Heritability is a population-level quantity in genetics that describes how much of the observed variation in a characteristic is statistically attributable to genetic differences among individuals. It applies to a specified phenotype, population, set of environments, and measurement procedure. It does not measure how much of an individual’s characteristic is “caused by genes,” nor does it indicate whether that characteristic can be changed.
The concept is central to animal and plant breeding, evolutionary biology, and the study of human variation. Different forms of heritability distinguish the contribution of all genetic differences from the contribution of genetic effects that predictably pass from parents to offspring. Estimates come from breeding experiments, comparisons among relatives, and analyses of measured genomic variation. Their interpretation requires attention to environmental variation, relatedness, statistical assumptions, and the distinction between explaining population differences and predicting individual outcomes.
Conceptual origins and historical development
The problem that heritability addresses predates the term: relatives resemble one another, but the extent of that resemblance varies among characteristics and circumstances. Nineteenth-century studies of inheritance encountered both sharply differentiated traits and continuously varying measurements such as height. Gregor Mendel demonstrated particulate inheritance through experiments published in 1866, while Francis Galton investigated resemblance among relatives and developed the idea of regression toward the mean. Neither family resemblance nor regression alone, however, separates inherited effects from shared living conditions.
A decisive theoretical connection appeared in Ronald A. Fisher’s 1918 paper on correlations between relatives under Mendelian inheritance. Fisher showed how many inherited factors, each following Mendelian rules, could produce continuously distributed traits. His analysis helped reconcile Mendelian genetics with biometric studies of variation. Rather than treating continuous inheritance as an alternative to particulate inheritance, the resulting framework explained it through the combined effects of many loci and environmental influences.
During the 1930s and 1940s, researchers including Sewall Wright and Jay L. Lush developed and applied the statistical framework that became quantitative genetics. Lush’s work on livestock breeding made the distinction between transmissible genetic merit and other sources of resemblance especially important. Breeders needed not merely to identify outstanding animals, but to predict whether their offspring would also perform well. A high-performing animal could owe part of its superiority to favorable feeding, dominance effects, or combinations of genes that would not be reproduced intact in its descendants.
Quantitative genetics therefore developed around the analysis of differences within populations. Its central quantities are means, variances, covariances, and expected responses to selection. This approach differs from identifying a particular gene responsible for a recognizable inherited condition. A trait can involve thousands of genetic variants yet still have a well-defined heritability under a statistical model. Conversely, a known genetic mechanism does not by itself establish a population-level heritability value.
The historical connection between heredity research and eugenics also matters to interpretation. Genetic claims were used to support coercive social policies, often by confusing familial resemblance with genetic causation or treating population statistics as fixed properties of individuals and social groups. Modern quantitative genetics distinguishes these questions explicitly: estimating variation within a sampled population does not establish the causes of differences between populations, determine individual worth, or supply a scientific justification for a political policy.
Variance components and the main definitions
Heritability is usually expressed as a ratio of variances. For a quantitative phenotype , a simple model writes its deviation from the population mean as the sum of a genetic component and an environmental component :
Here “environmental” is a statistical category, not merely a synonym for climate, upbringing, or external surroundings. Depending on the model, it can include nutrition, exposures, developmental variation, unmeasured biological processes, and measurement error. The genetic component describes differences associated with inherited genetic constitution in the population being analyzed.
If genetic and environmental components are uncorrelated, phenotypic variance is
Broad-sense heritability is then
This quantity includes all genetic contributions represented in . It is often especially relevant to experiments with genetically replicated organisms, such as inbred lines or clonally propagated plants. If genetically identical replicates differ in phenotype, those differences help estimate non-genetic variation under the experimental conditions. Such experiments still require suitable replication, allocation to environments, and control of spatial or temporal effects.
Genetic variance can be partitioned further:
where is additive genetic variance, is dominance variance, and represents interaction variance among loci. Dominance concerns interactions between alleles at the same locus; epistasis concerns interactions involving different loci. Narrow-sense heritability is
The additive component represents variation in breeding values: the sum of average effects of inherited alleles in the population. “Additive” does not require every molecular process to operate independently. Average effects are statistical quantities that depend on allele frequencies and the genetic backgrounds in which alleles occur. Biological interactions can therefore contribute to additive variance.
The distinction between and is essential because total genetic resemblance is not identical to predictable transmission through sexual reproduction. Recombination and segregation alter combinations of alleles, whereas average additive effects provide the standard basis for predicting short-term response to selection. Under a valid nonnegative variance-component model, both ratios lie between zero and one. Sampling procedures can nevertheless produce estimates outside those limits if estimation is unconstrained; such results indicate uncertainty or model inadequacy, not biologically negative inheritance.
