Why Adjustment Needs Diagnostics
Batch adjustment aims to reduce technical differences among samples while retaining biological differences relevant to the analysis. These goals can conflict: an adjustment may remove some real biological variation if the signal being “corrected” also tracks a biological variable. Conversely, an adjustment can leave systematic batch differences in the data. A third result that can look alarming is a negative adjusted value, even when the original measurements were nonnegative counts.
These outcomes are not diagnosed by a single plot or threshold. The interpretation depends on the measurement scale, the study design, and what the adjustment actually did. The earlier material on expression transformations and batch diagnostics provides the relevant background: in particular, a value on a log or centered scale is not interchangeable with a count.
A Model for Separating the Three Problems
For one gene, consider the additive model
Here \(Y_i\) is the transformed expression for sample \(i\), \(C_i\) is a biological condition indicator, and \(B_i\) is a batch indicator. The coefficient \(\beta\) is the condition contrast at a fixed batch, and \(\gamma\) is the batch contrast at a fixed condition. In this simple model, an adjustment that subtracts the batch contribution aims to remove \(\gamma B_i\) while retaining \(\beta C_i\).
This model is a diagnostic reference, not a guarantee that every gene follows the same pattern. Its interpretation assumes that the additive terms are appropriate on the analyzed scale, batch and condition are correctly recorded, and the design contains enough overlap to estimate their effects separately. A single batch coefficient may not describe batch-by-condition interactions or other nonlinear effects. If condition and batch are perfectly confounded, their separate effects are not identifiable from these data: adjustment cannot determine which part of their shared pattern is technical and which is biological.
A useful comparison is between the biological contrast before and after adjustment, together with batch contrasts before and after. Attenuation of a biological contrast is a warning, not proof of overcorrection: it might reflect removal of technical variation that had inflated the original contrast. Evidence is stronger when the attenuation is inconsistent with the design, known positive controls, or biological patterns supported by samples measured across batches.
Why a Correction Can Remove Biology
Suppose an adjustment subtracts a fitted nuisance component \(Z_i\hat{\gamma}\) from expression, producing \(Y_i^{\mathrm{adj}}=Y_i-Z_i\hat{\gamma}\). If the nuisance variable \(Z_i\) also varies with condition, the subtracted component can include condition-related variation. In contrast, a model that estimates condition and batch jointly can preserve a condition term in its fitted values, provided the design allows the two effects to be separated.
The basic issue can be seen directly from a condition contrast. For two samples or group means with conditions \(C=1\) and \(C=0\),
The adjusted contrast changes whenever the nuisance score differs by condition. If that score contains biological signal, the second term subtracts part of the biological contrast. If \(Z\) is a valid technical batch indicator and the design is balanced, its condition means may be equal, so this particular source of attenuation is absent.
Worked Example: Detecting Attenuation From a Contaminated Correction
Consider four hypothetical samples for one gene. Let \(C=0\) denote control and \(C=1\) denote treatment; \(B=0\) and \(B=1\) denote two batches. Suppose the observed transformed expression follows \(Y=10+2C+B\), but an adjustment subtracts the score \(Z=B+0.8C\), as if all of that score were technical.
| Condition | Batch | \(Y\) | \(Z=B+0.8C\) | \(Y^{\mathrm{adj}}=Y-Z\) |
|---|---|---|---|---|
| Control | 0 | 10 | 0 | 10 |
| Treatment | 0 | 12 | 0.8 | 11.2 |
| Control | 1 | 11 | 1 | 10 |
| Treatment | 1 | 13 | 1.8 | 11.2 |
Before adjustment, the treatment-minus-control contrast is \((12+13)/2-(10+11)/2=12.5-10.5=2\) expression units. After adjustment, it is \((11.2+11.2)/2-(10+10)/2=11.2-10=1.2\) units. The batch contrast before adjustment is \(11-10=1\) unit within either condition; after adjustment it is \(10-10=0\). Thus the adjustment removed the batch difference but also attenuated the biological contrast by \(2-1.2=0.8\) units, exactly the condition-related part of \(Z\). This is a constructed illustration of overcorrection, not evidence that every reduction in a contrast is harmful.
A practical diagnostic is to compare the relevant biological contrasts on the same scale before and after adjustment and ask whether the correction variable or fitted component tracks the biology. If batch and condition do not overlap in the design, the data alone cannot establish which part of a shared pattern should be retained.
Finding Batch Structure That Remains
Residual batch structure can be checked by examining adjusted expression against batch labels. For a gene, compare adjusted group means within condition where possible; across many genes, inspect distributions, sample-level summaries, and a low-dimensional representation such as principal components. A visible separation by batch after adjustment is a warning, but a plot alone does not establish that the remaining pattern is technical: batches may differ in biological composition.
Also check whether the adjustment was applied on the intended scale, whether batch labels and sample-to-batch assignments are correct, and whether the fitted batch pattern is plausibly additive. A method can leave structure because the effect is not well represented by its model, because the estimated effect is imprecise, or because some relevant technical source was not included. Conversely, apparent residual separation can arise from a biological variable distributed unevenly across batches. Compare within-condition samples or otherwise account for the study design before labeling a pattern residual batch effect.
Worked Example: Quantifying a Remaining Batch Difference
Suppose a hypothetical gene has the following transformed expression values, with one control and one treatment sample per batch. Batch A has values 8 and 10; Batch B has values 14 and 16. The condition contrast is 2 units in both batches. An adjustment subtracts estimated batch offsets of \(-1.5\) for A and \(+1.5\) for B.
