When Correction Removes More Than Batch Variation
Batch correction is intended to reduce technical differences among processing batches while preserving expression differences associated with biological condition. These goals conflict when the batches contain different condition compositions. If nearly all control cells were processed in one batch and nearly all treated cells in another, a difference attributed to batch may also be the condition effect of interest. A correction procedure that removes the batch-associated component can therefore remove genuine biology as well.
The key issue is not whether corrected cells appear better mixed. It is whether the corrected expression values preserve the condition contrast that the design can support. When batch and condition are completely aligned, the observed data alone cannot determine how much of their difference is technical and how much is biological. When there is some overlap, adjustment can be possible, but a correction that ignores condition can still distort the contrast.
A Model for Separating the Two Sources of Difference
For a single gene, let \(Y_i\) be the normalized expression for cell \(i\), \(C_i\) be an indicator equal to 1 for the condition of interest and 0 otherwise, and \(b_i\) identify the batch. A simple additive model is
Here, \(\mu\) is a reference mean, \(\beta\) is the condition-associated expression difference, \(\gamma_{b_i}\) is the batch effect, and \(\varepsilon_i\) is residual variation. The model assumes that the condition effect is the same across batches, that batch effects are additive shifts for this gene, and that the residuals have mean zero conditional on condition and batch. It also assumes the expression scale is suitable for this additive description. These are modeling assumptions, not guarantees supplied by a correction algorithm.
The condition effect \(\beta\) can be estimated separately from batch effects only if the observed design contains enough overlap to distinguish them. In particular, at least some batches must include both conditions, or there must be a connected pattern of condition overlap across batches that supports the comparison. When condition is determined by batch, the columns representing condition and batch are redundant: multiple combinations of \(\beta\) and \(\gamma_b\) describe the same fitted values. The condition contrast is then not identifiable from these data.
Why Blind Batch Offsets Can Contain Biology
A condition-blind correction may estimate a batch offset from the difference between each batch's overall mean and a grand mean. That estimate is valid as a batch effect only if the condition composition does not contribute to the difference in batch means, or if the estimation procedure otherwise accounts for condition. To see the source of the problem, write \(q_b=P(C_i=1\mid b_i=b)\), the treated proportion in batch \(b\). Under the additive model, the expected batch mean is
Thus the expected difference between batch means A and B is
The second term is biological signal arising from different condition proportions. A procedure that treats the entire batch-mean difference as technical variation subtracts that component along with the actual batch shift.
Worked Example: Complete Confounding and an Erased Difference
Suppose a hypothetical gene has normalized expression values 4.0 and 4.2 in two control cells from batch A, and 10.0 and 10.2 in two treated cells from batch B. There are no control cells in batch B and no treated cells in batch A. The control mean is \((4.0+4.2)/2=4.1\), the treated mean is \((10.0+10.2)/2=10.1\), and the observed condition difference is \(10.1-4.1=6.0\) expression units.
The grand mean is \((4.0+4.2+10.0+10.2)/4=7.1\). A correction based on batch means estimates offsets of \(4.1-7.1=-3.0\) for batch A and \(10.1-7.1=3.0\) for batch B. Subtracting each offset gives corrected values 7.0 and 7.2 in both batches. The corrected control and treated means are both 7.1, so the corrected condition difference is \(7.1-7.1=0\) expression units.
This correction creates perfect overlap in the corrected values, but it does not establish that the original difference was technical. The same data are consistent with a 6.0-unit condition effect and no batch effect, a 6.0-unit batch effect and no condition effect, or intermediate combinations. There is no within-batch condition comparison to resolve the ambiguity.
Partial Overlap Does Not Make Every Correction Safe
With condition overlap, a condition effect may be estimable, but the way batch offsets are calculated still matters. In particular, the overall mean difference between batches is not generally the batch effect when treated proportions differ. A model that includes both condition and batch can use within-batch contrasts to separate the effects, provided the additive assumptions are reasonable and the design has adequate overlap. A blind batch-mean correction does not make that separation.
Worked Example: Imbalance Attenuates a Condition Contrast
Consider a hypothetical gene measured in two batches. In batch A, three control cells have mean expression 10 and one treated cell has expression 12. In batch B, one control cell has expression 14 and three treated cells have mean expression 16. These means follow the additive pattern \(\mu=10\), a batch B shift of 4 expression units, and a condition effect of 2 expression units.
The batch A mean is \((3(10)+1(12))/4=10.5\). The batch B mean is \((1(14)+3(16))/4=15.5\). With equal batch sizes, the grand mean is \((10.5+15.5)/2=13.0\), so centering each batch on the grand mean gives offsets of \(-2.5\) for A and \(2.5\) for B. Subtracting these offsets adds 2.5 to every A value and subtracts 2.5 from every B value.
After correction, the control mean is \((3(12.5)+1(11.5))/4=12.25\). The treated mean is \((1(14.5)+3(13.5))/4=13.75\). The corrected condition difference is \(13.75-12.25=1.50\) expression units, smaller than the 2.00-unit difference present within each batch. The batch mean difference was 5.00 units, not the true 4.00-unit batch shift: the additional 1.00 unit came from the 0.50 difference in treated proportions multiplied by the 2.00-unit condition effect.
