A day always contains 24 hours. When someone sleeps more, they must spend less time sitting, walking, or exercising. Health researchers have learned to treat daily time use as one budget split four ways: sleep, sitting, light activity, and moderate-to-vigorous activity. They have also learned that one set of health cut-offs does not fit every population. Asian adults carry metabolic risk at lower body mass index, and sitting appears protective rather than harmful for people with epilepsy. Nobody has checked whether these population-specific movement "sweet spots" are stable facts or artifacts of how long a cohort was followed. Over a short window, people who are already becoming ill sit more and move less, and that alone can make a behavior look harmful or protective. We propose to take one large retrospective cohort with wrist accelerometry and linked death and hospital records, estimate subgroup-specific optimal 24-hour time budgets separately within successive follow-up windows, and measure how much a threshold derived from one to three years of follow-up differs from one derived from five or more years.
In two sentences, the claim is this. Population-specific movement thresholds should be reported as functions of the follow-up horizon, because short-horizon estimates mix true behavioral effects with the behavior changes of people who are already sick. By deliberately truncating a long cohort at the follow-up lengths typical of existing studies, we can show how far those studies' conclusions would move with more time. That tells future guideline work which of its subgroup thresholds to trust.
Where the gap is
Three bodies of work each hold part of the answer, and none holds all of it. Compositional data analysis (CoDA) maps the four-part time budget through log-ratio transformations into ordinary coordinates, so that "move 30 minutes from sitting to exercise" becomes a well-defined regression contrast. The method is established: Li et al. applied isometric log-ratio coordinates and simulated time reallocations in 3,974 Chinese students, and Fu et al. found reallocation effects that were asymmetric and larger in overweight preschoolers. Both are cross-sectional, and both use fitness or body-fat surrogates rather than clinical events. We reuse their method and move it onto hard outcomes over time.
The motivating finding comes from Fan et al.. In 98,561 UK Biobank participants with a median 7.1 years of follow-up, they reported that more sedentary time was associated with lower mortality in people with epilepsy (hazard ratio 0.87, 95% CI 0.78–0.97). Their restricted cubic splines, which are smooth curves fitted to find bends in a dose–response relationship, placed a protective zone at 7–13 hours per day of sitting. They modeled each behavior separately rather than as a closed budget. They also never asked whether the protective association holds across early and late follow-up. That second omission is our entry point. A subgroup-specific reversal is exactly the kind of result that reverse causation, or depletion of frail people early in follow-up, could produce.
Two further studies set the constraint and the population dimension. In MECH-HK, a combined lifestyle score predicted 47% lower metabolic syndrome incidence (OR 0.53) after an average of only 1.16 years of follow-up, while no single component did. It is unknown whether that benefit persists or only delays onset. The Singapore Myocardial Infarction Registry analysis found that obesity carried no age-adjusted mortality penalty overall, yet carried a 42–58% excess long-term mortality in female and Malay patients classified as metabolically healthy obese. Risk thresholds therefore split along intersections of sex, ethnicity, and metabolic status. That registry has no movement data, so its subgroups motivate our stratification but cannot be studied there directly. The Asian BMI literature (NFPHEP, Wuhan COVID-19 cohort) confirms that cut-offs shift by population. It also shows the standard response: swap one rigid binary threshold for another. We aim instead for a continuous optimal region in time-budget space with uncertainty attached.
Core Claim
For subgroups that differ in baseline disease or metabolic phenotype, the time budget that minimizes hazard depends on the follow-up horizon, and the dependence is largest precisely where subgroup-specific "reversals" have been reported. Our prediction is that the epilepsy sedentary-protective association is concentrated in the first two years and weakens or disappears after year five. If instead it holds steady across windows, that is strong evidence that the reversal is real and should inform guidance for people with epilepsy.
