Does Methamphetamine Use Disorder Reduce Directed Exploration? A 114-Person Decision Experiment With an Anxiety Test

Short answer: A complete primary experiment compared 56 treatment-seeking adults with methamphetamine use disorder (MUD) and 58 healthy comparison adults as they chose between uncertain rewards. The MUD group showed less model-estimated directed exploration (6.06 versus 4.58; p = .021), a lower probability of being in the high initial-learning-rate group (32% versus 67%; p < .001), and lower asymptotic learning rates (0.22 versus 0.30; p = .003). Breathing resistance increased reported anxiety, but the predicted group-by-resistance effects on these computational measures were not significant. Some other contrasts became nonsignificant after outlier or working-memory checks. This study measured task choices and mathematical model parameters, not methamphetamine concentrations in venous whole blood, a person’s present intoxication, driving ability, or the cause of an actual incident.

The source analyzed here is the complete investigator-authored February 24, 2025 medRxiv research manuscript by Carter M. Goldman, Toru Takahashi, Claire A. Lavalley, Ning Li, Samuel Taylor, Anne E. Chuning, Rowan Hodson, Jennifer L. Stewart, Robert C. Wilson, Sahib S. Khalsa, Martin P. Paulus, and Ryan Smith: Individuals with Methamphetamine Use Disorder Show Reduced Directed Exploration and Learning Rates Independent of an Aversive Interoceptive State Induction, original complete empirical preprint, DOI 10.1101/2024.05.17.24307491. The investigator cohort was later described in a June 2025 peer-reviewed version of record; those publications are versions of the same cohort, not independent replications. This analysis uses the supplied complete preprint’s original numbers and separates established results from additional interpretations.

What was the actual scientific question?

When facing uncertain options, a person can repeatedly select what presently appears most rewarding, or sample a less familiar option to learn whether it could be better. Directed exploration describes information-seeking that responds to differences in uncertainty. The researchers asked whether adults receiving treatment for MUD changed exploration strategically when additional choices would make information useful. They also asked whether an uncomfortable bodily sensation changed this strategy.

This is narrower than asking whether a person with a drug-use history has impaired judgment in general. The experiment used an artificial reward-learning task with known mathematical structure, not a comprehensive neuropsychological examination. It compared groups with preexisting characteristics rather than randomly assigning a history of methamphetamine exposure. The scientific value lies in separating several mechanisms that ordinary labels such as impulsivity, decision-making, and attention often obscure.

How many independent adults were studied?

The independent sample included 114 adults: 56 in an inpatient, treatment-seeking MUD group and 58 healthy comparisons. They were not 228 independent people because each adult performed a run without breathing resistance and a separate run with resistance. Nor do thousands of trial-level choices create thousands of newly randomized participants. The two groups differed notably in sex distribution: 20 women and 36 men in MUD, versus 42 women and 16 men in controls. Their average ages were 36.75 and 35.41 years, respectively; that age comparison was nonsignificant (p = .500).

Some questionnaires were added after recruitment started. For example, the working-memory descriptive comparison included 55 MUD participants and 57 controls, not all 114; a baseline craving questionnaire had only 39 affected-group responses. Reported mixed-model denominator degrees of freedom reflect model structure and repeated choices, not an independent biological sample of 2,612 people. The generalization unit remains the adult, with additional uncertainty from the design and group selection.

Was methamphetamine administered in the study?

No. Participants were identified by current or historical clinical diagnosis and treatment status. The MUD group reported an average of 50.04 days since methamphetamine use and 34.47 days since starting treatment. Those were group historical averages, not objectively witnessed last doses or pharmacokinetic sampling times. The investigators did not randomize any person to take methamphetamine, establish a drug dose in milligrams, or assign an oral, smoked, intranasal, or injected drug route.

The controlled experimental manipulation was inspiratory resistance through a breathing mask. Drug testing and a breathalyzer were used in eligibility procedures, but the source does not report quantitative methamphetamine or amphetamine parent-drug concentrations in a legal venous whole-blood matrix. The absence of controlled drug administration must not be disguised as an acute methamphetamine challenge.

How did the Horizon Task create a reason to explore?

People chose between two on-screen slot-machine options yielding uncertain points. Before free choice, they observed four forced-choice outcomes. Sometimes the information was balanced, with two prior observations of each option. Sometimes it was unbalanced, with three observations of one option and one of the other. Then a game permitted either one free choice (Horizon 1, or H1) or six free choices (Horizon 6, or H6).

In H1, choosing an uncertain option offers little opportunity to use newly learned information because the game is about to end. In H6, testing the uncertain option may improve later choices. That contrast helps separate information-seeking for future advantage from choice variability that occurs even when exploration cannot pay off. The task’s points and scheduled outcomes are not drug rewards, real-world financial losses, chemical quantities, or clinical performance thresholds.

