Why the Design, Not the Software, Chooses the Strategy
A batch-correction method cannot compensate for a study design that fails to distinguish biological signal from technical variation. The practical question is not simply whether samples differ by batch. It is whether the study contains enough overlap, pairing, or other information to estimate the biological contrast of interest separately from technical effects.
Let \(y_{gi}\) denote the transformed expression for gene \(g\) in sample \(i\), after an appropriate normalization and transformation, and let \(b_i\) denote its known batch. A common working model is
where \(x_i\) encodes a biological condition, \(\beta_g\) is its effect for gene \(g\), \(\gamma_{g b_i}\) is the batch effect, and \(\epsilon_{gi}\) is residual variation. This additive model assumes that the condition and batch effects are adequately represented on the chosen expression scale and that residuals are suitably modeled for the intended analysis. If biological effects vary across batches, an interaction may be needed, and a single additive batch adjustment is inadequate.
Normalization addresses differences in measurement scale or composition, such as sequencing depth or other distributional features. It does not, by itself, remove every technical difference between processing batches. Conversely, batch adjustment is not a substitute for suitable normalization. The earlier material on normalization and batch diagnostics provides the background for assessing both steps.
Match the Strategy to the Design
| Study situation | Usually appropriate | Key qualification |
|---|---|---|
| Scale or composition differences, with no evidence or design reason for a separate batch effect | Normalization alone, followed by diagnostics | Normalization does not guarantee that other technical effects are absent. |
| Known batches with biological groups represented across batches | Include batch as a covariate in the analysis model | The design matrix must identify the biological and batch effects separately. |
| Repeated or paired samples, with each pair processed under the same batch conditions | Block on subject or pair; model batch as well if it is separately estimable and relevant | Within-pair comparisons control stable subject-level differences; they do not solve a batch change that occurs between paired measurements. |
| Known batches and a need for an adjusted expression matrix | Consider ComBat with biological variables protected in the model | Requires adequate overlap and a suitable batch model; it cannot resolve perfect confounding. |
| Important technical variation not captured by recorded batch labels | Consider latent-factor methods, such as SVA or RUV | Factors must represent nuisance variation rather than the biological signal of interest. |
For inference, a known batch is often best included directly as a covariate in the model used to estimate the biological effect. This makes the adjustment explicit and retains uncertainty in the model. ComBat is particularly relevant when an adjusted expression matrix is needed for downstream analyses, but it also requires a defensible design and decisions about which biological variables to preserve. Latent-factor methods address unmeasured variation; they should not be selected merely because a known-batch adjustment is inconvenient.
Worked Example: A Balanced Design With a Known Batch
Worked Example: A Balanced Design With a Known Batch
Suppose a hypothetical study measures one gene on a transformed expression scale. There are two batches, A and B, and both control and treated samples occur in each batch. The observed values are:
| Batch | Control values | Treated values |
|---|---|---|
| A | 8, 9 | 10, 11 |
| B | 10, 11 | 12, 13 |
The control means are \(8.5\) in A and \(10.5\) in B; the treated means are \(10.5\) in A and \(12.5\) in B. The observed treatment contrast is \(10.5-8.5=2\) in A and \(12.5-10.5=2\) in B. The batch contrast is \(10.5-8.5=2\) among controls and also \(12.5-10.5=2\) among treated samples.
Thus the additive model estimates a treatment effect of \(2\) expression-scale units and a B-versus-A batch effect of \(2\) units. Because the two conditions are equally represented in both batches, the unadjusted overall means also give a treatment contrast of \(11.5-9.5=2\). In this balanced example, normalization alone would not create a treatment-effect bias from this additive batch difference. However, adjustment for batch can still be useful for modeling and may improve precision when batch contributes residual variation. The exact benefit depends on the full data and error structure; these values alone do not establish it.
