Why Can Methamphetamine-Dependent Adults Show Similar Average Stop Times but Weaker Neural Error Updating? A 96-Person fMRI Study

Answer: In a study of 62 recently abstinent methamphetamine-dependent adults and 34 comparison participants, the groups did not differ significantly in mean go reaction time, mean stop-signal reaction time, or their overall stop-error rate. Nevertheless, certain stop-signal difficulty interactions and adjustments after a stopping error differed, as did the functional-MRI responses associated with a modeled Bayesian surprise signal in the caudate, orbitofrontal cortex, thalamus, and inferior parietal lobule. This is evidence about selected laboratory behavioral patterns and brain-hemodynamic responses, not a quantitative test of methamphetamine in blood, a diagnosis of impairment in an individual, or proof of why a driving incident occurred.

The primary research is Harlé, Zhang, Ma, Yu, and Paulus, Reduced Neural Recruitment for Bayesian Adjustment of Inhibitory Control in Methamphetamine Dependence, Biological Psychiatry: Cognitive Neuroscience and Neuroimaging (2016), DOI 10.1016/j.bpsc.2016.06.008. Its distinctive contribution is to analyze what happens before a stop cue, immediately after a failed stop, and when brain activity is aligned with the expected-versus-observed difference. Those questions are more specific than asking whether people with a methamphetamine-use history are simply “slower.”

What is the main finding, and what is not established?

The central contrast is between ordinary task averages and adaptation to trial history. Group-average speed and an overall rate of failed stops did not show the simple, significant deficit one might have expected. Yet the probability of failing on difficult stop trials interacted with diagnostic group, and the time adjustment made after a stop error differed from the adjustment after a successful stop. This matters because a person may maintain a similar average speed while reacting differently to particular cues or errors. Conversely, a significant brain-imaging interaction does not require a significant difference in every behavioral aggregate.

The investigators also examined an inferential process: how much a person should anticipate encountering a stop signal on a new trial after seeing the recent sequence. A computational model translated the sequence into a trial-by-trial estimated stop probability. It then assigned an unsigned prediction error to the magnitude of the difference between predicted and observed events. Some brain regions tracked this model-derived surprise measure differently across the existing groups. The experiment did not measure an isolated neural molecule or prove that the model was the only mental process available to participants.

Who were the 96 independent participants?

The primary comparison involved 62 methamphetamine-dependent individuals and 34 healthy control subjects. These are 96 independently recruited human beings. The dependent group was evaluated after entry into residential treatment, with task testing in the third or fourth treatment week. It was not a randomized experiment in which one set of otherwise equivalent volunteers received methamphetamine and the other received placebo. Nor was it a longitudinal observation of the same people before use, during use, and after complete recovery.

The controlled laboratory blocks supplied many repeated behavioral observations per individual, but did not expand the biological sample. In the source demographic comparison the methamphetamine-dependent group averaged 38.0 years of age, versus 36.1 among controls, a nonsignificant age comparison. Education, in contrast, differed: averages were 13.0 versus 14.8 years, respectively, with p < .001. Background differences in past use of substances other than methamphetamine also existed. Such differences are material when deciding whether a group association can be assigned a single causal explanation.

Did the study administer methamphetamine or measure an acute dose?

No experimental methamphetamine challenge was administered. The methamphetamine-dependent volunteers were recruited based on clinical history and treatment status; the investigators used screening procedures, including urine toxicology, in connection with treatment sobriety. That is not the same as administering a known mass of methamphetamine or measuring a postdose concentration series. The article does not provide a validated quantitative methamphetamine concentration, its d- or l-isomer distribution, a corresponding amphetamine metabolite concentration, or a blood elimination slope for each participant.

This distinction changes the question that the study can answer. Its evidence concerns existing groups with different use histories, abstinence and treatment circumstances. It cannot estimate a measured milligram dose that produced a given fMRI contrast. It also cannot determine the exact final use time of someone outside the study. A positive historical classification or a negative clinical screening result is not a substitute for a quantitative, specimen-specific forensic determination.

