Does eating less really extend life?

01Our answer, up front

The best-supported statement today: in genetically diverse mice, caloric restriction (CR) extends lifespan proportionally to dose — the 40% CR group lived a median ~9 months longer than ad libitum (+36%), a conclusion we reproduced from the raw data. Yet in the same study, genetic background explained three times more lifespan variance (23.6%) than diet (7.4%). Intermittent fasting (1–2 fasting days per week, with little reduction in total calories) also extended mouse lifespan, but only in animals that were not heavy before the intervention. In humans, two years of moderate restriction (~12% achieved) genuinely improved metabolic markers and slowed one "pace of aging" biomarker by 2–3%, at the cost of bone density and lean mass. For human lifespan itself there is no direct evidence, and there will not be for a long time.

What each layer of evidence answers:

Scope: the state of evidence on "eating less (caloric restriction and intermittent fasting) and lifespan/aging." This is not dietary advice, does not address weight-loss therapy in people with obesity, and does not cover the separate evidence on time-restricted eating.

02Why this mouse study deserved a reanalysis

In October 2024, Di Francesco et al. published the DRiDO study in Nature: 960 Diversity Outbred (genetically heterogeneous) mice randomized to five arms — ad libitum (AL), fasting one day per week (1D), two consecutive days per week (2D), 20% CR, or 40% CR — starting at 6 months of age, with the 937 animals alive at intervention start followed until natural death. Original paper

It differs from textbook CR experiments in three ways: genetically diverse mice rather than inbred strains (closer to human population variation); intervention begun in midlife rather than at weaning; and over 200 health traits measured alongside lifespan. Its headline — genes matter more than diet — was widely covered.

The authors released individual-level data on GitHub (MIT license). We re-ran the analysis independently in R: first 17 consistency checks of the released data against the paper (sample sizes, group counts, censoring, the authors' saved trait–lifespan coefficients), all passing; then an independent survival reanalysis. Reproduction package

03Reanalysis 1: the dose gradient holds, matching the paper

Our recomputed median lifespans (Kaplan–Meier): ad libitum 765 days, 1-day fast 855, 2-day fast 846, 20% CR 904, 40% CR 1043 days. The 40% CR group lived 278 days longer than ad libitum — 9.1 months, +36.3%, exactly the paper's "about 9 months (36.3%)." Overall log-rank P=1.3×10⁻²³ (test statistic). Group estimates with CIs

Three statistical lenses agree. Restricted mean survival time (common horizon τ=41 months): +182 days for 40% CR (95%CI 131–234), +100 for 20% CR (51–148), +64 for 2-day fasting (14–112), +37 for 1-day fasting (−10–85, crossing zero). A Cox model stratified by birth generation gives hazard ratios vs AL: 0.33 (0.26–0.41) for 40% CR, 0.55 for 20% CR, 0.70 for 2D, 0.82 for 1D (0.67–1.00, marginal). RMST with bootstrap Cox results

One more layer is the rate of ageing itself: fitting Gompertz mortality models, the mortality doubling time is 150 days on ad libitum but ~230 days on 40% CR, 20% CR and 2-day fasting — the rate at which mortality accelerates was more than halved; the 1-day fast group (165 days) was not distinguishable from control. Gompertz fits

Two honest footnotes. First, the paper's ordering "40 > 20 > 2D > 1D > AL" comes out slightly differently in our medians — 1D (855 d) edges 2D (846 d), with heavily overlapping CIs; on the whole-curve RMST measure the paper's ordering is restored (2D 840 > 1D 813). The data cannot rank the two fasting arms. Second, the proportional-hazards test is significant (p=0.009): hazard ratios vary over time, so each Cox HR should be read as an average effect.

