Does Caffeine Speed Reactions by Improving Attention Rather Than Motor Speed? An EEG Study of 15 Women

Quick answer: In a randomized, blinded crossover study of 15 healthy women, a weight-adjusted oral caffeine dose was associated with faster responses on a visual oddball task. Single-trial EEG analyses suggested that changes in a stimulus-locked P3b signal explained approximately one-third of the modeled reaction-time difference. Caffeine did not show the corresponding change in response-locked P3b latency. The result favors an attentional or stimulus-evaluation explanation for faster responses in this particular task; it does not prove that caffeine cannot alter motor performance in other tasks, and it is not a clinical or forensic test of an individual’s impairment.

The primary source is Saville, de Morree, Dundon, Marcora, and Klein’s controlled experiment, Effects of caffeine on reaction time are mediated by attentional rather than motor processes, published in Psychopharmacology (2018). Its most useful contribution is methodological: it separates stages of performance that an ordinary stopwatch collapses into a single number.

What question did the researchers actually test?

A fast response to a light depends on more than how quickly a hand can move. A person must register the light, discriminate its meaning, select an appropriate response, prepare that response, and execute it. A shorter total reaction time could arise from a change in any of these stages. Merely observing that someone reacted faster after caffeine cannot identify the stage that changed.

The investigators therefore recorded brain electrical activity during a three-stimulus task and studied the timing of the P3b, a component of the event-related potential commonly associated with stimulus processing and decision-related activity. Their two candidate measurements were a P3b latency aligned to stimulus presentation and another aligned to the participant’s response. They asked whether a statistical mediation model linked caffeine’s effect on total reaction time more strongly to one timing measurement than to the other.

That is a narrower question than whether caffeine makes a person a safer driver, improves every type of attention, treats fatigue, or prevents mistakes. Each of those would require outcomes and validation not supplied by this experiment.

Who participated, and how many independent people were analyzed?

Eighteen women were initially recruited. Two records were excluded because fewer than 20 clean trials remained, and a third was excluded because of a technical collection problem. The final analytic sample was 15 independent participants, not 18 completed usable people and not thousands of separate subjects. The reported initial average age was 21.6 years, with a standard deviation of 4.1 years; the paper does not claim this age summary is a separate clinical cohort.

Participants reported no neurological or psychiatric diagnoses and were nonsmokers. Their average habitual caffeine intake was 124 ± 109 mg per day. That substantial variability is relevant: the experiment did not recruit a uniform group of people with identical caffeine tolerance. All analyzed participants were female, so extrapolating the magnitude of the effect to men, older adults, adolescents, sleep-deprived drivers, or patients with medical conditions would require additional evidence.

Each woman completed two sessions. The multiple trials, electrodes, extracted signal peaks, and model observations generated within those sessions are repeated observations nested in the same people; they must not be treated as independent participants.

How much caffeine was given, and what was the comparison?

The experiment used a randomized, counterbalanced, double-blind crossover design. During one visit, the participant swallowed capsules containing 6 mg of caffeine per kilogram of body weight, together with 6 mg/kg dried milk. During the comparison visit, she swallowed capsules containing 12 mg/kg dried milk, without caffeine. The sessions were separated by exactly one week.

To reduce participants’ ability to identify placebo, investigators told them that one session involved caffeine and the other taurine; in fact, the comparator was the dried-milk placebo. The experimenters were also blinded to the assigned condition. This design makes a within-person comparison more useful than comparing two unrelated groups, but it does not remove every possible expectancy or carryover concern.

The cognitive task began an average of two hours and six minutes after capsules were given, with a reported range from one hour 50 minutes to three hours two minutes. Participants were advised to retain their usual caffeine habits, obtain a good night’s sleep, avoid alcohol and strenuous exercise before sessions, and eat a light meal roughly two hours before testing. The dose was an administered mass per kilogram, not a measured blood concentration.

What did the visual oddball test ask people to do?

The task displayed colored circles, each appearing for half a second. White circles constituted 70% of presentations and required a left-hand keyboard response. Green targets made up 15% and required a right-hand response. Red distractors made up the remaining 15% and required no response. Participants completed three blocks of 340 trials per session, with variable intervals between presentations.

This arrangement helps create a measurable P3b response to relatively infrequent target events while giving the investigators a precisely timed record of each correct target response. Although the task involves hands, vision, stimulus identification, and response selection, it is not a steering, hazard-avoidance, divided-attention driving, or courtroom eyewitness task.