How did the stop-signal task work?

Participants completed six functional-MRI blocks of a stop-signal task. A majority of trials instructed them to act promptly on a displayed stimulus. On a minority of trials a later auditory stop cue asked them to withhold the prepared response. The original design specifies 216 go trials across blocks—36 in each block—and stop trials constituted the remaining one-quarter of the scheduled task. Thus an individual’s many button presses and stop attempts are nested within that person; they should never be counted as separately randomized human subjects.

Stop-signal delay, abbreviated SSD, is the interval from presentation of the action stimulus to the later stop cue. A stop cue close to the response is generally harder to obey than an early one. Researchers calibrated delay levels to performance, allowing comparisons of stop-cue difficulty rather than relying on a single delay for all people. A failed stop is an operational event in this experiment: a response occurred when stopping had been requested. It is not an observation of a driver failing to brake in traffic.

What did the Bayesian model actually calculate?

The model estimated P(stop), a trial-specific expectation that the next event would require stopping. Each go or stop observation was treated as information that could revise that expectation. The procedure drew on a probabilistic learning framework rather than asking participants to report, in their own words, their perceived likelihood of a stop cue. A stronger expectation can rationally lead somebody to pause before initiating a response; this may improve readiness to stop at a cost in go-response speed.

The model’s signed prediction error retains the direction of the discrepancy: whether a stop event was more or less likely than expected. The unsigned prediction error represents the absolute magnitude, irrespective of direction. The latter is often described as “surprise.” In this task it is a numerical quantity generated from the programmed trial sequence and an observer model. It is not a laboratory measurement of drug dose, dopamine concentration, actual synaptic signaling, or a direct reading of an individual’s beliefs.

For the brain analyses, trial-by-trial model values were linked to functional-MRI signal changes. This permits a test of whether hemodynamic responses covary with computational quantities under specified assumptions. Correlation with a model regressor, however, is not proof that each brain region performed the model’s calculations in the same way a software program does.

Were average go and stop reaction times different?

No statistically significant overall group difference was established. Mean go reaction time was 590 milliseconds in the methamphetamine-dependent group and 624 milliseconds among controls. The source’s main group test for go reaction time was χ²(1) = 0.48 with p = .49, a genuinely nonsignificant finding. Notice that the dependent group even had the numerically shorter mean. It would therefore be incorrect to summarize this source as demonstrating slower mean responding in all people with methamphetamine dependence.

Estimated stop-signal reaction times likewise averaged 231 milliseconds in the methamphetamine-dependent group and 239 milliseconds among controls. The reported between-group test was p = .61. Stop-signal reaction time is an inference from a response-inhibition paradigm, not the duration of a measured braking maneuver in a vehicle. A nonsignificant difference does not prove complete equivalence between groups, but it does mean this experiment did not support the asserted average difference at its tested level.

How did expected stopping probability affect response speed?

Across the sample, higher modeled P(stop) was associated with longer go response time. The reported estimated effect was 274 milliseconds per unit change in the modeled probability, with t(94) = 4.6 and p < .001. This slope is defined over a full unit change on the model’s probability scale; it should not be described as an observed 274-ms slowdown after every stop cue. The actual trial-by-trial probability differences are usually smaller, and participants’ response distributions vary.

The investigators reported positive probability–latency slopes within both groups. One interaction between group and probability approached the conventional significance boundary, χ²(1) = 3.8, p = .05. Because this is marginal, and because other neural and behavioral effects were tested, it deserves more caution than a robust replicated finding. It does not justify asserting that one participant has a particular adaptive-learning impairment or that a specific case event arose from a deficient estimate of stopping probability.

What did stop-error rates show?

The overall proportion of failed stops was approximately 0.49 in the dependent group versus 0.44 in the controls. The main effect of group did not reach significance: χ²(1) = 2.3, p = .13. Nevertheless, the interaction between group and the actual delay of the stop cue was statistically significant, χ²(1) = 4.8, p < .05. The authors interpreted that interaction as higher failure likelihood for the dependent group as stopping became more difficult. In other words, the evidence is conditional on trial difficulty, not an unqualified global error-rate difference.

