The verdict
Adequate sleep is a real health lever, but the popular reading — that any deviation from 7-8 hours shortens your life — is half wrong. The evidence splits the question in two, and only one half is a lever.
Chronic short sleep is a genuine, causally-footed harm. Sleeping under ~7 hours, with most of the penalty below 5 h, carries roughly 12% higher all-cause mortality, higher type-2-diabetes risk, and higher coronary risk — the coronary link now supported by genetics, not association alone. For a chronically short sleeper, moving toward ~7-8 hours is the lever worth pulling. (Cappuccio et al., 2010) (Wang et al., 2022)
Long sleep is mostly a marker of underlying illness, not a cause of it. Sleeping 8+ hours does not shorten a healthy person’s life, and cutting sleep to avoid death or dementia is not supported. A habitual long sleeper is a candidate for a check-up, not a sleep-reduction target. (Cappuccio et al., 2010) (Livingston et al., 2024)
When you sleep may matter as much as how long: irregular day-to-day timing predicts death at least as strongly as short duration, and is easier to fix. The newest evidence ties sleep disorders — apnea, restless legs, daytime sleepiness, insomnia — to dementia and cognitive decline, so the cognition lever is multi-pronged rather than a duration target (association-grade; no trial yet shows that treating them prevents dementia).
For insomnia, CBT-I comes first — larger, more durable, and harm-free — while hypnotics carry a harm signal and melatonin is small but benign. For sleep apnea, CPAP relieves symptoms but has not been shown to prevent heart attacks or death. Throughout, 7-8 hours is a reference range, not a precise target or an upper cap.
Lengthen short sleep, and stop reading long sleep as a cause
Both short and long sleep sit above a ~7-8 h reference on the mortality curve, but the two arms are not the same kind of finding. Cappuccio’s meta-analysis (16 studies, 27 cohorts, 1.38 million people, 112,000 deaths) pools short sleep at RR 1.12 (1.06-1.18, I2 39%) and long sleep at RR 1.30 (1.22-1.38, I2 71%) against a 7-8 h referent. (Cappuccio et al., 2010) The larger association is the weaker claim.
The short-sleep RR 1.12 is a robust direction, not a precise dose. A 12% excess is significant and holds up, but it is small for an observational metric — self-reported, single-timepoint sleep, near the range residual confounding alone can produce. It is believed over the long arm not because 12% is itself compelling, but because it carries a named mechanism, Whitehall II’s cause-specific direction (curtailed sleepers died predominantly of cardiovascular causes, increased sleepers of non-cardiovascular causes), and now genetic support. (Cappuccio et al., 2010)
The long arm has none of that footing. No study has demonstrated a mechanism by which sleeping longer causes death; Cappuccio reads long sleep as «a powerful additional marker of ill-health» (Cappuccio et al., 2010). Three tells mark it as artifact-shaped, each an instance of The U-Shaped Association Artifact: it is the larger association yet the mechanism-free one; its heterogeneity is far higher (I2 71% vs 39%); and it strengthens with age and with a longer cutoff (>=60 y RR 1.46; >=10 h RR 1.54) — a frailty gradient, not a physiological dose-response. The short arm, by contrast, is flat across every subgroup (age, cutoff; heterogeneity P=0.89). (Cappuccio et al., 2010)
Because Cappuccio bundles all short sleep into one pooled and flat category, it cannot say whether the 6-7 h band carries the penalty or only chronic short sleep does. Yin’s per-hour dose-response spline (67 articles, 3.58 million participants, 7 h reference) resolves the shape: 6 h sits at 1.01, barely above the nadir; 5 h at 1.04; the arm steepens only below ~5 h, reaching 1.12 at the 3 h extreme. (Yin et al., 2017) So the pooled 1.12 is driven by the chronic-short (sub-5 h) studies, not the 6-7 h band — the penalty is a chronic-short-sleep (<5-6 h) phenomenon, and the lower bound is arguably 6+ h, not a hard 7. The long arm, disaggregated, is steep and asymmetric (9 h 1.15 -> 10 h 1.32 -> 11 h 1.53), reconfirming that the larger arm is the mechanism-free one.
