Can Reward Prediction Errors Explain Learning Differences in Methamphetamine-Use History? A 168-Person Study

Short answer: In one cross-sectional behavioral experiment involving 168 adults, reward predictions and recognition differed in selected comparisons between 82 people with a history of methamphetamine-use disorder and 86 substantially younger nonusers. The clearest group-by-gender interactions occurred in the size of reward-prediction errors and in a calculated learning-rate measure. In women, the previously dependent group made larger average prediction errors and showed a weaker relationship between error size and next-trial updating; the analogous group comparison in men was not significant. Recognition accuracy also differed by overall group, but its formal group-by-gender interaction and its interactions with prediction error were not statistically significant. These are experimental task results, not a direct measurement of methamphetamine or dopamine in human blood, and not a determination of any individual’s impairment or responsibility for an event.

The original primary report is YuQing Liu and colleagues, Reward Prediction Errors Drive Reinforcement Learning and Recognition Memory: Gender Differences in Methamphetamine Abusers, Brain and Behavior 16 (2026), e71630, DOI 10.1002/brb3.71630. Its useful forensic contribution is a methodological distinction: an error signal computed from an on-screen reward estimate is not a measured drug concentration, and a difference in a selected behavioral endpoint is not proof of acute intoxication. The distinction matters most when the original study reports both statistically supported comparisons and closely related null results.

What precise scientific question did the investigators test?

The experiment asked whether an unexpected reward is converted into a subsequent change in expectation differently among adults with a recorded history of methamphetamine-use disorder and nonusers. A person predicts a reward, receives a value assigned by the task, and then has an opportunity to revise the next prediction. The gap between forecast and feedback is a reward-prediction error. The study also asked whether images previously paired with feedback were recognized after a delay. These are separate questions: changing a numerical forecast is one type of learning, while identifying an image as previously encountered is a recognition decision.

The main sex-specific hypothesis concerned whether the relation between calculated surprise and the size of an adjustment varied across existing groups of men and women. The paper includes several layers of testing: descriptive task averages, analyses of covariance that adjusted for age, follow-up subgroup comparisons, and generalized linear mixed models. They do not all yield the same significance pattern. A sound reading therefore reports the model and outcome alongside its numerical result rather than declaring a global effect on all cognition.

How many genuinely independent people were studied?

Researchers initially recruited 180 people: 90 with a methamphetamine-use history and 90 nonusers. Records were excluded for incomplete task data or task accuracy below the specified 50% threshold. The analyzed sample was 82 people in the methamphetamine-history group (43 men and 39 women) and 86 controls (45 men and 41 women). The independent biological denominator is therefore N = 168, not 180 for the final analyses and not thousands of person-equivalents created by repeated task trials.

The four sex-by-history cells matter because a significant interaction can arise from different directions or sizes of group contrasts rather than from one constant effect common to all adults. In the history group, male and female participants also had different average reported durations of earlier drug use. The sample was not a random population sample of drivers, patients attending emergency departments, or witnesses giving evidence. Its denominators are most informative for the specific laboratory tasks and the selected recruitment settings.

Why are the control and methamphetamine-history groups not interchangeable?

The methamphetamine-history participants came from compulsory rehabilitation facilities in Zhejiang, China, whereas comparison volunteers were recruited separately. The overall mean age was 32.37 years in the history group and 19.86 years among controls. The approximately 12.5-year gap is striking. Male history-group participants averaged 37.05 years versus 20.63 years in male controls; female history-group participants averaged 27.08 versus 19.03 years. Recruitment location, age, education, health characteristics and circumstances of supervised abstinence could be associated with task performance.

Age was entered as a covariate in the reported analyses. Adjustment helps address the specified measured variable but does not make two differently recruited groups experimentally exchangeable. It cannot eliminate all unmeasured confounding, explain differences in how rehabilitation subjects and community controls approach incentives, or recreate the same participants under alternate lifetime drug histories. The findings describe observed associations, not a controlled test of an assigned history of substance exposure.

Was anyone given methamphetamine as part of this experiment?

No. The study did not administer an acute experimental dose of methamphetamine, randomize people to a stimulant or placebo, or record postdose concentrations over time. The relevant exposure classification was an existing history and rehabilitation status, with the history group already more than a month into supervised rehabilitation. The article expressly notes the lack of detailed frequency, administration route and abstinence-duration data. Reported years of previous drug use are historical characteristics, not administered milligrams or independently observed last-use times.

