EVIDENCE LIBRARY
Does Daily Caffeine Worsen Working Memory Without Changing Task-Load fMRI? A 20-Man Crossover Trial
A 20-man randomized crossover found worse 3-back performance with daily caffeine and withdrawal, but no change in task-load fMRI. Saliva testing was not a blood impairment test.
A controlled 2023 crossover experiment in 20 healthy men found worse high-load working-memory performance during sustained daily caffeine administration and during acute withdrawal than during placebo. The caffeine condition also produced lower right-hippocampal blood-oxygen-level-dependent (BOLD) signal. Yet the classic 3-back-versus-0-back task-load BOLD pattern did not differ significantly by treatment. Crucially, the researchers directly measured salivary caffeine to check exposure—not forensic venous whole-blood caffeine, not person-specific driving performance, and not a causal mechanism of intoxication in a particular case. These distinctions control what the experiment can establish. The original primary experiment is Lin and colleagues, Scientific Reports 13:1002 (2023), original research DOI.
The answer in brief: worse task performance did not mean a larger load-dependent BOLD effect
On each condition, the same volunteers completed four versions of a computerized letter N-back task. A 0-back task indexed simpler recognition and motor response; a 3-back task added a substantial working-memory demand. The researchers used mixed models to adjust 3-back performance for the corresponding 0-back measure. This was more informative than a single unadjusted reaction-time number because caffeine and withdrawal can change elementary speed separately from memory operations. The analysis reported a condition effect on adjusted 3-back errors, F(2,197)=5.0, p=.008, and on adjusted 3-back response time, F(2,196)=6.6, p=.002. Those comparisons belong to repeated task measurements within twenty people, not hundreds of separately randomized volunteers.
The study does not support the claim that every caffeine user has impaired memory. It tested a deliberately narrow population of young, already high-consuming, nonsmoking men under a controlled 10-day exposure schedule. It also does not establish that changing caffeine will alter the accuracy of any eyewitness account, produce a diagnosable clinical disorder, or explain a particular collision. Its value is methodological: separate an observed task result from a neurovascular signal and from an assayed chemical exposure in a different biological matrix.
Who participated, and why does the true sample size matter?
The independent biological sample was 20 healthy male volunteers, with reported mean age 26.4 years (SD 4.0), body mass index 22.7 kg/m² (SD 1.4), and habitual self-reported caffeine consumption 474.1 mg/day (SD 107.5). Each person passed through three randomly ordered conditions. One caffeine-condition data set was excluded because its participant did not comply with treatment, while a scanner malfunction prevented one participant’s placebo-condition imaging data. Task trials, imaging voxels, saliva samples, and repeated visits are nested observations. Treating those as newly recruited people would be pseudoreplication and could dramatically exaggerate certainty.
The all-male enrollment was a design choice, not evidence that the effects necessarily apply identically to women. Participants were habitual high consumers, so the experiment speaks more directly to the physiology of repeated exposure and cessation in that context than to a caffeine-naïve person taking one coffee, a hospitalized patient, a child, or someone combining multiple drugs. The authors identified sample size, neurovascular interpretation, and population generalizability as concerns. The original power planning was tied to brain-activity measures, not to forensic validation of behavioral thresholds.
What were the three blinded exposure conditions?
In the continuous-caffeine condition, volunteers took 150 mg caffeine capsules three times daily, a nominal 450 mg/day, during a ten-day treatment interval. In the placebo condition, outwardly identical capsules contained mannitol. For the withdrawal sequence, caffeine was given through the first dose on day nine and then replaced with placebo for the remaining administrations, creating the specified cessation window by day ten. This schedule should not be silently simplified into one missed morning coffee or a single-dose pharmacokinetic challenge.