The genetic check settles the short-arm verdict. Wang’s meta-review pools 11 Mendelian-randomization studies — which use inherited genetic variants to fix lifetime sleep, immune to reverse causation — and finds genetically-predicted short sleep raises coronary artery disease (pooled OR 1.24, 1.15-1.34) and heart failure: «short sleep duration is a causal risk factor for CAD and HF». (Wang et al., 2022) Long sleep gets no genetic support, matching its marker reading — but the long-sleep instruments are few and underpowered, so this is insufficient evidence, not a demonstrated null. (Wang et al., 2022)
The action follows the arm, not the number. “7-8 h” is the referent the curve is measured against, not a target to hit precisely and not an upper cap: shortening a long sleeper’s sleep optimises a symptom, and there is no evidenced harm from sleeping past 8 h to cap against. For a chronically short sleeper, extending toward ~7-8 h is a lever with mechanism, direction, and genetic support behind it.
The same split shows up for diabetes and heart disease
The arm-asymmetry is not a quirk of the mortality data. It repeats on every outcome sleep touches, and the next two are the ones that carry most of the health stakes: diabetes and the heart.
Shan pooled ten prospective cohorts and found the identical U-curve on type-2 diabetes — risk bottoms out at 7-8 hours and climbs on both sides, about 9% more diabetes per hour below 7 h and 14% per hour above 8 h (Shan et al., 2015). And the two arms split the way they did for death:
- Short sleep is causally footed. The short arm carries a worked, human-tested mechanism — lab sleep-restriction lowers glucose tolerance and insulin sensitivity — and shows no nonlinearity, a clean dose relationship (Shan et al., 2015).
- Long sleep is a marker, left unadjudicated. Shan calls the long-arm mechanism «considered more speculative», the arm runs hotter (higher heterogeneity), and it goes borderline the moment one study is dropped (Shan et al., 2015). No strong check — no genetic test, no corrected referent — was run, so the long-sleep-diabetes link is suspected artifact, not established harm.
Shan can only watch the association; it cannot fix which way cause runs, because illness lowers sleep as readily as sleep raises illness. Capers closes that gap. Pooling randomized trials that manipulated sleep, Capers found that «sleep restriction increases food intake and total energy expenditure» — the causal step the cohorts could only assume (Capers et al., 2015). Short sleep really does drive people to eat more.
But the popular reason for it is wrong. The leptin-down, ghrelin-up appetite-hormone story is not the pathway: across the same trials, sleep restriction moved neither hormone (leptin p=0.84, ghrelin p=0.83). Capers, citing Chaput and St-Onge, concludes «hedonic factors are likely stronger» — that, plus simply having more waking hours in which to eat (Capers et al., 2015). So hold leptin/ghrelin as a discounted mechanism, never a demonstrated one; the extra eating is real, its hormonal explanation is not.
The heart tells the same story, and Wang 2022 finally brings the strong check the sleep literature had been arguing for without. On incidence, the observational curves echo mortality — coronary disease is U-shaped, stroke J-shaped, per extra hour RR 1.13 (1.07-1.20) (Wang et al., 2022). Then Wang pools eleven Mendelian-randomization studies:
- The short arm survives. The same genetic check that settled the mortality verdict extends to the heart’s own endpoints: genetically-predicted short sleep raises heart attack (OR 1.20, 1.12-1.28) alongside the coronary-disease and heart-failure results already shown (Wang et al., 2022). Genes converge with the mechanism and the Whitehall direction.
- The long arm gets no genetic support — «long sleep duration has no causal associations with stroke and CAD in the MR studies» (Wang et al., 2022). But the long-sleep instruments are few and underpowered, so this is insufficient evidence, not a demonstrated null.
When you sleep may beat how long
Duration is only one knob. A person can sleep seven hours at ragged, drifting times, and how regular the timing is may be the bigger lever — the newly-visible one that duration-centric sleep advice misses entirely.
Windred measured both dimensions objectively, by wrist accelerometry rather than self-report, in one cohort — roughly 61,000 UK Biobank participants — and raced them head-to-head for mortality. Day-to-day regularity, scored as the Sleep Regularity Index, predicted death more strongly than how many hours someone slept: the most-regular fifth had a hazard ratio of 0.70 (0.59-0.83) against the least regular, versus 0.76 (0.65-0.89) for duration (Windred et al., 2023). Adding duration to a regularity model bought nothing — «sleep duration does not explain significant additional variance in mortality risk beyond the variance explained by SRI scores» (nested likelihood-ratio test, p=.20) (Windred et al., 2023). Unlike the duration U, the regularity curve is monotonic: steadier is simply better, with no upper turn.