If instead the batch effect is estimated while accounting for condition, the within-batch differences are \(12-10=2\) in A and \(16-14=2\) in B. The condition contrast supported by both batches is therefore 2.00 expression units under the additive model. The assumptions are plausible for this constructed example because the same condition difference occurs in each batch; in real data, that pattern must be assessed rather than presumed.
Why the Visual Result Is Not the Statistical Result
A low-dimensional embedding or other visualization can help reveal batch separation, but it is not a direct test that a correction preserved the condition effect. If condition and batch are aligned, removing the feature that distinguishes batches may also remove the feature that distinguishes conditions. Cells can then mix in the plot precisely because the relevant signal was removed. Conversely, residual visual separation does not by itself prove that correction failed: biological differences, technical variation, or both may contribute to the displayed geometry.
Worked Example: Perfect Mixing Is Not Effect Recovery
Return to the complete-confounding example. The control values in batch A average 4.1, and the treated values in batch B average 10.1, for an observed contrast of 6.0 expression units. After batch-mean centering, the values in both conditions are 7.0 and 7.2, so their corrected means are equal. A plot of these corrected values would show no separation by batch or condition.
The calculation demonstrates visual mixing, but it cannot tell whether the biological contrast was 0, 6.0, or another value. The correction has imposed a result; it has not supplied the missing comparison. Valid effect recovery would require information that separates condition from batch, such as condition overlap across batches, or external assumptions or measurements that identify the batch effect. The plot alone provides none of that information.
A Short Justification for the Identifiability Limit
In the complete-confounding example, set \(C_i=0\) for every cell in A and \(C_i=1\) for every cell in B. The expected difference in group means under the additive model is
For any constant \(d\), replacing \(\beta\) by \(\beta+d\) and \(\gamma_B-\gamma_A\) by \(\gamma_B-\gamma_A-d\) leaves this sum unchanged. Therefore the expected group difference is unchanged by that substitution. Since the observed design supplies no other condition-by-batch comparison, the data cannot distinguish these parameter values. A correction that sets the corrected difference to zero selects one decomposition; it does not prove that this decomposition is the true one.
Diagnostics, Alternatives, and Limits
Before interpreting corrected expression, examine the condition-by-batch structure and the condition contrasts within batches. Useful checks include:
- Cross-tabulate condition and batch. Empty cells in the table indicate missing comparisons; severe imbalance warns that batch summaries may incorporate condition composition.
- Compare within-batch contrasts. If both conditions occur in a batch, calculate their expression difference there. Contrasts that disagree substantially may challenge the common-effect assumption or indicate noisy estimates.
- Inspect before-and-after condition contrasts. Record how much a proposed correction changes the condition difference, not only whether batch labels mix better in an embedding.
- Check the adjustment model. A batch-adjusted condition estimate requires overlap and a defensible model for the condition and batch effects. It cannot recover a comparison absent from the design.
The main alternatives answer different questions. Leaving expression uncorrected avoids mechanically removing a condition-associated component, but leaves technical batch variation in the data. Blind batch centering can be useful for a descriptive display when its limitations are clear, but it may distort condition contrasts under imbalance. A model that includes both condition and batch can estimate a condition contrast when the design supports it and its assumptions are reasonable. No method can identify separate condition and batch effects from complete confounding using the same expression measurements alone.
Correction also has gene-specific consequences: a shift that is appropriate for one gene need not describe another. That issue is separate from the central design problem here. Even a perfectly implemented procedure cannot determine which aligned component is biological without identifying information. A corrected embedding should therefore not be treated as evidence that a biological effect has been preserved or that confounding has been resolved.
Check Your Understanding
Try each question, then compare your reasoning with the solutions.
- In a study where all controls are in batch A and all treated cells are in batch B, what combination of effects does the observed group difference identify?
- Two batches have treated proportions \(q_A=0.2\) and \(q_B=0.7\). If the condition effect is 3 expression units, how much of the batch-mean difference is contributed by differing condition composition?
- In the imbalanced two-batch example above, why does a within-batch comparison preserve a 2-unit contrast while blind batch-mean centering yields 1.5 units?
- A corrected embedding shows control and treated cells mixed. What additional evidence is needed before concluding that the biological condition effect was recovered?
Solutions. (1) It identifies \(\beta+\gamma_B-\gamma_A\), not the condition effect and batch difference separately. (2) The contribution is \(\beta(q_B-q_A)=3(0.7-0.2)=1.5\) expression units. (3) Each within-batch condition difference is 2 units, while the raw batch means also reflect unequal condition proportions; centering those means subtracts some condition signal, giving a pooled corrected contrast of 1.5 units. (4) The design must support a condition comparison separately from batch, for example through condition overlap across batches, and the effect should be assessed with an appropriate model or within-batch contrasts. Visual mixing alone is insufficient.