Covariate Roles in Genetic Association Models

Why Covariate Roles Matter

A genotype–phenotype regression often includes variables beyond genotype. The reason for including each covariate matters: controlling a common cause can reduce confounding, while controlling a mediator changes the effect being described, and controlling a collider can create bias. A covariate that predicts the phenotype may also improve precision, but that alone does not establish that adjustment is necessary or harmless.

Let \(G\) denote genotype, often coded as an allele count, and let \(Y\) denote a phenotype. A regression coefficient for \(G\) describes an association conditional on the other model variables. It is not automatically a causal effect. To interpret it causally requires assumptions about the data-generating process, the covariates, and the model. The same variable can have different roles in different scientific settings.

Definition: A confounder is a common cause of genotype and phenotype, or a variable that adequately controls a relevant common-cause pathway. A precision variable predicts the phenotype but, under the conditions of the analysis, is not needed to control confounding and does not change the target genotype coefficient. A mediator lies on a causal pathway from genotype to phenotype. A collider is a common effect of two variables. Conditioning on a collider can induce an association between its causes.

These roles are causal, not merely descriptive. For example, a variable correlated with both \(G\) and \(Y\) is not necessarily a confounder: the correlations may arise for other reasons. A causal diagram helps express the relevant structure. The arrows \(G \leftarrow C \rightarrow Y\) depict a possible confounder \(C\); \(G \rightarrow M \rightarrow Y\) depicts a mediator \(M\); and \(G \rightarrow K \leftarrow U \rightarrow Y\) depicts a collider \(K\), with \(U\) an additional cause of the phenotype.

Confounders: Adjust to Control a Common Cause

In genetic association, ancestry can contribute to confounding: ancestry may be related to allele frequencies and also to a phenotype through environmental, social, or other genetic pathways. If the analysis does not adequately account for this structure, the genotype coefficient can partly reflect differences between ancestry groups rather than the association of interest within comparable ancestry backgrounds. A covariate used to control this structure may be a measured cause or a proxy; its usefulness depends on how well it captures the relevant variation.

In a linear model, the omitted-variable relationship makes the issue explicit. Suppose the correctly specified conditional mean is \(E(Y\mid G,C)=\beta_GG+\beta_CC\), with an intercept understood, and the residual has conditional mean zero. If \(C\) is omitted and a simple linear regression of \(Y\) on \(G\) is fitted, the population slope is:

\beta_{\text{omit}} = \beta_G+\beta_C\frac{\operatorname{Cov}(G,C)}{\operatorname{Var}(G)}.

Thus, if \(C\) affects the phenotype and is associated with genotype, the omitted-variable term need not be zero. This identity concerns linear projection slopes; the causal interpretation of \(\beta_G\) additionally requires appropriate causal assumptions, including control of relevant confounding and a suitable model.

Worked Example: A Confounder Changes the Genotype Slope

Consider hypothetical centered variables measured in a sample from a population. Let \(G\) be allele dosage, and let \(C\) be a centered ancestry-related covariate. Suppose the conditional mean is \(E(Y\mid G,C)=1.2G+0.8C\), with a mean-zero error uncorrelated with \(G\) and \(C\). For illustration, assume \(\operatorname{Var}(G)=0.5\) dosage-squared and \(\operatorname{Cov}(G,C)=0.2\) dosage-covariate units. The conditional coefficient of \(G\) is 1.2 phenotype units per allele.

If \(C\) is omitted, the population slope is:

\beta_{\text{omit}} = 1.2+0.8\frac{0.2}{0.5} = 1.2+0.32 = 1.52.

The omitted-variable term is positive because \(C\) both predicts \(Y\) positively and covaries positively with \(G\). In this hypothetical linear model, adjusting for \(C\) targets the conditional genotype slope of 1.2, whereas omitting it gives a slope of 1.52. The calculation illustrates how confounding can operate; it does not show that any particular ancestry adjustment is sufficient in a real analysis.

Precision Variables: Predictive Does Not Mean Confounding

A baseline covariate can predict the phenotype without being a confounder. If it is independent of genotype in the target population and is included in a correctly specified linear model, it need not change the population genotype slope. It can nevertheless reduce residual variation and thereby make the estimate more precise. Examples might include a pre-existing measurement that predicts a quantitative trait but has no association with genotype in the analysis population.

