Choosing Covariates Without Overadjustment

Why Covariate Choice Is a Design Decision

A covariate is not automatically appropriate merely because it is measured or associated with the outcome. In a multi-donor single-cell study, the choice should follow the planned comparison and the way each variable arose. Adjusting for a pre-condition donor characteristic may address imbalance or improve precision. Adjusting for a variable caused by condition can instead remove part of the effect being studied; conditioning on a common consequence of condition and another cause of the outcome can create a spurious association.

The unit relevant to the comparison is determined by the study design and estimand. In the examples below, \(Y_d\) denotes a donor-level outcome for a specified cell population, such as a donor-level expression summary, and \(C_d\) denotes the condition assigned to or observed for donor \(d\). The discussion concerns choosing covariates for that comparison, not treating the many cells from one donor as independent experimental units.

“Pre-treatment” is a statement about timing and causal possibility, not simply the date a field was entered into a spreadsheet. Age at enrollment is ordinarily pre-treatment; a post-exposure cell-state proportion is not. A processing batch may be a useful technical covariate if its assignment is not affected by condition and it is not a consequence of condition. The study timeline and allocation process determine which description is defensible.

A Framework for Choosing Covariates

Begin with the target comparison and estimand established during study planning. Then draw the temporal order: donor characteristics and allocation decisions, condition, technical processing, and outcome measurement. For each candidate covariate, ask whether it precedes condition, whether it predicts the outcome, and whether it is associated with condition through the design or recruitment process. These questions help distinguish confounding control from adjustment that changes the target quantity.

For an observational comparison, a baseline covariate may be needed to control confounding: a common cause, or a proxy for a common cause, of condition and outcome. A causal interpretation additionally requires assumptions such as conditional exchangeability (no unmeasured confounding after conditioning on the chosen baseline covariates), consistency, and positivity. Positivity means that within each relevant covariate pattern, both conditions have a nonzero probability of occurring. Covariate adjustment cannot guarantee these assumptions.

In a randomized study, random assignment makes baseline covariates independent of assignment in expectation. Their adjustment is not generally required to remove confounding from assignment, although pre-specified baseline adjustment can improve precision or account for chance imbalance in a realized sample. Randomization does not make post-condition variables safe to adjust for: their causal position is unchanged.

A simple additive model makes the role of a candidate covariate explicit:

Here \(X_d\) is the selected covariate vector and \(\varepsilon_d\) represents remaining variation. In this model, \(\beta_C\) is the condition contrast at a fixed value of \(X_d\), under the model's assumptions. If the condition effect varies across covariate levels, an additive model may not represent the intended comparison; a planned interaction or a defined standardized contrast may be needed. Adjustment does not itself establish that the model is correctly specified or that a causal interpretation is justified.

A useful justification for adjustment is standardization. If \(X\) is a pre-condition covariate and the conditional mean outcome under condition \(c\) is \(\mu_c(x)\), then an average contrast for a specified target distribution of \(X\) is:

The weights \(w(x)\) define the population or sample over which the comparison is averaged. If the conditional contrast is constant across \(x\), this average equals that common contrast. If it varies, the choice of target weights matters. For a causal interpretation, the assumptions stated above must hold and the covariate must be pre-treatment; standardization is not a remedy for conditioning on a mediator or collider.

Variables That Should Not Be Adjusted For Automatically

A mediator lies downstream of condition. If the target is the total effect of condition, adjusting for a mediator generally removes some of the pathway through which that total effect operates. A covariate can also be a collider. Conditioning on a collider may open a noncausal path between its causes, even if those causes were independent before conditioning. These are different problems: mediator adjustment can change the estimand, while collider adjustment can introduce bias.

Technical Covariates, Diagnostics, and Alternatives

Technical factors such as processing site, operator, platform, or batch can contribute systematic outcome variation. Their usefulness depends on how they were assigned and whether they are separable from condition. A preplanned technical factor that affects the measured outcome and is not caused by condition may be adjusted for, especially when allocation has created imbalance. By contrast, a technical measurement recorded after condition that could itself respond to condition may be a mediator or collider. A label such as “technical” does not settle the causal question.

Before analysis, use a causal diagram or a written timeline to make the proposed roles explicit. Then check whether condition groups have meaningful overlap in each proposed adjustment variable. If all treated donors occur at one site and all reference donors at another, the data cannot distinguish condition from site without additional design information; adding both variables to a formula does not create the missing comparison. Earlier material on rank and alias checks provides the relevant design-matrix test: the model must be estimable.

Useful diagnostics include balance tables for pre-condition covariates, plots or tables showing condition overlap within covariate levels, and checks for sparse or empty combinations. Inspect whether technical levels are represented in both conditions and whether a proposed variable is actually measured before the relevant causal process. Large changes between crude and adjusted estimates are a prompt to examine composition, model form, overlap, and the estimand; they are not proof that either estimate is correct.

The main alternatives each answer a slightly different need. An unadjusted comparison is transparent and may be suitable in a well-randomized, balanced design, but can be inefficient or misleading with important imbalance. Blocking or stratifying at design or analysis stage makes comparisons within known levels explicit, but can become impractical with many covariates or sparse strata. Regression adjustment can accommodate several covariates compactly, but depends on model specification and estimability. Matching or restriction can improve comparability in some observational settings, but may discard donors and does not control unmeasured confounding. None of these approaches makes a mediator or collider safe to condition on.

Do not select covariates by searching for those with small p-values in the observed outcome data. Such selection can make estimates unstable and their uncertainty understated, and it confuses predictive association with causal appropriateness. Prefer a small, pre-specified set justified by the design, subject-matter knowledge, and target estimand. Report what was adjusted for and why, and distinguish a total-effect target from a conditional or direct-effect target.

Common Mistakes and Limits

• Adjusting for every available variable. More covariates are not automatically better. A post-condition variable may block a causal pathway or create collider bias.
• Calling a variable “baseline” because it appears in baseline metadata. Verify when it was measured and whether condition or study selection could have affected it.
• Treating a technical label as sufficient justification. A batch variable is useful only when its causal timing and relation to condition support adjustment and the comparison remains estimable.
• Interpreting the adjusted coefficient as the total effect by default. The coefficient is conditional on the specified model and covariates; its target depends on the design and assumptions.
• Using adjustment to solve complete confounding. If condition and a technical factor are perfectly aligned, the observed data do not contain the independent contrasts needed to separate their effects.

Covariate selection has limits. It cannot repair poor measurement, unobserved confounding, absent overlap, or selection processes that are not understood. A careful covariate plan improves the interpretability of a comparison, but it is not a substitute for a design that creates informative comparisons across donors.