This does not overturn the duration U-curve — it sits beside it. Windred’s objective duration range is truncated, its longest quintile reaching only past 7.56 h, so it never touches the
9-10 h long-arm where the illness marker lives, and it confirms the short-sleep associations (Windred et al., 2023). Regularity is a complementary axis for the person who already sleeps enough but at chaotic times — shift-adjacent schedules, social jetlag — not a rival that unseats how-much-sleep. The proposed mechanism is circadian disruption: irregular timing scatters the body’s clock signals, a when-not-how-much pathway, held directionally because the design is correlational.
Two things keep this a candidate lever rather than a settled one. The finding rests on this single cohort, correlational, with no trial yet raising regularity and measuring survival — so it ranks below the established sleep levers. But it is concrete and cheap: the top fifth for regularity fall asleep and wake within roughly 1-hour windows most days, the bottom fifth within ~3-hour windows, and Windred notes regularity «may also be an easier dimension to target through interventions» than extending sleep, which is biologically and socially hard to do (Windred et al., 2023).
For the aging brain, target the disorder — not a duration number
Duration and regularity are the quantity and timing of sleep. The newest evidence points somewhere else entirely — at named sleep disorders, and at a different outcome: the aging brain. A gold meta-analysis of 76 longitudinal cohorts (Zhang 2025) tracked eight disorder types against four cognitive endpoints — Alzheimer’s disease, vascular dementia, all-cause dementia, and cognitive decline — and the result reframes the cognition question. This is not a duration target you hit or miss. It is a map of separate levers, each with its own size and its own endpoint (Zhang et al., 2025).
Ranked by strength, the disorders that carry a real signal are:
- Restless-legs and other sleep-related movement disorders — vascular dementia RR 2.53 (1.30-4.93), the single strongest association in the map.
- Excessive daytime sleepiness — vascular dementia RR 1.85 (1.39-2.47).
- Obstructive sleep apnea — Alzheimer’s disease RR 1.39 (1.16-1.68).
- Poor sleep quality — all-cause dementia RR 1.17 (1.03-1.32).
- Insomnia — all-cause dementia RR 1.13 (1.04-1.23).
- Circadian-rhythm disturbance and REM-sleep behaviour disorder — no significant association.
Two refinements sharpen the map. Insomnia’s whole signal comes from trouble falling asleep — difficulty initiating sleep carried RR 1.11 (1.01-1.20), while trouble staying asleep and early waking were null. And the movement and daytime-sleepiness disorders load onto vascular dementia, which places them on the cerebrovascular route the fabric already counts among the cardiometabolic big rocks — so reducing them is partly a way of pulling the vascular rock, not a wholly separate lever (Zhang et al., 2025).
Sleep duration still shows a U here, but its shape depends on which endpoint you ask about. Short sleep (under 7 h) predicted cognitive decline only — RR 1.27 (1.12-1.42) — and did not reach dementia or Alzheimer’s. Long sleep (over 8 h) did the opposite: «sleep duration > 8 h mainly elevates the risk of AD, dementia, and cognitive decline» (Zhang et al., 2025). So for dementia the U collapses to a one-sided rise at the long end.
That long arm reads the same way it does for mortality. It is a preclinical marker, not a cause: Zhang describes long sleep as «a preclinical marker driven by the APOE ε4 carrier gene» (Zhang et al., 2025), the early illness showing up as extra sleep rather than the sleep causing the illness. The same arm-level rule the fabric applies to mortality and diabetes applies again here, one outcome at a time: adjudicate the arm, not the curve The U-Shaped Association Artifact. Telling a healthy 8.5-hour sleeper to cut back has no evidenced benefit Sleep Duration and Mortality.
One check runs the right way. If the whole map were a reporting artifact — people with early cognitive change misjudging their own sleep — then a cleaner instrument should weaken it. It does the reverse. Objectively measured insomnia carried RR 1.26 (1.15-1.40) with low heterogeneity (I2 26.1%), stronger than the self-reported pooled estimate of 1.09 and far more consistent (I2 77.7%) (Zhang et al., 2025). The better measurement sharpens the insomnia signal instead of erasing it — evidence against a pure measurement artifact on that arm Measurement Error in Dietary Assessment.