The analysts retained correctly responded target trials within a specified 120–1,000 millisecond reaction-time window, subject to further signal-cleaning and peak-identification rules. The repeated events improved the precision of within-person estimates, but they do not change the sample size of 15 people.

What was physically measured: a chemical specimen or an EEG signal?

Investigators used 59 scalp electrodes and two infra-orbital electrodes to record electrical potentials, then applied artifact removal, component analysis, and filtering. The P3b latency reflects a timing feature of an electrophysiological waveform. It is not a concentration of caffeine in brain tissue. A voltage-derived signal is not interchangeable with the concentration of a drug in a tube of blood.

No caffeine or metabolite concentration was quantified in venous whole blood, separated plasma, serum, urine, oral fluid, hair, vitreous humor, breath, sweat, or tissue for the reported mediation analysis. The paper does not provide a caffeine LC–MS/MS or GC–MS assay, a calibration curve for bodily specimens, a drug-specific limit of detection, or a quantitation limit in ng/mL. These details are not omissions to be filled by assumptions from another study; they were not measured endpoints here.

It would consequently be incorrect to translate a reported 6 mg/kg oral dose into a particular individual’s forensic blood concentration, or to use an EEG-derived P3b latency as proof of a legally relevant toxicological level.

What changed in reaction time and P3b latency?

The printed descriptive statistics show a mean reaction time of 436.4 milliseconds during caffeine and 448.2 milliseconds during placebo. The direct difference between these two reported condition averages is 11.8 milliseconds. The mean stimulus-locked P3b latency was 439.6 versus 448.6 milliseconds, respectively, a nine-millisecond descriptive difference. The corresponding response-locked averages were approximately −2.8 versus +2.5 milliseconds, but the relevant model found no reliable caffeine condition effect on that response-aligned measure.

These values summarize a within-person study; the accompanying ± values are the source’s descriptive dispersion measures, not 95% confidence intervals for the treatment effect. They also should not be confused with the separately estimated total effect from the trial-level mediation model.

Why distinguish stimulus-locked from response-locked P3b?

A stimulus-locked waveform is aligned to the instant a circle appears. Its latency can illuminate the timing of processes related to recognizing and evaluating that circle. A response-locked waveform is aligned instead to the button press; it can reflect later decision-to-action or response-related processing. Those interpretations are useful working models, not maps assigning one unique biological operation to an isolated electrode peak.

Here the treatment condition significantly predicted stimulus-locked P3b latency. The source’s nested-trial model comparison reported a chi-square statistic of 10.27 and p = .001. Stimulus-locked latency also predicted reaction time in a separate model comparison (p < .001). Both relationships were prerequisites for the investigators’ mediation calculation.

The response-locked condition comparison was p = .273, a nonsignificant result. Importantly, response-locked latency itself remained associated with reaction time; it simply was not significantly shifted by caffeine in this experiment. The authors therefore did not run the proposed response-locked mediation analysis. That is more precise than saying the motor system was irrelevant or experimentally proven unaffected in every possible sense.

How much of the faster response was statistically mediated?

For the directly extracted stimulus-locked signal, the model estimated an average causal mediation effect (ACME) of −3.48 milliseconds, with a 95% interval of −5.30 to −1.85 milliseconds and reported p < .01. Its average direct effect was −6.75 milliseconds, and the estimated total treatment effect was −10.23 milliseconds. The estimated proportion statistically mediated was 0.34, with a 95% interval of 0.19–0.57.

In practical language, approximately one-third of the model’s average reaction-time improvement was associated with the measured change in this particular EEG processing marker. This does not mean precisely 34% of all caffeine’s neurological action occurs in attention pathways or that exactly one-third of one person’s individual response is explained. The value depends on the model, measurement rules, study population, and assumptions needed to interpret mediation.

What makes the mediation result stronger than a simple correlation?

The treatment was assigned experimentally in a crossover protocol, the participants served as their own controls, and the timing features were examined at the single-trial level while accounting for observations nested within individuals. The investigators did more than correlate one participant’s overall speed with another participant’s average brain signal. They checked that caffeine was associated with the proposed mediator and that the proposed mediator was associated with reaction time before estimating the mediated share.

Nevertheless, randomization of caffeine is not the same as randomization of the P3b mediator. A causal interpretation of a mediation coefficient still depends on assumptions about measurement, temporal interpretation, and possible mediator–outcome confounding. The measured signal also reflects distributed neural activity rather than a single experimentally isolated attentional circuit.

