Does daily regularity have independent value, or reflect activity and health?

01What the scores measure

Wrist devices record movement, with algorithms classifying sleep, wake and nonwear. NHANES minute MIMS is a processed movement summary involving interpolation, dynamic-range handling, filtering and integration. It measures neither energy expenditure nor melatonin or an endogenous clock directly. Raw acceleration frequency and the daily envelope of movement summaries are different levels; filtering raw acceleration does not establish that the daily pattern has been removed. Official minute documentation MIMS method

We retain the nonparametric concepts, preserve calendar gaps, and construct hourly means and clock profiles from usable minutes. SRI is 200 × agreement proportion − 100, with theoretical range −100 to 100. Nonparametric method Original SRI

A probability condition matters for SRI. If the two states are independent with equal sleep probability p, agreement is p²+(1−p)², giving expected SRI 100(2p−1)². It is zero at p=1/2; our chosen p=1/3 gives 100/9. Constant wake also scores 100. This algebra supplies no population reference value: one score cannot simultaneously establish regularity, duration and health.

02Identical movement curves can have different latent explanations

We construct seven-day signal examples: an observed sum of background, a latent 24-hour component and a schedule component. Parameters are chosen, with no participant fitting or physiological calibration. Two different latent phases combined with different schedule components give identical observable waveforms. Within this model's allowed parameters, the activity curve cannot uniquely identify latent phase; this is not a general proof about all physiological models.

Doubling the waveform doubles mean signal while preserving IS, IV and RA. Adding background 8 preserves the original periodic variation but changes RA from 0.594 to 0.245. Varying the daily schedule while keeping the latent component fixed changes IS from 1 to 0.854 and SRI from 100 to 41.7. These are results within chosen examples, not human effect sizes.

Missingness also changes the scores. Removing three early-morning hours on two days, then incorrectly filling them with zero, raises RA to 0.648. Retaining five complete days under our rules gives RA 0.594. The latter does not recover missing truth; it defines a different observed calculation domain. For an independent Gaussian hourly-noise example, the finite seven-day IS expectation is 23/167≈0.138, not a clinical threshold.

Another measurement can add information. The original SRI study measured dim-light melatonin onset in extreme regularity groups of students and observed a phase relationship. These were selected subgroups with a single phase assessment and observational contrasts. The light model also aligned mean phase to observed data; matching does not validate a causal regularity intervention. Original SRI study

03What the two-cycle reanalysis asks

We use NHANES 2011–2014 public minute data and the 2019 public mortality linkage, retaining every death cause. The parent universe contains 10,907 examined adults aged ≥20 with positive MEC weights; 10,886 are linkage eligible. Some public follow-up or cause values are perturbed, while vital status is retained. Public follow-up remains distinct from restricted 2022 follow-up. Age ≥80 is topcoded. Official data Mortality guidance

The complete sources contain 166,350,335 minute records. Fixed server versions, all byte ranges, local full-file hashes, identities, calendar coordinates and official daily aggregates are checked. This checks acquisition identity and parsing consistency, not physiological truth in wrist measurements or algorithm sleep. Eight bearing fields with published tables also undergo complete frequency checks before health models are fitted.

The primary measurement rule selects full protocol days 2–8. A minute requires 60 data seconds, zero quality flags, finite nonnegative MIMS, valid wear sleep/wake and no ambiguous transition. Each day requires ≥95% usable minutes and ≥45 in every hour; at least four days include a weekend. Every clock minute needs two contributing days; SRI requires two genuine adjacent day pairs and 2,592 paired minutes. Missingness and nonwear never become zero movement. The strict measurement domain contains 4,796 adults; these fixed conditions are not clinical standards.

Within each coverage family, four indices are fitted separately, with the same complete-covariate people in all three models. M1 controls an age spline, sex, race/ethnicity, cycle and season. M2 adds mean usable-minute MIMS, observed algorithm sleep hours, coverage and day count. M3 adds BMI, smoking, general health, recorded cardiovascular disease/stroke, cancer, diabetes, education and income. Sleep hours per selected day count valid observed states, including naps and fragments, without imputing full-day sleep; they are not main-night sleep-period duration. Mean MIMS is a wrist-movement proxy, with no energy or MET conversion.

The design starts from the full adult universe and subsets to analysis domains, retaining original MEC weights/2 and cycle-qualified PSUs/strata: 61 PSUs and 29 strata. Survey Cox models use Efron ties, Taylor covariance and normal-Wald intervals. An unavailable high-dimensional all-coefficient inversion is not used for individual exposure inference. IS/IV/RA use weighted primary common-sample IQR increments; SRI uses 10 points higher. Every sensitivity retains these units. Index terms are linear; full nonlinear shapes were not assessed and specified adjustment functions do not guarantee removal of confounding. Public nominal clock time cannot reconstruct light, shift work, timezone/DST or long-term endogenous rhythms. Four primary M3 tests use Holm adjustment; other fixed families are sensitivity diagnostics.