Hold the grade honestly. Every estimate here is observational and heterogeneity is high; the authors tested for publication bias and found it for some outcomes — sleep-disordered breathing on Alzheimer’s carried Egger’s p=0.001 — while a trim-and-fill left the pooled estimates unchanged (Zhang et al., 2025). The map marks a sleep disorder as a candidate lever and a screening flag — a person with one sits at higher baseline risk Dementia Prevention and Modifiable Risk Factors, Baseline Risk and the Relative-Absolute Split — but no trial in it tests whether treating the disorder actually lowers dementia incidence . Whether fixing broken sleep protects the brain is exactly the untested question — which raises the treatment question head-on: once sleep is broken, what actually works?
For insomnia, start with therapy, not pills
Three treatments answer that question, and the honest ranking runs opposite to how most people reach for them — drug first, behaviour last. The order that follows the evidence is set by type, durability, and harm, not by the effect size on any single sleep number.
Cognitive behavioural therapy for insomnia (CBT-I) comes first. A meta-analysis of 87 randomized trials found large effects on insomnia severity (ISI g=0.98) and sleep-onset latency (g=0.57), and its smallest effect on total sleep time (g=0.16) — CBT-I consolidates sleep rather than adding hours (van Straten et al., 2018). Two properties earn it the top slot. It holds up across patients — «quite robust (similar for patients with or without comorbid disease, younger or older patients, using or not using sleep medication)» (van Straten et al., 2018) — so it works even for chronic hypnotic users, which makes it the exit route from pills. And it is guideline first-line: «the American College of Physicians recently recommended CBTI as the initial treatment for all adults with insomnia» (van Straten et al., 2018).
Prescription hypnotics buy real sedation against a real harm flag. A matched cohort found a dose-dependent mortality signal (hazard ratios rising from 3.6 to 5.3 across dose tiers) plus a cancer signal (HR 1.35) (Kripke et al., 2012). But this is one observational cohort, and the live confounder is confounding by indication — the distress and occult illness behind a hypnotic prescription independently raise mortality. So the hazard is a decision-relevant harm flag, not a demonstrated death toll, which moves hypnotics to a short-course, lowest-effective-dose, last-resort role.
Melatonin is small but benign. It cut sleep-onset latency by about 7 minutes and raised total sleep time by roughly 8 (Ferracioli-Oda et al., 2013) — a modest effect. It earns a place anyway because it shows no tolerance and a «relatively benign side-effect profile» (Ferracioli-Oda et al., 2013), making it a low-stakes adjunct, especially for timing complaints, never a substitute for CBT-I.
So do drugs give real sleep? CBT-I restores continuity and efficiency; melatonin nudges timing; hypnotics produce drugged sedation whose long-run safety is unestablished — three different things, not interchangeable routes to one outcome. (Sedating antihistamines, the other over-the-counter reach, are a named gap — no systematic review is held.) Apnea is the remaining treatable sleep disorder, and it is the sharpest lesson in why fixing a symptom is not the same as changing an outcome.
Treat sleep apnea for its symptoms, not to prevent heart attacks
Sleep apnea tracks with cardiovascular disease in observational data, and CPAP — the mask that splints the airway open each night — is the standard fix. The obvious hope is that fixing the breathing fixes the heart. Pooled trials say it does not. Yu’s meta-analysis of 10 randomized trials (7,266 adults) found positive airway pressure left hard cardiovascular events flat: major adverse cardiovascular events RR 0.77 (95% CI 0.53-1.13, not significant), all-cause death 1.13. Even the one cardiovascular surrogate the trials measured, systolic blood pressure, did not move (-0.20 mm Hg, 95% CI -2.29 to 1.89). (Yu et al., 2017) That null surrogate is the mechanism of the null outcome — the pressure never fell, so the events it was supposed to prevent never budged.
What the mask does buy is real, and sits on a different axis. Daytime sleepiness eases (Epworth -1.92 points), and disease-specific quality of life, mood, and depression scores all improve. (Yu et al., 2017) These are patient-important outcomes in their own right, and reason enough to treat a symptomatic person. They are simply not heart-attack prevention.
