ComBat When Batch and Biology Are Confounded

Why Confounding Changes the Meaning of an Adjustment

ComBat estimates feature-specific batch location and scale effects while borrowing information across features. That borrowing can stabilize estimates when batch parameters vary across features, but it cannot create information absent from the sample design. If cases occur mainly in one batch and controls mainly in another, an observed difference can reflect biology, batch, or both. ComBat must rely on the design and its model assumptions to distinguish them.

The central question is therefore not simply whether ComBat returns adjusted values. It is whether the available data can support the separation those values imply. With perfect confounding, some effects are not separately identified. With partial confounding, they may be estimable but imprecise. In either case, an adjustment is conditional on the model; empirical-Bayes estimation does not turn an ambiguous comparison into an unambiguous one.

Definition: For feature $i$ and sample $j$, let $Y_{ij}$ be the measured value, $b(j)$ the batch containing sample $j$, and $x_j$ the vector of biological covariates. A standard ComBat location-scale model can be written $$ Y_{ij}=\alpha_i+x_j^{\mathsf T}\beta_i+\sigma_i\{\gamma_{i,b(j)}+\delta_{i,b(j)}\epsilon_{ij}\}, $$ where $\alpha_i$ is a feature baseline, $\beta_i$ describes biological covariate effects, $\sigma_i$ is a feature scale, and $\gamma_{i,b}$ and $\delta_{i,b}$ are batch-specific location and scale effects. The errors $\epsilon_{ij}$ are typically assumed to have mean zero and variance one. In the usual empirical-Bayes approach, batch parameters across features are modeled as draws from common distributions and are estimated with information pooled across features.

This model is useful only insofar as its components can be learned from the design. In a conventional ComBat analysis, one first fits the biological component and standardizes feature values, then estimates batch-specific location and scale effects on the standardized values. After adjustment, the fitted baseline and biological component must be retained when returning to the original measurement scale.

Identification: What the Design Can and Cannot Separate

Consider first the additive part of the model for a single feature: $Y=\alpha\mathbf{1}+X\beta+Z\gamma+\varepsilon$, where $X$ contains biological covariates and $Z$ contains batch indicators under a chosen reference coding. Full column rank of the combined design matrix is required to uniquely estimate all its coefficients. But it is not required for every particular contrast to be estimable.

A contrast $c^{\mathsf T}\theta$, where $\theta$ collects the model coefficients, is estimable if it has the same value for every coefficient vector that gives the same fitted values. Equivalently, $c^{\mathsf T}v=0$ for every $v$ in the null space of the design matrix. This distinction matters in rank-deficient batch designs: some coefficient combinations may be estimable even though the individual coefficients are not.

Key identification result: If a biological indicator and a batch indicator are identical across samples, only their combined contribution to the fitted values is identified. No estimation method, including empirical-Bayes pooling across features, can determine how much of that combined contribution is biological and how much is a batch effect without additional information or assumptions.

Worked Example: Perfect Confounding Leaves the Cause Unresolved

Suppose a hypothetical study measures one feature in two batches. Every control is processed in batch A and every case in batch B. The observed mean is 10 units for controls and 16 units for cases. Let $D_j=1$ for a case and $B_j=1$ for batch B. Here $D_j=B_j$ for every sample.

In the additive model $Y_j=\alpha+\beta D_j+\gamma B_j+\varepsilon_j$, the fitted mean is $\alpha+(\beta+\gamma)D_j$. The data identify the combined difference $\beta+\gamma=16-10=6$ units, not $\beta$ and $\gamma$ separately. For example, both $(\beta,\gamma)=(6,0)$ and $(\beta,\gamma)=(0,6)$ fit the same two group means.

Consequently, a batch correction that removes the batch-associated six-unit difference also removes the observed case-control difference in these data. That result does not demonstrate that the biological effect is zero; the design cannot distinguish the two explanations. A different allocation of the same six units between biology and batch would require an external source of information or a stated modeling assumption.

This is why the earlier material on estimable biological contrasts is relevant: a request to estimate a biological coefficient is not automatically answerable just because a software routine can produce a number. In a rank-deficient design, check the estimability condition for the particular biological contrast. An arbitrary choice of reference coding or a dropped design column may yield a computational solution, but it does not supply the missing scientific information.