The plan was fixed before our indices and health outputs, with published results and structural preflight already known. This is neither prospective preregistration nor blinded analysis. Originals, participant tables and design matrices stay private; public code and aggregate results are traceable. Fixed plan and implementation package

04Which associations remain on the same people?

The primary common sample contains 4,362 people and 312 deaths. Entries are all-cause mortality HRs (95% intervals) per stated increment. Each index is fitted separately, so these are not mutually adjusted independent components. People, event counts and units are unchanged across M1–M3.

RA changes little with movement/sleep adjustment, then reaches 0.854 with recorded-health/other adjustment; primary Holm p=0.0356. Here, prognostic information means a conditional association, without validated incremental discrimination, calibration, external prediction or decision benefit. An association remains beyond these proxies, so average movement alone does not account for every association. SRI changes from 0.807 to 0.922, with interval 0.823–1.033 still allowing some favorable association. Primary M3 intervals for IS and IV also include 1. Nondetection is not proof of no effect, and differently scaled HRs do not rank the scores.

Weighted descriptive correlations are −0.020 for RA–mean MIMS, −0.375 for IV–mean MIMS and 0.203 for SRI–mean MIMS. These are not interchangeable movement-volume measures. Attenuation is not a percentage of confounding explained: added variables can be common causes, mediators or conditioning factors, and Cox coefficients are noncollapsible. Measurement error and specified functional forms limit adjustment. No score was directly intervened upon.

05Coverage, selection and follow-up change the scope

Relaxed coverage changes allowed missingness, days and people. Its M3 HRs are 0.817 (0.768–0.868) for RA, 0.867 (0.805–0.933) for SRI and 1.166 (1.041–1.305) for IV. The five-day family gives RA 0.824 (0.723–0.938) and SRI 0.918 (0.818–1.031). These sensitivity intervals do not receive the primary Holm adjustment. Choosing the most favorable domain is not robustness; stronger relaxed-domain associations are not automatically “poor-quality false positives.”

The all-seven-day family has 96 deaths and a nuisance race/ethnicity category with no events, generating an infinite-coefficient/separation warning. All 12 rows across three models are marked unstable under the prespecified implementation rule, not displayed as reliable HRs. Categories and models were not changed to rescue estimates. This does not establish that another day of wear always improves reliability.

Restricting to absence of recorded cardiovascular disease/stroke, cancer and diabetes retains 3,265 people/129 deaths: RA 0.874 (0.748–1.021), SRI 0.987 (0.844–1.153). RA remains favorably directed with a wider interval; this does not establish that disease explains everything or that subclinical disease is absent. A genuine 24-month landmark retains 4,305 survivors/255 later deaths: RA 0.850 (0.739–0.977), SRI 0.925 (0.814–1.051). It changes the target population and cannot eliminate reverse causation.

Replacing the age spline with ≤50/>50 categories gives SRI 0.932 (0.838–1.037), without a markedly stronger favorable association. Thus our sample does not establish coarse age as the main explanation for disagreement with the 2026 paper. Sleep algorithms, wear domains, future accident-death exclusions and covariates also differ. This is not a stepwise reproduction of that paper. Original 2026 study

Neither of the two parent-universe zero-month deaths enters the primary common sample, which also has no follow-up ≤1 month. The one-month landmark therefore gives identical main estimates, not an independent replication. Monitoring follows MEC; public month resolution and perturbation cannot reconstruct exact monitoring completion. No half-month was invented, and future follow-up length never defines a baseline group.

The common sample represents 40.0% of parent people and 41.1% of original weight. Retained versus other adults have weighted BMI means 28.43 versus 29.31, known cardiovascular disease/stroke 7.44% versus 9.59%, and ages 47.63 versus 47.28. Similar mean age does not establish negligible selection. Only 434 excluded people still have qualifying primary indices; their movement/sleep means do not describe all 6,545 other adults. Retention comparisons are original-weight descriptions without survey intervals or inverse-selection repair.

On the same 4,796 people and selected primary days, admitting technically valid nonwear/unknown movement changes mean RA by −0.00208. SRI retains original state pairs, so its difference is zero. This small difference applies only to already ≥95%-wear days, not excluded people. Twelve people have a day with all valid states classified wake; 14 have a day classified all sleep; one has SRI 100. These diagnose algorithm states, not clinical sleep.

06A constant HR is a working summary

Actual risk intervals are split at 24/60 months, preserving original weights and survey clusters, with two exposure-by-period interactions per index. There are 4,362 people/12,776 intervals; no primary zero-month record needs exclusion. Units match the main table. This never groups baseline people by future survival length.

