Why Covariate Choice Matters in QTL Regression
In an expression QTL analysis, the regression coefficient for genotype describes the association between genotype and molecular expression, conditional on the covariates in the model. That conditioning is not a neutral housekeeping step: it can remove confounding, change the effect being estimated, improve precision, or introduce bias. The aim is therefore not to include every available variable, but to decide which variables are needed for the scientific question and whether their adjustment is justified.
A covariate can be measured on the donor, such as age or an ancestry measure, or on the molecular sample, such as an RNA-quality measure. Its name or measurement source does not determine whether it should be adjusted for. Its role in the relationship among genotype, expression, and other causes of expression does.
The Covariate-Adjusted QTL Model
For donor \(i\), let \(Y_i\) be a processed molecular-expression phenotype and \(G_i\) the dosage of the allele being tested, coded as 0, 1, or 2 copies. Let \(\mathbf{C}_i\) be a vector of selected donor-level or molecular covariates. A standard linear model is
Here \(\beta_G\) is the expected expression difference per additional allele, conditional on the covariates in the model. If expression is on a standardized scale, its units are standard deviations of expression per allele; otherwise, the units follow the expression scale. This conditional association is not automatically a causal effect.
The usual regression interpretation requires a correctly specified conditional mean, \(E(\varepsilon_i\mid G_i,\mathbf{C}_i)=0\), and enough independent donors for the chosen standard errors and tests to be appropriate. The predictors must also have full rank: the genotype column cannot be an exact linear combination of the covariates and intercept. If donors are related or measurements are repeated, ordinary independent-error regression may not be adequate; those settings require other model structures.
Covariates commonly considered include ancestry measures, age, sex, and relevant assay-quality measurements. A variable should not be called a confounder merely because it predicts expression. For confounding control it must also help explain why genotype and expression are associated in the analyzed sample. Conversely, a variable can matter even if it is not a conventional demographic variable: a sample-processing measure may be related to both genotype allocation and measured expression.
How Confounding Changes the Genotype Coefficient
Population structure provides a standard example. An ancestry variable can be associated with allele frequencies and with expression, for example because expression differs across ancestry groups for reasons unrelated to the tested variant. If ancestry is omitted, part of its expression association can be attributed to genotype. The following short derivation shows the direction and size of this omitted-variable bias for one covariate.
Suppose the true model includes a centered covariate \(C\), and errors have zero covariance with genotype:
The slope from regressing \(Y\) on \(G\) alone is \(\operatorname{Cov}(G,Y)/\operatorname{Var}(G)\). Substituting the true model gives
Thus, the unadjusted slope equals the conditional genotype coefficient plus a bias term. The bias is zero if the covariate has no association with genotype, has no effect on expression, or both. With multiple covariates, the same principle applies to the part of genotype associated with those omitted variables.
Worked Example: Adjusting for an Ancestry-Related Covariate
Hypothetical data. In 400 unrelated donors, let \(G\) be allele dosage, with \(\operatorname{Var}(G)=0.50\) dosage-squared. Let \(A\) be a centered ancestry measure with \(\operatorname{Cov}(G,A)=0.20\) dosage-units, and suppose ancestry has an expression coefficient of \(1.5\) expression standard deviations per ancestry unit. The hypothetical conditional genotype coefficient is \(0.30\) expression standard deviations per allele. Assume the errors are uncorrelated with genotype and ancestry.
If ancestry is omitted, the expected genotype-only slope is
The ancestry-adjusted model estimates \(0.30\) expression standard deviations per allele under the stated model, whereas the unadjusted slope is \(0.90\). The difference arises because the genotype is associated with ancestry, and ancestry predicts expression. It would be incorrect to interpret the unadjusted \(0.90\) as the variant’s conditional association in this example.
Confounders, Mediators, and the Target Effect
Whether adjustment is appropriate depends partly on the target. If the goal is the total effect of a genotype on expression, adjusting for a mediator removes the part of the association that operates through that mediator. If the goal is a direct effect not operating through the mediator, adjustment may be relevant, but causal interpretation then requires additional assumptions about confounding of the mediator–expression relationship. The distinction between total and direct effects is central to genetic mediation as well as to QTL analysis.
Worked Example: A Molecular Mediator
Hypothetical model. A variant changes the expression of a transcription factor \(M\), which in turn affects the expression \(Y\) of a target gene. Suppose the mediator model is \(M=0.5G+u\), where \(M\) is measured in standardized units, and the target-expression model is \(Y=0.2G+0.8M+e\), where \(Y\) is in expression standard deviations. Assume the errors have conditional mean zero and that the simple pathway model is appropriate.
Substituting the mediator equation into the target-expression equation gives
The total genotype association under this model is \(0.6\) expression standard deviations per allele: \(0.2\) is the direct component and \(0.8\times0.5=0.4\) is the component through \(M\). Including \(M\) in the regression gives a genotype coefficient of \(0.2\), not \(0.6\), under the stated assumptions. That adjustment answers a different question. It would be a mistake to adjust for \(M\) automatically and then describe \(0.2\) as the total genotype effect.