The strongest defensible conclusion is therefore a study-specific mechanistic interpretation: the experimental caffeine effect on this oddball task was statistically consistent with faster perceptual-attentional processing, whereas the response-aligned latency lacked the necessary caffeine effect. The paper does not prove an exclusive mechanism for caffeine’s behavioral effects.

Did the EEG amplitude change significantly?

No significant condition effect on the P3b amplitude was established. The source reported a chi-square test of 1.40 and p = .236 for the amplitude model comparison. This negative result matters because amplitude differences could change the signal-to-noise properties of the waveform and thereby affect the accuracy of single-trial peak detection. A large condition-specific amplitude change could complicate the interpretation of an apparent latency difference.

Absence of statistical evidence for an amplitude effect is not proof of exact physical equality in all EEG components. The result pertains to the P3b signal extraction used here, with the available power and preprocessing choices. It should not be generalized to every scalp electrode, brain region, or other electrophysiological component. Nor does a non-significant amplitude test establish an analytic sensitivity or clinical exclusion threshold for caffeine.

Did alternative analysis methods agree with the headline result?

The researchers checked whether choosing stimulus-locked versus response-locked peak windows caused the contrast. When they constructed inferred rather than separately measured latencies, the stimulus-locked mediation proportion was 0.47, with a 95% interval of 0.28–0.74 and reported p < .01. In contrast, the inferred response-locked signal still failed the condition-effect prerequisite. That consistency strengthens the claim that the difference was not simply an artifact of using separate peak-picking windows.

A further reanalysis using a conventional Pz electrode rather than a principal-component-derived virtual channel estimated a mediated proportion of 0.33. A version without within-person latency centering estimated 0.36. These are reassuring sensitivity analyses, but they do not substitute for a large independent replication across different samples and tasks.

The conventional grand-average ERP approach was notably less decisive. Its reported mediation proportion was 0.30, but p = .13, so the mediation was not statistically significant in that analysis. The headline should preserve both findings instead of implying every analysis independently proved the effect.

Does this establish that caffeine leaves motor function unchanged?

No. The task involved selecting and pressing keyboard keys in response to colored circles. The negative response-locked result tells us that this particular EEG timing measure was not significantly shifted under the tested conditions. It does not test a broad range of motor demands, including grip force, sustained muscular work, coordination under fatigue, balance, steering maneuvers, or rapid braking during real traffic.

The authors themselves noted that the same recording sessions also included a separate leg-contraction experiment. Different motor tasks recruit different processes, and a signal interpreted as response-linked in one laboratory setting cannot be taken as a universal motor performance biomarker.

A careful account should say that the data favor a perceptual-attentional mediation model for faster oddball responses. It should not claim that caffeine is incapable of acting on the motor system or that an individual has normal motor coordination because a response-locked P3b condition test was nonsignificant.

Can faster reaction time be called better attention in all situations?

Reaction time is one dimension of performance. Accuracy, error rates, false alarms, missed responses, maintenance of attention over long periods, and response strategy can matter just as much. The trial’s central comparison uses correctly completed target trials with enough artifact-free EEG data to estimate latencies. It does not deliver a comprehensive real-world vigilance profile for every event a person encountered.

The source also did not study intentional distraction, intoxication with another compound, chronically sleep-deprived professional drivers, multi-vehicle decisions, or abrupt conflicts between simultaneous tasks. A mean modeled improvement near ten milliseconds in one experimental paradigm should not be treated as a fixed improvement in traffic safety, fault avoidance, or eyewitness reaction.

The choice of task and event selection are particularly important if a forensic report attempts to compare this study with performance on a very different test. Similar terminology—attention, vigilance, reaction time—does not establish similar construct validity or applicability.

What do the sample and preprocessing limitations mean?

Fifteen analyzed participants is a small biological sample, notwithstanding the large number of individual trials. Three of the 18 recruits could not be included, two due to too few clean observations and one because data collection failed. Those exclusions matter for reproducibility because EEG mediation depends on obtaining sufficient artifact-free recordings, not merely on enrolling volunteers.

The sample was all women, participants were young and generally healthy, habitual caffeine use varied, and only a single visual oddball paradigm was used. The investigators examined a hypothesized P3b component using a particular component-extraction and peak-picking workflow. A different laboratory or task could generate a different mixture of perceptual and response-related neural events.

One should distinguish this finite sample’s significant finding from a population-wide quantitative rule. The interval around the proportion mediated—0.19 to 0.57—is wide enough to show substantial uncertainty about the size of the modeled contribution.