Research Program
We use the UK Biobank accelerometry subsample described by Fan et al.. Each participant's seven-day wrist recording is converted into an average daily composition of sleep, sitting, light activity, and moderate-to-vigorous activity. Outcomes come from linkage to death registries and hospital inpatient records: all-cause mortality as the primary outcome, and incident type 2 diabetes and cardiovascular events as secondary outcomes. We predefine three subgroup axes, each tied to a finding in the file. The first is epilepsy versus no epilepsy, a direct replication target. The second is sex crossed with metabolic phenotype (normal-weight versus obese, metabolically healthy versus unhealthy), mirroring the SMIR categories. The third is self-reported ethnic background, with South Asian participants analyzed separately where numbers allow and ethnicity-appropriate BMI cut-offs applied. The file does not report UK Biobank ethnic counts, so the feasibility of the third axis is itself something we will check first and report.
The estimator is deliberately simple: a compositional Cox model fitted separately within follow-up windows. For participant $i$ in subgroup $g$, with time budget $x_i$ transformed into three isometric log-ratio coordinates $\mathrm{ilr}(x_i)$ and covariates $z_i$, the hazard at time $t$ in window $w$ is
$$h_i(t) = h_{0,w}(t)\,\exp\!\big(\beta_{g,w}^{\top}\,\mathrm{ilr}(x_i) + \beta^{(2)\top}_{g,w}\,\mathrm{ilr}(x_i)^{\otimes 2} + \gamma^{\top} z_i\big).$$
Here $h_{0,w}$ is a baseline hazard for window $w$ (0–2, 2–5, and more than 5 years, with landmarking so that each window includes only people alive at its start). The coefficients $\beta_{g,w}$ and the quadratic terms $\beta^{(2)}_{g,w}$ let the curvature of risk over the time budget differ by subgroup and window. The optimal composition $x^{*}_{g,w}$ is the budget that minimizes the fitted hazard. We report a bootstrap region around it rather than a single point.
The quantity of interest is the horizon shift, defined as the Aitchison distance (the natural distance between compositions) between $x^{*}_{g,\text{early}}$ and $x^{*}_{g,\text{late}}$. We also translate it into minutes reallocated between behaviors so clinicians can read it. A smooth time-varying-coefficient version serves as a sensitivity analysis, not the headline.
The second product is a truncation experiment. We refit the pooled model with follow-up artificially censored at 1.16 years (the MECH-HK average), 2, 3, 5, and the full available time. At each truncation we record the subgroup optimum and the reallocation hazard ratios. This yields a curve showing how much a study of a given length would misstate the long-run threshold. It directly quantifies the concern raised by limited follow-up in Direction 3, using real rather than simulated data.
The epilepsy subgroup, with 1,167 participants, will have few deaths per window. We will report the minimum detectable hazard-ratio change per window before fitting. If early-window deaths are too sparse, we will collapse to two windows rather than overinterpret noise.
Baselines
The naive reference point is guideline-count scoring. Each participant receives the number of movement guidelines met, and a Cox model is fitted on that count, as in the adherence approach of Chung et al. and the additive score of MECH-HK. It uses the same cohort, covariates, and outcome, and it has no tuning. The proposal beats it if the subgroup-by-window compositional model gives better time-dependent concordance and calibration in held-out folds. It must also show subgroup optima that the count score cannot express. The file reports no number for this approach on UK Biobank mortality.
The strongest existing method on this exact question is the per-behavior spline analysis of Fan et al.. Its reported figures are a sedentary hazard ratio of 0.87 (95% CI 0.78–0.97) in epilepsy, protective zones of 7–13 h/day sitting and 0.1–0.2 h/day moderate-to-vigorous activity, and a Random Forest mortality AUROC of 0.87 ± 0.08 against 0.78 for demographics alone. We rerun their spline specification on our cohort extraction with identical covariates. We first check that we reproduce the 0.87 hazard ratio within its interval, since failing to reproduce it would itself be informative. Their model is then fitted within the same windows as ours. Beating it does not mean higher AUROC. Our aim is interpretive, and a flexible forest may well discriminate better. The criterion is that our model reveals a window-dependent sign or magnitude change that the pooled spline hides, and that the change holds up in bootstrap resamples.