One caveat keeps the null honest. CPAP adherence in these trials was low — nightly use averaged 1.4 to 6.6 hours — so this is an average-adherence null, not a high-adherence one. (Yu et al., 2017) Whether someone who wears the mask all night, every night, for years would see a cardiovascular benefit is genuinely unknown: insufficient evidence, not a refutation. The honest reading holds both halves — no benefit at the adherence people actually manage, and an open question at the adherence they rarely reach.
A different kind of lever attacks the apnea at its source. Most obstructive sleep apnea is driven by obesity, and the SURMOUNT-OSA trials (Malhotra) tested weekly tirzepatide — a weight-loss drug — in people with moderate-to-severe apnea and obesity. It cut the apnea-hypopnea index (AHI, the events-per-hour severity score) by roughly half, measured off the machine, and moved the very surrogates CPAP left flat: systolic blood pressure fell 3.7 to 7.6 mm Hg, and inflammation (hsCRP) dropped as well. (Malhotra et al., 2024)
But moving a surrogate is not preventing an event, and the caution that punctured CPAP binds tirzepatide too. AHI, blood pressure, and CRP are all markers; neither trial measured heart attacks, strokes, or deaths, and the outcome trial (SURMOUNT-MMO) is still running. (Malhotra et al., 2024) So the hard-outcome question is unproven for both the mask and the drug, and a surrogate advantage over CPAP is not a demonstrated outcome advantage -> Surrogate Outcomes.
The decision this leaves is clean. Do not treat sleep apnea in order to prevent a heart attack or death — that expectation is not supported for either the mask or the drug. Treat it to relieve the symptoms and the daytime toll, which it genuinely does. And treat the cardiovascular risk directly, through the levers that carry the outcome evidence: blood pressure, lipids, and weight.
Rank sleep honestly, and act on the arms that hold
Where does sleep rank among the levers? Below the biggest rocks. A person who smokes, carries excess weight, or barely moves gains far more by pulling those first, and for someone already sleeping seven or eight regular hours the remaining sleep gains are small. But for two strata sleep is a genuine big rock, not a refinement: the chronic short sleeper, and the person whose sleep-wake timing swings widely from day to day. A candidate third arm — the map of sleep disorders onto cognitive decline — is real but still association-grade, with no trial yet showing that treating a disorder prevents dementia. -> Dementia Prevention and Modifiable Risk Factors, Sleep Regularity and Mortality
The everyday sleep-hygiene advice reaches for more — a consistent light and circadian schedule, exercise to deepen sleep, protecting enough time in bed, avoiding late meals, timing caffeine. Each is plausible, but none is yet backed by the systematic-review evidence this deliverable is built on. They are candidate levers, named honestly as gaps rather than graded recommendations. Two the wiki treats elsewhere: late-evening eating on Meal Timing and Fasting, and caffeine timing on Coffee Consumption and Health.
Several questions a careful reader will ask, this evidence base cannot yet answer, and naming them is part of the answer. Whether sleep quality, architecture, or within-night continuity change hard outcomes at the same total hours is not held. A systematic review of light and circadian scheduling, and a meta-analysis of exercise’s effect on sleep, are not held. Nor is a Mendelian-randomization test of chronotype genetics or social jetlag. And whether treating a sleep problem lowers dementia incidence — CBT-I or CPAP measured against cognitive outcomes — is the open interventional question the association map most needs. The nearest landmark, an updated review of sleep problems and cognitive decline, is not yet held.
So the honest close is not a single number to hit. If you sleep too little, move toward seven or eight hours — that arm is causal, and the gain is real. If your nights land at wildly different times, steady them; the fix is cheap and the signal is strong. And if you already sleep enough, on a regular schedule, this is a lever you have largely pulled — the questions that remain are open, not urgent, and your attention is better spent on the bigger rocks.
Evidence box
Question ’What is the effect of sleep (duration, quality/architecture, continuity, regularity, timing/chronotype) on each patient-important outcome, what is the dose-response shape (is the duration-mortality curve U-shaped, and is the upper arm real or an artifact?), what do sleep aids do to those outcomes, and how does sleep rank as a lever?‘ Evidence included 13 sources — 8 gold, 4 high, 1 moderate Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note 1 source(s) below the gold evidence bar feed this page: Kripke (cohort, moderate). Each labelled by tier; none load-bearing for the core claims. Last updated 2026-09-04 · Independently reviewed: No · Full edit history