The three sequences were randomized and semibalanced. The researchers standardized usual sleep–wake timing, monitored activity, avoided dietary caffeine, and used laboratory conditions designed to reduce variability in light, posture, meals, and outside stimulation. On day ten, N-back assessments occurred approximately one, five, nine and thirteen hours after habitual awakening. The fMRI acquisition occurred in the final session, about five hours after the last capsule. Consequently an imaging contrast is a task-state measurement at a selected time, not an assay of instantaneous cerebral caffeine molecules.
What was actually collected and measured?
The physical chemical matrix was saliva. Repeated samples were analyzed for caffeine with high-performance liquid chromatography coupled to tandem mass spectrometry (HPLC–MS/MS), with a reported analytical detection threshold of 20 ng/mL. The manipulation check found a condition-by-sampling-time interaction, F(2,91)=17.7, p<.001, consistent with caffeine rising during administration, declining during withdrawal, and not showing a similar increase under placebo. The article directs readers to supplementary Table S1 for individual concentration details; the main article does not furnish a universal saliva concentration that corresponds to a specific N-back error or individual impairment judgment.
The cognitive instrument was a computerized visual-letter N-back task. The physiological outcome was 3-tesla functional MRI BOLD signal, with cerebral blood-flow adjustment. Neither EEG waves nor PET receptor binding nor direct dopamine or adenosine-receptor concentrations were measured here. A BOLD response depends on blood oxygenation and vascular dynamics; it is not a direct neuron count or a chemical mass concentration. As a result, a lower hippocampal signal cannot by itself be described as destroyed neurons, clinical amnesia, or a validated degree of legal intoxication.
How were 0-back attention and 3-back working memory separated?
During 0-back participants pressed a key for a target letter, demanding vigilance and a simple stimulus–response rule. During 3-back they judged whether each new letter matched one presented three positions earlier, demanding updating and monitoring of short-term information. Sessions contained repeated high- and low-load blocks; each trial included many stimuli. The authors evaluated both response errors and reaction time for correct responses. They statistically adjusted 3-back outcomes for corresponding 0-back performance rather than assuming that any slow response necessarily represented a memory failure.
That adjustment matters when interpreting stimulants. A drug may speed pressing a key without improving retention or manipulate visual attention without increasing useful task capacity. Conversely, a relatively longer high-load response time after adjustment should not be reclassified as poor simple motor speed. The report provides raw 3-back medians and means as descriptive checks but the inferential estimates described as net working-memory findings arise from models that use the lower-load outcomes as covariates. Readers should not confuse a displayed figure ratio with the precise model input: the paper expressly notes that the figures depict ratios while inferential analyses regress out lower-load variance.
What did the original 3-back performance results show?
Adjusted high-load errors differed among conditions: F(2,197)=5.0, p=.008. Relative to placebo, the caffeine condition had more errors (post hoc p<.031, Cohen’s d=.40), and the withdrawal condition also had more (p=.013, d=.45). Caffeine and withdrawal were not distinguishable on this particular error comparison (p=.960). Descriptive unadjusted 3-back error-rate medians were 0.05 under placebo, 0.07 under caffeine, and 0.07 under withdrawal, with interquartile ranges 0.03–0.09, 0.03–0.14, and 0.03–0.14, respectively. These fractions use the paper’s unusual error-rate definition, so treating 0.07 as a simple fraction of all administered stimuli would be unwarranted.
After controlling for 0-back response-time variance, high-load latency had an overall condition effect F(2,196)=6.6, p=.002. Caffeine was slower than placebo (p=.003, d=.57) and slower than withdrawal (p=.023, d=.47). The paper did not identify a clear withdrawal-versus-placebo net response-time difference. The descriptive unadjusted 3-back reaction-time means were 647.5 ms (SD 128.2) for placebo, 688.9 ms (SD 179.0) for caffeine, and 671.2 ms (SD 123.6) for withdrawal. Those means are not effect sizes and should not be presented as a diagnostic cutoff.