Partial Confounding: Estimable Does Not Mean Precise

When each biological group occurs in more than one batch, the model may be identifiable, but scarce cross-classified samples can make biological and batch effects difficult to distinguish. Under a linear model with independent errors of common variance $\sigma^2$, the variance of a coefficient estimate increases as its covariate becomes more predictable from the other design columns. In a two-indicator example, the inflation is summarized by the variance inflation factor $1/(1-r^2)$, where $r$ is the correlation between the indicators after accounting for the intercept.

Worked Example: Strong Overlap Inflates Uncertainty

Consider a hypothetical study with 20 samples classified by batch $B$ and disease status $D$. The four cell counts are: 9 controls in batch A, 1 case in batch A, 1 control in batch B, and 9 cases in batch B. All four cells are represented, so the additive design with an intercept, $B$, and $D$ has full rank. Suppose the observed cell means for one feature are 10, 13, 14, and 17 units in that order, and the fitted residual standard deviation is 1 unit.

The two indicators each have 10 ones among 20 samples, so their centered sums of squares are $S_{BB}=S_{DD}=20(0.5)(0.5)=5$. Their centered cross-product is $S_{BD}=9-20(0.5)(0.5)=4$, because 9 samples have both indicators equal to one. Thus $$ r=\frac{S_{BD}}{\sqrt{S_{BB}S_{DD}}}=\frac{4}{\sqrt{5\cdot5}}=0.8. $$

After adjusting for batch, the information for the disease coefficient is $$ S_{DD\cdot B}=S_{DD}-\frac{S_{BD}^2}{S_{BB}}=5-\frac{16}{5}=1.8. $$ The disease-coefficient standard error is therefore $SE(\hat\beta)=1/\sqrt{1.8}=0.7454$ units. If the disease and batch indicators were uncorrelated with the same $S_{DD}=5$, the corresponding standard error would be $1/\sqrt{5}=0.4472$ units. The variance inflation factor is $1/(1-0.8^2)=2.7778$. The observed cell means give an additive disease difference of $13-10=3$ units and a batch difference of $14-10=4$ units, but the sparse off-diagonal cells make those separate estimates substantially less precise than in a well-overlapped design.

This calculation assumes the additive model is appropriate and the residual variance is common across the four cells. The four-cell design makes the two effects estimable; it does not guarantee a small standard error or rule out other model failures. The standard error quantifies uncertainty under the stated model, not uncertainty about whether that model is scientifically adequate.

What ComBat Adjusts—and What It Assumes

For clarity, write the standardized residual for feature $i$ and sample $j$ as $$ Z_{ij}=\frac{Y_{ij}-\widehat\alpha_i-x_j^{\mathsf T}\widehat\beta_i}{\widehat\sigma_i}. $$ ComBat estimates a batch location $\widehat\gamma_{i,b(j)}$ and scale $\widehat\delta_{i,b(j)}$ for these residualized, standardized values. Its adjusted value on the original scale is $$ Y^{\mathrm{adj}}_{ij} =\widehat\alpha_i+x_j^{\mathsf T}\widehat\beta_i +\widehat\sigma_i \frac{Z_{ij}-\widehat\gamma_{i,b(j)}}{\widehat\delta_{i,b(j)}}. $$ The last term is the adjusted residual, rescaled to the feature’s original units; the first terms retain the fitted baseline and biological component. If a formulation defines its standardized residual on a different scale, its back-transformation must be changed consistently. Adjusting a residual and then omitting the fitted biological component would not preserve that component.

The familiar ComBat interpretation relies on assumptions such as an adequate biological design, a location-scale representation of batch effects, and sufficiently well-behaved standardized errors for the chosen estimation procedure. Empirical-Bayes pooling further relies on a useful distributional description of batch effects across features. These assumptions can stabilize estimates, but they do not establish that the fitted biological component is correct or that a batch-associated shift is technical rather than biological.

Worked Example: Back-Transforming a Batch-Adjusted Value

Suppose, hypothetically, that for one feature the fitted baseline is $\widehat\alpha_i=10$ units, the fitted biological contribution for a particular sample is $x_j^{\mathsf T}\widehat\beta_i=4$ units, and $\widehat\sigma_i=2$ units. The sample is in batch B, with fitted standardized location $\widehat\gamma_{i,B}=0.5$ and scale $\widehat\delta_{i,B}=2$. Its observed value is $Y_{ij}=20$ units.

