Trait Measurement Error And Genetic Correlation

Why Measurement Error Matters for Genetic Correlation

A genetic correlation describes the correlation between the additive genetic effects on two traits. It is not, by definition, the correlation between the recorded measurements of those traits. A recorded outcome can contain biological variation not explained by the additive genetic effect, as well as noise introduced by the measurement process. Keeping these sources distinct is essential: independent measurement noise can reduce precision without changing the genetic-correlation estimand, while correlated or incorrectly modeled errors can be confused with genetic covariance.

The consequences depend on where the error enters the data. Noise that is independent across individuals adds variance to each person's measurement but does not create covariance between relatives. Error shared across traits for the same person can contribute to their observed within-person covariance. Error correlated across individuals can also affect the covariance patterns used to estimate genetic components. Thus, an analyst must specify the error structure rather than assume that every non-genetic term is harmless.

A Two-Trait Model With Measurement Error

For individual \(i\) and trait \(t\), write the observed outcome as an additive genetic effect, a biological residual, and a measurement error, in addition to its mean. The biological residual represents variation in the phenotype not attributed to the modeled additive genetic effect; measurement error represents noise in recording or measuring the outcome.

Here, \(\mu_t\) is the mean for trait \(t\), \(A_{it}\) is the additive genetic effect, \(e_{it}\) is the biological residual, and \(M_{it}\) is the measurement error. For the covariance formulas below, assume the genetic effects have covariance \(\operatorname{Cov}(A_{i1},A_{j2})=K_{ij}\sigma_{A12}\), where \(K_{ij}\) is the genetic relatedness entry and \(\sigma_{A12}\) is the additive genetic covariance. Assume genetic effects are uncorrelated with biological residuals and with measurement errors. Assume measurement errors are uncorrelated with biological residuals. Also assume the biological residual covariance is \(\operatorname{Cov}(e_{i1},e_{j2})=\mathbf{1}(i=j)\sigma_{E12}\), and the measurement-error covariance is \(\operatorname{Cov}(M_{i1},M_{j2})=\mathbf{1}(i=j)\sigma_{M12}\). The indicator \(\mathbf{1}(i=j)\) equals 1 for the same individual and 0 otherwise. These assumptions specify that neither residual source has cross-individual covariance; the two sources may each have cross-trait covariance within an individual.

Under the stated assumptions, covariance between the observed outcomes is the sum of the genetic covariance pattern and the same-individual residual covariances. To see why, expand the covariance of the sums. The covariance of the genetic effects contributes \(K_{ij}\sigma_{A12}\); the biological residuals contribute \(\mathbf{1}(i=j)\sigma_{E12}\); and the measurement errors contribute \(\mathbf{1}(i=j)\sigma_{M12}\). All cross-covariances between these components are zero by assumption. Therefore:

The assumptions matter. If biological residuals have covariance across different individuals, their covariance must be added for those pairs. If genetic effects and biological residuals are correlated, cross-terms must also be included. Omitting such terms is not justified simply because the components have different names.

For a single trait, under analogous assumptions and with measurement errors independent across individuals, the observed variance is \(\sigma_{At}^2+\sigma_{Et}^2+\sigma_{Mt}^2\). Independent measurement error adds to the diagonal variance, while the covariance across distinct individuals remains \(K_{ij}\sigma_{A12}\) in the two-trait model above. This is the basic reason independent outcome noise usually makes genetic covariance harder to estimate rather than adding genetic covariance itself.

Reliability and the Observed Phenotype Correlation

Define the underlying phenotype before measurement error as \(T_{it}=\mu_t+A_{it}+e_{it}\), and suppose for this paragraph that measurement error is uncorrelated with \(T_{it}\). The reliability of a single measurement is the fraction of its observed variance attributable to variation in \(T_{it}\):

If measurement errors for the two traits are also uncorrelated with one another, then they add variance but no cross-trait covariance. Let \(\rho_T\) be the correlation between the underlying phenotypes. Because \(\operatorname{Cov}(Y_{i1},Y_{i2})=\operatorname{Cov}(T_{i1},T_{i2})\) under these assumptions, dividing by the observed standard deviations gives:

This derivation concerns the observed phenotypic correlation. It does not say that the genetic correlation is multiplied by \(\sqrt{R_1R_2}\). Under the model, independent measurement error does not alter \(\sigma_{A12}\), \(\sigma_{A1}^2\), or \(\sigma_{A2}^2\), and so it does not alter their ratio \(r_g\). It does increase outcome variance, however, so estimates based on finite data will generally be less precise. The exact change in standard error depends on the design, sample size, relatedness structure, and estimation procedure; there is no universal multiplier based only on reliability.

Correlated Errors and Genetic-Covariance Estimation

Errors correlated across traits within the same person contribute to observed within-person covariance. In a model that distinguishes the residual terms, this covariance is assigned to \(\sigma_{M12}\), not to \(\sigma_{A12}\). In contrast, genetic covariance produces the relatedness pattern \(K_{ij}\sigma_{A12}\), which varies across pairs according to their genetic relatedness. Estimation uses these different covariance patterns to separate components.

That separation depends on the model being appropriate. If correlated measurement error is omitted or incorrectly constrained, the model may not represent the covariance actually present in the data. The resulting estimates can be biased, and the direction and magnitude of bias depend on the design and the misspecification. In particular, the clean distinction between a diagonal residual term and a genetic-relatedness term relies on the stated absence of cross-individual error covariance. If an error source is correlated across individuals, it can generate off-diagonal covariance that may resemble or distort genetic covariance.

Repeated Measurements and Precision

When repeated measurements are available, averaging them can reduce independent measurement noise. Suppose each of \(m\) measurements has error variance \(\sigma_M^2\), and the errors are independent across repeats and independent of the underlying phenotype. The average error is the sum of the errors divided by \(m\); its variance is therefore \(m\sigma_M^2/m^2=\sigma_M^2/m\). The variance of the underlying phenotype is not divided by \(m\), because the repeated measurements concern the same underlying trait value.

Interpreting Estimates and Avoiding Common Errors

A reported genetic correlation should be interpreted as an estimate of shared additive genetic effects on the traits as modeled, along with its uncertainty. Measurement quality matters to that uncertainty and to whether the model can separate genetic covariance from other covariance. It does not license replacing the genetic correlation with a phenotypic correlation, or assuming that a correction based on phenotype reliability automatically recovers the genetic correlation.

• Mistake: “Reliability is 0.5, so the genetic correlation must be halved.” This applies an attenuation idea for observed phenotypic correlations to a different estimand. Under independent measurement error, genetic covariance and genetic variances are unchanged; precision is generally reduced.
• Mistake: “Any same-person cross-trait covariance is genetic.” Biological residual covariance and correlated measurement error can also contribute. The covariance model must include those terms when they are plausible.
• Mistake: “Measurement error is independent, so it cannot cause bias.” This is only a safe conclusion under a specified error structure and an appropriate fitted model. Correlation across individuals, correlation with modeled components, or omitted cross-trait error covariance can violate the assumptions.

In practice, document what counts as measurement error, whether repeated readings share error sources, and whether errors can be correlated across traits or individuals. Sensitivity analyses using plausible residual covariance structures can help assess whether the genetic-correlation estimate depends strongly on these modeling choices. The key is to separate changes in the target quantity from changes in the information available to estimate it.