Methamphetamine-induced adaptation of learning rate dynamics depend on baseline performance

Reviewer #1 (Public review):

The authors examine how probabilistic reversal learning is affected by dopamine by studying the effects of methamphetamine (MA) administration. Based on prior evidence that the effects of pharmacological manipulation depend on baseline neurotransmitter levels, they hypothesized that MA would improve learning in people with low baseline performance. They found this effect, and specifically found that MA administration improved learning in noisy blocks, by reducing learning from misleading performance, in participants with lower baseline performance. The authors then fit participants' behavior to a computational learning model and found that an eta parameter, responsible for scaling learning rate based on previously surprising outcomes, differed in participants with low baseline performance on and off MA.

Questions:

(1) It would be helpful to confirm that the observed effect of MA on the eta parameter is responsible for better performance in low baseline performers. If performance on the task is simulated for parameters estimated for high and low baseline performers on and off MA, does the simulated behavior capture the main behavioral differences shown in Figure 3?

(2) In Figure 4C, it appears that the main parameter difference between low and high baseline performance is inverse temperature, not eta. If MA is effective in people with lower baseline DA, why is the effect of MA on eta and not IT?

Also, this parameter is noted as temperature but appears to be inverse temperature as higher values are related to better performance. The exact model for the choice function is not described in the methods.

Reviewer #2 (Public review):

Summary:

Kirschner and colleagues test whether methamphetamine (MA) alters learning rate dynamics in a validated reversal learning task. They find evidence that MA can enhance performance for low-performers and that the enhancement reflects a reduction in the degree to which these low-performers dynamically up-regulate their learning rates when they encounter unexpected outcomes. The net effect is that poor performers show more volatile learning rates (e.g. jumping up when they receive misleading feedback), when the environment is actually stable, undermining their performance over trials.

Strengths:

The study has multiple strengths including large sample size, placebo control, double-blind randomized design, and rigorous computational modeling of a validated task.

Weaknesses:

The limitations, which are acknowledged, include that the drug they use, methamphetamine, can influence multiple neuromodulatory systems including catecholamines and acetylcholine, all of which have been implicated in learning rate dynamics. They also do not have any independent measures of any of these systems, so it is impossible to know which is having an effect.

Another limitation that the authors should acknowledge is that the fact that participants were aware of having different experiences in the drug sessions means that their blinding was effectively single-blind (to the experimenters) and not double-blind. Relatedly, it is difficult to know whether subjective effects of drugs (e.g. arousal, mood, etc.) might have driven differences in attention, causing performance enhancements in the low-performing group. Do the authors have measures of these subjective effects that they could include as covariates of no interest in their analyses?

Reviewing Editor (Public Review):

Summary:

In this well-written paper, a pharmacological experiment is described in which a large group of volunteers is tested on a novel probabilistic reversal learning task with different levels of noise, once after intake of methamphetamine and once after intake of placebo. The design includes a separate baseline session, during which performance is measured. The key result is that drug effects on learning rate variability depend on performance in this separate baseline session.

The approach and research question are important, the results will have an impact, and the study is executed according to current standards in the field. Strengths include the interventional pharmacological design, the large sample size, the computational modeling, and the use of a reversal-learning task with different levels of noise.

(i) One novel and valuable feature of the task is the variation of noise (having 70-30 and 80-20 conditions). This nice feature is currently not fully exploited in the modeling of the task and the data. For example, recently reported new modeling approaches for disentangling two types of uncertainty (stochasticity vs volatility) could be usefully leveraged here (by Piray and Daw, 2021, Nat Comm). The current 'signal to noise ratio' analysis that is targeting this issue relies on separately assessing learning rates on true reversals and learning rates after misleading feedback, in a way that is experimenter-driven. As a result, this analysis cannot capture a latent characteristic of the subject's computational capacity.

(ii) An important caveat is that all the drug x baseline performance interactions, including for the key computational eta parameter did not reach the statistical threshold, and only tended towards significance.

(iii) Both the overlap and the differences between the current study and previous relevant work (that is, how this goes beyond prior studies in particular Rostami Kandroodi et al, which also assessed effects of catecholaminergic drug administration as a function of baseline task performance using a probabilistic reversal learning task) are not made explicit, particularly in the introduction.

(iv) In the discussion, it is stated that the existing literature has, to date, overlooked baseline performance effects, but this is not true in the general sense, given that an accumulating number of studies have shown that the effects of drugs like MA depend on baseline performance on working memory tasks, which often but certainly not always correlates positively with performance on the task under study.