Covariate-Preserving Linear Batch Adjustment

Why Preserve Biological Covariates During Batch Adjustment?

Technical batches can shift measured values even when the underlying biology is similar. But a difference between batches may also reflect a real biological difference: perhaps cases were disproportionately processed in one batch, or a biological covariate such as age varies across batches. Adjusting for batch without accounting for those covariates can remove biological signal along with technical variation.

Covariate-preserving linear batch adjustment addresses this by fitting batch and biological covariates in the same model. It then subtracts the fitted batch contribution, leaving the fitted biological component in the adjusted measurements. This is a feature-wise procedure: the model is fitted separately for each feature, such as a metabolite or transcript, using the same sample-level design.

Definition: For feature $j$ measured on $n$ samples, let $y_j$ be its $n$-vector of observed values, $X$ the design matrix of biological covariates (including an intercept), and $Z$ the design matrix of batch indicators. A covariate-preserving linear adjustment fits $$ y_j = X\beta_j + Z\gamma_j + \epsilon_j $$ and forms $$ y_j^{\mathrm{adj}} = y_j - Z\hat{\gamma}_j. $$ Here $\hat{\gamma}_j$ contains the estimated batch coefficients. The adjustment removes those fitted batch terms, not the fitted biological component $X\hat{\beta}_j$.

With reference coding, one batch is the reference and its batch coefficient is zero; the other columns of $Z$ indicate non-reference batches. Subtracting $Z\hat{\gamma}_j$ therefore moves each sample toward the fitted level of the reference batch, conditional on the biological covariates. The adjusted values retain the original measurement scale and units. The choice of reference changes the overall level assigned to adjusted values, but not the estimated biological contrasts when the model is identifiable.

Model Conditions and Identifiability

The linear model assumes that the expected feature value is additive in the specified biological covariates and batch. In particular, it assumes a common biological effect across batches unless batch-by-biology interactions are included. For ordinary least squares, the combined design matrix $[X\;Z]$ must have full column rank: no column can be exactly explained by the other columns. Otherwise, some coefficients cannot be uniquely estimated.

These assumptions have practical implications. There must be enough overlap to distinguish batch from biology. If every case is measured in one batch and every control in another, then the observed difference cannot be separated into a batch effect and a biological effect using these data alone. With imperfect but strong association between batch and biology, coefficients may be identifiable but imprecise. Inference also commonly assumes errors with conditional mean zero and, for standard least-squares standard errors, independent errors with constant variance. Those inferential assumptions are distinct from the basic calculation of fitted values.

Inspect the sample allocation across batches and biological groups before fitting. Then check that the model is full rank and that the fitted structure is scientifically plausible. If the design is not identifiable, software may drop columns or return an arbitrary parameterization; that does not recover information absent from the data.

Worked Example: Adjusting a Two-Batch Case-Control Feature

Suppose a hypothetical feature is measured in two batches, A and B, with two controls and two cases per batch. Values are in arbitrary assay units.

Batch	Group	Observed values
A	Control	10, 12
A	Case	14, 16
B	Control	16, 18
B	Case	20, 22

Fit the model with an intercept, a case indicator $x$, and a batch-B indicator $z$:

y = \beta_0 + \beta_1 x + \gamma z + \epsilon.

The mean for batch A controls is $(10+12)/2=11$, and the mean for batch A cases is $(14+16)/2=15$. Thus the estimated case contrast is $15-11=4$ units. Batch B has means $17$ for controls and $21$ for cases, giving the same case contrast, $21-17=4$. The batch contrast is $17-11=6$ units among controls, and also $21-15=6$ among cases. Therefore, $\hat{\beta}_0=11$, $\hat{\beta}_1=4$, and $\hat{\gamma}=6$.

Adjustment subtracts $6z$: batch A values remain unchanged, while each batch B value is reduced by 6. The adjusted batch B values are 10, 12, 14, and 16 in the same group order. The case-control difference remains 4 units. The data have both groups represented in both batches, and the matching contrasts support the additive model in this small example.

Why the Biological Coefficient Is Preserved

The preservation property follows directly from the least-squares fit. Let $\hat{\beta}_j$ and $\hat{\gamma}_j$ be the joint least-squares estimates, and let $e_j=y_j-X\hat{\beta}_j-Z\hat{\gamma}_j$ be the fitted residual vector. The least-squares normal equations make the residual orthogonal to every column of the design matrix, including the columns of $X$. Consequently, $X'e_j=0$.

y_j^{\mathrm{adj}} = y_j-Z\hat{\gamma}_j = X\hat{\beta}_j+e_j.

If the adjusted values are then regressed on the same $X$ using the same samples and ordinary least squares, the coefficient estimate is

(X'X)^{-1}X'y_j^{\mathrm{adj}} = (X'X)^{-1}X'(X\hat{\beta}_j+e_j) = \hat{\beta}_j+(X'X)^{-1}X'e_j = \hat{\beta}_j,

provided $X$ has full column rank. Thus the fitted biological coefficients are exactly preserved by this algebraic adjustment under those conditions. This does not mean that they are necessarily unbiased estimates of causal effects, or that their uncertainty can be ignored. It means the batch subtraction, as defined, leaves the joint model’s fitted biological component in the adjusted data.

Worked Example: Continuous Covariate With Overlapping Batches

Consider a hypothetical feature measured in two batches. Let $x$ be a biological covariate measured in years, and let $z=1$ for batch B and $z=0$ for batch A. Values are in assay units.