No period-difference test crosses conventional 0.05, but this does not establish time-invariant effects. RA is more favorably estimated at 24–60 months and near 1 after 60; the late IS interval excluding 1 is not promoted to a new main conclusion. These secondary estimates have precision/multiplicity limits; the checks validate neither every confounder nor whole-model PH. Main Cox HRs are constant-coefficient working summaries, not death probabilities, years gained or healthspan effects.

07Favorable cohort findings answer specific questions

MESA is a cohort distinct from NHANES and UK Biobank. Among 1,992 participants free of known cardiovascular disease, 111 cardiovascular events occurred over median 4.9 years. Fully adjusted HR was 2.14 (1.24–3.68) for sleep-duration SD >120 versus ≤60 minutes. Multiple cardiovascular and sleep factors were controlled, supporting prognostic information in timing variability. The measure is SD; residual confounding and actigraphy error remain, and comparison bins are not intervention prescriptions. MESA cardiovascular study

Another analysis of the same MESA sleep cohort, with 1,759 participants and 176 deaths, found HR 0.61 (0.45–0.83) for the composite labelled “regular-optimal” versus “irregular-insufficient.” Duration and two variability measures enter the clustering together, so this does not attribute the association to changing regularity alone. The papers are not two independent cohorts. MESA mortality study

UK Biobank findings also retain favorable associations. In SLEEP, after adding sleep duration, highest versus lowest SRI quintile had mortality HR 0.74 (0.62–0.89), with device-derived average activity adjusted. It overlaps the eLife population, with different sleep algorithms, nap handling and SRI standardization. Failure to detect added duration information in a model comparison establishes neither no duration benefit nor a comparison of intervention effects. SLEEP study eLife report

NHANES 2022 reports favorable RA associations, including a recorded-health restriction. Activity adjustment used self-report and sleep definitions differ from ours. The exact quartile contrast for its single HR is incompletely specified; we do not invent a raw-unit or highest-versus-lowest interpretation. The recent 2026 paper reports favorable SRI associations, but age uses only ≤50/>50 categories and subsequent accident deaths are excluded. Its estimand and residual age confounding require scrutiny. Our analysis retains every death cause; these NHANES papers are not independent participant replications. NHANES 2022 NHANES 2026

08Direct schedule interventions remain short

A 2025 randomized crossover abstract reports 16 adults and a 12-night regular condition, with reduced duration variability and favorable resting heart rate/RMSSD changes. This supports some short-term physiological changeability. Nonsignificant mean-duration difference does not establish equivalent duration. Only the abstract was obtained; paired data, full safety reporting and all endpoints remain unverified. Registered primary outcomes are not interchangeable with the abstract's HRV readouts. Trial abstract Registration

Early findings differ. A 1996 prospective two-condition study reports better subjective alertness with added regularity instructions; a 1982 before/after study of 12 male students over 38 nights detected no significant EEG, performance or mood changes. Only the original abstracts were read: we invent neither randomization, missing sample sizes nor confidence intervals, and nondetection does not prove no effect. 1996 abstract 1982 abstract

Circadian misalignment poses another question. A controlled 14-person crossover changed behavioral timing and raised two-hour post-meal glucose AUC by approximately 6%, while actual sleep fell 56 minutes. The overall manipulation can change short-term physiology; nonsignificant sleep covariates do not exclude a sleep contribution, and it is not a habitual-regularity or longevity trial. The 2026 cell experiment also has single-time sampling, time-since-last-stimulus, viability and normalization limits, with no clock-specific rescue separating mechanisms. Controlled misalignment Cell experiment and limits

09What conclusions are supported?

Timing pattern and volume are different measurement targets, and at least RA retains a conditional association under the declared conditions. Adjustment does not identify an independently manipulable causal effect or validate clinical incremental prediction and decisions. Primary SRI uncertainty and favorable relaxed-domain SRI findings both belong in the account. The coarse-age diagnostic does not confirm a large expected change, and the seven-day family supplies no usable estimates.

Intervention value requires separate measurement of schedules, activity opportunities, duration and physiological phase, recording work, medication, apnea and subclinical disease, and feasible randomized schedule changes with function and disease outcomes. Changing an algorithm score alone cannot answer these questions. Healthspan, lifespan gains, optimal thresholds and individual treatment plans are not estimated here.

The public package retains all 64 fixed main/sensitivity rows, including 12 unstable rows; 12 period estimates; 56 successful-fit independent weighted-Efron score/derivative checks; selection, correlation, missingness and measurement diagnostics; and eight chosen signal inputs. Score checks do not independently validate Taylor covariance. Originals, participant rows and matrices remain private. The 2025/early-trial full texts, some supplements, complete author pipelines and clinical causality remain unresolved. Reproduction does not remove those gaps. Code, aggregate results and figures