Interpreting SNP Heritability Without Overclaiming

What an SNP Heritability Estimate Describes

An estimate such as “SNP heritability is 0.40” can sound like a statement about how much of a person’s trait comes from genes. It is not. SNP heritability describes a proportion of variation among individuals in a specified population, for a specified phenotype and measurement scale, under a model that uses a specified set of genetic markers. Its interpretation depends on those choices.

The term “SNP heritability” is often written \(h^2_{\mathrm{SNP}}\). In an additive model, it refers to the fraction of phenotypic variance attributed to the component captured by the analyzed SNPs. “Captured” matters: a marker can tag genetic variation through linkage disequilibrium even when it is not itself the causal variant. The estimate is therefore not automatically a measure of the variance caused by the typed SNPs alone.

For example, write a covariate-adjusted phenotype as \(y_i=g_i+e_i\), where \(g_i\) is the additive component captured by the SNPs and \(e_i\) is the remaining component. If the components have mean zero and are uncorrelated, then

This expression is a model-based variance ratio, not a universal property of the trait. In particular, “phenotypic variance” must be understood on the analysis scale and under the analysis’s treatment of covariates. For a binary phenotype, for instance, a proportion on the observed scale is not automatically the same as one on a liability scale; those scales require explicit distinction.

Worked Example: Interpreting a Variance Ratio

Population Variance Is Not Individual Determinism

Heritability is a comparison of variation across individuals. It does not partition one individual’s observed value into a genetic percentage and an environmental percentage. The model’s variance components describe how values differ across the study population, not the composition of a particular person’s phenotype.

A high value also does not mean that the trait is fixed or cannot respond to environmental change. Heritability can be high in one setting because people in that setting experience relatively similar environments, while the same trait could show lower heritability in a setting with more varied environmental exposures. The estimate concerns variation under the conditions represented in the analysis; it is not a measure of how much a trait can change.

Captured Variance Is Not the Same as Causal Attribution

An SNP can be statistically associated with a trait because it is correlated with a causal variant. In that case, it can help tag genetic variation without being the biological cause of the trait difference. More generally, an SNP heritability estimate summarizes a model’s captured variance; the estimate alone does not identify which variants are causal, which biological mechanism is involved, or whether an association is free of confounding.

The distinction can be seen in a simple hypothetical example. Let \(C\) be a standardized causal genotype and \(S\) a different standardized marker, with correlation 0.8. Suppose the phenotype follows \(y=2C+e\), where \(e\) is uncorrelated with both genotypes. The marginal regression coefficient of \(y\) on \(S\) is

Thus \(S\) has a nonzero association in this example even though the phenotype model assigns the direct effect to \(C\), not \(S\). The association illustrates tagging through correlation; it does not demonstrate that changing \(S\) would change the phenotype. Real analyses require appropriate design and assumptions to address confounding, and a heritability estimate by itself cannot supply a causal interpretation.

SNP Heritability Is Not Total Genetic Variance

SNP heritability is limited by what the analyzed markers and model capture. If the SNP set does not tag some genetic variation well, that variation need not contribute fully to the estimate. Depending on the study and model, relevant differences can arise from marker coverage, allele frequencies, linkage disequilibrium, the variants included, and the assumed effect structure. Genetic contributions not represented by the fitted additive component are not automatically included.

For that reason, \(h^2_{\mathrm{SNP}}\) should not be read as “all genetic influence.” It is also not automatically equal to narrow-sense heritability, which refers to the additive genetic contribution as defined for a population and model. The two quantities may differ because the SNP-based estimate reflects the genetic variation tagged by the analyzed markers under its estimation assumptions. Neither quantity is a fixed percentage that applies to every population or environment.

SNP Heritability Is Not Prediction Accuracy

A heritability estimate describes a variance ratio. Prediction accuracy describes how well a predictor, such as a polygenic score, predicts phenotypes for individuals. They are related in principle but are not interchangeable. Prediction depends not only on the amount of genetic variance available to capture, but also on the quality of the estimated effects, the sample size and design used to estimate them, the target population, and the phenotype measurement.

To make the distinction concrete, suppose a hypothetical phenotype has total variance 100 units squared and the SNP-captured component has variance 40 units squared. The resulting SNP heritability is 0.40. Now suppose an independently evaluated polygenic score has correlation 0.45 with the phenotype in the target sample. For a simple linear prediction, the squared correlation is the fraction of phenotype variance accounted for by the score’s linear predictions:

A Reliable Interpretation Checklist

Before interpreting an SNP heritability estimate, identify what the number refers to. The following sequence keeps the conclusion tied to the estimand rather than extending it beyond the analysis.

Common Overclaims

A frequent mistake is to write, “40% of this person’s trait is determined by their SNPs,” after observing \(h^2_{\mathrm{SNP}}=0.40\). That reasoning confuses variation across people with the makeup of one person’s phenotype. The correction is: “In this population, on this phenotype scale, the model attributes 40% of phenotypic variance to the additive component captured by the analyzed SNPs.”

Other conclusions that do not follow from the number alone include:

• “The associated SNPs caused the trait.” Association can reflect linkage disequilibrium or other sources of association; heritability does not establish causation.
• “This is the total genetic contribution.” The estimate concerns the component captured under the specified SNP set and model, not necessarily every genetic contribution.
• “A score will predict 40% of the trait.” Heritability and prediction accuracy are different quantities; score performance must be assessed directly.
• “The trait is genetically fixed.” A variance ratio does not measure whether the trait can respond to environmental change.
• “A higher estimate proves a stronger genetic effect in one population.” The ratio can change when total variance, tagging, measurement, or other population features change.