This absence is central to toxicological interpretation. A controlled 20-mg oral stimulant challenge, an unverified lifetime history, and a chemically measured blood concentration are three different evidence types. The present study contains the second type, not the first or third. It cannot generate a drug-specific dose-response curve for a contemporaneous offense, estimate a person’s methamphetamine elimination rate, or establish the pharmacological state of someone whose blood was collected at a legal event.

What did the 64 feedback trials actually require?

Each participant completed eight practice trials, followed by 64 learning trials. A geometric cue appeared for two seconds; the volunteer then had up to five seconds to estimate an average reward value. A task-assigned feedback value followed. The possible feedback values were 30, 35, 40, 45, 55, 60, 65 or 70 points. These figures are task points, not measured drug doses, human serum concentrations, clinical scores or money that establishes an actual loss. The investigators drew from a pool of 80 geometric images to construct learning and later recognition stimuli.

The procedure asked for an estimate of the overall mean reward value across images. It did not require someone to learn a personal drug reward, gamble with actual legal damages, or identify an intoxicated individual. Because the images and trials recur within a participant, 64 observed estimates supply repeated information about that person’s behavior. They do not turn 168 independently observed people into 10,752 independent study subjects. Statistical analysis must respect this clustering.

How was reward-prediction error calculated?

For a learning trial, the investigators defined the signed reward-prediction error as the assigned feedback value minus the participant’s prediction. For example, predicting 40 task points and receiving 60 produces a signed error of +20; predicting 60 and receiving 40 produces −20. The absolute error is 20 points in both cases. Taking the absolute value removes direction: it measures the size of surprise in that task, not whether the result was better or worse than anticipated.

That distinction makes the term “reward prediction error” easy to misuse in legal writing. It is a computed behavioral variable built from a person’s typed estimate and a programmed feedback number. It is not a dopamine-neuron recording, a brain chemical assay, a measured pharmacokinetic error, or a toxicology instrument’s measurement uncertainty. The task-generated number can be statistically related to learning without demonstrating a unique neurological mechanism. The authors explicitly describe dopamine-system explanations as hypotheses requiring further direct study.

What did the investigators call a learning rate?

The study estimated a trial-related updating coefficient by dividing the change in successive reward estimates by the preceding signed prediction error. Conceptually, a larger update for a given error reflects a greater adjustment in the next prediction. The article presents learning rate on a nominal zero-to-one scale; the mathematical quantity is a derived behavioral parameter rather than a physiological percentage of neurons or a drug-response magnitude. Its units and calculation differ from response time, recognition accuracy, and the absolute prediction error in task points.

The equation depends on sequential estimates and error denominators, so trial-level peculiarities and near-zero errors require care in interpretation. The published article does not supply a full participant-level audit of these edge cases. The most defensible use of the result is to describe what the specified task and analysis reported, including its age adjustments and follow-up models. It would be unsupported to translate the fitted updating parameter into a validated generalized mental-capacity or impairment scale.

Numerical evidence table 1: the participants and descriptive results

These rows describe the original numerical distributions; the formal claims of statistical significance depend on the analyses below. In particular, a difference between two displayed means does not automatically prove a statistically significant contrast after adjustment. The large age separation and differing recruitment sources remain present even where an adjusted test yields a low p value. None of the table’s values has mass-per-volume units, and none should be reported as a human whole-blood methamphetamine result.

Were prediction errors significantly larger in the history group overall?

The age-adjusted main effect of history group on absolute prediction error was not significant: F = 0.939, p = .334. The main effect of gender was also nonsignificant (F = 1.087, p = .299). However, the group-by-gender interaction was statistically significant (F = 13.126, p < .001). These findings should be stated together. The significant interaction does not retrospectively make either nonsignificant main effect significant.

The follow-up group contrast among women showed larger absolute errors in women with a methamphetamine history than in women without one (F = 11.053, p = .001). In men the history-versus-control difference was not statistically significant (F = 0.882, p = .349). This specific pattern supports an interaction in the fitted model. It does not say that every female participant was more error-prone than every female control, or that no individual male could make large prediction errors.

Which learning-rate comparisons were significant, and which were genuinely null?

