Target trial emulation  ·  NZAVS 2020–2023  ·  Lab report

Work Hours: The Body, Not the Mind

How work hours affect well-being: a target trial emulation in a national New Zealand cohort

Ballerina X. S. Chong · Chris G. Sibley · Joseph A. BulbuliaPLOS One · 2026

Longer work hours travel with poorer well-being, but an association cannot say what would happen if hours changed. We emulate the experiment we would like to run: among 24,579 pre-retirement New Zealanders, we estimate how 28 well-being outcomes would differ if the same cohort worked ten more or ten fewer hours each week than observed.

24,579pre-retirement adults, NZAVS
3annual waves, 2020–2023
28well-being outcomes
±10hours per week, the two policies

The emulated trial

What is our question?

For pre-retirement New Zealand adults, how would 28 well-being outcomes differ one year later if weekly work hours shifted ten hours up, or ten hours down, from what would otherwise occur?

What is our population?

24,579 NZAVS participants of working age at baseline (2020/2021), followed across three annual waves to 2022/2023, with censoring adjustment carrying inference to the baseline cohort.

What interventions do we compare?

Shift up · +10 hours

Weekly hours shifted ten upward, versus the natural course — the censoring-adjusted hours actually worked.

Shift down · −10 hours

Weekly hours shifted ten downward, versus the same natural course.

The central result

Shifting work hours affects recovery — fatigue, sleep, body mass — more clearly than well-being as a whole.

Under the upward shift, fatigue rises and sleep falls; body mass index and perceived physical health also worsen, though these estimates are more sensitive to residual confounding. Under the downward shift, the clearest fall is in fatigue. Most of the 28 outcomes change little under either policy, and the data better support the downward shift. Naive baseline associations are broader, larger, and sometimes reversed in sign; the causal shifts that survive adjustment concentrate in recovery and embodied health.

+10 hours · the cost

Recovery erodes

Fatigue rises by 0.08 standard deviations and nightly sleep falls by roughly 3.4 minutes. Body mass index increases by roughly 0.19 kg/m² and perceived physical health declines by roughly 0.13 points, though both estimates are more sensitive to residual confounding. Perceived support rises slightly and carries the same caution.

−10 hours · the relief

Fatigue lifts first

Fatigue falls by 0.04 standard deviations — the clearest change under the downward policy. Body mass index (−0.13 kg/m²) and perceived physical health (+0.07 points) also improve, and both estimates remain sensitive to residual confounding.

The estimates

least vulnerable to unmeasured confounding interval excludes zero; E-value modest -0.10 -0.05 0 +0.05 +0.10 +10 HOURS PER WEEK fatigue: 0.079 [0.035, 0.123], E-value bound 1.22 FATIGUE +0.079 · E 1.22 perceived physical health: -0.055 [-0.100, -0.010], E-value bound 1.10 PERCEIVED PHYSICAL HEALTH -0.055 · E 1.10 perceived support: 0.047 [0.002, 0.092], E-value bound 1.04 PERCEIVED SUPPORT +0.047 · E 1.04 −10 HOURS PER WEEK fatigue: -0.042 [-0.069, -0.015], E-value bound 1.13 FATIGUE -0.042 · E 1.13 body mass index: -0.020 [-0.034, -0.006], E-value bound 1.08 BODY MASS INDEX -0.020 · E 1.08 perceived physical health: 0.030 [0.003, 0.057], E-value bound 1.06 PERCEIVED PHYSICAL HEALTH +0.030 · E 1.06

Standardised average treatment effects with Bonferroni-corrected 95% confidence intervals, as reported in the paper's results. Green marks the effect the paper identifies as least vulnerable to unmeasured confounding under each policy; gold marks estimates whose corrected intervals exclude zero but whose E-value bounds are modest. Full 28-outcome tables appear in the paper's supplement.

Inference & robustness

01

The natural course comparator

Each policy is compared with the censoring-adjusted hours the cohort actually worked — not with a hypothetical zero-hours world, and not with each other.

02

Adjustment

Modified treatment policy estimators, implemented in the lmtp R package, adjust for baseline differences and for dropout using machine learning.

03

Uncertainty, stated plainly

Uncertainty remains substantial. Many point estimates attenuate towards zero after confounding control and censoring adjustment, and many Bonferroni-corrected intervals include zero.

04

Scope

Estimates are defined for these shift policies in this population, and people gain and lose hours for many reasons. Both hours and outcomes are self-reported; random measurement error would likely attenuate true effects.

Citation

Chong, B. X. S., Sibley, C. G., & Bulbulia, J. A. (2026). How work hours affect well-being: A target trial emulation. PLOS One. doi:10.1371/journal.pone.0350816

Data: the New Zealand Attitudes and Values Study. Estimation: modified treatment policies with the lmtp R package.