What distinguishes directed exploration from random exploration?

Directed exploration (DE) is operationalized through a model’s information bonus: a tendency to favor an option because it is relatively less known, especially when future choices remain. Random exploration (RE) concerns noisier selection that is not specifically directed at learning about an uncertain option. The model also estimates an initial learning rate and an asymptotic learning rate, which summarize how strongly observed outcomes alter beliefs under its assumptions.

All four are estimated computational parameters, not direct neurochemical measurements. A fitted value is not a blood level, not a dopamine concentration, not a direct brain recording, and not proof of conscious deliberation. Different choices, priors, parameterization, outlier rules and modeling assumptions could affect estimates. A person may make noisy decisions for reasons unrelated to pharmacological impairment, including familiarity, strategy, fatigue or task comprehension.

How was an aversive bodily state induced?

A silicone mask was used during both task runs. In one run, an inspiratory resistor produced additional breathing effort at 40 cm H₂O/L/second; the other run had no added resistance. The order varied, allowing the researchers to compare the same adult’s choices under the two conditions. The intervention produced discomfort and higher reported anxiety, not a drug dose. It was deliberately tolerable and cannot be equated with severe panic, oxygen deprivation, acute withdrawal or a medical respiratory emergency.

The investigators reported a significant anxiety-state manipulation and a group-related difference in its size. In the discussion, average anxiety ratings were 3.05 with resistance versus 1.46 without resistance on a scale with a maximum of 10. An experimental manipulation can change a subjective feeling without necessarily changing the model parameter whose change was predicted in advance. The results require distinct tests of anxiety, choice accuracy, exploration and learning.

Which differences in task accuracy were observed?

During the six-choice horizon, healthy comparisons had an estimated marginal accuracy of 0.81 and the MUD group 0.69. The model comparison was reported in the results as t(109) = 5.42, p < .001. Accuracy improved with successive choices and was higher when prior information was balanced across the two options. These are proportions of selecting the task’s higher average reward option, not percentages of traffic hazards recognized or factual claims remembered.

The model also reported a group-by-breathing-resistance accuracy interaction, F(1,2612.0) = 4.15, p = .042. Yet the descriptive within-group resistance effects did not reach significance: controls showed 0.82 versus 0.81 (p = .129), and MUD participants 0.68 versus 0.69 (p = .172). A supported interaction does not prove that resistance changed accuracy significantly within either group on its own.

Evidence table 1: actual sample and task outcomes

The proportions and model summaries represent distinct quantities. In particular, an estimated marginal learning-rate score is not a response time, a probability of impairment, a dose-response slope, or an assay result. Exact denominators and which analysis produced the value should travel with any quotation of the number.

How strong was the directed-exploration comparison?

For the primary DE parameter, the original group model gave F(1,109) = 5.45, p = .021. The difference was compatible with less information-directed selection in MUD participants. A separate horizon-by-group interaction on the underlying information bonus was stronger: F(1,340) = 12.60, p < .001. Comparisons showed a difference under H6 (p = .010), where learning could affect later choices, but not H1 (p = .843), where that opportunity was absent.

The central interpretation is therefore about adapting exploration to the number of future task opportunities, not a claim of universal inability to learn. When working memory was included as a covariate in a smaller subset, the original primary DE group term became nonsignificant (p = .082), although the more specific horizon-by-group interaction remained significant (p < .001). This distinction must be preserved; the two tests answer related but not identical questions.

Did the random-exploration result survive sensitivity tests?

Not consistently. The initial group comparison for RE lay at the conventional threshold, F(1,109) = 3.93, p = .050. When two potential outliers were removed, the group difference was nonsignificant (p = .209). The analogous horizon-by-group interaction in decision noise was also nonsignificant before adjusting working memory (p = .056). It became nominally significant in that subset after adjustment (p = .043), but the supplementary Bayesian interaction analysis did not support it (reported Bayes factor 0.18).

Consequently, the RE finding should be treated as fragile rather than presented with the confidence attached to the clearer information-bonus and initial-learning-rate tests. Reporting the original positive boundary result without the actual null sensitivity analyses would give readers an inflated impression of consistency. The null findings are substantive evidence about limitations of this model-specific result.

What were the genuine learning-rate findings?

The initial-learning-rate parameter had a bimodal distribution, so the investigators classified people into high- and low-estimated-value clusters for a logistic analysis rather than treating the parameter as conventionally normal. The reported high-value proportions were 0.67 in controls and 0.32 in MUD, with a significant group test, χ²(1) = 11.02, p < .001. The asymptotic-rate parameter was also lower in MUD, F(1,118.4) = 9.42, p = .003.