Longer stop delays were strongly related to greater failure probability across people, which confirms that the manipulation affected the intended challenge. Higher modeled P(stop) was, in turn, associated with lower probability of error, with an overall odds-ratio estimate of 0.10 and p < .001. The study did not detect significant group interactions for several other modeled probability terms. These genuine nulls matter, because selective differences cannot legitimately be expanded into a claim that every Bayesian adjustment was abnormal.

Evidence table 1: the behavioral tests and real null results

This table deliberately keeps milliseconds, a unitless error proportion, a statistical interaction, and a modeled probability slope distinct. It does not convert the absence of statistical significance into proof of identical neurological function, and it does not reinterpret probability as a drug-testing threshold.

What happens immediately after a stopping error?

The trial following a stop signal can reveal how participants adjust their behavior. In the comparison group, response time on a go trial was not shown to slow significantly after a successful stop relative to a preceding go trial (estimated 6 milliseconds, p = .18). After a failed stop, however, controls showed an estimated 16-ms slowing, with p < .01. Such a pattern is consistent with a response adjustment particularly associated with an error, rather than a uniform pause whenever a stop signal appeared.

The methamphetamine-dependent group showed different sequencing. Go responses after a successful stop were slower than after a go trial by an estimated 31 milliseconds, p < .001. The incremental slowing after a failed versus successful stop was estimated at 10 milliseconds and was not significant, p = .19. A group interaction involving previous-trial type was significant, χ²(1) = 41, p < .001. It is therefore more precise to describe a changed pattern of poststop adjustment than to suggest that the dependent group simply lacked the ability to delay a response.

These measurements do not establish whether a person’s response strategy would remain the same outside the scanner, after additional recovery, under severe fatigue, or during a life-critical decision. The delay is a computer-key response to abstract stimuli. The inferential leap from that response to actual road braking requires another, independently validated study.

Which imaging regions differed in the stop-versus-go model contrast?

The authors tested how functional-MRI signal modulation by P(stop) interacted with trial type and group. Seven spatial clusters passed their specified whole-brain statistical procedures for that contrast. Two especially relevant regions were the left posterior caudate and left orbitofrontal cortex. In controls, their direction of signal modulation across go versus stop trials resembled tracking of the unsigned difference between predicted and observed events; in the dependent group the corresponding model-related modulation was attenuated.

In the left posterior caudate, the table reports a Talairach peak at (−15, −14, 26), peak Z = 3.57, and a 16-voxel cluster. The left orbitofrontal region (BA 11) has a peak at (−38, 47, −11), Z = 3.42, and 14 voxels. The authors also reported group contrasts in frontal and anterior-cingulate regions, including a right anterior cingulate peak Z = 4.71. An imaging contrast within selected voxels must not be relabeled a demonstrated whole-brain dopamine abnormality or an individual’s neurological diagnosis.

The study’s Table 2 prints a peak voxel p-value alongside each Z statistic. Its image-level multiple-comparison control used a separate cluster-size and threshold procedure; the printed voxel p-value is not itself a familywise-corrected probability for the entire effect. That technical distinction prevents apparently very small voxel p-values from being presented as if they were universally corrected clinical certainty.

How did unsigned prediction-error responses differ?

An unsigned prediction error (UPE) is high when a trial is more surprising under the model, whether it is unexpectedly a stop or unexpectedly a go event. The authors compared this model regressor with signals in regions showing relevant group interactions. They reported attenuated UPE-related responses in the dependent group in posterior caudate and orbitofrontal regions, with group effect-size estimates of Cohen’s d = 0.45 and 0.42, respectively. These summarize group-level signal differences, not individual concentration–response coefficients or guaranteed features of every participant.