A variable can also be a collider rather than a confounder. For example, if a sample-selection or measurement variable is influenced by both genotype-related processes and other causes of expression, conditioning on it can connect those causes. This is why “adjust for every variable associated with the outcome” is not a safe rule. A causal diagram or an explicit account of how variables arise can help distinguish common causes from mediators and colliders.
Worked Example: How Collider Adjustment Can Create an Association
Hypothetical data-generating system. Let \(G\) and an unmeasured donor factor \(U\) be independent binary variables, each taking 0 or 1 with equal probability. Let expression be \(Y=U\), and let a recorded sample-selection score be \(C=G+U\). Thus \(C\) is influenced by both \(G\) and \(U\), while \(U\) affects expression.
Without adjustment, \(G\) and \(Y\) are independent, so the genotype–expression slope is 0. But the identity \(Y=U=C-G\) holds for every donor. A regression of \(Y\) on both \(G\) and \(C\) therefore has genotype coefficient \(-1\) and covariate coefficient \(1\), since
The adjusted model has produced a genotype association where none existed marginally. Conditioning on \(C\), a common effect of \(G\) and \(U\), induces an association between them. In a real QTL study the relationships would rarely be this exact, but the example shows why the role of a covariate matters more than its predictive strength.
Precision, Overadjustment, and Practical Selection
A covariate that predicts expression but is not associated with genotype can improve precision. By explaining variation in \(Y\), it may reduce the residual variance while leaving the genotype coefficient’s target unchanged. However, covariate adjustment is not guaranteed to improve precision: estimating extra coefficients costs degrees of freedom, and a covariate strongly associated with genotype can leave little independent genotype variation from which to estimate \(\beta_G\).
Worked Example: A Prognostic Covariate and Precision
Hypothetical regression summaries. In 100 donors, suppose the centered genotype values have \(\sum_i(G_i-\bar G)^2=80\) dosage-squared. In the genotype-only fit, the residual variance estimate is \(s^2=1.44\) expression-standard-deviation-squared. A measured covariate \(C\) is uncorrelated with genotype but predicts expression; after adding it, suppose the residual variance estimate is \(s^2=0.64\) in the adjusted fit. For illustration, use the standard-error calculation \(SE(\hat\beta_G)\approx\sqrt{s^2/\sum_i(G_i-\bar G)^2}\), which applies when genotype is orthogonal to the added covariate.
Without \(C\),
With \(C\),
In this hypothetical comparison, the covariate improves precision because it reduces residual variation without reducing the genotype variation available for estimation. The figures are illustrative model summaries, not a claim that every outcome-predictive covariate will help. If \(C\) were correlated with \(G\), the denominator would depend on genotype variation remaining after accounting for \(C\), and the standard error could increase instead.
A defensible selection process begins with the QTL estimand and a causal account of the variables, not a search for whichever covariate makes a p-value smallest. In practice:
Decide whether the aim is a genotype–expression association conditional on specified covariates, a total genetic effect, or a direct effect excluding a proposed pathway.
Ask which measured variables could be related to both genotype and expression in the analyzed donors. Include justified confounders, such as ancestry measures when population structure could affect both.
Determine whether the variable is a confounder, mediator, collider, or useful predictor unrelated to genotype. Do not infer its role solely from a regression p-value.
Confirm the covariates are measured and coded appropriately and that the regression can estimate their coefficients alongside genotype. Report the covariates and interpret the genotype coefficient conditionally on them.
Common Mistakes
- Adjusting for every available measurement. This may condition on a mediator or collider, reduce precision, or make the model poorly supported. Include a covariate for a stated reason.
- Calling every expression predictor a confounder. A variable associated with expression is not necessarily associated with genotype. Predicting the outcome alone does not establish confounding.
- Interpreting a mediator-adjusted coefficient as the total effect. Adjustment for a mediator changes the target by removing the pathway through that variable.
- Choosing covariates by significance in the same data. Data-driven selection can make estimates and uncertainty difficult to interpret. Prefer scientific justification and a clearly specified analysis plan.
- Ignoring genotype–covariate overlap. If genotype is nearly determined by a covariate, the adjusted coefficient may be imprecise even when the regression technically runs.
Check Your Understanding
For each question, explain the reasoning rather than naming a covariate without context.
- In the omitted-variable formula, what happens to the genotype-only slope if a covariate predicts expression but is uncorrelated with genotype?
- A variant changes an intermediate molecular trait that affects target-gene expression. What does adding the intermediate trait to the regression do to the target being estimated?
- Why is a variable that predicts expression not automatically a confounder?
- In the collider example, why does adjustment produce a genotype coefficient even though genotype and expression were independent before adjustment?
- Give one reason a covariate can reduce the standard error of a QTL coefficient and one reason adjustment can instead make that standard error larger.