Those parameters are not exact percentages of synapses modified or a direct rate of biological recovery. Moreover, the supplementary model that additionally conditioned asymptotic rate on the initial rate did not consistently support an independently distinct remaining group effect. The modeling choices and covariance between parameters are significant interpretive limitations, even though the original simple comparisons were statistically supported.

Did breathing resistance change the primary computational endpoints?

The original Table 3 reports nonsignificant group-by-resistance interactions for every displayed primary model parameter: DE p = .214; RE p = .177; initial learning rate p = .685; asymptotic learning rate p = .674. Resistance main effects also were nonsignificant on these endpoints. These are tests actually performed, not outcomes simply left unmeasured.

The correct conclusion is not that discomfort never affects human judgment. This particular tolerable breathing manipulation raised reported anxiety yet did not establish the hypothesized differential changes in these computational estimates. The intensity may have been too modest; sample size and model sensitivity also matter. Conversely, the small accuracy interaction cannot legitimately be substituted for absent interactions in directed exploration or learning rate.

Evidence table 2: significant tests alongside source-reported nulls

The table includes results of mixed models, a logistic classification and a sensitivity analysis. It does not pool them into one clinical diagnostic accuracy statistic or suggest that repeated trials were separate patients. The preprint contains additional supplemental analyses; not all exploratory comparisons were independently replicated.

Did people with MUD simply avoid all negative outcomes?

The researchers attempted to distinguish avoiding uncertainty from choosing whichever option looked immediately more rewarding. When the option with higher observed mean reward was also less well known, the relevant choice probability was 0.82 in healthy comparisons versus 0.74 in the affected group; the group term was significant (p = .012). But adjustment for available working memory yielded p = .054, just outside the usual threshold. It would be inaccurate to present this secondary association as entirely covariate-robust.

An analogous test of preferring the higher-reward option when that option was also less uncertain showed no group difference: estimated probabilities 0.67 and 0.66, p = .895. This null contrast helped the authors frame the central behavioral distinction as uncertainty handling, not a generalized refusal of all unpleasant outcomes. The experimental task, however, is not a direct observation of a person’s decisions to abstain, enter treatment, or commit a real-world act.

Did trait cognitive reflectiveness explain the differences?

Cognitive Reflection Test scores were lower on average in the MUD group (0.84 versus 3.33 correct). Across groups, higher reflectiveness was associated with DE at a borderline p = .050 and with initial learning-rate classification at p = .008. But after adding working memory, the DE association was nonsignificant (p = .054), while the initial-rate relationship remained nominally supported (p = .007). Within MUD participants alone, the relevant relationships did not reach conventional significance.

The investigators explored mediation using cross-sectional measures. Even a statistically estimated indirect path does not demonstrate that past methamphetamine exposure caused less reflectiveness and thereby produced a particular choice. No intervention randomized cognitive reflectiveness, and mediator-outcome confounding cannot be excluded. The present results invite questions about how people update beliefs; they do not prove a mechanism or an effective treatment.

Were anxiety ratings, craving or abstinence timing linked to parameters?

Across the fitted computational analyses, self-reported anxiety and its interaction with resistance were not significantly associated with parameters (reported ps > .266 and > .117, respectively). Separate models using withdrawal, craving and drug-use symptom questionnaires generally found no supported symptom associations. Within affected participants, exploratory models using time since last methamphetamine use, treatment duration, medication or comorbidity also found no significant parameter effects (reported ps > .153).

These analyses are not proof that abstinence never changes cognition. They are cross-sectional tests in a selected, already treatment-engaged group, involving variable self-report timing and covariate availability. In particular, a nonsignificant relationship with reported days of abstinence is not a validated drug-elimination time course or evidence about any single person’s recovery trajectory.

Why are selection and setting important?

The clinical group came from two recovery programs, while healthy comparisons had no psychiatric history or psychotropic medication use. Several psychiatric and substance-use comorbidities were permitted in MUD, and some participants were receiving medication. Sex balance and trait anxiety, impulsivity, depression and working-memory performance differed between groups. These are potentially meaningful alternative explanations for cross-sectional group associations and reasons not to attribute all differences to one pharmacological cause.

Model-based random-effects methods can handle repeated trial dependence under their assumptions. They cannot supply randomization of life histories. Residual confounding, participant motivation, task experience, clinical environment, different treatment circumstances, and imperfect fit of the computational model remain possible. Nor does a source manuscript with 114 participants establish sensitive individual classification of all people with the same diagnosis.

What human forensic specimen and analyte did the researchers actually measure?

The behavioral endpoint was a computerized choice response, transformed into accuracy and computational parameters. Self-reported anxiety and psychological questionnaires were other outcomes. The source describes eligibility drug testing and breathalyzer procedures; these are distinct from a reported quantitative method identifying methamphetamine parent drug or amphetamine metabolite in legally collected human venous whole blood. No such whole-blood concentration, chromatographic calibration, uncertainty interval, analyte-specific limit of quantitation, or contemporaneous mass-per-volume result appears in the reported study.