For the calculated updating rate, the investigators reported a significant group main effect (F = 8.368, p = .004), a gender main effect (F = 7.739, p = .006), and a group-by-gender interaction (F = 12.485, p < .001). The adjusted history-versus-control difference was prominent among women (F = 27.666, p < .001) but not significant in the male comparison (F = 0.273, p = .602). The latter is a genuine published null test, not an endpoint omitted from the experiment.

Within the methamphetamine-history participants, the male-versus-female learning-rate contrast was itself nonsignificant (F = 0.085, p = .771). Among nonusers, the male-versus-female contrast was significant (F = 27.732, p < .001). Thus a group-by-gender interaction should not be retold as a proven difference between women and men inside the history group. The relevant statistical questions have different comparators and cannot be substituted for one another.

Did larger errors lead to more efficient learning among the women?

The follow-up female-only generalized mixed model estimated a negative association between absolute reward-prediction error and learning rate in the control group (slope approximately −0.026) and a less negative corresponding slope in the methamphetamine-history group (approximately −0.006). The group-by-error term was +0.020 with p = .002. The main terms in that female model were also significant: group estimate −0.382, p < .001, and absolute-error estimate −0.026, p < .001.

These signs matter. A less negative slope is not the same thing as an observed increase in chemically measured dopamine, and the model does not directly demonstrate that an error “caused” learning in a specific woman. In men, the absolute-error term was nominally significant (β = −0.004, p = .048), but the group main term (p = .085) and group-by-error interaction (p = .302) were nonsignificant. Repeated observations from the same participants were incorporated in these models, but the model coefficients remain observational associations.

What did the recognition-memory test count as a correct answer?

After a 10-minute break, volunteers received 32 image judgments: 16 images already seen during learning and 16 new images. They indicated old or new and rated confidence on a four-point scale. The main published recognition accuracy concerned old-image hits, and the investigators excluded correct old-image responses scored as complete guessing. That is a specific operational definition of recognition performance, not a comprehensive measure of everyday episodic memory. It leaves separate questions about lure false alarms, confidence calibration and eyewitness identification outside the principal reported test.

The test contained relatively few studied images, and its short retention interval cannot represent a multiday witness interview, a months-old recollection or an emotionally charged incident. Feedback-point surprise might influence subsequent recognition in theory, but this study evaluated a narrow recognition relationship using its own images, criteria and statistical model. The published recognition findings need to be assessed independently of the updating-rate findings rather than presumed to share the same error mechanism.

Was there a verified gender interaction in recognition accuracy?

Recognition accuracy showed an overall adjusted group difference (F = 5.993, p = .015). The gender main effect was nonsignificant (F = 0.747, p = .389). Critically, the formal group-by-gender recognition interaction was also nonsignificant (F = 2.324, p = .129). A significant subgroup contrast between female history participants and female controls (F = 12.338, p < .001) does not override the nonsignificant omnibus interaction.

The male history-versus-control recognition comparison was nonsignificant (F = 1.236, p = .268). Nor was the sex comparison significant within the methamphetamine-history group (p = .144) or within controls (p = .643). It is therefore more accurate to report an overall group association, a selected female subgroup contrast and a null formal moderation test than to announce a generalized proven difference in recognition effects between the sexes.

Did the error size statistically predict recognition memory?

The recognition mixed models did not replicate the female learning-rate interaction. Among women, absolute prediction error was nonsignificant (p = .169), as were group (p = .550) and their interaction (p = .748). Among men, error size (p = .569), group (p = .920), and error-by-group (p = .911) were likewise nonsignificant. These are directly tested null results, not merely an absence of reported blood testing or a statement that a clinical endpoint was never assessed.

The contrast matters scientifically: the study associates reward prediction errors with selected updating-rate patterns but does not establish that those same errors enhanced old-image recognition under its models. The small memory sample of 16 old images and scarcity of extreme prediction errors may have reduced the experiment’s ability to identify subtle links. But limited power is not a license to reverse the actual null result or claim that a neural memory mechanism was experimentally proven.

Numerical evidence table 2: supported tests beside their real null counterparts

Reported p values are measures of model-based compatibility under the analyses used, not the probability that a specific driver was impaired or the probability a recalled event occurred. Several related outcomes and subgroup comparisons were considered; the article does not establish one single multiplicity-corrected forensic classification rule. An outcome that was not chemically measured must never be described as a statistically negative chemical assay.

How serious is the large age imbalance for causation?