A covariate model is the direct choice for estimating the treatment contrast while accounting for the known batch. ComBat could be considered if the goal were an adjusted expression matrix across many genes, with treatment protected as a biological covariate. Since both conditions occur in both batches, the design has the overlap needed to distinguish their main effects. Diagnostics should still check whether residual batch patterns remain or whether the additive model is questionable.
Confounding: What the Data Can and Cannot Identify
A design is confounded when biological condition and batch are so aligned that their effects cannot be separately estimated. Consider two batches with controls only in A and treated samples only in B. With batch A as the reference, set \(\gamma_{gA}=0\). The model then gives a control mean of \(\mu_g\) and a treated mean of \(\mu_g+\beta_g+\gamma_{gB}\). The observed difference is therefore
Only the sum of the treatment effect and the B-versus-A batch effect is identified. Different values of \(\beta_g\) and \(\gamma_{gB}\) can produce the same observed difference. Adding a batch covariate, applying ComBat, or estimating latent factors cannot recover separate effects from this design without additional information or assumptions.
Worked Example: Perfect Batch-Condition Confounding
In a hypothetical experiment, control samples are measured only in batch A, with values \(8\) and \(9\), and treated samples only in batch B, with values \(12\) and \(13\). The observed group means are
The observed difference is \(12.5-8.5=4\) expression-scale units. Under the model with batch A as reference, that difference estimates \(\beta_g+\gamma_{gB}=4\), not \(\beta_g=4\). For example, a treatment effect of \(1\) and a batch effect of \(3\), or a treatment effect of \(3\) and a batch effect of \(1\), are both compatible with the same group means. No calculation from these observations alone distinguishes those explanations.
The appropriate conclusion is that the combined treated-versus-control difference is \(4\) units, but the separate biological treatment effect is not identified. More sophisticated adjustment cannot create the missing overlap.
Blocking and Paired Samples
Blocking uses the design structure directly: samples are compared within blocks that share relevant characteristics, such as subject, matched pair, or experimental unit. In a paired study, a model can include a subject-specific intercept. If both measurements from a subject are processed in the same batch, differencing the paired observations removes that subject intercept and any stable batch contribution shared within the pair.
Subtracting the first measurement from the second gives \(y_{g,s,2}-y_{g,s,1}=\beta_g+\epsilon_{g,s,2}-\epsilon_{g,s,1}\). This removes the shared subject-specific level \(\mu_{gs}\). The interpretation depends on the design: if the two measurements are processed in different batches, the difference may also include a batch change.
Worked Example: Blocking by Subject
Suppose four hypothetical subjects each have a baseline and follow-up measurement, and both measurements for a subject are processed in the same batch. The within-subject follow-up-minus-baseline differences for one gene are \(1, 2, 2,\) and \(1\) expression-scale units. The paired estimate is
The sample standard deviation of the four differences is \(s_d=\sqrt{[(1-1.5)^2+(2-1.5)^2+(2-1.5)^2+(1-1.5)^2]/3}=\sqrt{1/3}=0.5774\). The standard error of the paired mean difference is \(SE=s_d/\sqrt{4}=0.2887\) expression-scale units. The estimated average within-subject change is \(1.5\) units. This calculation uses the four pair differences as its observations; it does not treat the eight measurements as independent.
Because each pair shares its batch, a stable batch offset cancels in the difference. If baseline samples were all run in A and follow-up samples all in B, batch and time would be confounded and this paired contrast would combine both effects. Pairing controls subject-level variation, not every technical difference.
ComBat and Latent-Factor Methods
ComBat models known batches using gene-specific batch effects, commonly allowing both location and scale differences, and uses empirical Bayes estimation to stabilize batch estimates across genes. Its use assumes that batch effects are meaningfully represented by that model and that biological variables to be preserved are included appropriately. With inadequate overlap, protected biological effects and batch effects may be poorly identified; perfect confounding remains unresolvable. ComBat is therefore a candidate for a suitable design, not a repair for any design.