04Reanalysis 2: genes explain three times more than diet

The paper's most-quoted number: among mice alive at 6 months, genetic background explained 23.6% of lifespan variance (h²=0.236, 95%CI 0.106–0.360, kinship-matrix model), versus 7.4% for diet. Original paper

We could independently verify half of that. Our variance decomposition on the same data: diet explains 7.0% of lifespan variance, rising to 8.8% and 11.1% in mice surviving past 12 and 18 months — point-for-point agreement with the paper's 7.4%→8.4%→11.4%, including the direction "diet's share grows with age." Variance decomposition

The genetics half cannot be recomputed: the released data contain no per-mouse genotypes. Using birth generation as a crude proxy explains 0.4% — unsurprising, since DO genetic diversity sits at the individual level. So we flag "genes three-to-one over diet" as cited from the paper, not independently verified; its confidence interval is wide (11–36%), and the exact ratio should not be taken literally, though the direction (genetics larger than diet) is supported by the paper's model.

The practical reading: under identical housing, two mice differ in lifespan more because of their genomes than their bowls. That does not weaken the diet effect — it bounds its size.

05Reanalysis 3: body weight, dose, and why "it's just weight loss" fails

Splitting at median 6-month body weight (469 light, 468 heavy) and recomputing: CR extends lifespan significantly in both halves (20% and 40% CR vs ad libitum: light half p=1.8×10⁻⁴ and 3.2×10⁻⁷, heavy half p=3.9×10⁻⁵ and 7.4×10⁻¹⁵; median gains 115–292 days). Fasting's effect concentrates in the lighter half — light-half 2-day fasting p=0.006, while in heavier mice 1-day fasting p=0.21 and 2-day fasting p=0.065, both without statistical support. This matches the paper: fasting did not extend life in mice with high pre-intervention weight. Stratified reanalysis

A counterintuitive detail: among ad libitum mice, lighter animals at 6 months lived longer (840 vs 755 days, p=0.059); inside the 40% CR group, baseline weight no longer predicted lifespan at all (p=0.90) — severe restriction erased the head start of being naturally lean. Within-group contrasts

Most worth remembering is the paradox the paper itself stresses, which our reanalysis supports: CR groups lost weight and lived longer, yet within each group, the individuals that retained more weight lived longest. Weight retention across the 10–11-month phenotyping period ("PhenoDelta") was the strongest lifespan correlate among all 200+ traits — our Spearman ρ=0.256 (p=7×10⁻¹⁵) against the paper's adjusted r=0.287 — and weight loss at any age tracked shorter lifespan. The paper explicitly rejects "CR merely counteracts obesity": the benefit of eating less does not fit inside the word "slimming." Trait–lifespan associations

06The monkey fight: two 20-year trials, two answers

Only two long-term primate CR trials exist, superficially contradictory; both checked at primary source.

Wisconsin (WNPRC), reported 2009: adult-onset (7–14 y) 30% CR cut age-related mortality roughly threefold (HR=3.0, p=0.03; age-related deaths 37% vs 13%) and halved diabetes, cancer and cardiovascular disease — but all-cause mortality was not significantly different (p=0.16). Colman 2009 A 2014 update reported all-cause mortality reaching significance. Colman 2014

National Institute on Aging, reported 2012: the same 30% CR produced no survival benefit — age-related deaths 24% vs 20% (control vs CR) in young-onset animals, Cox p=0.975 on that measure (all-cause p=0.255); old-onset all-cause p=0.934. Metabolic measures improved; lifespan did not move. Mattison 2012

A 2017 joint analysis attributed the divergence to enumerable design differences: NIA's control diet was 3.9% sucrose versus 28.5% in Wisconsin's (Wisconsin controls were effectively junk-food-fed, with high baseline mortality); NIA controls were portion-fed rather than truly ad libitum (themselves mildly restricted); and ages of onset differed. Joint analysis Pooled, the primate direction still leans positive, but the magnitude is far less stable than in mice. One more within-species warning: back in 2010, about a third of 41 recombinant inbred mouse strains lived shorter lives under CR, and wild-caught mice showed no overall benefit. Liao 2010

07Humans: CALERIE's three-layer answer

Nearly all direct human evidence comes from CALERIE phase 2: of 238 eligible volunteers, 220 were randomized (145 CR, 75 control), prescribed 25% restriction for 2 years; 11.9% was actually achieved (19.5% in the first 6 months, ~9% thereafter), with 218 in the intention-to-treat analysis. Ravussin 2015 Three layers, stated separately.