This conclusion depends on the conditions just stated. If the covariate is associated with genotype, the conditional coefficient and the marginal coefficient need not be the same, even if the covariate is not a confounder in a causal sense. In nonlinear models such as logistic regression, conditional and marginal coefficients can differ even without confounding. Therefore, “it predicts the outcome” is not by itself a complete justification for adjustment.

Worked Example: A Prognostic Covariate Can Improve Precision

Suppose a hypothetical quantitative-trait model is \(Y=2G+3C+\varepsilon\), where \(G\) is centered genotype dosage, \(C\) is a centered baseline predictor, and \(\varepsilon\) is mean-zero noise. Assume \(G\) and \(C\) are independent, \(\operatorname{Var}(G)=1\) dosage-squared, \(\operatorname{Var}(C)=1\), and \(\operatorname{Var}(\varepsilon)=4\) phenotype-units-squared. Also assume the errors are independent of \(G\) and \(C\), and that the linear conditional mean is correct.

Without \(C\), the population slope for \(G\) is 2, since \(C\) is independent of \(G\). The residual is \(3C+\varepsilon\), whose variance is:

\operatorname{Var}(3C+\varepsilon) = 3^2\operatorname{Var}(C)+\operatorname{Var}(\varepsilon) = 9(1)+4 = 13.

With \(C\) included, the residual is \(\varepsilon\), with variance 4, and the genotype coefficient remains 2. For a large sample of \(n=100\) independent observations with genotype variance near 1, the approximate standard errors are \(\sqrt{13/(100\times1)}=0.3606\) without adjustment and \(\sqrt{4/(100\times1)}=0.2000\) with adjustment. These are illustrative large-sample calculations, not exact sample standard errors. Under the stated assumptions, adjustment improves precision without changing the population slope. In observational data, one must still check whether the independence and model assumptions are plausible.

Mediators: Adjustment Changes the Effect Being Described

A mediator is downstream of genotype and carries some part of the pathway to the phenotype. A regression that adjusts for a mediator does not generally estimate the same quantity as a regression that leaves it out. In a linear model with no genotype–mediator interaction, the genotype coefficient without mediator adjustment can represent a total effect, while the coefficient controlling for the mediator can represent a direct effect holding the mediator fixed—provided the causal assumptions and model conditions support those interpretations.

In particular, a direct-effect interpretation requires more than fitting \(Y\) on \(G\) and \(M\). It requires an adequate model and assumptions that address confounding of the genotype–outcome, genotype–mediator, and mediator–outcome relationships. The conditional error assumption matters too: if the outcome error is associated with the mediator after conditioning on genotype, the adjusted regression coefficient need not equal the direct-effect parameter.

Worked Example: Total and Direct Effects in a Mediated Pathway

Consider hypothetical centered genotype dosage \(G\), mediator \(M\), and phenotype \(Y\) generated by \(M=0.4G+u\) and \(Y=2+1.0G+2M+e\). Assume \(G\) is independent of \(u\) and \(e\), and \(u\) and \(e\) are independent, mean-zero errors. In particular, \(e\) is independent of \((G,u)\), so \(E(e\mid G,M)=0\). Assume the linear equations are correctly specified and that there is no unmeasured confounding of the relevant causal relationships.

Substituting the mediator equation into the outcome equation gives:

$$ Y = 2+1.0G+2(0.4G+u)+e = 2+1.8G+2u+e. $$

Because \(G\) is independent of \(u\) and \(e\), the total-effect slope is \(1.8\) phenotype units per allele: \(1.0+2(0.4)=1.8\). In the regression of \(Y\) on both \(G\) and \(M\), the stated conditional error assumption gives the model coefficient 1.0 for \(G\), holding \(M\) fixed, and 2 for \(M\). Thus the adjusted genotype coefficient describes the direct component in this model, not the total pathway effect. The independence assumption for \(e\) is essential here; independence of the errors from \(G\) alone would not guarantee the adjusted coefficient is 1.0.

Colliders: Adjustment Can Create an Association

A collider is caused by two or more variables. Its causes can be independent before conditioning but associated among individuals with the same collider value. If one cause is genotype and another cause predicts the phenotype, conditioning on the collider can therefore induce a spurious genotype–phenotype association. This is sometimes called collider bias or selection bias, especially when the collider represents inclusion in the analyzed sample.

