Abstract
There is abundant evidence that humans and other animals employ mul-
tiple control systems to solve complex decision-making tasks. Conven-
tional dichotomies such as model-free or habitual systems, versus model-
based or goal-directed systems are under debate and revision but remain
useful for designing and analyzing tasks and can be exploited to cast
light on psychiatric disorders. One particular issue that has attracted less
study, although it is likely to be of great importance in the control that we
exert over the deployment of these control systems, is the meta-cognitive
question of self-monitoring and self-evaluation, i.e., control over control.
Given the known involvement of the neuromodulator’s dopamine and
norepinephrine in aspects of both control and metacognition, we used
neuropharmacological perturbations to examine interactions between
control and control over control. In particular, we used the systemic
administration in healthy human volunteers of Propranolol which sup-
presses the operation of norepinephrine and putatively enhances meta-
cognitive accuracy, and Levodopa-B, which boosts dopamine and,
although there are conflicting results in the literature, has been reported
to boost the influence of model-based control.
We examined the effects of the drugs on choice and confidence
ratings in two tasks: a conventional perceptual decision-making
task used to study to confidence judgements, and a two-outcome
task that offers an exquisitely fine decomposition of model-free
and model-based choice and credit assignment. Using hierarchical
Bayesian fitting, we found that Propranolol significantly decreased
meta-cognitive ability, particularly among individuals with lower
weight (and so a higher effective dose), while there was a very weak
trend for Levodopa-B to improve it (see Fig. 1). However, perfor-
mance was not significantly affected by drugs in either task. In the
two-outcome task, Propranolol increased model-based behavior but
had no effect on model-free behavior, while Levodopa-B had no
effect on either. Regarding control over control, when control systems disagree, meta-
control might naturally be exerted to determine which one should be
favored. For instance, when decision-makers lack confidence, the model-
based controller should be preferred because it is statistically superior.
However, if decision-making is uncertain, the model-based system may
not be reliable. In support of this hypothesis, we found that model-based
behavior was less likely to increase after low confidence, as an effect
of Levodopa-B. We assessed the validity of our findings by comparing
results from randomly permuted drug conditions with our empirical drug
conditions. Overall, we suggest that our study sheds new light on the role
of noradrenergic and dopaminergic systems in different levels of control
and points to potential avenues for mitigating dysfunction within and
between these systems.