Latent-factor methods such as SVA and RUV target structured nuisance variation that is not adequately described by recorded batch labels. SVA estimates surrogate variables associated with unwanted variation; RUV uses control genes or samples when suitable controls are available. Their validity depends on the factors capturing nuisance structure rather than the biological contrast of interest. Inspect estimated factors against known conditions and batches, and compare biological conclusions with and without their inclusion. A factor that tracks the condition of interest may represent biology, confounding, or both; automatic removal is not justified by correlation alone.
Normalization alone is preferable when the problem is adequately addressed by the normalization model and diagnostics do not indicate a separate batch structure requiring adjustment. For differential expression from RNA-seq counts, use an appropriate count-model approach that includes relevant design variables rather than treating a batch-corrected transformed matrix as a replacement for that model.
Diagnostics, Limitations, and Common Mistakes
Before and after adjustment, examine expression distributions and diagnostic plots, including whether known batches or conditions remain structured in sample-level summaries. Check that the design matrix is full rank for the effects being estimated, that each relevant condition is represented across batches when separate effects are desired, and that the proposed model reflects the actual pairing or repeated-measure structure. After adjustment, verify both that unwanted batch structure has been reduced and that the biological signal of interest has not been erased.
Methods can fail in different ways. Normalization alone can leave batch patterns that are unrelated to library scale. A covariate model can be unstable or unidentified in a confounded design. Blocking is ineffective if samples are assigned to blocks incorrectly or if the paired measurements experience different unmodeled technical shifts. ComBat can overadjust, underadjust, or obscure biological differences if its design variables are inappropriate. Latent-factor estimates can absorb real biology or miss nuisance variation. No single diagnostic plot proves that adjustment is correct.
Common mistake: “The treated samples were all measured in batch B, so I will run ComBat and then compare treated with control.” This reasoning assumes an algorithm can distinguish the treatment effect from the batch effect despite perfect confounding. It cannot: with batch A as reference, the observed contrast is \(\beta_g+\gamma_{gB}\). The correction requires an untestable allocation between those terms unless independent information or new data supplies it.
The main limitation of a strategy-selection rule is that real studies can combine several problems: scale differences, known batch effects, repeated measurements, and unmeasured technical factors. Start from the experimental design and target estimand, fit only identifiable effects, and use diagnostics to evaluate the assumptions. A technically adjusted result is not automatically a biologically valid result.
State which biological comparison is wanted and on what expression scale.
Tabulate condition by batch and identify pairs, subjects, and other blocks.
Determine whether the desired biological and technical effects can be separated from the observed design.
Use covariates or blocking for modeled effects, consider ComBat for a suitable adjusted matrix, and consider latent factors for unmeasured nuisance structure.
Check Your Understanding
Answer each question, then compare with the solutions below.
- Both conditions are represented in every batch, with similar numbers in each. What is the main advantage of including batch in the analysis model rather than relying only on normalization?
- Controls occur only in batch A and treated samples only in batch B. What quantity is identified by the observed group difference under the additive model with batch A as reference?
- Why can a paired analysis control stable subject differences but fail to separate condition from batch if every baseline is in A and every follow-up is in B?
- When might ComBat be considered instead of only adding batch as a covariate to a gene-wise analysis?
- What is one important risk of adding latent factors to a model?
Solutions. (1) The model estimates the biological contrast while accounting explicitly for known batch variation, and may improve precision when batch contributes variation. Normalization alone does not necessarily remove that structure. (2) The group difference identifies \(\beta_g+\gamma_{gB}\), not the separate treatment and batch effects. (3) Pair differencing removes stable subject-level effects, but if condition and batch change together, the difference includes both changes. (4) ComBat may be considered when a batch-adjusted expression matrix is needed for downstream analyses, provided the design has adequate overlap and biological variables to preserve are specified. (5) Estimated factors may capture and remove genuine biological signal, especially when they track the condition of interest.