Why Use Negative Controls?
A batch adjustment can make expression values look more comparable across processing groups, but that appearance does not by itself show that technical differences have been removed. A negative control provides an outcome expected not to respond to the biological condition of interest. If that control still varies with batch after adjustment, it is evidence of residual technical structure. Examples include a carefully selected set of genes with no expected condition response, or technical measurements such as spike-in abundance or library-quality metrics.
The diagnostic is informative only in relation to its assumptions. A control gene must genuinely be unaffected by the biological condition, and its batch response must be relevant to the technical effects of interest. A technical measurement can reveal a processing difference without showing how every gene responds to that difference. Most importantly, observing little residual batch association in controls does not create missing comparisons between conditions and batches.
What Is Being Compared?
Let \(R_{gi}\) denote the residualized or adjusted value for control gene \(g\) in biological sample \(i\). Let \(B_i\) denote that sample's batch and \(C_i\) its condition. A typical question is whether the adjusted control values still differ by batch after accounting for condition where the study design supports that comparison. The unit of replication should usually be the biological sample, not the individual cell: cells from one sample share biological and technical history and are not independent biological replicates.
A simple prespecified summary for sample \(i\) is a control score \(S_i\), such as the average adjusted expression over a fixed set of control genes. This reduces many measurements to one diagnostic outcome per sample. The genes and any transformations or weights should be selected before inspecting which choices give the most reassuring result. Otherwise, selecting controls after seeing their batch patterns can conceal residual signal.
The model used to compare batches must respect the study design. If condition and batch overlap sufficiently, a sample-level model can include both. If every sample in one condition belongs to one batch and every sample in another condition belongs to another batch, their effects are aliased: the data alone cannot separate them. The earlier material on estimability and rank deficiency gives the formal design-matrix perspective; here the practical point is that a control diagnostic does not manufacture the absent comparisons.
A Sample-Level Batch Contrast
For a control score observed in two batches, estimate the residual batch difference as the difference between the sample means. A Welch standard error allows the sample variances to differ. The confidence interval describes uncertainty in the mean control-score contrast under the sampling model; it is not a probability statement about a fixed parameter.
Here \(\bar{S}_b\), \(s_b\), and \(n_b\) are the sample mean, sample standard deviation, and number of independent biological samples in batch \(b\). A test of \(H_0:\delta=0\) asks whether the data provide evidence against a zero mean contrast for this score. The test does not ask whether all genes are free of batch effects.
Worked Example: Testing a Residual Control-Score Difference
Suppose a hypothetical balanced study has both conditions represented in each of two batches. After the chosen adjustment, the prespecified control score (in log-expression units) is summarized once per independent biological sample. The four values in batch 1 are 0.10, 0.12, 0.08, and 0.10; those in batch 2 are 0.28, 0.30, 0.26, and 0.28. Assume the adjustment and sampling design justify comparing the batches after accounting for condition, and use Welch's two-sample procedure.
The means are \(\bar{S}_1=0.10\) and \(\bar{S}_2=0.28\), so the estimated batch-2-minus-batch-1 contrast is \(0.28-0.10=0.18\) log-expression units. In each batch, the deviations from the mean are \(0, 0.02, -0.02, 0\), whose squared sum is \(0.0008\). Thus the sample variances are
The standard error is
The test statistic is \(t=0.18/0.01155=15.59\). The Welch degrees of freedom are 6 for these equal sample variances and sample sizes. The two-sided \(p\)-value is below 0.001, and a 95% interval using \(t_{0.975,6}=2.447\) is
Under the stated assumptions, this is strong evidence that the chosen control score retains a positive mean batch difference after adjustment. The interval estimates the mean contrast for this score; it does not quantify every gene's batch effect or establish that condition effects are recoverable.
Summaries Can Hide Heterogeneity
An average across control genes is useful when a shared technical shift is plausible, but it can mask controls that behave differently. A control set may contain one subgroup with a large positive batch contrast and another with little or an opposite contrast. Therefore, inspect both the prespecified aggregate and gene-level patterns. Gene-to-gene variability is descriptive unless a suitable model accounts for dependence and the intended gene-level population; treating a list of genes as independent biological replicates is not a substitute for biological replication.
Worked Example: A Control Average Masks Uneven Responses
In a hypothetical dataset, four prespecified control genes have adjusted batch-2-minus-batch-1 contrasts of 0.20, 0.18, 0.02, and -0.01 log-expression units. Their average contrast is
The mean is positive, but the four values are not close to a common shift: two are near 0.2, while one is near zero and one is slightly negative. This pattern warns against reporting only the aggregate score as if it represented every control. It could reflect heterogeneous technical responses, an unsuitable control gene, or noisy estimates. The contrasts alone do not distinguish these explanations. Replicate-level uncertainty and technical context are needed before making a stronger claim.
A related diagnostic uses a technical measurement, such as a spike-in summary, rather than a gene set. Its advantage is that it may be less directly affected by endogenous biological expression. Its limitation is that the measurement may not capture the technical mechanisms affecting endogenous genes. The measurement must also be interpreted in its own units: a residual shift in library complexity, for example, is evidence about that metric, not a direct estimate of a gene-expression correction.