The age separation is the most conspicuous observed limitation. Even within the female cells, the history group averaged roughly eight years older than controls; in the male cells the difference was more than sixteen years. A covariate correction assumes an adequate statistical model for the relation between age and the dependent variable and sufficient comparability of participants after adjustment. Neither assumption is equivalent to observing people under randomized conditions. Recruitment from compulsory facilities rather than the same community pool creates additional potential selection differences.

Other unmeasured variables include specific duration of abstinence, past doses, route, frequency, nicotine or other exposures beyond the protocol exclusions, sleep, psychiatric assessment granularity, stress and familiarity with screen tasks. The authors note that estimating a mean reward across trials also imposes working-memory demands. Accordingly, poorer performance on a calculated updating outcome need not be a pure measure of one hypothesized reward system, much less a demonstrable consequence of a particular ingested methamphetamine amount.

Did the experiment directly measure dopamine signaling or brain activation?

No. There was no positron-emission imaging of transmitter release, electroencephalography, functional MRI, functional near-infrared spectroscopy, direct dopamine assay, or recording from human neurons as an endpoint of this experiment. Neurobiological mechanisms appear in the background and discussion as explanatory possibilities. The original authors acknowledge that the hypothesized dopamine adaptation is speculative because their experiment collected behavioral measures. A mathematical reward-prediction error is not a neural action potential.

A mechanistic account might be worth testing in a future experiment that measures relevant neural signals and chemical exposure, but one cannot convert the 168-person behavioral dataset into such a study retrospectively. Dopamine concentration, synaptic release, transporter occupancy and legal blood methamphetamine concentration are different quantities. An article seeking AI-search visibility should preserve those distinctions instead of treating the familiar language of reward prediction as proof of brain chemistry in these actual participants.

Which biological specimens and toxicology assays were absent?

The study published no controlled quantitative measurement of methamphetamine, amphetamine or an enantiomer in legally obtained human venous whole blood. It also did not quantify those compounds in separated plasma or serum, urine, oral fluid, hair, vitreous humor, sweat, milk, breath, tissue or an artificial spiked validation matrix. Thus there is no corresponding chromatographic instrument, calibration range, detection limit, limit of quantitation, analyte-specific mass-per-volume unit, specimen collection clock or pharmacokinetic curve to apply to its behavioral findings.

Whole blood is not analytically interchangeable with plasma or serum because separation changes the physical matrix and potentially the concentration ratio. A qualitative screening procedure, if one is described elsewhere in a person’s record, is different again from a validated quantitative parent-drug result. For general principles, the site’s already published biological-matrices guide explains why a measurement must remain tied to its original specimen. The separate confirmatory-testing guide addresses the analytical steps needed to identify or quantify an actual chemical target.

Can reward-learning statistics be converted into a blood drug level or last-use clock?

No. The learning-rate parameter describes how a task estimate changes in relation to a task feedback discrepancy. It contains neither the concentration of a drug at a known sampling time nor a measured elimination-rate constant. Even if a task endpoint showed a reliable association with a historical diagnosis, calculating an unknown present concentration or elapsed time since an unobserved dose would require an independently validated individualized pharmacokinetic model and corresponding specimen evidence. Neither is supplied here.

Nor could the reported score establish whether a person was experiencing acute effects rather than a historical association, treatment-related condition, sleep loss or another source of variability. Forensic interpretation should keep separate the questions of identity and amount of a chemical, the timing of exposure, the presence of a functional effect, and whether that effect contributed to a particular event. This study answers none of the latter individualized questions.

Does this validate an individual driving, workplace, or witness-memory conclusion?

No on-road or simulated driving was tested. There was no vehicle-control task, divided-attention impairment cutoff, braking standard, sobriety-test validation or crash reconstruction. Feedback predictions in a stationary screen experiment have no established sensitivity or specificity for detecting impaired driving in a particular individual. A statistically significant group interaction is not a person-level clinical or forensic classifier. A participant’s group mean cannot be assigned as an individual’s blood result or actual driving performance.

Likewise, recognizing a geometric image after a ten-minute interval is not a study of eyewitness testimony, the accuracy of an assault or crash recollection, legal competence, deception or a particular witness’s credibility. Those are different constructs with different evidence requirements. Using the study as background on behavioral research may be appropriate, but an expert evaluating a case must analyze the actual collected specimens, validated methods, incident evidence, timing and independently assessed functional endpoints.

What would a stronger follow-up study need?