Worked Example: Conditioning on a Collider Creates a Slope

Suppose, hypothetically, that \(G\) and an unmeasured variable \(U\) are independent, centered variables with variances 1. Let \(K=G+U\) be a collider, and let \(Y=U+e\), where \(e\) is independent mean-zero noise. The true causal effect of \(G\) on \(Y\) is zero, and without conditioning on \(K\), \(\operatorname{Cov}(G,Y)=0\).

For the regression of \(Y\) on \(G\) and \(K\), the coefficient of \(G\) can be calculated from the two-predictor regression formula. Here, \(\operatorname{Var}(K)=2\), \(\operatorname{Cov}(G,K)=1\), \(\operatorname{Cov}(Y,G)=0\), and \(\operatorname{Cov}(Y,K)=1\). Therefore:

\hat{\beta}_{G\mid K} \text{ in the population} = \frac{\operatorname{Cov}(Y,G)\operatorname{Var}(K) -\operatorname{Cov}(Y,K)\operatorname{Cov}(G,K)} {\operatorname{Var}(G)\operatorname{Var}(K)-\operatorname{Cov}(G,K)^2} = \frac{0(2)-1(1)}{1(2)-1^2} = -1.

The adjusted association is negative despite the true genotype effect being zero. Indeed, \(U=K-G\), so \(Y=K-G+e\), which has coefficient \(-1\) for \(G\) when \(K\) is held fixed. Holding the collider fixed makes higher \(G\) imply lower \(U\), and \(U\) predicts \(Y\). The example shows why a covariate should not be included just because it is available or associated with the outcome.

A Practical Way to Reason About Adjustment

Before adding a covariate to a genetic association model, state the target quantity and the causal role you believe the variable has. A coefficient conditional on a covariate answers a conditional question; it is not automatically the total genetic effect or a less-biased estimate. Use subject-matter knowledge and the study design to consider whether the variable is upstream of genotype and phenotype, downstream of genotype, or a common effect.

Name the target.
Specify whether the goal is a conditional association, a total causal effect, or a direct effect holding a mediator fixed.

Map the causal role.
Ask whether the covariate is a common cause, a prognostic variable independent of genotype, a mediator, a collider, or something else.

State the assumptions.
Check that the model, error conditions, and confounding assumptions needed for the desired interpretation are plausible.

Interpret the fitted coefficient accordingly.
Report the conditioning set and avoid describing an adjusted association as causal unless the causal assumptions justify it.

Common Mistakes

Adjusting for every available variable. More covariates do not automatically mean less bias. Conditioning on a collider can create bias, and conditioning on a mediator can remove part of the total effect.
Calling every predictor a confounder. A variable that predicts \(Y\) may be useful for precision, but it is not thereby a common cause of \(G\) and \(Y\).
Calling the adjusted coefficient “the genetic effect.” State what is held fixed. With a mediator in the model, the coefficient may target a direct rather than total effect; with other covariates, it is a conditional association.
Assuming a mediator-adjusted coefficient is automatically causal. The required error and confounding assumptions must hold. In particular, independence of errors from genotype alone does not ensure the outcome error has mean zero conditional on both genotype and mediator.
Ignoring the model scale. In nonlinear models, conditional and marginal coefficients can differ even without confounding. The meaning of adjustment depends on both the causal structure and the regression model.

Key takeaway: Covariate adjustment is a choice about the comparison and estimand, not a mechanical correction. Adjust for a confounder to control a common-cause pathway when assumptions support it; a suitable precision variable may improve precision; mediator adjustment changes the target; and collider adjustment can induce bias.

Check Your Understanding

For each question, explain the role of the covariate and the consequence of adjustment.

An ancestry variable affects allele frequencies and a quantitative phenotype. What potential role does it play, and what assumption is needed before adjustment can be said to control the relevant confounding?
A baseline measurement predicts the phenotype but is independent of genotype, and the linear model assumptions hold. What can adjustment change, and what need not change?
In a correctly specified linear mediation model, what is the difference between the genotype coefficient from a regression that includes the mediator and one that omits it?
Why can conditioning on a collider associate genotype with another cause of the phenotype even if those causes were initially independent?
Why is it not enough to assume that the outcome error is independent of genotype when interpreting a mediator-adjusted coefficient?