What a Results Statement Can—and Cannot—Claim
When condition and batch overlap incompletely, a model may separate their associations to some extent. When they are perfectly confounded, the data do not distinguish a condition effect from a batch effect. Reporting must preserve that distinction: a difference between the observed condition groups is a valid description of those groups, but it is not automatically an estimate of what condition would have changed if batch had been held constant.
This matters in single-cell RNA sequencing because a large number of cells does not create independent replication of a condition assigned at the sample level. For a gene-level comparison, it is often useful to summarize expression for each biological sample and cell population, then regard those sample summaries—not the individual cells—as the replicates for the condition comparison. The reporting issue remains even with careful summaries: if every control sample is in one batch and every treated sample is in another, adding cells cannot separate the two effects.
The Model and the Aliasing Problem
Consider one gene and one cell population. Let \(Y_s\) be its expression summary for biological sample \(s\), \(C_s\) indicate condition, and \(B_s\) indicate batch. A simple additive model is
Here \(\mu\) is a baseline, \(\beta\) is the condition contrast at a fixed batch, \(\gamma\) is the batch contrast at a fixed condition, and \(\varepsilon_s\) is sample-level error. For this interpretation, the model assumes the chosen expression scale is appropriate, the mean contributions of condition and batch are additive, and the error terms have mean zero. Standard model-based uncertainty calculations additionally require a suitable error distribution and independence across biological samples, or an alternative method that accounts for dependence. These assumptions do not repair a rank-deficient design.
If all controls are in batch A and all treated samples are in batch B, then the two binary indicators are identical: \(C_s=B_s\) for every sample. Substitution gives
Only the sum \(\beta+\gamma\) is informed by the condition-group contrast. For any number \(a\), replacing \(\beta\) by \(\beta+a\) and \(\gamma\) by \(\gamma-a\) leaves every fitted mean unchanged. Thus there is no unique data-based value of \(\beta\). A software package may omit a coefficient, select a constraint, or return a numerical result under a particular parameterization; none of those actions makes the separate condition effect identified by the observations.
Worked Example: Showing Why the Coefficient Is Not Identified
Worked Example: Showing Why the Coefficient Is Not Identified
Suppose a hypothetical study has three control samples, all processed in batch A, with gene-expression summaries \(4.8,\ 5.0,\ 5.2\), and three treated samples, all processed in batch B, with summaries \(6.7,\ 6.9,\ 7.0\). The summaries are on a log-expression scale.
The control mean is \((4.8+5.0+5.2)/3=15.0/3=5.000\). The treated mean is \((6.7+6.9+7.0)/3=20.6/3=6.867\). The observed treated-minus-control contrast is therefore
In the model above, the fitted group contrast corresponds to \(\beta+\gamma=1.867\), not to \(\beta=1.867\) alone. For example, \(\beta=1.867,\ \gamma=0\) and \(\beta=0,\ \gamma=1.867\) both reproduce the same difference in fitted group means. So do other pairs with the same sum. The data identify the observed group difference, but do not identify how much of it is due to condition versus batch.
A defensible result is: “Mean expression was 1.867 log-expression units higher in the treated samples than in the control samples. Because all treated samples were processed in batch B and all controls in batch A, the observed difference cannot be separated into condition and batch effects.” The statement describes the contrast and the design limitation without assigning the contrast to one cause.
Report the Estimable Contrast and Its Uncertainty
A non-identifiable condition effect does not mean that every numerical comparison is unavailable. The observed difference between the groups can still be summarized, and uncertainty for that descriptive comparison can be reported if its sampling assumptions are stated. The interval concerns the contrast between the observed groups; it does not become a batch-adjusted condition interval merely by being calculated from biological replicates.
Worked Example: A Descriptive Interval Does Not Remove Confounding
In another hypothetical study, four control samples in batch A have summaries \(4.9,\ 5.0,\ 5.0,\ 5.1\), and four treated samples in batch B have \(6.6,\ 6.7,\ 6.7,\ 6.8\). Treat the biological samples as independent and use a two-sample \(t\) interval for the difference in group means. This calculation is illustrative: it assumes the sample-level observations are approximately normal within groups, and it does not adjust for batch.
The means are \(\bar{x}_{C}=20.0/4=5.000\) and \(\bar{x}_{T}=26.8/4=6.700\), so the observed contrast is \(6.700-5.000=1.700\) log-expression units. In each group, the sum of squared deviations from its mean is \(0.020\), giving \(s_C^2=s_T^2=0.020/(4-1)=0.006667\). The standard error for the independent-group difference is
With equal variances, the pooled \(df\) is \(4+4-2=6\). Using \(t^*=2.447\) for a two-sided 95% interval gives
Under the stated sampling assumptions, this interval describes uncertainty in the difference between the observed group means. It is not a 95% interval for a condition effect separated from batch: condition and batch remain perfectly aliased. A suitable report would name the interval’s target and immediately note this design limitation.
