Similar total earnings do not necessarily imply identical decision strategies. In a 2015 methamphetamine study, participants and controls earned similar amounts on a computerized task, while computational models suggested differences in how they learned about rewards and chose between options.
Who participated?
Harlé and colleagues compared 16 sober treatment participants with methamphetamine dependence and 16 healthy comparison participants. Testing occurred during the third to fourth week of a 28-day inpatient treatment program. Urine screening supported abstinence. This was not an acute dosing experiment.
Participants completed 20 games of 16 choices between two lottery options. Reward probabilities were fixed within each game but initially unknown. Choices therefore produced both potential rewards and information about future opportunities.
Preserve the overall null result
The groups did not significantly differ in total earnings. Nor did their earnings show a significant group-by-game interaction. The study therefore did not demonstrate that the methamphetamine group simply performed worse by the task’s overall monetary outcome.
The researchers instead asked whether different learning and decision policies could generate similar aggregate performance. They compared five candidate policies combined with a Bayesian learning model.
What did the fitted models suggest?
A simple win-stay/lose-shift policy best fit eight of 16 treatment participants and three of 16 controls. That proportion comparison was a nonsignificant trend, p = 0.06. It should not be described as a confirmed categorical difference between the populations.
Among participants best described by a Softmax policy, the methamphetamine group had a lower fitted reward-maximization parameter. The groups also differed in model-estimated prior reward expectations, while their estimated environmental-stability parameters were similar.
These are inferred quantities conditional on the candidate models and assumptions. Researchers did not directly read a participant’s belief or measure a fixed personal trait called pessimism. A different model set could alter the interpretation of the same choices.
What did structural imaging add?
An exploratory analysis linked a choice parameter with gray-matter volume in a thalamic region among a smaller model-defined subgroup. It did not establish that a regional difference caused the strategy or that the strategy caused substance dependence.
Small samples, subgroup selection, multiple candidate models and exploratory anatomical correlations limit individual prediction. Statistical mediation in cross-sectional data does not establish a biological causal chain.
Implications for expert interpretation
The study supports separating outcome from process: similar scores can coexist with different modeled strategies. It does not establish generalized irrationality, inability to learn, acute intoxication or driving impairment in an individual.
A defensible discussion identifies the actual task, the significant and nonsignificant comparisons, and the assumptions connecting observed choices to inferred learning mechanisms.
Primary scientific source
Harlé KM, Zhang S, Schiff M, Mackey S, Paulus MP, Yu AJ. Altered Statistical Learning and Decision-Making in Methamphetamine Dependence: Evidence from a Two-Armed Bandit Task. Frontiers in Psychology. 2015;6:1910. doi:10.3389/fpsyg.2015.01910. Historical observational and computational research.
Leave a Reply