An ML study of 50,188 dogs in the Dog Aging Project — then a rigorous self-audit that threw out everything that didn't survive honest, leak-free analysis. Here's what's left standing.
A dog's lifespan is dominated by two things you can't change — its age and its body size (smaller dogs live longer; it traces to the IGF1 gene). The one factor you can change that survives rigorous testing is physical activity. Supplements, diet fads, and "genetic longevity scores" did not hold up.
Findings that survived leak-free re-analysis, cross-fitting, and (for exercise) a longitudinal test.
The only modifiable factor that survived every test. Dogs whose activity increased over time had 2.7 percentage points lower subsequent mortality (24.7% vs 31.2%). Cross-fit causal estimate ≈ +1 month.
Honest, modest, and real — though some of it is "healthier dogs move more."
The single biggest factor — and it's genetic. Smaller dogs live years longer; the signal traces to the IGF1 region (validated out-of-sample). Not changeable, but it sets realistic expectations.
Judge a dog as "doing well for its size," not against other sizes.
A gradient-boosted model (XGBoost-AFT) gives well-calibrated survival curves (Brier skill +0.35 over baseline). Its discrimination is mostly age — useful for individual risk, not for ranking lifestyle tips.
Beat deeper neural models; honest about what it can and can't do.
An earlier version of this study (and this page) reported much bigger, flashier claims. A 19-finding self-audit showed most were artifacts. In the spirit of honest science, here's what fell:
Reality: ~98% of that is just age. Knowing a dog's age alone scores 0.831; all the lifestyle and genetic features together add about 0.01. The model is good at "old dogs die sooner," which we already knew.
It was circular — scored on the same dogs used to build it. Done honestly (held-out dogs), the polygenic score predicts lifespan no better than a coin flip. There is no usable DNA longevity test here.
Didn't replicate. In the full cohort the effects shrank from +200 days to ~+5 days (not significant), and vanished entirely under proper statistics. "Bone meal," the old #1 factor, was a survey-format artifact, not a real effect.
Overstated 5–14×. The honest causal estimate is ~+1 month (and −2.7 points of mortality longitudinally). Still real, still the best lever — just not magic.
Full audit (19 findings, before/after numbers, methods) is summarized in the Methods section. The corrections came from an automated, adversarial re-analysis of the original pipeline.
More-active dogs clearly live longer in the raw data (below). But raw gaps overstate cause — sicker dogs move less. After proper causal adjustment the effect is smaller, but it survives, including a longitudinal test where increasing activity predicted lower later mortality.
Raw association — not the causal effect. This gap is inflated by reverse causation.
The biggest single factor in canine lifespan — and it's largely genetic (IGF1). Smaller dogs simply live longer.
Median age at death by body-size class, Dog Aging Project cohort.
An honest, deliberately simple estimate: your breed's observed median lifespan, with the one adjustment the data supports — activity. No supplements, no genetics, no false precision.
Based on — dogs of this breed. The activity adjustment (±~1 month) reflects the honest causal estimate, not the inflated quintile gap. A statistical average, not a veterinary prediction for your individual dog.
Observed lifespan and mortality by breed (descriptive data — real, not modelled).
| Breed | Median age | Size | n | Mortality |
|---|
Data: Dog Aging Project 2024 curated open-access release (50,188 dogs, 15,056 deaths, 7,688 genotyped). Survival models: XGBoost accelerated-failure-time on a clean, leak-free feature set with native missing-value handling, evaluated by calibration (integrated Brier) and age-stratified C-index. Causal effects: cross-fit (double-ML) AIPW adjusting for confounders only; exercise additionally tested with a longitudinal landmark design on the follow-up survey. Genetics: censoring-aware GWAS and a kinship-aware polygenic score.
Why this page changed: the original analysis contained data leakage (a circular genetic score), single-imputation that understated uncertainty, survey-format artifacts, and effect sizes inflated by reverse causation. An adversarial 19-point self-audit corrected each one. Where a finding didn't survive, we removed it rather than soften it.
Observational study — associations, not proof. The exercise effect, though it survives a longitudinal test, can only be fully confirmed by a randomized activity trial. Not veterinary advice.
Analysis & self-audit by Nick Friesen · corrected edition 2026.