A stronger causal design would prospectively recruit appropriately age-matched groups using comparable settings, characterize education and relevant co-exposures, and measure precise prior exposure and abstinence histories. If the scientific question concerns an acute pharmacological effect, ethically approved randomized administration or another causal identification strategy would be needed; a historical group comparison cannot substitute for one. Trial-level data and explicit handling of near-zero prediction-error denominators would improve reproducibility of the learning-rate estimator.

To evaluate a proposed chemical-behavioral relation, investigators would need validated drug assays in specified real physical matrices at documented times, with prespecified analytes and quality controls. To test a claimed neural mechanism they would need actual neural or neurochemical measurements rather than modeled screen scores. To make an individual-driving claim they would need an independently validated driving outcome. None of these desiderata should be mistaken for measurements made in the 2026 paper.

What is the narrow scientifically defensible conclusion?

Within the original 168-person laboratory sample, past methamphetamine-group membership and gender interacted for absolute reward-prediction errors and calculated feedback learning. Female subgroup contrasts and a female-only error-by-group learning model were significant; corresponding selected male comparisons were not. Recognition showed a group association and a female subgroup difference, but its formal group-by-gender interaction was nonsignificant, and neither sex showed a significant error-by-group recognition interaction. The unequal ages, recruitment settings and unmeasured exposure characteristics constrain causal interpretation.

That precise combination of positive and null findings is more informative than the vague claim that methamphetamine simply “damages reward learning” or “impairs recognition in women.” This was not an acute methamphetamine challenge, was not a confirmed chemical exposure series, and did not measure a person’s legal blood sample, clinical impairment, real-world driving or the cause of an event. Careful use of the study requires keeping group-level computational behaviors distinct from molecular toxicology and from an individual forensic conclusion.

Frequently asked questions about reward prediction and methamphetamine history

Does the study prove methamphetamine acutely caused a learning change?

No. The participants were classified by an existing methamphetamine-use history and different recruitment settings. No controlled study dose or blinded acute stimulant-placebo comparison was conducted. Confounding, including marked group-age differences, prevents a person-specific causal attribution.

How many human subjects contributed to the primary analysis?

There were 168 independently analyzed adults, comprising 82 history-group and 86 nonuser participants. A larger recruited total of 180 was reduced after exclusions. The 64 learning trials and 32 subsequent image judgments were repeated observations, not additional independent people.

Were absolute prediction errors higher in all methamphetamine-history participants?

No significant overall group main effect was reported for absolute prediction error (p = .334). The formal group-by-gender interaction was significant (p < .001), and the female group comparison was significant (p = .001); the male group comparison was nonsignificant (p = .349).

Was memory impairment significantly different by gender?

The formal recognition group-by-gender interaction was nonsignificant (p = .129). A significant female subgroup comparison does not establish that the effects of history were statistically different between sexes. Recognition error-by-group terms were also nonsignificant in both sex-specific models.

Did investigators chemically measure methamphetamine in venous whole blood?

No. Reward feedback points, learning-rate coefficients and image recognition are behavioral measures. The study did not provide chemically quantified parent methamphetamine or amphetamine concentrations in human venous whole blood, plasma, serum, urine or other forensic specimen matrices.

Can the learning-rate number identify someone who recently used methamphetamine?

No. No validated sensitivity, specificity, concentration-response model or pharmacokinetic conversion from this screen task to individually verified recent drug use was established. The study also lacked reliable detailed routes, amounts and specific elapsed abstinence times.

What would be needed to connect these results to a particular legal incident?

Case-specific specimens and chain of custody, analytically validated assays, documented timelines, relevant behavioral observations and an appropriate incident-specific causal analysis would be needed. A group-level laboratory association cannot independently establish intoxication, driver error, witness truthfulness or event causation.

Original source and scope of this analysis

Primary empirical research: YuQing Liu, Die Wang, LePing Wang, SongSong Wang, ZhiYing Chen, Tiancheng Wang, Shengyi Jin, Qiang Zhou and Liping Yang (2026), Reward Prediction Errors Drive Reinforcement Learning and Recognition Memory: Gender Differences in Methamphetamine Abusers, Brain and Behavior 16(9):e71630, original published study DOI 10.1002/brb3.71630. The source is a published full-text original empirical study. This independently composed explanatory account is limited to its reported methods and aggregate numerical findings. It neither supplies a clinical assessment nor substitutes for a validated forensic laboratory report or incident-specific analysis.