Why doesn’t this paper establish a blood-caffeine impairment threshold?

The investigators administered a weight-adjusted oral caffeine dose and recorded EEG and behavior. They did not collect a corresponding quantitative forensic caffeine concentration series. Without validated paired chemical measurements, the article supplies no conversion from 6 mg/kg to venous whole-blood, plasma, or serum ng/mL for a particular person. Even if a different paper supplies pharmacokinetic measurements, that would be a different study and population, not a missing value that can be grafted onto this one.

The distinction is central in forensic toxicology: biological matrices must be interpreted according to the specimen actually measured. Direct confirmation of a drug and interpretation of a behavioral or EEG endpoint are analytically different questions. See also the site’s guide to confirmatory drug testing and its discussion of timed drug-concentration interpretation.

There is no validated individual classification cutpoint in this study for intoxication, unsafe driving, time since caffeine ingestion, compliance with a dosing regimen, or collision causation. A measured EEG latency is not a substitute for an independently valid toxicological specimen result.

How should this study be used in a forensic report?

The paper is useful when explaining why observed response speed cannot be attributed automatically to stronger muscles or faster motor execution. It gives an empirical example in which a controlled exposure, behavioral endpoint, and candidate neurophysiological mediator were measured within the same people. It also illustrates that statistical mediation can depend on whether individual trials or grand-averaged signals are used.

Appropriate expert use would identify the actual exposure protocol, limited sample, task demands, model-estimated timing differences, null condition effects, and absence of quantified postdose toxicology. These facts can support a general scientific discussion of what reaction time means. They cannot establish whether a particular person used caffeine at a specific time, whether another drug caused that person’s error, or whether that person could have avoided a specific roadway event.

Any case-specific inference would need the applicable specimen collection and analytical records, the relevant behaviors, event timing, corroborating circumstances, and task-specific validity evidence. Absent those inputs, the scientifically supportable answer is bounded by what the primary experiment actually tested.

Frequently asked questions

1. Did caffeine actually shorten measured reaction times?

Yes. The reported mean on analyzed correct target trials was 436.4 milliseconds during caffeine and 448.2 milliseconds during placebo, while the trial-level total treatment effect was modeled at −10.23 milliseconds. Those figures summarize different calculations within the same small crossover study; neither quantifies a universal change for every person.

2. Was the study conducted in 15 or 18 people?

Eighteen women were recruited, but only 15 had usable datasets for the analysis. The two conditions and repeated oddball trials do not double or multiply the independent participant count.

3. What does a 34% mediated effect mean?

It is the proportion of the modeled average reaction-time effect attributed statistically to changes in stimulus-locked P3b latency under the authors’ mediation assumptions. It is not a percentage of measured caffeine molecules, a brain concentration, or an individualized diagnostic threshold.

4. Did caffeine significantly change response-locked P3b latency?

No statistically significant treatment effect was established for that latency (p = .273). The response-locked measure was associated with reaction time, but it was not significantly altered by caffeine under the model used, preventing the intended mediation analysis.

5. Was caffeine measured in a blood or urine sample?

No quantitative caffeine or metabolite value in venous whole blood, plasma, serum, urine, oral fluid, or another forensic drug specimen is reported for the mediation analysis. The assigned oral dose and EEG timings are not substitute specimens.

6. Did every statistical approach confirm the mediation finding?

No. Several single-trial sensitivity analyses supported stimulus-locked mediation, but conventional averaged-ERP mediation was nonsignificant (p = .13). Both findings should be reported when comparing methodological approaches.

7. Could this study prove that a driver was unimpaired after caffeine?

No. The experiment measured keyboard responses and scalp EEG in healthy volunteers, not driving, collisions, forensic blood concentrations, or individual impairment classification. A case-specific opinion requires separate evidence appropriate to the particular legal and scientific question.

Primary study and bottom line

Bottom line: This controlled study supports a task-specific association between faster caffeine-condition responses and earlier stimulus-related P3b timing, with roughly one-third of the model’s average reaction-time effect statistically mediated through that signal. Response-locked latency and P3b amplitude did not show corresponding significant treatment effects. It is meaningful evidence about experimental mental chronometry—not a validated measure of a person’s toxicological status or forensic driving ability.

Primary empirical research: Saville CWN, de Morree HM, Dundon NM, Marcora SM, Klein C. Effects of caffeine on reaction time are mediated by attentional rather than motor processes. Psychopharmacology. 2018;235:749–759. Full original study and DOI.