Metabolism and weight — genuine improvement. CR lost 7.5 kg (maintaining −10.4%), including −2.2 kg of lean mass; LDL, total-cholesterol:HDL ratio, blood pressure, C-reactive protein and insulin sensitivity all improved, by more than weight loss alone would predict. Kraus 2019 Resting metabolic rate adjusted for body composition fell significantly at 12 months (p=0.04, suggesting "metabolic adaptation") but not at 24; core temperature did not differ; thyroid hormone T3 stayed lower. Ravussin 2015

Pace of aging — one small signal. Two years of CR slowed the DNA-methylation pace-of-aging measure DunedinPACE by 2–3% (the control group's pace was unchanged); the PhenoAge and GrimAge clocks showed no between-group difference over the same period. This is the only "slower pace of aging" signal ever seen in a human RCT, but it is ~2% per year on a surrogate marker — not lifespan evidence. Waziry 2023 Immunologically, CR preserved thymic function and pointed to a druggable inflammatory gene, PLA2G7, whose suppression reproduced part of the benefit in mice. Spadaro 2022

Safety — real costs. Bone mineral density fell significantly at lumbar spine, total hip and femoral neck (about −0.013 to −0.017 g/cm², between-group p≤0.03) — in young, non-obese adults; lean mass fell 2.2 kg while strength was preserved; overall adverse-event rates did not differ significantly (serious events 1 CR vs 7 control); dropout was higher on CR (18.2% vs 5.3%). Villareal 2016 Safety analysis

Lifespan — no direct evidence. No trial has randomized people to eat less and counted the years; none can, practically. Observational data offer an intriguing hint: in the WHI cohort, biomarker-calibrated lower energy intake was associated with lower all-cause mortality in younger postmenopausal women, with the direction reversed in the oldest group. Prentice 2024 Observational association is not causation, but it points the same cautious direction as the lean-mass and immune trade-offs seen in 40% CR mice.

08The four competing explanations, judged

09What we know, and what we don't

Known: in genetically diverse mice CR extends lifespan proportionally to dose (independently reproduced; +36% median at the highest dose); genetics outweighs diet; CR's benefit does not depend on baseline leanness while fasting's does; CR is not a synonym for weight loss; moderate restriction in humans improves metabolism, slightly slows a pace-of-aging marker, and costs bone; primate results lean positive but unstable.

Unknown: any human lifespan effect at any dose (no direct evidence); male mice (this study was all-female); the long-term human safety window of CR (the bone findings' downstream meaning is unsettled); any lifespan effect of fasting in humans (even the animal evidence is conditional); whether the "better metabolism → more years" bridge can be built in humans.

10Methods, reproduction and updates

This is an evidence audit of one primary question, combining an independent survival reanalysis of a public animal dataset with layered extraction of primary human and primate literature; it is not a systematic review and includes no new wet-lab work. Search cutoff 2026-09-18, covering the original paper and its public data, the CALERIE primary literature, both primate trials and their joint analysis; search strings and coverage gaps are in the research record.

Download the reproduction package: the pre-registered analysis plan (with decision rules), all R/Python code, input manifest with SHA256 hashes, and all output tables. Individual mouse data come from the authors' GitHub repository (MIT license) and are not redistributed per repository policy; the paper text is not redistributed. Known gaps: the authors' full Figshare archive was unreachable from our network on the search date (HTTP 403), so we anchored on the GitHub release cross-checked against paper-reported numbers; the release contains no sex field (the study was all-female by design) and no genotypes, so sex stratification and heritability cannot be independently recomputed — both flagged accordingly.

First published: 2026-09-18; next scheduled evidence review: 2027-03-18; earlier if CALERIE long-term follow-up, male DRiDO data, or a new large animal trial appears.

Authorship, computation, translation checking and editing were all performed by the same Kimi AI agent, with no independent agent, human, or clinical review. Per-source reading depth, conflicts and remaining limitations are in the record below.