Worked Example: Interpreting a Technical Control
Suppose a hypothetical study records a prespecified spike-in score for each biological sample. After adjustment, the sample-level means are 1.40 units in batch 1 and 1.46 units in batch 2. The estimated contrast is \(1.46-1.40=0.06\) units. If the estimated standard error is 0.04 units, then \(t=0.06/0.04=1.50\). With sufficient degrees of freedom, this is not compelling evidence of a nonzero mean contrast at a two-sided 5% level; for example, a 95% interval using a critical value of 2 is \(0.06\pm2(0.04)=[-0.02,0.14]\) units.
The interval includes zero, but it also includes positive differences as large as 0.14 units. The result is therefore inconclusive about whether a practically important residual shift remains unless an acceptable shift was specified in advance and the interval is sufficiently narrow relative to it. It does not prove that the batches are equivalent, nor that endogenous genes have no residual batch effects.
Why a Clean Control Does Not Resolve Confounding
Negative controls can diagnose residual technical association, but they cannot by themselves establish condition-effect identifiability. Consider a target gene measured in two groups where condition A occurs only in batch 1 and condition B only in batch 2. Let the observed group means be 5 and 7. In a model with baseline \(\mu\), condition effect \(\beta\), and batch effect \(\gamma\), the difference in observed means is only
The data are consistent with \((\beta,\gamma)=(2,0)\), with \((0,2)\), and with infinitely many other pairs whose sum is 2. A negative control that shows no residual batch association does not tell us which decomposition applies to the target gene. If one additionally assumes that the target gene has exactly the batch response measured by the control, that assumption may supply information; it is an extra transfer assumption, not a conclusion established by a reassuring control result.
Diagnostics, Alternatives, and Limitations
A careful diagnostic checks more than whether a \(p\)-value exceeds 0.05. Examine the size and uncertainty of the control contrast, the distribution across control genes, and whether the result changes materially across reasonable prespecified summaries. Check that the controls are not condition-responsive, that sample labels and batch assignments are correct, and that the adjustment did not use information in a way that removes the signal the diagnostic is meant to detect. For a sample-level analysis, confirm that independent biological samples—not cells—provide the replication.
The main alternatives answer related but different questions. Cross-tabulations and design-matrix checks reveal whether batch and condition overlap sufficiently for contrasts to be estimated. Directly modeling batch effects for each gene examines gene-specific patterns, but remains limited by the design and model. Technical quality metrics can flag specific processing differences, while negative-control genes can probe endogenous expression patterns under their control assumptions. These approaches complement one another; none converts absent batch-condition comparisons into observed evidence.
| Approach | What it can show | Main limitation |
|---|---|---|
| Negative-control genes or a control score | Residual batch association for specified controls | Depends on control validity and may not represent all genes |
| Technical measurements | Residual association in a measured technical property | The property may not capture all relevant expression effects |
| Design overlap and estimability checks | Whether the requested condition contrast is supported by the design | Do not diagnose every gene-specific technical response |
| Gene-specific batch modeling | Variation in estimated batch responses across genes | Cannot separate aliased effects without additional assumptions or data |
A non-significant test is not evidence of no residual association when the sample is small or the interval is wide. Conversely, a statistically detectable difference may be too small to matter practically. Define a meaningful residual shift in context when possible, and report estimates and intervals as well as tests. The diagnostic also inherits uncertainty from normalization, control selection, and any fitted correction; treating estimated residual values as error-free can make uncertainty appear smaller than it is.
Common Mistake
Wrong reasoning: “The negative-control genes have no significant batch effect after correction, so the biological condition effect is now identifiable and the correction succeeded.” This confuses failure to reject a control contrast with evidence of equivalence, generalizes from controls to all genes, and ignores whether the study design separates condition from batch.
Correction: State what control was tested, estimate its residual contrast and uncertainty, and limit the conclusion to that control under its assumptions. Separately assess whether the target condition contrast is estimable from the design. If batch and condition are aliased, say so even when the controls look reassuring.
Explain why the gene or technical measurement is expected not to respond to condition, and what technical signal it can plausibly represent.
Use biological samples as the replication units and determine whether the batch contrast can be adjusted for condition.
Report the control contrast, its uncertainty, and gene-level heterogeneity rather than relying on a single significance threshold.
Describe evidence for residual association in the controls, then report target-effect identifiability separately.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A study has many cells but only two independent biological samples per batch. Why is it generally inappropriate to treat every cell as an independent replicate for a batch test?
- A prespecified control score has an estimated residual batch contrast of 0.03 units with a 95% interval from -0.20 to 0.26. What can be concluded?
- Condition A occurs only in batch 1 and condition B only in batch 2. Does a control with no detected batch difference make the target condition effect identifiable? Explain.
- Why should a control-score average be accompanied by gene-level inspection?
Solutions. (1) Cells from the same biological sample are correlated and share sample-level history; treating them as independent can understate uncertainty and inflate evidence. The biological samples, or a suitable sample-level analysis, should supply replication. (2) The estimated contrast is small, but the interval is wide and includes meaningful differences in either direction. The result is inconclusive, not proof of no residual association or equivalence. (3) No. The target condition and batch effects remain aliased in that design. The control result concerns the control under its assumptions and does not create the missing comparisons or establish the target's batch response. (4) Opposing or heterogeneous gene responses can cancel in the average, so the score may hide important residual patterns or unsuitable controls.