Evidence table 1: observed behavioral endpoints, effects and real nulls
| Original endpoint | Source observation | Interpretive boundary |
|---|---|---|
| Adjusted 3-back errors | F(2,197)=5.0; p=.008. Caffeine vs placebo p<.031; withdrawal vs placebo p=.013. | Higher task error under both; the caffeine–withdrawal comparison itself was p=.960 (not significant). |
| Adjusted 3-back correct-response latency | F(2,196)=6.6; p=.002. Caffeine vs placebo p=.003; caffeine vs withdrawal p=.023. | Behavioral response-time outcome, not quantitative blood concentration or individual impairment. |
| Descriptive raw 3-back RT | Placebo 647.5±128.2 ms; caffeine 688.9±179.0; withdrawal 671.2±123.6 (mean±SD). | Raw descriptive values, not covariate-adjusted mean differences. |
| Simple 0-back errors | Condition F(2,198)=1.9; p=.152; condition×session F(6,198)=1.4; p=.216. | Neither comparison was significant; do not describe an observed clinical equivalence margin. |
| Adjusted 3-back error condition×session | F(6,195)=0.3; p=.936. | No detected differential trend across the four repeated sessions. |
Did simpler psychomotor response speed improve?
On the 0-back task, error rates did not show a significant overall condition effect, F(2,198)=1.9, p=.152, or condition-by-session effect, F(6,198)=1.4, p=.216. The simpler response-time endpoint did vary by condition, F(2,198)=9.6, p<.001, chiefly because withdrawal was slower than placebo (p<.001) and slower than caffeine (p=.003). Caffeine did not significantly differ from placebo on 0-back reaction time (p=.815). The descriptive means were 531.7 ms under placebo, 536.1 ms during caffeine, and 560.3 ms during withdrawal.
This is an example of why different tasks cannot be collapsed into a single label of “caffeine improves reaction time.” Daily caffeine was not associated with a significant advantage over placebo for this simpler response-time comparison, while the cessation condition was slower. The researchers also observed repeated-session effects unrelated to treatment that reflected practice, timing or other day-level influences. Those changes are relevant for a laboratory model, but a criminal or civil event requires its own timeline, dose evidence, specimen collection facts, co-exposures, clinical state and objective behavior.
What did fMRI show about the working-memory load contrast?
Across all three regimens, heavier 3-back versus lighter 0-back load recruited cortical attention and motor-related regions, while activity in several default-mode and limbic regions was lower under high load. Many task-load clusters passed family-wise-error-corrected thresholds. However, the authors reported that these load-dependent BOLD patterns did not significantly differ across placebo, daily caffeine and withdrawal. A significant task-load contrast measured across everybody is not itself a significant interaction between the task load and drug condition. This is a common but consequential error in neuroimaging summaries.
A distinct analysis irrespective of workload identified lower right-hippocampal BOLD activity under caffeine versus placebo, cluster-corrected p=.035, Cohen’s d approximately −1.3. The left middle frontal region difference compared with placebo reached p=.057, which is a trend, not p<.05 evidence. Caffeine also differed from withdrawal in the right middle frontal gyrus, p<.001. By contrast, withdrawal versus placebo did not show a significant regional activation difference in the report. The selected clusters are not equivalent to a global brain lesion or an individual capacity test.