Neither a participant’s history nor urine screening establishes a measured concentration in whole blood, serum, plasma, oral fluid, hair, sweat, vitreous humor, milk or genuinely exhaled breath. Equally, a fitted directed-exploration number of 4.58 has no ng/mL units. For foundational guidance on why one specimen cannot substitute for another, the same-domain published biological-matrices guide provides the relevant distinctions.

What does this mean for individual impairment and causation?

No one drove in a validated on-road or simulator impairment trial, and the study did not measure a collision, witness identification accuracy, legal consent, criminal intent, workplace injury, or a particular person’s reliability under oath. Participants were not challenged with experimentally controlled doses. The experiment reports neither diagnostic sensitivity and specificity for acute stimulant effects nor a validated behavioral threshold applicable to a legal case.

It would be a category error to take a group difference in uncertain reward choice and assert that an identified driver was impaired, an identified witness was unreliable, or a particular incident was caused by methamphetamine. A case-specific analysis requires actual contemporaneous clinical and behavioral evidence, authenticated specimens where relevant, analytical method and uncertainty, co-exposures, timing, alternatives and event reconstruction. The group study can inform a narrow scientific hypothesis but cannot replace those facts.

What is the proportionate overall interpretation?

Within the supplied original research, treatment-seeking adults with MUD on average showed less strategic exploration when learning about uncertain options could benefit later task choices. Initial model-based learning rate and accuracy also differed, while anxiety induction failed to generate the hypothesized effects on computational parameters. Several related results were fragile under covariate or outlier checks. This is more informative than either the overbroad claim of global drug-induced incapacity or the opposite claim that no behavioral differences were observed.

The manuscript analyzed here was a February 2025 preprint, subsequently associated with a June 2025 peer-reviewed publication of the same cohort. This provenance deserves clear labeling rather than treating a preprint and a later journal article as separate confirmatory studies. The primary findings are task-specific associations and controlled effects of mild breathing resistance, not directly measured methamphetamine pharmacokinetics or case-level causation.

Frequently asked questions

1. Was the independent sample 114 people or several thousand trials?

It was 114 adults: 56 treatment-seeking participants with MUD and 58 healthy comparisons. Each person performed repeated games under two breathing conditions. Trial-level denominator degrees of freedom are properties of fitted models and do not increase the biological sample.

2. Did the experiment administer methamphetamine or measure a blood dose-response?

No. Methamphetamine-use history was a group characteristic. The experimentally altered condition was breathing resistance. There was no reported quantified legal human venous whole-blood parent-drug concentration series and no randomized methamphetamine dose.

3. Is directed exploration simply another name for random choices?

No. Directed exploration concerns preferentially sampling uncertain options when future choices make information valuable. Random exploration describes less reward-determined variability without the same directed information incentive. Both are inferred from a model, not directly observed drug effects.

4. Did the breathing mask make the MUD group significantly worse at directed exploration?

No. Although resistance increased anxiety, the DE group-by-resistance interaction was nonsignificant at p = .214. The corresponding initial-learning and asymptotic-learning interaction p values were .685 and .674; the RE interaction was .177.

5. Were all cognitive-model group comparisons significant?

No. The initial RE result at p = .050 became nonsignificant after removing potential outliers (p = .209), and the primary DE group effect became nonsignificant in a smaller working-memory-adjusted sample (p = .082). The information-bonus horizon interaction remained positive after that adjustment, addressing a different, more specific question.

6. Can these parameters identify when someone last used methamphetamine?

No. Neither directed exploration nor a fitted learning rate has a validated conversion to last-use time, whole-blood ng/mL, dose, intoxication or legal impairment. Self-reported abstinence averages and eligibility testing do not establish individualized pharmacokinetics.

7. Does the study prove the cause of a driver’s crash or a witness’s testimony?

No. It did not measure a validated real-world driving, collision, eyewitness or legal-event endpoint, and participants were not randomized to a methamphetamine-use history. A separate case-specific record is necessary before making any individual causal inference.

Primary source and scope

Goldman CM, Takahashi T, Lavalley CA, et al. Individuals with Methamphetamine Use Disorder Show Reduced Directed Exploration and Learning Rates Independent of an Aversive Interoceptive State Induction. medRxiv, complete original empirical preprint, posted February 24, 2025. Original preprint DOI 10.1101/2024.05.17.24307491. This independent explanatory article preserves the source’s positive and nonsignificant tests. It is not a novel experiment, a clinical diagnostic protocol, a case-specific toxicology opinion, or a claim of legal impairment.