Batch	$x$ (years)	Observed value
A	0	5
A	1	7
A	2	9
B	1	10
B	2	12
B	3	14

Fit $y=\beta_0+\beta_1x+\gamma z+\epsilon$. In batch A, the values follow $5+2x$, so the intercept and slope are 5 assay units and 2 assay units per year. At $x=1$, the batch-A fitted value is $5+2(1)=7$, whereas the batch-B value is 10; the difference is 3 units. At $x=2$, the values are 9 and 12, again a difference of 3. Hence $\hat{\gamma}=3$, and the fitted model is $y=5+2x+3z$.

Subtracting the batch term leaves batch A unchanged and reduces each batch B value by 3: the adjusted batch-B values are 7, 9, and 11 at $x=1,2,3$. The biological slope remains 2 assay units per year. The batches overlap at $x=1$ and $x=2$, so the batch offset is distinguishable from the linear covariate trend in this example. The exact fit also agrees with the assumed additive model.

Failure Modes and Alternatives

The most serious failure is complete confounding. For example, if all controls are in batch A and all cases are in batch B, the case indicator and batch indicator are identical. The design matrix is rank deficient, so the separate case and batch coefficients are not identifiable. A correction cannot determine which part of the observed difference is technical and which part is biological.

Worked Example: When Batch and Biology Cannot Be Separated

Suppose a hypothetical study observes two controls in batch A with values 10 and 12, and two cases in batch B with values 20 and 22. With $x=1$ for case and $z=1$ for batch B, every observed sample has $x=z$. The model is $y=\beta_0+\beta_1x+\gamma z+\epsilon$, but its fitted mean is $\beta_0+(\beta_1+\gamma)x$. Only the sum $\beta_1+\gamma$ is identified.

The group means differ by $21-11=10$ units. That 10-unit contrast could be represented as a 10-unit biological effect and zero batch effect, as zero biological effect and a 10-unit batch effect, or as many other combinations adding to 10. The data do not choose among them. Subtracting an estimated batch contribution here would require an unsupported assumption or external information, such as technical controls measured across batches.

Other warning signs include unequal biological slopes across batches, influential observations, and residual patterns inconsistent with the fitted mean structure. If the batch effect changes with a biological covariate, the additive model is misspecified; a batch-by-covariate interaction may be needed, although adjustment then requires a clearly defined target for what should be retained. If batches differ in variance rather than only mean, subtracting batch offsets does not equalize that variance. These are reasons to inspect design and model adequacy, not reasons to automatically apply a more complicated correction.

The main alternatives differ in their assumptions and targets. A batch-only regression that omits known biological covariates can attribute biological differences to batch and remove them. Stratifying by batch avoids pooling measurements across batches but does not itself produce a common corrected scale. Covariate-preserving linear adjustment estimates separate feature-wise batch terms without borrowing information across features. Empirical-Bayes ComBat instead uses cross-feature information to stabilize batch estimates under its modeling assumptions. No method can resolve complete batch-biology confounding from the same data without additional assumptions or information.

Practical Sequence and Common Pitfall

Define the target and covariates.
Specify which biological effects must remain, the batch variable, and the feature scale on which the additive model is being fit.

Check overlap and rank.
Tabulate biological groups by batch, examine covariate ranges across batches, and verify that the combined design can estimate the intended coefficients.

Fit and inspect the joint model.
For each feature, estimate biological and batch coefficients together; check whether additive mean shifts are plausible and whether unusual residuals dominate the fit.

Subtract only the batch contribution.
Compute $y_j^{\mathrm{adj}}=y_j-Z\hat{\gamma}_j$, retaining the fitted biological contribution and residual variation.

Verify what was retained.
Confirm that biological contrasts of interest remain on the adjusted scale and that the correction has not been interpreted as proof that all batch structure is gone.

A common mistake is to fit batch and biology jointly, then replace each measurement by only its fitted biological value $X\hat{\beta}_j$. That operation also discards the residual variation, not merely the batch contribution. Covariate-preserving batch adjustment instead subtracts $Z\hat{\gamma}_j$ from the observed value. A different mistake is to subtract fitted biological effects while aiming to preserve biology; that reverses the intended operation.

The adjusted data are estimated, not newly observed measurements. Their values depend on the fitted model, coding choice, and samples used. Standard downstream analyses that treat adjusted values as if batch coefficients were known exactly may understate uncertainty. Use the adjustment to address the stated technical nuisance effect, and retain the design and modeling decisions needed to interpret later analyses.

Key takeaway: Fit technical batch and biological covariates jointly, check that their effects are identifiable, and subtract only the fitted batch term. This preserves the fitted biological effects under the model; it cannot resolve confounding or repair a misspecified batch structure.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

In the model $y=\beta_0+\beta_1x+\gamma z+\epsilon$, what quantity is subtracted to adjust for batch while retaining the fitted biological covariate effect?
Why can’t a model distinguish a case effect from a batch effect when every case is in batch B and every control is in batch A?
If the fitted batch coefficient is 2.5 assay units and a sample is in the reference batch, what batch contribution is subtracted under reference coding?
Give one reason that subtracting batch-specific fitted means alone is not the same as covariate-preserving linear adjustment.

Solutions. (1) Subtract $z\hat{\gamma}$, or more generally $Z\hat{\gamma}$, from the observed value; do not subtract $X\hat{\beta}$. (2) The case and batch indicators coincide, so the design cannot separately identify their coefficients; only their sum is estimable. (3) Zero, because the reference batch has $z=0$; the fitted contrast is subtracted only for samples in the non-reference batch. (4) Replacing observed values with fitted means discards residual variation as well as batch contributions, whereas covariate-preserving adjustment subtracts only the fitted batch term.