For calibration and flexibility we include Deep Cox Mixtures. That model mixes several Cox models over latent groups, which lets it capture crossing survival curves. It reported an expected calibration error of 0.0200 against 0.0992 for a standard Cox model on FLCHAIN, and good calibration across minority groups in SEER. We give it the same inputs (ilr coordinates plus covariates), the same folds, and a hyperparameter budget matched to our model-selection effort. If it calibrates much better within subgroups, our parametric windows are missing structure and we will say so.
Two ablations isolate the key ingredient, which is subgroup-by-window estimation. The first is a pooled compositional Cox with subgroup interactions but no windows. The second is a windowed model with no subgroup terms. If the horizon shift appears only when both are present, the proposal's central claim is supported.
Why not just…?
Why not just drop the first two years of follow-up, the standard fix for reverse causation? This is the most serious alternative, and it is partly contained in our design: our late windows are a lag exclusion. What lag exclusion alone cannot provide is the comparison. It discards early events without showing how different they were, so it cannot tell a reader of a 1.16-year study how far off their estimate is. If the early and late optima turn out statistically indistinguishable in every subgroup, then simple lag exclusion was sufficient. That would be a publishable null in its own right, and we will report it as such.
Why not just add subgroup interaction terms to one compositional Cox model? That is our first ablation. It assumes the effect of time use is constant over follow-up. Deep Cox Mixtures showed crossing survival curves in SUPPORT and poor calibration of a single Cox model when proportionality fails. We expect the same failure here, and the ablation comparison will settle it on our data.
Why not just fit a flexible survival learner and read off partial dependence? A Random Forest already reaches an AUROC of 0.87 in this cohort (Fan et al.). Partial-dependence curves from such models do not respect the closed 24-hour budget, however: they vary sitting while holding everything else fixed, which is impossible. They also do not separate horizons unless windows are imposed anyway. We include Deep Cox Mixtures to keep this alternative honest on calibration.
Proposed Experiments
The first decisive check is the epilepsy reversal. We preregister that the sitting-to-activity reallocation hazard ratio in epilepsy moves toward or past 1 between the 0–2-year and more-than-5-year windows. We also preregister that the Aitchison distance between window optima exceeds its bootstrap 95% range under a permutation null, in which window labels are shuffled among events. Confirmation implies the published protective zone largely reflects early deaths among sicker, more mobile or more sedentary patients. A stable reversal implies a real population-specific threshold.
The second is the truncation curve. For each subgroup we plot the reallocation hazard ratio and optimum against the truncation length, from 1.16 years to full follow-up. We report the follow-up length beyond which the estimate stops moving by more than its standard error. This gives a concrete minimum follow-up recommendation for studies like MECH-HK.
The third is the metabolic-phenotype intersection. We test whether metabolically healthy obese women show a different optimal budget from normal-weight women, and whether that difference emerges only late, as the late-emerging mortality excess in SMIR would suggest.
Because this is retrospective observational data, confounding by unmeasured factors remains. We therefore add a negative-control outcome, accidental external-cause death, which movement should affect little. We also report E-values for every headline reallocation effect.
Across all analyses, the metrics are time-dependent concordance and subgroup expected calibration error in five-fold cross-validation, reallocation hazard ratios with bootstrap intervals, and horizon-shift distances. Every model is run on identical folds and covariates.
References
- The impact of 24-h movement behaviors on college students’ physical fitness and its isotemporal substitution benefits: a compositional data analysis approach
- Relationship between 24-hour Activity Behavior and Body Fat Percentage in Preschool Children: Based on Compositional Data and Isotemporal Substitution Analysis
- Physical activity levels may impact on the risk of premature mortality in people with epilepsy
- Combined Influence of Eight Lifestyle Factors on Metabolic Syndrome Incidence: A Prospective Cohort Study from the MECH-HK Study
- Association of body mass index, metabolic health status and clinical outcomes in acute myocardial infarction patients: a national registry-based study
- Prevalence of overweight and obesity in 15.8 million men aged 15–49 years in rural China from 2010 to 2014
- Association of body mass index with severity and mortality of COVID-19 pneumonia: a two-center, retrospective cohort study from Wuhan, China
- Associations between 24-hour movement guidelines and health-related quality of life among urban preschool children in Singapore, Japan, and China
- Deep Cox Mixtures for Survival Regression