Assumptions That Permit Batch Adjustment

Why Batch Adjustment Needs Assumptions

Batch adjustment aims to account for technical differences between batches while retaining biological differences associated with condition. The distinction cannot be made from the expression values alone: it depends on a model that says how batch and condition relate to those values, and on a design with enough overlap to distinguish their effects. An adjustment can produce a complete-looking matrix even when those requirements are not met.

Consider one gene at a time, and let \(y_{ig}\) be the transformed expression for gene \(g\) in cell \(i\), measured on a specified scale such as normalized log expression. Let \(b_i\) denote the batch and \(c_i\) the biological condition of that cell. The basic additive model is:

Here \(\mu_g\) is a reference level, \(\tau_{g,c_i}\) is the condition effect, \(\beta_{g,b_i}\) is the batch effect, and \(\varepsilon_{ig}\) is the remaining deviation. One condition and one batch are reference categories with effects set to zero; this constraint makes the parameterization interpretable. The central assumptions are that condition effects are shared across batches, batch shifts are additive on the chosen scale, and the observed design contains adequate overlap between batches and conditions.

What the Assumptions Mean

Shared condition effects. For two conditions, say control and treated, the model says the treated-minus-control difference is a single value \(\tau_g\), not a different value in each batch. This is a substantive modeling assumption, not a consequence of having fitted a model. It is most plausible when the samples represent comparable biological populations and the condition is expected to have a consistent effect across batches. If the condition contrast varies by batch, one common adjusted contrast may conceal that variation.

Additive batch shifts. The model assumes the batch contribution is a constant offset on the scale of \(y_{ig}\), regardless of condition. For example, a batch difference of 1.2 log-expression units is treated as 1.2 units in both control and treated cells. This is scale-dependent: an additive shift on a log scale corresponds approximately to a multiplicative factor on the original expression scale. It does not mean that batches differ by the same number of raw counts. A correction must be interpreted on the scale on which its assumptions were stated.

Adequate overlap. To separate condition from batch, the data need comparisons that distinguish the two. A simple, strong sufficient design is for every batch being adjusted to contain observations from every condition being compared. Then each batch supplies a within-batch condition comparison, rather than relying only on differences between batches. If one condition appears only in one batch, the corresponding condition and batch effects cannot generally be separated without additional assumptions or information.

These statements concern expected expression. For inference, the remaining errors also need appropriate treatment. In particular, cells from the same donor or biological sample are not automatically independent experimental replicates. A model that treats thousands of cells from a few donors as independent can understate uncertainty, even if its mean structure is reasonable. Comparisons should also target a clearly defined cell population: a change in the proportion of cell types can change an overall mean without indicating a within-cell-type expression change.

A Short Justification for Within-Batch Comparisons

The key reason overlap helps is that subtracting condition means within the same batch removes the batch offset. Write \(m_{bc}\) for the expected expression of gene \(g\) in batch \(b\) and condition \(c\), suppressing the gene index for readability. Under the additive model:

Subtract the control expectation from the treated expectation within batch \(b\). Both the reference level and the batch offset cancel:

Thus, if the assumptions hold, each batch with both conditions targets the same condition contrast. Likewise, subtracting batch means within a condition removes the condition effect and isolates the batch difference. These cancellations justify adjustment only when the relevant comparisons are present and the additive structure is a reasonable description of the expected values.

Overlap and Unsupported Comparisons

Overlap is not merely a convenience for estimating a more precise adjustment. It determines whether the data contain information that separates the effects. If control is observed only in batch A and treated only in batch B, a difference between their means could be caused by condition, batch, or both. An additive model can assign numerical values to both effects only by imposing constraints or extra assumptions; the observed comparison itself cannot tell which explanation is right.

A design with some, but not all, batch-condition combinations observed may support some comparisons and not others. For the basic interpretation in this tutorial, a cautious practical requirement is to verify that each batch used for the contrast contains both conditions. This is a transparent sufficient condition for direct within-batch comparisons. It is not a claim that every incomplete design is mathematically unidentifiable; more complex designs require examining which contrasts the observed design actually supports.

Checking the Assumptions in Practice

Assumption checks should use the design and the data on the scale that will be analyzed. They cannot prove a model true, but they can reveal obvious incompatibilities or warn that an estimate depends on a fragile assumption.

• Inspect the design. Tabulate the number of biological samples and cells in each batch-condition combination. Identify empty combinations and distinguish cell counts from independent biological replicates.
• Compare within-batch contrasts. Where both conditions occur, estimate their difference separately by batch for a targeted gene or a suitable summary. Large discrepancies may reflect sampling noise, different cell composition, measurement problems, or a condition effect that is not shared. A single pooled contrast can hide the discrepancy.
• Inspect batch differences within condition. If batch shifts appear very different between conditions, a constant additive offset may be a poor description on the selected scale. Remember that apparent patterns can also result from noise, sparse observations, or changes in which cells are represented.
• Check the target population and scale. Ensure that comparisons refer to the same cell type or other defined population, and that the correction is being interpreted on the modeled expression scale. A shift appropriate for one scale need not be appropriate on another.
• Assess sensitivity. Ask whether the conclusion changes substantially under reasonable alternative specifications or after excluding a poorly supported batch. Such changes indicate that the result relies on modeling choices or limited overlap.

No visual mixing of cells after correction establishes that the biological condition effect was preserved. Likewise, similar corrected distributions do not demonstrate that the raw batch shifts were purely technical. Adjustment is a model-based operation; it is not an independent validation of the assumptions used to perform it.

Common Pitfall and Limits

Wrong reasoning: “The correction algorithm returned adjusted values, so the batch effect has been removed and the condition comparison is now valid.” The output only shows that an algorithm produced values. It does not show that condition and batch were separable, that shifts were additive, or that condition effects were shared.

Correction: State the target contrast and the scale, check that the design supplies relevant overlap, and assess whether the within-batch comparisons are compatible with a shared effect. Treat an unsupported comparison as unsupported rather than interpreting an adjusted number as recovered biology. Even when assumptions are plausible, the result is an association in the analyzed data, not by itself evidence that the condition caused the expression difference.

The additive model is useful because it gives a clear interpretation and a direct reason for adjustment: batch offsets cancel in within-batch condition comparisons. Its limitations follow from the same simplicity. It does not automatically handle condition-specific batch behavior, changing cell composition, unmeasured sample-level structure, or a mismatch between the model scale and the intended interpretation. More elaborate approaches can represent more structure, but they cannot manufacture missing comparisons or remove the need for a defensible design.