Confounding and Covariate Choice in Methylation Studies

Why Covariate Choice Matters

A difference in methylation between two groups can reflect the group contrast of interest, differences in the groups’ characteristics, or differences in how their samples were measured. For example, an age imbalance can create a methylation difference if age is related both to group membership and to methylation. A difference in cell composition can change a bulk-tissue measurement even if methylation within each cell type is unchanged. A technical factor can make one group’s measurements systematically higher because its samples were processed differently.

Covariate adjustment is therefore not a mechanical step of adding every available variable to a regression. The analyst must decide what each variable represents, whether it precedes or follows the exposure of interest, and whether the data contain enough overlap to separate its contribution from the group contrast. The same variable can be an appropriate adjustment variable for one scientific question and an inappropriate one for another.

A CpG-Level Model and What Adjustment Means

Let \(M_{ik}\) be the M-value for sample \(i\) at CpG \(k\), and let \(D_i\) indicate the group or exposure being compared, with \(D_i=1\) for the exposed or case group and \(D_i=0\) for the comparison group. A basic adjusted model for one CpG is

Here, \(\mathbf{C}_i\) contains measured covariates such as age, sex, cell-composition summaries, and technical variables. The coefficient \(\beta_{Dk}\) is the model’s adjusted difference in mean M-value between groups for samples with the same modeled covariate values. It is measured in M-value units. It is not automatically a causal effect: that interpretation requires additional assumptions about the study design, confounding, measurement, and model specification.

A regression adjustment is useful only to the extent that its model is suitable for the data. For the displayed linear model, the conditional mean must be adequately represented by the included terms; the observations must provide information to estimate the group contrast; and the standard errors must be appropriate for the study’s dependence and variance structure. A curved relationship with age, for example, may not be handled by a single linear age term. If samples are clustered or repeatedly measured, an independent-error analysis may also be unsuitable.

How Confounding Creates a Misleading Contrast

The following simple result shows why a group imbalance matters. Suppose, for one CpG, the conditional mean follows an additive model with a binary covariate \(C_i\):

Taking the mean within each group and subtracting gives

The unadjusted group difference therefore contains the conditional group contrast \(\beta_D\) plus a term due to the covariate imbalance. If the covariate has no relationship with methylation in this model (\(\gamma=0\)), or its distribution is the same in both groups, this particular source of distortion is absent. In real data, these conditions are not established merely by seeing a small change in a regression coefficient.

Choosing Among Age, Sex, Cell Composition, and Technical Factors

Age and sex are often plausible pre-exposure covariates in observational comparisons, but they should be assessed in the context of the study rather than included by habit. Check how they are distributed across groups and whether their relationships with methylation are plausibly represented. If age has a nonlinear association with methylation, a linear term may leave residual confounding. If some sex categories occur almost exclusively in one group, the data may not support a stable within-category comparison.

Cell composition needs particular care in bulk methylation data. A bulk measurement reflects methylation across a mixture of cells, and differences in the proportions of cell types can contribute to a group difference. Earlier material on cell-type composition explains why these proportions matter for interpretation. Whether to adjust for them depends on the target question: composition may be a pre-existing confounder, a consequence of the exposure, or part of the biological response of interest.

If group membership can change cell composition, then adjusting for composition removes the portion of the observed difference that operates through that change. The adjusted coefficient is not simply a “better corrected” version of the total difference; it targets a comparison conditional on composition. Interpreting it as a direct causal effect requires strong assumptions, including appropriate control of confounding of the exposure–outcome and mediator–outcome relationships. Cell fractions also sum to one, so a model with an intercept should not include every fraction as an unrestricted separate predictor; one fraction must be omitted or a suitable composition representation used.

Technical variables such as processing batch or array position can affect measured values rather than the underlying biology. If technical assignments are associated with group membership, they may create a spurious group contrast and are commonly considered adjustment variables. However, adjustment works only when the design contains enough overlap to distinguish group from technical factor. A covariate cannot resolve perfect aliasing between group and batch.

Diagnostics, Alternatives, and Failure Modes

Before fitting a CpG-level model, inspect covariate distributions by group. For categorical factors, use counts or a cross-tabulation. For continuous factors, compare distributions rather than relying only on means. Check overlap: are there cases and controls at comparable ages, in each relevant sex category, and across technical levels? A large imbalance or an empty group-by-factor cell warns that the adjusted contrast may depend heavily on model assumptions or may not be estimable at all.

Then consider whether the model represents the covariate relationships plausibly. Examine age trends, influential observations, residual patterns, and whether the group contrast appears to vary across covariate levels. A single group coefficient assumes a common contrast unless interactions are included. For example, if group differences differ by sex, a model with only additive sex adjustment reports a restricted average conditional contrast, not separate sex-specific contrasts. Standard errors should reflect the design and residual variance; a well-chosen covariate list does not compensate for an inappropriate uncertainty calculation.

Adjustment is not the only way to address an imbalance. Stratified comparisons can display within-stratum contrasts directly when there are few relevant strata and adequate data in each. Standardization can compare groups after averaging model-based predictions over a specified covariate distribution. These approaches make different assumptions and can target different summaries. In high-dimensional methylation analysis, regression adjustment is often convenient across many CpGs, but convenience does not validate the covariate set or the model.

Several warning signs deserve particular attention:

• Near or perfect confounding: group and technical factor have little or no overlap. Estimates may be unstable or not separately identifiable.
• Overadjustment: a variable on the exposure-to-methylation pathway is included even though the target is the total effect.
• Incorrect functional form: a linear age term is used despite a substantially nonlinear age relationship.
• Compositional redundancy: all cell fractions are entered with an intercept, producing exact dependence among predictors.
• Coefficient change treated as proof: a change after adjustment is evidence that the model’s conditional comparison differs from the marginal one, but does not by itself prove that the removed component was confounding.

A common mistake is to adjust for every variable that differs between groups and then call the remaining coefficient the “true” methylation effect. This reasoning ignores whether a variable is a confounder, mediator, collider, or merely a proxy, and it ignores overlap and model assumptions. Instead, define the scientific contrast first, draw or state the plausible temporal relationships, inspect the design, and explain what conditional comparison the chosen model estimates. If technical factor and group are inseparable, report that limitation rather than claiming adjustment has solved it.