Evidence table 2: imaging, exposure verification and uncertainty
| Domain | Reported original result | What was and was not measured |
|---|---|---|
| Salivary caffeine HPLC–MS/MS | Condition×sample time F(2,91)=17.7; p<.001; analytical detection threshold 20 ng/mL. | Actual saliva caffeine trajectory supports dosing manipulation; not human venous whole-blood caffeine or clinical impairment threshold. |
| Task-load BOLD contrast | 3-back vs 0-back shows robust attention/default-mode patterns; no significant between-treatment alteration of load-dependent activity. | Functional MRI hemodynamic surrogate; not measured synaptic caffeine or impairment diagnosis. |
| Right hippocampal activity | Lower under caffeine vs placebo independent of workload; cluster pFWE=.035; d≈−1.3. | Regional BOLD statistical group contrast; not histologic tissue damage or individual disease. |
| Left middle frontal activity | Caffeine vs placebo pFWE=.057; not significant at .05. | Near-threshold result, not proof of a positive regional effect. |
| Exploratory hippocampus–frontal connectivity | No significant between-condition functional correlation difference. | Null exploratory imaging analysis; no measured neuronal wiring change. |
Why the lower hippocampal BOLD signal does not prove brain injury
BOLD contrast arises from changes in deoxygenated hemoglobin and vascular physiology related to task activity. Caffeine can influence resting cerebral blood flow and the relationship between vascular and metabolic demand. The authors adjusted imaging models for individual whole-brain perfusion, but acknowledged that simultaneous cerebral blood-flow and BOLD acquisition would better address neurovascular coupling. Consequently the right-hippocampal statistical difference cannot be translated directly into fewer neurons, damaged memory cells, a reduced anatomical hippocampus on that day, or a specific chemical dose reaching brain tissue.
The team discusses adenosine receptors, glutamate-related plasticity and earlier structural MRI analyses from the same people as possible explanatory context. Those mechanisms were not directly measured by receptor-binding assays or electrophysiological recordings in this experiment. Likewise, findings described elsewhere using rodents with disease or aging models are not automatically evidence of equivalent treatment effects in healthy humans. Interpreting an observed fMRI cluster as a causal molecular pathway would go beyond the tested methods.
Was the experimental withdrawal condition chemically verified?
Yes, in the limited sense that serial saliva analysis provided evidence that caffeine concentrations followed the intended dosing versus cessation trajectories. Treatment-by-sampling-time differences supported the experimental manipulation, while the main article referred detailed concentration numbers to supplementary material. A detectable saliva value must be interpreted in the correct sample and method. The threshold of 20 ng/mL describes the study’s assay detection level; it is not a clinical tolerance threshold, a DUI threshold, a time-since-use calculator or a statement about how reliably a different laboratory would identify caffeine in blood.
Withdrawal was a scheduled switch from regular capsules to placebo within the trial. It was not a test of a worker spontaneously skipping a beverage with unknown caffeine quantity, nor did it use an objective measurement of a particular person’s neurobehavioral safety during vehicle operation. The randomized within-person design does strengthen comparisons among the controlled regimens. It does not remove the need to document exact treatment adherence, sex, habitual consumption, concurrent medicines and sleep conditions before extrapolating to a new person.
Why 20 people—not repeated sessions or brain voxels—form the biological sample
The mixed-model denominator can contain many observations because the same participant provided performance at four sessions and many task blocks. Spatial fMRI analyses additionally process voxels and clusters that have strong internal correlations. Neither F-test degrees of freedom nor the number of voxels should be substituted for twenty independent participants. One participant’s caffeine-condition data were excluded for noncompliance and another placebo scan was lost to malfunction; these outcome-specific gaps constrain further assumptions about a single constant paired N in each comparison.
The repeated-measures crossover gives more efficient within-person information than three unrelated samples, but can still be affected by period, carryover, unblinding due to noticeable withdrawal symptoms, and limited diversity. Readers should seek the protocol, analytic scripts and underlying data when adjudicating an exact reanalysis rather than reconstructing a personalized effect size from three published condition means.
How does the experiment affect forensic toxicology interpretation?
It principally illustrates three distinct evidence types. First, a chemically specific saliva caffeine assay establishes something about exposure in that fluid. Second, a behavioral task measures response accuracy or latency under particular instructions, rewards and timing. Third, fMRI generates a statistical blood-oxygenation signal that responds to vascular and neural activity. Concordance between all three can improve mechanistic hypotheses, but one measurement never automatically validates another as a legal test of impairment. The study neither sampled the separate compartment of human venous whole blood nor established a validated functional cutoff or an individualized time-since-dose formula.