More elaborate models add maternal effects, permanent environmental effects, sex-specific effects, or other components. These must be defined before a numerical estimate can be interpreted. A reported “heritability of 0.5” is incomplete without specifying which genetic variance is in the numerator, what phenotype was measured, and which sources of variation are represented in the denominator.
Dependence on populations, environments, and measurement
Heritability is not an intrinsic constant of a trait. It changes when either genetic variance or non-genetic variance changes. Suppose one population has additive genetic variance of and environmental variance of , with other components absent. Its narrow-sense heritability is . If environmental variance falls to while additive variance remains , heritability becomes . No increase in the biological importance of genes is required; the denominator has changed.
Environmental uniformity therefore often increases heritability, while environmental heterogeneity can reduce it. The reverse can occur if an environmental change exposes previously unexpressed genetic differences. Genetic variation also changes through selection, migration, inbreeding, and genetic drift. Two populations exposed to identical conditions may consequently have different heritabilities because their allele frequencies differ. Even within one population, age, sex, season, and developmental stage can alter the relevant variance components.
A more general variance expression includes covariance between genotype and environment:
Gene–environment correlation occurs when genetic differences are associated with differences in environmental exposure. Parents may transmit alleles while also providing environments correlated with their own genetically influenced characteristics. Individuals can also evoke responses from others or select environments partly in relation to their characteristics. These processes complicate a simple separation into independently acting genetic and environmental causes. Whether an estimated genetic component includes such mediated pathways depends on the design and estimand.
Gene–environment interaction is different: the effect associated with a genotype depends on environmental conditions. Crop varieties may differ in yield under irrigation but perform similarly under drought, or their rankings may reverse. A model can represent this by adding a term, but its interpretation depends on how environments are sampled and measured. Interaction can change genetic variance across environments and limit the transfer of heritability estimates from an experimental station to farmers’ fields.
Measurement also matters. Random measurement error usually enlarges residual variance and lowers estimated heritability. Averaging repeated observations can reduce that contribution, making the heritability of an individual’s mean performance higher than that of a single observation. Changing scales can also change the answer: heritability for raw concentration need not equal heritability for its logarithm. Ratios such as body mass index and thresholds such as a disease diagnosis are statistically different phenotypes from their component measurements.
These dependencies explain why a large heritability can coexist with a substantial environmental shift in the population mean. Average human height increased in many populations during the twentieth century alongside changes in nutrition, disease burden, and living conditions. Such changes are compatible with substantial genetic contributions to height differences among people within particular cohorts. A statistic describing variation around a mean does not, by itself, identify the causes of changes in that mean.
Estimation from relatives, twins, and breeding designs
Traditional estimates use the expected genetic resemblance of relatives. Under an additive model with random mating, no relevant shared-environment covariance, and other simplifying assumptions, the covariance between a parent and offspring is . The regression slope of offspring phenotype on one parent’s phenotype is therefore . For offspring regressed on the mid-parent value—the average of both parents—the corresponding slope is . These are model-dependent results, not universal properties of family data.
Different relatives provide different information. Paternal half-siblings share, on average, one quarter of their additive genetic variation, so their expected covariance is under the standard assumptions. Full siblings share an expected one half of additive effects and one quarter of dominance effects; they may also share prenatal, household, or rearing environments. Controlled breeding designs distribute offspring across parents and environments to separate these sources. In plants, replicated genotypes, crosses, and multigeneration pedigrees provide additional experimental options.
A twin study compares monozygotic twins, who share nearly all inherited nuclear DNA sequence variation, with dizygotic twins, whose expected additive genetic relatedness is the same as that of ordinary full siblings. In a simple ACE model, phenotypic variance is partitioned into additive genetic effects , shared environmental effects , and nonshared environmental effects . If variance is standardized to one, expected twin correlations are
The familiar approximation follows from these assumptions.
Twin estimates depend on the equal-environments assumption: environmental influences relevant to the phenotype must not create greater similarity in monozygotic twins in a way misattributed to genetic effects. Other complications include assortative mating, prenatal differences, interaction between siblings, and non-additive genetic effects. With only the standard monozygotic and dizygotic comparisons, shared environmental and dominance components generally cannot both be estimated freely without additional information. The nonshared environmental component also includes measurement error; it is not a direct inventory of individually experienced events.
Adoption studies and extended pedigrees provide complementary contrasts, but adoption placement, prenatal influences, and selective participation require attention. Modern pedigree analyses often use an “animal model,” a statistical mixed model in which expected additive covariance is determined by a relationship matrix. Such models can accommodate repeated records, multiple generations, maternal effects, and common rearing environments. Their usefulness depends on whether the pedigree and sampling design contain enough independent information to distinguish the proposed components. More observations do not automatically resolve confounding if genetic relatedness remains closely aligned with environmental similarity.