The adjusted values in Batch A are \(8-(-1.5)=9.5\) and \(10-(-1.5)=11.5\). In Batch B they are \(14-1.5=12.5\) and \(16-1.5=14.5\). The adjusted batch means are \((9.5+11.5)/2=10.5\) and \((12.5+14.5)/2=13.5\). Their difference is \(13.5-10.5=3\) units, so a batch difference remains.
The treatment-minus-control contrast is \(11.5-9.5=2\) in Batch A and \(14.5-12.5=2\) in Batch B. The adjustment has retained that contrast in this example, even though it has not removed the batch separation. The conclusion is therefore specific: the adjusted data still show a 3-unit batch difference for this gene, while this calculation does not show attenuation of the within-batch condition contrast. With only four samples, these means are descriptive; they do not establish whether the remaining pattern is systematic across genes or statistically precise.
When Negative Adjusted Values Are Expected
Many adjustment procedures subtract an estimated batch contribution or center values around a reference. Subtraction can produce values below zero even if the input was nonnegative. On a log scale, zero is a reference determined by the transformation and its offset, not a universal lower bound for expression. A negative log-scale or centered adjusted value does not mean a negative number of RNA molecules.
The interpretation changes if the intended data object is a raw count matrix. Raw counts are nonnegative integers; a negative value is not a valid count and should not be passed to a count model as though it were one. Adjusted continuous values can be suitable for some downstream descriptive or modeling tasks, but they are not reconstructed raw counts. Check both the input scale and the output scale before interpreting the sign.
Worked Example: Interpreting a Negative Log-Scale Value
Suppose a hypothetical sample has log-base-2 expression \(Y=0.4\), and the estimated batch contribution to subtract is \(0.7\) log-base-2 units. The adjusted value is
This is negative, but it is not a negative count. On the original ratio scale, \(2^{0.4}\approx1.320\) before adjustment, while \(2^{-0.3}\approx0.8123\) after adjustment, relative to the scale's reference. These calculations explain the sign; they do not recover a raw read count. If the analyzed values were centered log expression, a negative value simply indicates expression below the centering reference.
The appropriate checks are to confirm the scale and transformation, inspect the distribution of adjusted values, and verify that the operation was intended for those data. A few negative values on a centered or log scale may be entirely expected. Negative values in a purported raw-count matrix indicate a scale or data-handling problem.
A Diagnostic Workflow and Its Limits
Record whether the input and adjusted values are counts, transformed expression, or centered residual-like values. Interpret zero and negative values only in relation to that scale.
For biological contrasts supported by the design, calculate or plot the same contrasts on the same scale. Investigate substantial attenuation, especially when adjustment variables track condition.
Compare adjusted values across batches within biological groups when possible, and examine sample-level plots and gene-level summaries for persistent batch association.
Verify sample labels, overlap of conditions across batches, the fitted model assumptions, and whether the adjustment output is being used as the type of data it represents.
These diagnostics complement one another. A principal-component plot may reveal a batch pattern but cannot say by itself whether that pattern is technical or biological. A preserved average contrast does not guarantee that every gene's biological variation was preserved. A lack of visible separation does not prove that every batch effect was removed. Controls, replicate structure, and design information help interpret the evidence, but no diagnostic can resolve perfect batch-condition confounding from the expression data alone.
Adjustment should also be distinguished from including batch as a covariate in a downstream model. A covariate model estimates the biological contrast while accounting for batch in that model; an adjusted expression matrix is a transformed data product that may be used for other purposes. They are not automatically interchangeable, and negative adjusted values or a change in a plotted contrast do not alone determine which analysis is appropriate. The central task here is to recognize what the diagnostic evidence says—and what it cannot say—about the adjusted values.
Common Mistakes
- Calling every smaller biological contrast overcorrection. A smaller contrast can result from removing technical variation that had inflated the unadjusted value. Compare the design, adjustment components, and relevant controls before concluding that biology was removed.
- Assuming batch separation proves residual batch effect. If biological groups are unevenly distributed across batches, separation can reflect biology. Check comparisons within biological groups and the degree of overlap in the design.
- Treating negative adjusted expression as negative counts. On log or centered scales, negative values can be valid. Check the scale; do not treat adjusted continuous values as raw counts.
- Assuming no visible batch pattern proves success. Plots have limited resolution, and an absence of obvious separation is not proof that all technical structure is gone.
- Trying to resolve perfect confounding computationally. If all controls occur in one batch and all treated samples in another, the data do not identify separate batch and condition effects without additional information or assumptions.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A biological contrast is 1.8 units before adjustment and 0.6 units after. What can and cannot be concluded from this change alone?
- Adjusted means are 5.2 in Batch A and 5.3 in Batch B within each condition. What does this suggest, and what limitation remains?
- A log-base-2 adjusted value is \(-1\). Is it a negative count? What is its exponentiated value?
- All control samples were processed in Batch A and all treated samples in Batch B. Can the expression data alone separate batch and condition effects? Explain.
Solutions. (1) The contrast has attenuated by \(1.8-0.6=1.2\) units. This is a diagnostic warning, but by itself it does not establish overcorrection: technical inflation, biological signal removal, or both could explain the change. (2) The within-condition batch difference is \(5.3-5.2=0.1\) units, suggesting little remaining mean separation in this comparison. It does not prove that batch structure is absent for all genes or that the difference is technically interpretable. (3) It is not a negative count if the value is on an adjusted log scale. Its exponentiated value is \(2^{-1}=0.5\) relative to the scale reference. (4) No. Batch and condition are perfectly confounded, so their separate effects are not identifiable from these data alone.