In forensic work, matrix identity, collection time, assay calibration and direct case behavior must be kept separate. Even a real caffeine blood concentration in an external case would require interpretation specific to that method, co-intoxicants, habitual use, sleep state, health and observed task demands. This trial does not predict a defendant’s driving lane position, a clinician’s treatment decision, an eyewitness’s memory accuracy or the causal role of caffeine in an event. Nor should a negative imaging group contrast be used as proof that caffeine has no brain effect whatsoever: the reported task-load null coexisted with a different regional hippocampal result.
What are the central limitations of this original trial?
The investigation enrolled twenty healthy young men who were habitual, relatively high caffeine consumers, and it had no female study arm or caffeine-naïve comparator. The controlled dose was 450 mg/day, not an unmeasured cup size. There were a small number of condition-specific missing data sets. Supplementary source material contains details beyond the numerical main-text data; absent appendix figures should never be filled by imagined participant concentrations. Cognitive testing used an artificial N-back procedure rather than safety-critical real-world driving or workplace tasks. The imaging endpoint depends on hemodynamics, cerebral blood flow and selected statistical cluster criteria.
There are also inferential boundaries in the statistical reporting. A nonsignificant result is not a demonstration of exact equivalence, especially with small samples. Behavioral treatment effects and imaging regional clusters arise from different models, with different comparisons and correction thresholds. The study makes a useful contribution by showing that high-load behavior, lower-load processing, salivary exposure and task-load imaging can diverge. That is a rationale for careful evidence integration, not authority for an untested impairment classification.
Frequently asked scientific questions
Does 450 mg/day of caffeine necessarily damage working memory?
No. The 450 mg/day schedule was one controlled ten-day exposure in twenty already habitual high-consuming young men. The study observed worse laboratory 3-back performance relative to placebo, but did not test permanent damage, diagnosis, or a dose–response curve across the general population.
Did caffeine and withdrawal have identical cognitive outcomes?
No. Their adjusted 3-back error rates did not differ significantly (p=.960), but the adjusted response-time analysis identified a difference between caffeine and withdrawal (p=.023). Different endpoints must be read separately, and nonsignificance does not prove strict equality.
Was caffeine actually measured in volunteers?
Yes, caffeine was analyzed in saliva by HPLC–MS/MS as an exposure manipulation check, with a stated 20 ng/mL detection threshold. The researchers did not report chemically assayed forensic venous whole-blood caffeine in this experiment.
Does lower hippocampal BOLD activity demonstrate cognitive injury?
No. It is a group-level blood-oxygenation statistical finding subject to neurovascular influences. The study did not count neurons or confirm tissue injury, and it did not establish a patient-specific clinical diagnosis.
Did caffeine significantly change the 3-back versus 0-back fMRI load effect?
The researchers did not identify a significant condition difference in the load-dependent brain activation pattern. A separate, workload-independent right-hippocampal caffeine-versus-placebo contrast did reach the specified cluster-corrected threshold.
Can saliva caffeine be converted into a legal whole-blood level or use time?
Not from this paper. Saliva and venous whole blood are physically different matrices; the study was not a paired, validated saliva–blood conversion or forensic back-extrapolation experiment.
Can these group results establish that a particular driver was impaired?
No. There was no driving outcome, actual incident, impairment cutoff, subject-specific blood test or evidentiary case record. Population task findings should not be substituted for the measured behavior and circumstances of an individual.
Original source and related evidence guides
Primary empirical study: Yu-Shiuan Lin, Janine Weibel, Hans-Peter Landolt, Francesco Santini, Helen Slawik, Stefan Borgwardt, Christian Cajochen and Carolin Franziska Reichert (2023), Brain activity during a working memory task after daily caffeine intake and caffeine withdrawal: a randomized double-blind placebo-controlled trial, Scientific Reports 13:1002, complete original DOI and publisher article. The discussion here independently interprets reported measurements without copying source tables or figures.
For the distinctions behind individual cases, consult the existing published same-site guides on biological matrices, confirmatory drug testing and therapeutic drug monitoring. These pages provide general methods context and do not establish an individualized legal threshold.