Estimation from measured genomic variation
Genomic methods estimate genetic contributions using observed variation across the genome, rather than relying only on recorded pedigrees. A common approach constructs a genomic relationship matrix from many single-nucleotide polymorphisms (SNPs). Individuals who are nominally unrelated nevertheless differ slightly in their genomic similarity. If pairs with greater genomic similarity also tend to have more similar phenotypes, a mixed model can estimate the variance associated with the measured markers. In 2010, Jian Yang and colleagues demonstrated this approach for human height using common SNP data.
The resulting quantity is usually called SNP heritability. It measures the phenotypic variance attributable, under the model, to additive effects captured by the analyzed variants. It is not necessarily the same as total narrow-sense heritability. Genotyping arrays may poorly capture rare variants, structural variation, or variants weakly correlated with the typed markers. Linkage disequilibrium allows a measured marker to capture effects of nearby unmeasured variants, but this tagging is incomplete and varies across genomic regions and populations.
Summary statistics from a genome-wide association study can also be used. LD score regression, introduced in 2015 by Brendan Bulik-Sullivan and colleagues, relates association test statistics to the amount of variation each marker tags. Under its assumptions, the relationship helps estimate SNP heritability and assess contributions from confounding. Related methods estimate shared genetic effects across traits. Their performance depends on features such as sample composition, genetic architecture, and the suitability of the linkage-disequilibrium reference data.
The term missing heritability describes discrepancies between heritability inferred from some family designs and the variance accounted for by particular genomic analyses. An important distinction is whether the comparison concerns only individually significant associated variants or all variants modeled jointly. Thousands of effects too small to pass a discovery threshold can collectively explain appreciable variance. Other possible contributors include imperfect tagging, rare variants, non-additive effects, differences in phenotype definition, and biases in either family-based or genomic estimates. The discrepancy is not evidence for a single missing biological mechanism.
Population structure, relatedness, and family environments remain relevant even when DNA is measured directly. Genetic ancestry can correlate with geography and environmental conditions, creating associations that are not direct effects of an individual’s variants. Within-family analyses exploit differences in inherited alleles among relatives and can reduce some forms of confounding. They may also distinguish direct effects from indirect parental genetic effects, sometimes called genetic nurture, in which parental genotype influences offspring outcomes through the environment. Such analyses often have lower statistical power and do not eliminate every assumption; they answer a more specific question than a conventional between-family analysis.
Selection response and practical breeding
Narrow-sense heritability is especially useful because it connects observed selection to expected change across generations. In the simplest breeder’s equation,
is the selection differential—the difference between the mean phenotype of selected parents and that of the original population—and is the expected difference between the offspring-generation mean and the original mean, measured under comparable conditions. The equation describes a short-term expectation under assumptions including an approximately linear relationship between phenotype and breeding value.
For example, if a livestock population has mean body mass units and the selected parents average , then . With , the expected response is units. The predicted offspring mean is therefore , not . Environmental advantages and nontransmissible genetic combinations can make selected parents outperform what their offspring inherit on average. If males and females undergo different selection, their contributions must be combined appropriately rather than treating the selected mean as a single undifferentiated quantity.
Breeding programs use information beyond an individual’s phenotype. Records from parents, siblings, offspring, repeated measurements, and genomic markers can improve estimates of breeding value. Best linear unbiased prediction, widely associated with Charles Henderson’s work, combines records and relationship information in mixed models. Genomic selection extends this logic to genome-wide markers and can estimate breeding values before a trait is directly measured. This is particularly useful for traits expressed late in life, measured only in one sex, or costly to evaluate.
Heritability alone does not determine achievable genetic gain. Selection intensity, the accuracy of breeding-value estimates, additive genetic standard deviation, and generation interval all matter. A trait with low heritability can respond to selection when reliable information is accumulated from many relatives or repeated observations. Conversely, high heritability does not guarantee rapid progress if additive variation is small in absolute units, reproduction is slow, or selection is constrained by other traits. A dimensionless variance ratio cannot replace the scale of variation relevant to a breeding objective.
Long-term response may depart substantially from the initial breeder’s equation. Selection changes allele frequencies and can generate associations among loci, while inbreeding and drift can reduce variation. Changes in management can alter environmental and genetic variances. Selection for one trait may also produce correlated changes in fertility, survival, or disease resistance. Breeding programs therefore use multiple-trait evaluations and explicit constraints rather than assuming that increasing a single highly heritable measurement necessarily improves overall performance.
In wild populations, the connection with natural selection is additionally complicated by environmental covariance between a trait and reproductive success. An observed association with fitness need not imply that selection acts directly on that trait, and environmental trends can obscure a genetic response. Pedigree or genomic analyses can help separate these processes, but heritability by itself does not demonstrate that a population is undergoing evolutionary change.
