Batch-Condition Confounding as an Identifiability Problem

Why Batch and Condition Must Be Separated

A difference in measured expression between two groups of cells can reflect biological condition, technical batch, or both. For example, if control cells were processed on one day and treated cells on another, a measured difference in a gene cannot automatically be attributed to treatment. The central question is not initially which adjustment procedure to use. It is whether the study design contains enough relevant comparisons to distinguish the effects being discussed.

This is an identifiability question. An effect is identifiable when the available design and assumptions determine a unique value for the target quantity. If multiple different condition effects are equally compatible with the observed data, the condition effect is not identified. This differs from an effect that is identified but estimated imprecisely: in the latter case, the data allow a unique target, but sampling variability leaves substantial uncertainty about its value.

Specify the Estimand Before Comparing Groups

Let \(Y\) denote the measured expression of a particular gene on a chosen scale, such as normalized log expression. Let \(b\) index batch and \(c\) index condition. Write \(\mu_{bc}\) for the expected value of \(Y\) among the relevant biological units in batch \(b\) and condition \(c\). The unit used to define this expectation must be stated: it might be a donor-level pseudobulk measurement for a specified cell type, rather than an individual cell treated as an independent biological replicate.

A condition contrast within batch \(b\) is \(\Delta_b=\mu_{b1}-\mu_{b0}\), comparing condition 1 with condition 0 in that batch. A batch contrast within condition \(c\) is \(\Gamma_c=\mu_{1c}-\mu_{0c}\). The contrast is in units of the chosen expression scale. If \(Y\) is log base 2 expression, for example, a difference of 1 corresponds to a two-fold ratio on the original scale, subject to the interpretation of the normalization.

When the condition effect may differ across batches, there is no single condition contrast until a target is chosen. One possible target is a weighted average, \(\Delta=\sum_b w_b\Delta_b\), where the nonnegative weights sum to 1 and represent the batches or target population of interest. When a common effect across batches is scientifically plausible, that common contrast may instead be the target. These choices are substantive; the observed cell counts do not automatically determine which target is scientifically meaningful.

How Complete Confounding Prevents Identification

Consider a simple additive description of the expected expression: the condition contributes an amount \(\Delta\), and the batch contributes an amount \(\Gamma\). This is a simplifying assumption, not a fact guaranteed by the data. When every control sample is in batch A and every treated sample is in batch B, the observed difference combines the two contributions. If the two observed group means are \(m_0\) and \(m_1\), then:

The data determine the sum, not its separate components. For any proposed value \(d\) of the condition effect, a batch contribution equal to \(m_1-m_0-d\) gives the same observed difference. This is a structural lack of identification, not merely a small-sample problem. More cells measured under the same allocation can make the two group means more precise, but cannot reveal which part of their difference came from condition rather than batch.

The additive description also makes an assumption: that a single batch contribution and a single condition contribution are meaningful over the comparisons being made. If condition effects vary by batch, the target must instead specify which within-batch contrasts are to be combined. No calculation can make a contrast available in a batch where one of its conditions was never observed.

When an Effect Is Identifiable but Uncertain

A condition contrast is directly informed when both conditions are represented in the same batch, allowing a within-batch comparison. This does not by itself guarantee a reliable estimate. There may be few independent biological replicates, high biological variability, or measurement problems. Those limitations affect uncertainty or validity; they are conceptually different from the absence of any comparison that separates condition and batch.

Partial Overlap and the Target of Inference

A study need not have every condition represented in every batch for every possible condition contrast to be informed. What matters is which comparisons are observed and which target is sought. If both conditions occur in some batches, the condition differences in those batches can be computed. Extending those results to batches without both conditions requires additional assumptions; it is not supplied by the unobserved comparisons themselves.

Design Checks, Alternatives, and Limits

Before interpreting a condition comparison, tabulate the biological replicates by condition and batch. A cross-tabulation quickly reveals whether the study has within-batch condition comparisons, whether some combinations are absent, and whether apparent overlap is based on independent donors or only on many cells from a few donors. Inspect the allocation records as well as the cell-level metadata: mislabeled, missing, or inconsistently defined batch labels can make the apparent design misleading.

The main alternatives address different design situations. If both conditions occur within batches, one can estimate within-batch contrasts and combine them for an explicitly defined target. If condition and batch are completely aligned, the honest options are to obtain new data with overlapping allocation, report only the observed combined contrast, or state that a condition effect cannot be separated without untestable external assumptions. Restricting analysis to batches with both conditions can make a narrower target estimable, but it does not automatically answer a question about all batches or all samples.

Computational normalization or integration may reduce technical variation under its assumptions, but it cannot create an observed comparison that the design lacks. A method may produce adjusted values in a completely confounded study; those values do not establish that the condition effect has been recovered. Any separation then depends on modeling assumptions or external information, not on the confounded data alone.

This distinction has practical consequences for single-cell studies. A large number of cells can describe the measured cell populations in detail, but the number and allocation of biological replicates govern the evidence for a condition contrast. Also, a contrast for a selected cell type and expression scale is not automatically a contrast for all cells, all donors, or absolute transcript abundance. State the unit, scale, target population, and comparison before interpreting the result.

Common Pitfall

A second warning is to call any difference across batches a batch effect without specifying the condition being held fixed. A batch contrast requires a comparison at the same condition, just as a condition contrast requires a comparison at the same batch or an explicitly justified combination of within-batch contrasts. When those comparisons are absent, labels alone do not resolve the ambiguity.