What Did the Four-Day Week Trials Actually Measure?

Six cohorts of companies recruited by the nongovernmental organization 4 Day Week Global between February 2022 and June 2024 each ran a six-month trial on an 80% of hours for 100% of pay model, preceded by a work-reorganization phase. The analytic sample was 2,896 employees at 141 organizations, recruited through country cohorts in Australia, Canada, Ireland, New Zealand, the UK and the US, compared against 285 employees at 12 US companies that expressed interest but did not participate (Fan et al., 2025). Baseline survey completion was 80% in trial companies against 58% in controls, and 73% of trial baseline respondents completed the endpoint survey against 59% of controls.

The 80% figure is the design target, not the delivered reduction. Average company-level hours went from 39.21 to 34.01, a cut of 5.20 hours, and individual reported hours went from 38.98 to 34.26, a cut of 4.72 hours, both at p < 0.001 (Fan et al., 2025). The modal reported week moved from 40 hours to 32 at both levels. Control companies stayed flat, from 38.88 to 39.16 at company level and 39.87 to 39.95 at individual level, neither significant. Against a 39.21-hour baseline, a 5.20-hour cut is nearer 13% than 20%.

The label also covers more than one schedule. Only 82% of participating companies offered a standard four-day week, 12% combined it with other arrangements, and 6% used a different form of work-time reduction (Fan et al., 2025). Anyone citing the trial is citing a family of work-time interventions with a shared pay promise, not a single uniform policy that can be copied line for line.

Across trial companies, work-related burnout measured with seven items adapted from the Copenhagen Burnout Inventory fell from 2.83 to 2.38 on a 1-5 scale, with the mean change reported as -0.44 points (95% CI -0.47 to -0.42) and an effect size of 0.55. That change is the paper's own paired estimate, published with its own confidence interval rather than derived by subtracting one rounded wave mean from the other, which is why the two printed means do not resolve to it exactly. Quote the reported change and its interval instead of computing your own.

Job satisfaction rose from 7.07 to 7.59 on a 0-10 scale (effect size 0.27), self-rated mental health from 2.93 to 3.32 on a 1-5 scale (effect size 0.39), and self-rated physical health from 3.01 to 3.29 (effect size 0.29), all at p < 0.001 (Fan et al., 2025). The authors describe these as small to medium effect sizes rather than dramatic ones. Control employees showed nothing comparable: burnout went 2.90 to 2.94, job satisfaction 6.66 to 6.47, mental health 2.90 to 2.93, and physical health 3.06 to 3.09, none of them statistically significant. All four trial-versus-control differences in pre-post change were significant, three at p < 0.001 and physical health at p < 0.01, and the company-level hours difference against controls came to 5.48 fewer hours (95% CI -6.87 to -4.09), effect size 1.48.

That trial-versus-control comparison is the strongest part of the design, and the authors say it supports causal inference more than the within-trial change alone. It is still not random assignment. Companies self-selected into the trials through an advocacy organization, and the authors write that because those companies may be more supportive of flexibility and wellbeing initiatives, the findings likely overestimate the true effect. Random assignment was not feasible for an organization-wide change at mostly small firms, and the authors call for a government-sponsored randomized study to settle it. Every outcome here, including hours, is self-reported: no biomarker, productivity, revenue, or service-level measure was collected.

At the individual level, burnout improvement tracked the size of the hour cut. Measured against control employees and workers whose hours did not fall, those cutting 8 or more hours improved by 0.271 points (effect size 0.339), those cutting 5 to 7 hours by 0.203 (effect size 0.254), and those cutting 1 to 4 hours by 0.098 (effect size 0.123), the first two at p < 0.001 and the third at p < 0.01 (Fan et al., 2025). A dose-response gradient like that is normally read as a sign the exposure itself matters.

At company level the gradient disappears. Pairwise comparisons across companies cutting 8 or more, 5 to 7, and 1 to 4 hours were not statistically significant, and all three groups beat controls by comparable margins (Fan et al., 2025). Two readings survive that pattern: either hours are the active ingredient and individual data measure the dose more precisely, or a large share of the benefit comes from being inside an organization that publicly reorganized its work, whatever the arithmetic on hours. The study cannot separate those, and a fair summary should not pretend it did.

The mediation analysis narrows the mechanism without resolving that question. Work ability rose from 7.04 to 7.83 on a 0-10 scale, sleep problems fell from 2.36 to 2.02 on a 1-4 scale, fatigue fell from 2.68 to 2.26 on a 1-5 scale, and exercise frequency rose from 2.40 to 2.69, all at p < 0.001. Perceived job demands edged up from 3.51 to 3.59 and schedule control from 3.62 to 3.72 (Fan et al., 2025). In the decomposition of company-level hour reductions, individual fatigue accounted for 15.4% to 32.9% of the relationship with wellbeing and work ability for 10.7% to 30.3%, with health behaviors explaining more than work experiences in most outcomes.

Take-up was uneven, which matters because the individual-level gain scaled with hours actually shed. Women, non-white respondents, parents of school-aged children, Australian and New Zealand respondents, and employees at larger companies or in civil, social, and other services were less likely to reduce their hours, while supervisors appeared to gain more from a given reduction (Fan et al., 2025). At 12 months hours were still down, 33.91 against a 38.97 baseline at company level and 33.47 against 38.52 at employee level, and every wellbeing outcome remained significantly above baseline, with slight hedonic adaptation appearing only for job satisfaction. That durability is the authors' argument against a novelty effect, though nobody was blinded and participants knew they were in a widely publicized trial.

The sample is knowledge work in high-income English-speaking economies. Respondents were 65% women, 73% college-educated, 28% supervisors, 46% in professional services and 34% in civil, social, and other services, 29% at nonprofits, 34% fully remote and 48% hybrid (Fan et al., 2025). Where respondents lived spreads slightly wider than where the cohorts ran: 40% in the UK or Ireland, 32% in the US or Canada, 14% in Australia or New Zealand, and 14% in other countries, employees of participating organizations who live outside the country their cohort was organized in. The authors call the sample disproportionately high-income and Anglophone with many small organizations, and say this restricts how far the results generalize.

The nearest evidence for shift-based operations is older and weaker. Bambra et al. (2008) reviewed compressed working week interventions, identifying 40 observational studies from 27 electronic databases plus websites, bibliographies, and expert contacts. Most measured only self-reported outcomes, and the authors judged the methodological quality not very high. Among the five prospective studies that included a control group there were no detrimental effects on self-reported health, and work-life balance was generally improved.

That review reports no pooled effect size, odds ratio, or meta-analytic estimate, so there is no number to move into a business case from it. It also found no study reporting differential impacts by socioeconomic group, though most of the included populations were fairly homogeneous (Bambra et al., 2008). A compressed working week redistributes the same hours into fewer days, which is a different intervention from cutting hours while holding pay constant, so the two literatures answer different questions and neither covers coverage-driven frontline operations well.

One disclosure belongs with the newer paper. The trial research was supported in part by National Science Foundation Award #2241840, and the acknowledgments thank the founders and staff of 4 Day Week Global, Four Day Week Campaign UK, and Autonomy, the advocacy organizations that ran the trials and supplied the sample. That does not invalidate the analysis, and the authors are unusually plain about their own limits. It does mean the companies reached the researchers through groups with a public position on the answer, which is part of why the authors flag their own estimates as probably too high.