Multiple traits, development, and disease liability
Many applications require genetic covariances rather than a separate heritability for each trait. A genetic correlation measures how strongly genetic effects on two traits covary:
Such covariance can arise because variants influence more than one phenotype, through pleiotropy, or because alleles affecting different traits are associated through linkage disequilibrium. A genetic correlation is not necessarily evidence that one trait causes the other.
For several traits, the multivariate response to selection is often written
where is the additive genetic variance–covariance matrix and contains selection gradients. This formulation shows why a trait may change even when selection does not act directly on it: selection on a genetically correlated trait can produce an indirect response. It also explains constraints. Selection favoring two changes may achieve little if the available genetic variation lies mostly along a different combination of traits.
Development can be treated similarly. Height at age five and height in adulthood are related but distinct phenotypes, with potentially different heritabilities and genetic correlations. A rising heritability with age does not necessarily mean environmental influences disappear; it may reflect accumulating genetic variance, changing exposures, different measurement reliability, or genetically correlated selection of environments. Longitudinal models can distinguish stable genetic influences from age-specific effects, although their conclusions depend on measurement schedules and the covariance structure assumed.
Binary outcomes require additional care. A disease diagnosis is coded as present or absent, but researchers often analyze it using a liability-threshold model. This assumes an unobserved continuous liability, conventionally normally distributed, with diagnosis occurring above a threshold. Genetic and environmental components are assigned to liability rather than directly to the binary observation. Liability-scale heritability therefore describes a model of variation in susceptibility; it is not the probability that an affected person’s condition was genetically caused.
Conversion between observed-scale and liability-scale heritability depends on the assumed population prevalence and, in case–control studies, the sampling fractions. Two studies can report different observed-scale values despite being compatible on a common liability scale. Uncertain prevalence, heterogeneous diagnostic criteria, and differences in age at observation complicate comparisons. The model is useful for connecting binary records to quantitative-genetic theory, but its latent scale is an assumption, not a directly measured biological quantity.
Heritability also differs from penetrance, familial recurrence risk, and an individual’s predicted probability of disease. Penetrance concerns the probability of a phenotype conditional on a specified genotype; recurrence risk concerns occurrence among relatives; heritability concerns population variance. A trait can have substantial heritability while currently measured genetic predictors remain insufficiently accurate for individual prediction.
Interpretation, limitations, and responsible comparison
The most common interpretive error is to treat heritability as a division of an individual into genetic and environmental percentages. A value of does not mean that 70 percent of a person’s height, behavior, or disease susceptibility is genetic. It describes variation among individuals under a specified model. Individual development involves interacting biological and environmental processes that cannot generally be divided into portions corresponding to that variance ratio.
A second error is to equate high heritability with immutability. Phenylketonuria illustrates the distinction between genetic etiology and environmental modification: variants affecting phenylalanine metabolism can produce a condition whose developmental consequences are substantially altered by dietary management. More generally, a factor shared by nearly everyone can be causally essential yet explain little observed variation. If all organisms receive adequate water, differences in water supply may contribute little to variance even though water remains indispensable for development. Variance explanation and intervention effects are different questions in causal inference.
Heritability within populations also does not identify the cause of differences between population means. Two groups could each have substantial heritability while their mean difference arose entirely from environmental conditions. Alternatively, genetic differences could contribute, but that requires separate evidence and an appropriate design. Neither a within-group heritability estimate nor a comparison of such estimates resolves the question. This limitation is particularly important when social classifications are mistaken for biologically homogeneous genetic populations.
Prediction requires another distinction. A polygenic score combines estimated effects of measured variants, but its predictive performance depends on the discovery sample, effect-estimation precision, linkage disequilibrium, allele frequencies, phenotype definition, and target environment. High heritability does not ensure that a particular score captures most relevant variation. Prediction can deteriorate when transferred between populations or cohorts. Indirect genetic effects and population structure can also make between-family prediction differ from prediction among siblings.
Uncertainty should accompany interpretation. Sample size, uneven relatedness, imprecise phenotyping, ascertainment, publication bias, and model misspecification can all affect estimates. Confidence intervals may be broad, particularly for small samples or weakly identified components. Estimates from twin studies, pedigrees, unrelated-individual genomic models, and within-family analyses need not agree because they may target different combinations of direct, indirect, additive, and tagged genetic effects.
A scientifically interpretable report therefore specifies the population and sampling frame, phenotype and scale, age and environmental setting, statistical model, genetic component estimated, and uncertainty. Comparisons are most informative when these features align. A numerical difference between two published heritabilities can otherwise reflect differences in measurement or design rather than a biological change in inheritance.