A source figure shows that the dependent group’s UPE-related coefficient in one region was not significantly different from zero. This does not prove that the person’s brain failed to process surprise; it means the modeled coefficient did not meet the study’s significance test in that specified region and sample. A model coefficient, like a nonsignificant psychometric result, carries sampling uncertainty and depends on the design and assumptions used to obtain it.

What differed after successful and failed stops in the brain?

A second set of imaging contrasts addressed correct versus erroneous stop trials. Among the regions showing group-dependent P(stop) modulation, the right thalamus and right inferior parietal lobule showed patterns that the authors interpreted as selective UPE signaling after stop errors in control participants. The right thalamus peak appeared at (11, −18, 19), peak Z = 2.99; the right inferior parietal lobule at (54, −51, 41), Z = 3.39. Their respective group-related UPE effect-size estimates were d = 0.41 and d = 0.62.

In controls, the thalamic/parietal model-related error response was associated with fewer failures on difficult stop trials; for one reported relationship the coefficient was r = −0.47, p < .05. The association is exploratory and conditional on the experiment’s model. It is not proof that increasing activation would prevent a crash, and it does not establish pharmacological reversibility after methamphetamine abstinence. The dependent group did not exhibit the same significant modeled effect in the authors’ tests, but one non-significant coefficient alone is not a claim of total absence of neural error processing.

Evidence table 2: source-reported imaging peaks and effect sizes

These are coordinates and group-level imaging statistics, not normative thresholds for a driver’s brain. Voxel or cluster statistics cannot be converted into a probability that a particular person was methamphetamine-impaired. The researchers analyzed selected contrasts and estimated model-based associations; no individual forensic classification algorithm or external driving validation was provided.

Did this experiment quantify methamphetamine in blood?

No. The authors did not report postdose parent methamphetamine or amphetamine-metabolite concentration in legally collected human venous whole blood. They also did not report a quantitative parent/metabolite measurement in separated plasma or serum, oral fluid, hair, sweat, vitreous humor, exhaled breath, or milk. No calibrated drug-specific LC–MS/MS or GC–MS concentration series, limit of quantification, validated chain of custody, or concentration-to-neural-response conversion was supplied by this original behavioral-imaging experiment.

One must distinguish three things: a history consistent with substance dependence, a qualitative screening procedure connected with sobriety monitoring, and a chemically confirmed quantitative result in an explicitly identified specimen. These can answer different questions. None may be silently substituted for another. For general specimen distinctions, see the independently published discussion of biological matrices in forensic toxicology. For what an analytical confirmation means, consult the separate guide to confirmatory drug testing.

The fMRI contrast is particularly distinct from a blood chemistry result. Blood-oxygen-level-dependent signal reflects changes linked to hemodynamic and metabolic responses during a cognitive task. Its units and mechanisms are not those of a measured drug concentration. A larger, smaller, or nonsignificant BOLD model coefficient tells us nothing quantitative about how many nanograms of methamphetamine were present in someone’s venous blood at a legally relevant time.

What background differences limit causal interpretation?

The two groups entered the study with different experiences. Years of education and reported alcohol, tobacco, cocaine, and cannabis histories differed, with some covariate adjustment incorporated into statistical models. Treatment participants were in a structured residential environment, which can differ substantially from community life in stress, sleep, nutrition, social routines, and concurrent clinical needs. Neither matching on selected demographic variables nor adjustment for recorded covariates guarantees removal of all confounding.

Importantly, the experiment did not measure a before-use baseline in the dependent group. If a response pattern was present before long-term substance exposure, that would produce a different causal explanation from a drug-induced change. The cross-sectional comparison cannot decide between those possibilities. It also cannot establish whether the observed differences persist for years, resolve with abstinence, or vary according to treatment. Such conclusions require longitudinal and preferably replication studies with defined exposure and recovery measures.

Does altered fMRI error processing prove poor judgment or dangerousness?

No. The findings concern a particular model of response inhibition and adaptive monitoring during a computer task in a scanner. Real-world judgments involve other visual, motivational, interpersonal, physical, and environmental demands. There was no experimental measure of dangerousness, general honesty, truthfulness, eyewitness accuracy, negligence, or collision liability. The task did not reproduce sudden braking while simultaneously steering, and no operator-level mapping from caudate or thalamic activation to driving fitness was validated.

