How much of lifespan is heritable—and why do estimates differ so much?

01The denominator is variation, not years

In a simple additive model, narrow-sense heritability is additive genetic variance divided by total phenotypic variance. Broad-sense heritability also includes nonadditive genetic components such as dominance. The question concerns differences between people. An 80-year life is not divided into 40 genetic years and 40 environmental years. Genetic–environmental covariance and interactions require additional assumptions for variance decomposition. s1

Consider an illustrative calculation: genetic variance 144 years² and environmental variance 133.3 years² give 51.9% heritability. Shifting everyone's lifespan upward by 10 years leaves both variances and heritability unchanged. Reducing environmental variance to 33.3 years² instead raises heritability to 81.2%. These chosen values estimate no intervention effect. They show that changing a population mean and explaining existing variation are different questions.

Surviving to 100, age at death, and dying from infection before a specified age are also different phenotypes. Familial clustering of exceptional longevity cannot directly supply the genetic fraction of ordinary lifespan variance. A 2006 Nordic twin study found stronger genetic associations with survival at older ages, but its abstract reports age-specific regressions and recurrence risks, not a universal heritability. s13

02What makes relatives similar?

Twin studies compare lifespan correlations in monozygotic and dizygotic pairs. A common approximation is h² = 2 × (MZ correlation − DZ correlation). Interpretation requires assumptions about genetic sharing and comparable environmental similarity. Standardizing total variance to 1, let A denote additive genetic variance, D dominance variance and C shared-environment variance. A simplified model gives r_MZ = A + D + C and r_DZ = 0.5A + 0.25D + C. The formula then estimates A + 1.5D, rather than narrow-sense A or broad-sense A + D. Unequal environments, assortment and selection can complicate it further.

The often-cited 1996 Danish study included 2,872 twin pairs, both surviving to 15, excluding the small number of pairs with a member still alive. To address skewness, it fitted variance models to squared age at death. Its preferred DE model reported 26% for men and 23% for women; an AE model gave approximately 23% and 20%. These values depend on phenotype scale and variance model: they are not simply outputs of a later correlation-difference formula. s2

Pedigrees face another problem. Relatives share resources, geography and habits as well as genes. Partners may resemble each other in these traits before forming a household. Complete lifespan is observed after death, so assortment concerning lifespan generally means selection on related traits, not foreknowledge of a partner's death age.

Ruby and colleagues’ 2018 study found substantial lifespan correlations among in-laws. These relationships ordinarily lack close genetic relatedness, challenging an exclusively genetic reading of familial similarity. The authors combined genetic and transmitted sociocultural factors in a latent variable, using blood-relative and in-law correlations to estimate assortment and transferable variance, t². This is not purely genetic heritability. s3 s4

Our recalculation retained all sex combinations for first cousins and first cousins once removed. Across birth decades 1800–1910, average nominal estimates were 20.3% and 26.3%; the authors' assortment-adjusted transferable-variance estimates averaged 4.63% and 2.69%. These are descriptive averages from the same summaries under different assumptions. Subgroups share pedigrees and are not independent replications; we derived no new confidence intervals.

When in-law correlation approaches blood-relative correlation, attributing all familial similarity to unadjusted genetic transmission becomes harder. The below 10% finding nevertheless depends on the latent structure, transmission coefficients, relationship identification and selection. Missing death years cannot simply mean “still alive”; requiring recorded children also selects the population. Treating t² as an upper bound on genetic contribution requires the model's variance and covariance conditions—it is not a universal inequality for every genetic–environmental structure.

The three familiar numbers can be aligned as follows. Their differences matter, but do not certify the assumptions of any model.

Sources: s2 s3 s5.

03In what world is the 2026 estimate calculated?

Shenhar and colleagues did not directly measure 55% genetic variance in a contemporary population. They fitted historical mortality patterns, assigned identical individual parameters to MZ twins and parameter correlation 0.5 to DZ twins, and adjusted parameter dispersion to reproduce twin correlations. They then set extrinsic mortality to zero while keeping other parameters fixed and standardized minimum inclusion age to 15. They called this counterfactual quantity heritability of intrinsic lifespan, HIL. s5

The useful intuition is that independent early external deaths can truncate differences between genetic groups and add variation within them, masking lifespan correlations. Reducing this noise can raise heritability without changing the genetic structure.

In the published Table 1, zero-extrinsic-mortality estimates with minimum age 15 were 57%/47% for Denmark, 56%/56% for Sweden and 53%/47% for SATSA in the SR/MGG models. The plotted intervals are 67%, not 95%. Moving SATSA's minimum age from 61 to 15 changes the target as well. US centenarian-sibling results concern selected longevity families and differ across the models; they do not directly validate a general-population value. The two models also share cohorts and central genetic assumptions.

Twins reared apart help examine shared childhood environments. Later SATSA birth cohorts also showed higher correlation differences. However, birth cohorts simultaneously change medicine, living conditions, records and survival selection. This alone does not isolate the causal effect of extrinsic mortality. Agreement within one standard error in small samples does not establish that environmental effects vanish.

The study offers a valuable explanation, but 55% remains a conditional model result, not an established universal genetic fraction.

