One Espresso, One Night, One Number: How to Read Health Trends Without a Lab

On a Tuesday in February I drank a double espresso at 4:12 p.m. — call it 150 milligrams of caffeine — specifically because I wanted to catch my own sleep in the act of falling apart. I had a tracker on my wrist and a spreadsheet open on the kitchen counter. The next morning the app reported 6 hours and 51 minutes of sleep, 58 of them "deep." The night before, no espresso, it had given me 7 hours and 22 minutes, 71 minutes deep.

There it was. Thirty-one minutes, gone. A mechanism I could name, a number I could screenshot.

This is the exact raw material from which health trends are manufactured, and it is also the reason most of them deserve more scrutiny than they get. Because I had eleven weeks of data from that same wrist, and I knew what it did on nights when nothing happened at all. It wandered by forty minutes in either direction for no reason I could identify. My "finding" was smaller than my own noise.

The impulse behind optimization culture isn't stupid. Sleep really does affect nearly everything measurable about you, caffeine really does interfere with it, and the interference really is dose- and time-dependent. What goes wrong is almost never the mechanism. It's the size of the effect, and the confidence with which a single night gets promoted to a protocol.

What one night cannot tell you

Start with the measurement itself. Chinoy et al. (2021), in the journal Sleep, ran seven consumer sleep-tracking devices against in-lab polysomnography in healthy adults. The general pattern: the devices were decent at estimating total sleep time and considerably worse at staging — assigning minutes to "deep" or REM. Which means the number that felt most damning in my little experiment, the deep-sleep drop, was the number with the loosest connection to what my brain was actually doing.¹

Then there's the design. I knew which night was the espresso night. I had a hypothesis and a stake in it. And I ran the caffeine night second, after a good night — so simple regression to the mean predicts a worse number regardless of what I drank. If you handed that experiment to a reviewer, they would not ask you to revise it. They would ask you why you'd bothered.

This is the structural problem with almost every protocol that arrives via a fifteen-second video. It is somebody's n=1 with the noise cropped out, narrated backward from the outcome. The mouth tape went on, the morning felt good, the mouth tape gets the credit. Nobody films the eleven weeks of baseline that would tell you whether "good" was inside the normal range for a Thursday.

One useful habit: before you evaluate a claim about sleep, estimate your own night-to-night spread. If your total sleep time routinely varies by 40 minutes for no reason, then any intervention promising 20 minutes is undetectable to you personally, even if it's completely real. You would need weeks of averaging to see it. That single sentence disqualifies a startling fraction of wellness advice — not as false, but as unmeasurable at the individual level, which is not the same thing and is rarely distinguished.

What caffeine actually does, in the order it does it

The mechanism is worth walking through slowly, because it explains both why caffeine's effect on sleep is real and why the popular framing of it is slightly off.

From cup to receptor to liver

You swallow. Caffeine is absorbed mostly in the small intestine, and absorption is close to complete — roughly 99% within about 45 minutes. Plasma concentrations peak somewhere between 30 and 120 minutes later, depending on what else is in your stomach.

Because the molecule is small and lipid-soluble, it crosses the blood-brain barrier essentially unopposed. No transporter required, no gatekeeping.

What it finds there is adenosine, which has been accumulating since the moment you woke up. Adenosine is a byproduct of cellular energy use, and its buildup is the leading candidate for the physical substrate of what Borbély, in his 1982 two-process model, called Process S — homeostatic sleep pressure, the thing that makes hour sixteen feel different from hour four.

Caffeine is a structural mimic. It occupies adenosine receptors — principally A1 and A2A — without activating them. At A2A receptors in the striatum, this blockade also releases a brake on dopamine signaling, which is a decent part of why coffee feels good rather than merely feeling like not being tired.

Here is the part that matters and gets skipped: caffeine does not remove sleep pressure. It conceals it. The adenosine is still binding, still accumulating, still there. You have muted the alarm, not addressed the fire. Which is why the crash on the far side is not a mysterious side effect — it's the accumulated signal arriving all at once when the blockade lifts.

The lifting is done mainly by the liver, where the enzyme CYP1A2 handles roughly 95% of caffeine clearance. Median half-life in healthy adults is about five hours, with a genuinely wide range — something like 1.5 to 9.5 hours across individuals. Run the arithmetic on my espresso: 150 mg at 4:12 p.m. leaves about 75 mg at 9 p.m., about 37 mg at 2 a.m. Thirty-seven milligrams is not nothing. It is roughly a third of a cup of coffee, circulating during the hours when the deepest slow-wave sleep is supposed to be happening.

How late is too late to drink coffee?

For most adults, the defensible answer is about eight hours before bed for a standard cup of coffee, and considerably earlier for a large one. If you go to bed at 11 p.m., that puts the last ordinary coffee around 3 p.m. That's a starting point, not a law, and the spread around it is wide enough to matter.

That figure comes from Gardiner et al. (2023), a systematic review and meta-analysis in Sleep Medicine Reviews that pooled controlled trials of caffeine and subsequent sleep. Their modeled thresholds: roughly 107 mg — one normal cup — about 8.8 hours before bed, and roughly 217 mg about 13.2 hours before bed, if you want to avoid measurable disruption. Note what 13.2 hours implies. For an 11 p.m. bedtime, a large coffee would have to be finished by about 10 a.m.

The most-cited single trial is Drake et al. (2013), Journal of Clinical Sleep Medicine, which gave 12 participants 400 mg of caffeine at 0, 3, and 6 hours before bedtime. Sleep was disrupted at all three timepoints, including the 6-hour condition, by something on the order of an hour of lost sleep. Twelve people is twelve people, and 400 mg is a large dose — roughly four cups at once. The study is real and useful; it is also thinner than the certainty with which its headline gets repeated.