First calculate the standardized residual: $$ Z_{ij}=\frac{20-10-4}{2}=3. $$ Then remove the estimated batch location and scale: $$ \frac{Z_{ij}-\widehat\gamma_{i,B}}{\widehat\delta_{i,B}} =\frac{3-0.5}{2}=1.25. $$ Finally, return to the original scale and add back the fitted baseline and biological component: $$ Y^{\mathrm{adj}}_{ij}=10+4+2(1.25)=16.5\text{ units}. $$

The adjusted value is 16.5 units under the supplied model and fitted parameters. It retains the modeled 14-unit baseline-plus-biological contribution and changes the residual contribution from 6 units to 2.5 units. This arithmetic illustrates the transformation; it does not show that the fitted parameters are reliable when batch and biology overlap.

Diagnostics, Alternatives, and Limits

Before interpreting adjusted values, examine the design itself. Tabulate biological groups by batch, check rank and connectivity, and assess whether the biological contrast of interest is estimable. If all samples in one biological group occur in one batch, no feature-wise adjustment can separate those effects from the observed data alone. If overlap is limited but nonzero, report that the estimates may be imprecise and inspect standard errors or sensitivity to plausible model specifications.

After fitting, residual diagnostics can reveal departures from the assumed model: remaining batch structure may indicate an inadequate adjustment, while implausible loss of biological separation can indicate that the fitted components do not represent the intended scientific target. These checks are warnings, not proof that the alternative attribution is correct. Inspect relevant feature distributions and residual patterns, and compare conclusions with negative controls, technical replicates, bridge samples, or reference materials when available.

The main alternatives answer different questions. Modeling batch and biology jointly in a downstream analysis can retain the uncertainty in estimated biological effects, provided the design supports the contrast. A fixed-effect batch model estimates batch-specific offsets; random-effect models impose a distribution on batch variation. Neither resolves perfect confounding without additional information. With insufficient overlap, retaining batch in the analysis, seeking new bridge or replicated measurements, or declining to make a separated biological claim can be more defensible than presenting a corrected matrix as if it settled the attribution.

Common pitfall: “ComBat uses empirical Bayes, so it can distinguish disease from batch even when all cases are in one batch.” This is incorrect. Empirical-Bayes pooling can stabilize feature-wise batch estimates under its across-feature assumptions, but the sample design still determines which effects are identifiable. Under perfect confounding, the data identify only a combined effect; under partial confounding, they may identify separate effects with inflated uncertainty.

The adjustment is also limited by its model form. Feature-specific location and scale parameters may not represent every technical artifact, and a biological component that is misspecified can affect what is attributed to batch. Results may depend on which covariates are included, how they are encoded, and which samples define the overlap. ComBat output should therefore be described as an adjustment under an explicit model, not as a data-driven guarantee that all technical variation was removed while all biology was preserved.

Map the design.
Cross-tabulate biological groups and batches; identify empty or sparse cells and the biological contrasts of interest.

Check identification and precision.
Assess rank and contrast estimability. For estimable contrasts, quantify how overlap affects uncertainty rather than treating estimability as evidence of precision.

Fit and inspect conditionally.
Use ComBat only with a defensible biological design and location-scale assumptions; inspect residual batch signal and whether the biological component remains credible.

Limit the conclusion.
State which effects are supported by the design, which conclusions depend on assumptions, and when additional measurements or abstention are needed.

Check Your Understanding

Answer each question, then compare with the solutions below.

If a disease indicator is identical to a batch indicator, which combination of their coefficients is identified in an additive model?
For the 20-sample design in the partial-confounding example, verify the variance inflation factor from the indicator correlation.
Using the supplied parameters in the back-transformation example, what value would result if the adjusted residual were returned to the original scale but the fitted baseline and biological component were omitted? Why is that not the ComBat adjusted value defined here?
Name two assumptions or design features that should be checked before interpreting a ComBat adjustment under strong overlap.

Solutions. (1) Only the sum $\beta+\gamma$ is identified, because the two indicators have identical values for every sample. (2) With $r=0.8$, the variance inflation factor is $1/(1-r^2)=1/(1-0.64)=1/0.36=2.7778$. (3) The adjusted residual on the original feature scale is $2(1.25)=2.5$ units. It omits the 14-unit fitted baseline-plus-biological contribution, so it is only a residual contribution, not the adjusted value of $16.5$ units. (4) Examples include checking that the biological contrast is estimable and that there is adequate cross-classification of groups by batch; checking the location-scale form and the standardized-error assumptions is also appropriate.