlovian bias and diffuse learning bias parameters.

Model M6 one-step-ahead predictions (coloured) overlaid on real data (grey) for the subjects with the 33% strongest vs. weakest parameter estimates. The coloured bars at the bottom indicate the trial-by-trial correlation across all subjects, of the parameter estimate with the effect of MPH on Go responding per cue. The R value indicates the average correlation. (A) The effect of MPH on Pavlovian bias (π_MPH). Strong π_MPH estimates predict that MPH increases the motivational bias (increased Go to Win cues and decreased Go to Avoid cues), and vice versa for weak π_MPH estimates. The influence of π_MPH is present from the first trial onward and decreases over time as indicated by the correlation coefficients. (B) Effect of MPH on diffuse learning bias (κ_MPH-diffuse). Strong κ_MPH-diffuse estimates predict that MPH increases the motivational bias for Win cues specifically, whereas this effect is diminished for subjects with relatively weak κ_MPH-diffuse estimates. The effect of κ_MPH-diffuse is experience-dependent and evolves over time. These one-step-ahead predictions illustrate how each parameter results in an increased motivational bias under MPH, but with unique temporal dynamics, even though the parameter themselves are constant.

DOI: http://dx.doi.org/10.7554/eLife.22169.016

Figure 5—figure supplement 2. M6 subject-level parameters in model space (i.e. untransformed).

The diagonal panels contain the posterior densities for the subject-level parameter means. The off-diagonal panels contain the parameter correlations over subjects. Importantly, the parameters estimating the effects of MPH on Pavlovian bias (π_MPH) and diffuse learning bias (κ_MPH-diffuse) are not correlated any of the other parameters.

DOI: http://dx.doi.org/10.7554/eLife.22169.017