Covariate Adjustment in Family-Based Association Tests

Why Adjust a Family-Based Quantitative-Trait Test?

A quantitative trait can vary with age, sex, treatment, or other measured characteristics even when a tested allele has no effect on the trait. If those covariates are not accounted for, trait variation unrelated to the genetic effect can obscure a transmission signal or make a score harder to interpret. Covariate adjustment aims to separate this predictable variation from the part of the trait used to assess association.

In a family-based test, the central comparison is still based on transmissions that could have occurred under the conditional null. Adjustment does not replace that comparison. Instead, a covariate model can provide trait residuals that weight the transmission score. The conditional test remains calibrated only when the weights and the transmission randomization satisfy the assumptions used to calculate its null distribution.

For example, with a binary coding \(G_{ij}\in\{0,1\}\) and a heterozygous parent whose transmission is equally likely to be either allele, \(E(G_{ij}\mid S_i)=1/2\). A positive score means that transmissions coded as 1 tend to occur with positive adjusted trait values, relative to the conditional transmission means. The score alone is not an effect estimate in trait units; its scale depends on the transmission coding and the residual weights.

Trait Models and Residualization

A simple starting model for a quantitative trait is \(Y_{ij}=\beta_0+\boldsymbol{\gamma}^{T}\mathbf{C}_{ij}+\varepsilon_{ij}\), where \(Y_{ij}\) is the trait, \(\mathbf{C}_{ij}\) is a vector of measured covariates, \(\beta_0\) is an intercept, \(\boldsymbol{\gamma}\) is a vector of covariate coefficients, and \(\varepsilon_{ij}\) is the part not predicted by that model. Under a linear model, the adjusted value is the residual \(R_{ij}=Y_{ij}-\widehat{\beta}_0-\widehat{\boldsymbol{\gamma}}^{T}\mathbf{C}_{ij}\). The genetic association is then tested using these residuals as weights in the family score.

This procedure has a clear interpretation when the covariate model is appropriate: the test asks whether transmission is associated with trait variation beyond the variation predicted by the specified covariates. The model need not be a claim that the covariates cause the trait. It is a statistical description used to form adjusted values. Covariates should be chosen and modeled with care; nonlinear relationships or interactions may require terms beyond a simple linear predictor.

A residualization step is not automatically equivalent to every possible joint model of phenotype and genotype. In particular, a fully specified model may account for nuisance-parameter estimation, family correlation in traits, or a genetic effect that varies by covariate. The basic residual score is a useful and common construction, but its calibration and interpretation depend on how the residuals were formed and on the conditional null being tested.

Conditional Mean and Variance

Let \(A_{ij}=G_{ij}-E(G_{ij}\mid S_i)\), and regard the residuals \(R_{ij}\) as fixed weights in the conditional calculation. If each centered transmission has conditional mean zero, then linearity of expectation gives a zero conditional mean score. For the variance, transmission dependence must be included: the general expression is the variance of the weighted sum, not a sum of variances unless the relevant covariance terms are zero.

Here \(S\) denotes all conditioned family information, and \(R\) denotes the residual weights. For independent transmissions with conditional means \(1/2\) and variances \(1/4\), the covariance terms between distinct transmissions vanish and the variance reduces to \(\frac14\sum_{i,j}R_{ij}^{2}\). That reduction requires the stated conditional properties given both \(S\) and the fixed residuals. Independence given \(S\) alone is not enough if conditioning on \(R\) changes transmission means, variances, or dependence.

With an appropriate conditional variance, a large-sample test often standardizes the score as \(Z=U/\sqrt{\operatorname{Var}(U\mid S,R)}\) and compares it with a standard normal distribution. This approximation requires enough informative transmissions and a suitable conditional limit approximation. If transmissions are dependent within families, the covariance structure must be reflected in the variance. The earlier material on dependence and variance in family-based scores develops that issue generally; covariate residualization does not remove it.

What Adjustment Changes

Adjustment changes the trait quantity attached to each transmission, and therefore can change the score, its variance, and the scientific question answered. It may improve precision when the covariates explain substantial trait variation that is irrelevant to the genetic association. It may also change an apparent signal if the transmission pattern in the observed families happens to align with a covariate-related trait pattern.

The following small illustration uses a covariate model with an intercept and a binary covariate \(C\). It contrasts the unadjusted trait score with the score after removing the fitted covariate means. The transmissions are treated as conditionally independent with mean \(1/2\), and the values are purely hypothetical.

Assumptions, Interpretation, and a Conditional-Variance Pitfall

The conditional test relies on the family-based null distribution being appropriate for the study design and conditioning information. In particular, the centered transmissions must have mean zero under the null, and the variance calculation must use their correct joint conditional distribution. For a residual-weighted score, the residuals must also be treated in a way that is compatible with that conditional distribution. A common safe construction is to calculate residuals from phenotypes and measured covariates without using the transmissions being tested, then condition on those observed trait and covariate values under a null where transmission randomization is independent of them.

This is not a blanket guarantee for every adjustment procedure. If weights or residuals are constructed using the tested transmissions, or if the conditional null does not make residuals independent of transmission randomization, then the conditional distribution of transmissions given the residuals may differ from the distribution given \(S\) alone. The score mean and variance must then be derived under the joint conditional law, or a calibration method that accounts for the data-dependent procedure must be used.

Covariate adjustment should not be interpreted as a universal correction for bias. Family-based conditioning addresses a different issue from covariate modeling: it bases inference on within-family transmission randomness, whereas covariates describe measured sources of trait variation. Adjustment does not repair genotyping errors, guarantee validity under every ascertainment scheme, or turn an association into a causal effect. The interpretation is conditional on the tested transmission coding, the modeled covariates, and the assumptions supporting the family-based null distribution.

Common Mistakes

• Using raw traits when the stated test is covariate-adjusted. The score then answers a different question and may include trait variation predicted by the covariates.
• Assuming residualization guarantees validity. Residuals used as fixed weights still require a conditional null compatible with the transmission distribution and the way those residuals were obtained.
• Dropping covariance terms automatically. The simplified variance is justified only when transmissions have the specified conditional variances and zero conditional covariances given the residuals.
• Calling an adjusted association causal. A significant score supports association with trait variation beyond the fitted covariate pattern, subject to the test assumptions; it does not by itself establish causation.