04External causes can operate through inherited susceptibility

An infection originating outside the body need not have genetically independent mortality risk. The abstract of a 1988 Danish adoption study reports that a biological parent's infection death before 50 was associated with infection death in the adoptee between 16 and 58: RR 5.81, 95% CI 2.47–13.7. This is familial risk, not a lifespan variance ratio. A 2010 Danish twin study estimated liability-scale heritability of 40% (12%–50%) or 19% (3%–35%) under different infection definitions. The narrowest definition had too few events for such modeling. Misclassified causes, underlying disease and shared exposure remain limitations. s8 s9

This challenges treating all extrinsic mortality as genetically unrelated noise. It does not justify inserting infection heritability 40% directly into an all-cause lifespan model. Exposure, acquiring infection and dying after infection are distinct stages.

Two 2026 preprints raise different statistical challenges. Hamilton emphasizes infection genetics and survival selection; Kornilov examines whether calibration absorbs omitted heritable external susceptibility into an “intrinsic” parameter. The latter's bias numbers arise from an assumed data-generating process and cannot be subtracted from 55% as an empirical correction. s10 s11

Critiques require scrutiny too. Removing people who died externally and setting external mortality to zero in a model are different operations. A simulation of the former does not directly prove the same selection error in the latter. Adding conditional cause-specific hazards does not itself establish biological or genetic independence. The published paper also examined rising late-life external hazards linked to individual intrinsic-hazard slopes, so describing it as considering only a fixed risk in young adults is incomplete. s5

05Why matching twin correlations may not identify the intrinsic component

We tested the inferential step with a deliberately simple model. Two equally frequent genetic groups have intrinsic death ages centered at 68 and 92, each with uniformly distributed variation of ±20 years. Their intrinsic variance ratio remains 51.9%. MZ twins share the group label; DZ labels have correlation 0.5. Conditional on the labels, random death processes are independent between twins. These settings make DZ lifespan correlation exactly half the MZ correlation in this example, not necessarily in realistic polygenic inheritance.

External death begins at 15, with average unselected hazard 0.004/year. Parameter k describes the genetic pattern in external risk: at k=0.4, hazards are approximately 0.00552/year in the shorter-lived group and 0.00248/year in the longer-lived group, whose hazard is about 45% of the other's. Negative k reverses the direction. These hazards are chosen, not converted from infection liability heritability of 40%. If the longer-lived group also has lower external risk, a model that wrongly assumes identical external hazards needs a larger intrinsic difference to match the MZ correlation.

At the prespecified illustrative setting k=0.4, observed Falconer heritability is 29.0%. The misspecified model matches both MZ and DZ correlations, yet extrapolates to 68.7% after removing its external hazard, compared with the true illustrative 51.9%. Reversing the external genetic relationship, k=−0.4, yields 32.6%. When external risk is actually constant, 51.9% is recovered.

The example shows that two correlations alone do not uniquely identify the intrinsic component. It does not match all mortality curves, fit a human cohort, or estimate the actual bias in the 2026 paper. Its two genetic states do not represent narrow-sense variance in realistic polygenic inheritance. Cause-of-death information, joint twin survival surfaces and age-specific correlations can help distinguish the alternatives.

06What can be concluded, and what would distinguish the models?

A valid comparison must align population, birth cohort, minimum survival requirements, death outcome, transformed scale, genetic-variance definition and any intervention counterfactual. Different targets explain part of the numerical disagreement; they do not certify every model as correct.

Molecular genetics provides another line of evidence. A small number of significant loci does not, by itself, rule out high heritability. Timmers and colleagues’ 2019 study used offspring genotypes and around one million parental lifespans; locus discovery depends on proxy genotypes, power, small effects, censoring and cohort conditions. Variance explained by significant loci, all measured common variants, twin heritability and HIL are distinct quantities. s12

The most discriminating next step is to obtain data, supplements and code aligned with the final 2026 article, then examine external susceptibility genetics, intrinsic–extrinsic correlations, joint survival fit and age cutoffs within the same historical cohorts before transporting findings to modern populations. The causal links between immunity and lifespan require original evidence, beyond lists of shared pathways.

Available evidence supports genetic differences in lifespan and shows that environment and selection affect estimates. It does not provide an individual's fixed genetic percentage, a number of achievable extra years, increased healthspan, or the benefit of a particular intervention.

Materials and calculation limits

We recalculated the 2018 public summaries, without obtaining original pedigrees or individual records. We checked the 2026 published main text; its final supplement was inaccessible. The 2025 preprint and supplement were treated as older versions. The fixed Zenodo package retains SATSA minimum-age 50 metadata, compared with 61 in the published article. Its Danish/Swedish combined correlations give 27.2%/37.4% under the simple difference formula, versus 23%/35% in published Table 1. These discrepancies remain unresolved. We display the final article's printed estimates with explicit versioning and do not claim reproduction of the final 55%. s6 s7 s14

The public code, aggregate inputs and figure reproduction package contains our calculations and chosen model. It excludes original pedigrees, restricted twin records, third-party originals and authors' simulation caches. We did not rerun the critiques' complete simulations or treat their numbers as estimates from human data.