And then there is you. Rétey et al. (2007), in Clinical Pharmacology & Therapeutics, found that a common variant in the ADORA2A gene — the A2A receptor itself — tracked with how much caffeine disrupted EEG-measured sleep. Sensitive genotypes showed clear disruption; others barely budged. The sample was small, as genotype-stratified sleep studies tend to be, but the direction has held up.

Clearance varies just as much, and for reasons that have nothing to do with willpower. Oral contraceptives roughly double caffeine's half-life. Smoking induces CYP1A2 and cuts it roughly in half. In the third trimester of pregnancy it can stretch toward 15 hours. Two people can drink identical lattes at identical times and be in pharmacologically different situations by midnight.

So a universal cutoff hour is a category error dressed as precision. The useful version is a range plus a self-test, which is what the end of this piece is about.

Every dose-response curve flattens

The engine under most optimization advice is a hidden assumption: that the relationship between a variable and an outcome is a straight line. If seven hours of sleep beats five, ten must beat seven. If some protein builds muscle, more builds more.

Almost nothing in physiology works that way.

Take the study that quietly seeded sleepmaxxing. Mah et al. (2011), in Sleep, had 11 Stanford basketball players extend time in bed to 10 hours for five to seven weeks. Sprint times improved. Free-throw and three-point accuracy improved. It's a lovely result, and it is routinely cited as evidence that more sleep is better without limit. But those athletes started chronically underslept. The study measured a move from deficient toward adequate, which is the steep part of the curve. It says almost nothing about moving from adequate to extraordinary, which is the part everyone actually wants to buy.

Same shape elsewhere. Morton et al. (2018), a British Journal of Sports Medicine meta-analysis of 49 studies and roughly 1,863 participants, found that the benefit of added dietary protein for resistance-trained gains plateaued around 1.6 g/kg of body weight per day. Above that, the line goes flat. Not dangerous — just flat. Yet protein-maxxing content is built almost entirely on the region of the graph where nothing further happens.

And optimization is not free. Every variable you decide to maximize costs attention, money, and evenings. Baron et al. (2017), also in Journal of Clinical Sleep Medicine, described what they named orthosomnia: patients whose sleep got worse because they were pursuing perfect tracker scores, arriving at clinic with printouts and insomnia they hadn't had before. The pursuit had become the pathology. It is a small case series, not an epidemic, but it names something real — the anxiety of measurement can exceed the benefit of what's measured.

Three tiers, and how to sort a claim into one

Here is the sorting I use, and it costs nothing to apply.

Well-established means replicated across independent groups, with a plausible mechanism, a visible dose-response relationship, and an effect size bigger than ordinary noise. Caffeine delaying sleep onset and suppressing slow-wave sleep is well-established.

Plausible but thin means the mechanism checks out and the outcome hasn't really been tested. This is where most current optimization advice lives, and it's the most seductive tier, because a good mechanism story feels like evidence.

Folk wisdom means widely repeated, possibly true, never controlled.

The archetype of tier two is the advice to delay your first coffee 90 minutes after waking so you don't blunt your cortisol awakening response. The cortisol awakening response is real and well-characterized. Caffeine does raise cortisol — Lovallo et al. (2005), in Psychosomatic Medicine, showed that clearly, along with partial tolerance in habitual users. What has not been demonstrated, as far as I can find, is the actual claim: that delaying coffee by 90 minutes produces better afternoon alertness or better sleep that night. The mechanism is fine. The outcome trial doesn't exist. Those are different things, and the gap between them is where health trends do their best business.

None of these are lies. Two of them might help you. The honest summary is that the confidence in the average social-media post exceeds the confidence in the underlying literature by roughly an order of magnitude.

Four questions, before you buy anything

When the next protocol arrives, ask these in order.

• Compared to what? An improvement against nothing is not an improvement against the obvious cheaper alternative.
• How big, in units I can feel? Not "significantly improved sleep quality" — how many minutes, on what scale, measured how.
• Who was in the study? Twelve young men in a sleep lab, 46 older adults with diagnosed insomnia, 11 college athletes. Were they you?
• What would make this wrong? If the claim has no failure condition, it isn't a finding. It's a brand.

If a claim survives all four, it's probably worth two weeks of your attention. Most don't survive the second.

Back to the spreadsheet

I redid my experiment properly, which mostly meant making it boring. No more single-night comparisons. I ran two-week blocks: fourteen days with my last caffeine at 4 p.m., fourteen days with it at 1 p.m., everything else — bedtime, alcohol, room temperature — held as steady as an actual life permits.

The result was smaller and duller than my dramatic Tuesday. Averaged across the earlier-cutoff block, total sleep time went up by about 18 minutes. Time to fall asleep dropped by around 9 minutes, which is the number I actually noticed, because lying awake is subjectively expensive in a way that a missing sleep stage isn't.

And here's the part I can't clean up: I can't rule out that the earlier cutoff worked partly because I was paying attention, and paying attention nudged me to bed sooner. Two weeks each is enough to see through nightly noise. It is not enough to separate caffeine timing from the general effect of caring. I'd need a longer run, randomized block order, and ideally not knowing which block I was in — which, with coffee, is nearly impossible to arrange at home.

That's the honest state of my own data, and it's better than what most optimization advice offers, which is a testimonial with a stopwatch.

An honest rule of thumb

Change one variable, hold everything else, and give it two weeks before you decide anything. Not because two weeks is magic, but because it's roughly the shortest window in which the average beats the noise. If a change is real and worth keeping, it will still be there on day fourteen. If it needed a good night to show up, it was never yours.

So: tonight, move your last caffeine of the day two hours earlier, leave the rest of your life alone, and don't look at the numbers until the fourteenth morning — because anything you can detect in a single night is almost certainly noise wearing a lab coat.