It would be equally mistaken to infer the opposite: the absence of a significant average stop-signal reaction-time group difference does not certify unimpaired real-world performance. Proper interpretation retains both the statistical null and the reported interaction pattern. A scientific expert should identify which source endpoint addresses the proposed claim, the specimen or task involved, whether the statistical comparison actually tested that claim, and which person-specific evidence is absent.

What additional evidence would a case-specific assessment require?

For a toxicological question, one would need the actual case specimen, collection time, method-validation records, analytical identity and concentration, quality controls, and interpretation appropriate to the matrix and compound. For a neurological or cognitive question, one would need a validated assessment of the person under relevant conditions and a sound basis for relating that measure to the alleged function. For traffic reconstruction, scene and vehicle evidence, observations, timing, road conditions, and alternative explanations require independent evaluation.

These components cannot be manufactured by borrowing a group-average fMRI coefficient. In litigation, a scientifically narrow conclusion can still be valuable: the paper identifies ways in which task averages, adaptive responses, and modeled neural prediction errors can diverge in selected groups. It also highlights how tempting but invalid it would be to translate an imaging group comparison directly into a legal assessment of a person whose drug concentration and behavior were not studied.

Frequently asked questions

1. Did the methamphetamine-dependent group have a slower average stop-signal reaction time?

No statistically significant difference was reported. The derived mean was 231 ms in the dependent group versus 239 ms in the control group, with p = .61. That does not prove perfect equivalence; it prevents claiming that this study established a slower group average.

2. Were failed stops significantly more common overall?

The group main comparison of failed stops was not significant: 0.49 versus 0.44, with p = .13. A separate significant group-by-delay interaction indicated that difficulty of the stopping challenge changed the pattern. Both results should be retained.

3. What is an unsigned Bayesian prediction error?

It is the absolute difference between the model’s expected probability and whether a stop event actually occurred. It is a computational surprise measure, not a blood chemical, measured synaptic dopamine amount, or clinically validated individual impairment score.

4. Does fMRI detect methamphetamine or amphetamine concentration?

No. In this study fMRI measured blood-oxygen-level-dependent hemodynamic signals associated with a computer task. No quantitative forensic venous whole-blood methamphetamine concentration or separate plasma-drug concentration was reported.

5. Why discuss caudate, orbitofrontal, thalamic, and parietal areas?

Some group contrasts linking task history and model-derived surprise to BOLD signals involved those regions. Coordinates and effect sizes describe group statistical patterns. They do not establish a person’s diagnosis, precise mechanism, drug exposure or road-traffic behavior.

6. Were the comparison groups randomly assigned methamphetamine exposure?

No. Participants were recruited from existing treatment and community populations. Past methamphetamine history, education, and other substance-use differences could influence the comparison. The design cannot independently establish causal drug-induced neural injury.

7. Can these data establish that a named driver or witness was impaired at a specific event?

No. The investigation did not evaluate actual road driving, the event, or a linked forensic drug specimen and concentration for such a person. Any case-specific inference would require independent measurements and evidence appropriate to that case.

Bottom line: meaningful error-learning differences do not equal a forensic impairment test

This study is most informative when its significant interactions and genuine null group averages are presented together. In 96 selected adults, stop probability was related to response speed, stop errors varied with task difficulty, and modeled error-related fMRI patterns differed in specified regions. Mean go and stop-signal reaction times, however, were not significantly different between groups, and several probability-related terms also remained nonsignificant. None of these endpoints is an analytically measured methamphetamine concentration or a validated individual impaired-driver threshold.

The original study, with its full methods, statistical analyses and interpretation, is available through Harlé and colleagues’ primary research report (DOI 10.1016/j.bpsc.2016.06.008). The scientifically warranted distinction is between a carefully bounded population-level behavioral-imaging association and an untested conclusion about a particular person’s exposure, actions, or legal event.