Document the Limitation Precisely
A clear report lets readers reconstruct which comparison was made and see why stronger claims are unsupported. For a gene-level result, give the expression scale, the sample or cell-population summary used, the condition groups, the number of biological samples per group, and the batch allocation. State whether the contrast is descriptive or comes from an identifiable model. If an interval or test is reported, identify the target contrast and the assumptions behind its uncertainty calculation.
Be specific about the design diagnosis. “Batch may be a concern” is vague when the metadata show that every control is in one batch and every treated sample is in another. Prefer wording such as “condition was perfectly aliased with processing batch: all control samples were in batch A and all treated samples were in batch B.” If a model matrix is rank deficient, say so and describe which effects cannot be separately estimated. Do not suggest that a corrected embedding, larger cell count, or coefficient selected by software resolves this sample-level lack of information.
| What the result supports | What it does not establish |
|---|---|
| The observed expression summaries differ between the recorded condition groups. | The difference is due to condition rather than batch. |
| A descriptive group contrast and, under stated assumptions, its sampling uncertainty. | A separately estimated condition coefficient when the design is rank deficient. |
| A sensitivity calculation conditional on explicitly stated batch-effect assumptions. | That those assumed batch effects are true or learned from the aliased data. |
Worked Example: State the Assumption Behind a Sensitivity Calculation
Worked Example: State the Assumption Behind a Sensitivity Calculation
Return to the observed contrast of \(1.867\) log-expression units in the first example. Suppose, for illustration, an investigator considers batch B to raise expression relative to batch A by an amount between \(0.5\) and \(1.2\) units. This is an externally specified range, not an estimate from the perfectly confounded data. Under the additive model, the condition contrast is the observed difference minus the assumed batch contrast:
At \(\gamma=0.5\), the implied condition contrast is \(1.867-0.5=1.367\). At \(\gamma=1.2\), it is \(1.867-1.2=0.667\). Across the assumed range, the implied condition contrast ranges from \(0.667\) to \(1.367\) log-expression units.
This range is conditional on the additive model and the chosen range for \(\gamma\); it is not a confidence interval and does not show that the condition effect is positive under every plausible batch effect. For example, if the assumed batch effect were \(2.0\), the implied condition contrast would be \(1.867-2.0=-0.133\). A report must make the assumed range and its source explicit and keep these scenario results separate from estimates supported directly by the design.
Alternatives, Diagnostics, and Common Pitfalls
The main alternatives answer different questions. A descriptive comparison summarizes the observed condition groups but combines condition and batch under perfect aliasing. A covariate-adjusted model can estimate a separate condition contrast only when the design contains information to distinguish the condition and batch terms and the model assumptions are credible. A sensitivity analysis explores consequences of specified assumptions; it does not identify effects from the aliased observations. The strongest remedy for a future study is to collect samples so that conditions are represented across batches, which creates the comparisons needed to separate the effects.
Before finalizing a result, check sample-level cross-tabulations of condition and batch, and inspect the design matrix for duplicate or linearly dependent columns. A model that cannot return a coefficient, drops a term, or gives unstable estimates is a warning to investigate the design rather than a reason to select a preferred output. Also verify that the reported sample count refers to independent biological samples, not just cells. These checks diagnose the basis for the reporting limitation; they do not supply the missing comparisons.
- Overclaiming causation: “Condition increased expression by 1.867 units” assigns the entire contrast to condition. Replace it with a description of the observed groups and the aliasing limitation.
- Treating more cells as more independent samples: more cells can improve measurement of each sample’s expression summary, but cannot create overlap between condition and batch.
- Calling a software-selected coefficient an estimate: a constraint or parameter-dropping rule can yield a numerical output without making the separate effect data-identified. Report the rank deficiency and estimable contrast instead.
- Presenting a sensitivity range as a confidence interval: scenario values reflect chosen assumptions about batch effects. Label them as conditional calculations, not sampling uncertainty.
The limitation is not merely technical wording. If batch and condition are perfectly aliased, a gene may be reported as differing between the observed groups, but the study cannot support the claim that changing condition, rather than changing batch, would produce that difference. Honest reporting preserves the useful descriptive result while marking the boundary of what the design can establish.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Every control is processed in batch A and every treated sample in batch B. What expression contrast can be described, and which model effects cannot be separated?
- A study reports 10,000 cells per condition but has one biological sample per condition, each in a different batch. Does the cell count resolve condition-batch aliasing? Why?
- An observed contrast is \(2.4\) units. Under an explicitly assumed batch contrast of \(0.7\) units, what condition contrast follows from the additive model?
- What is wrong with describing a 95% interval for an observed group difference as a 95% interval for the condition effect when condition and batch are aliased?
Solutions. (1) The observed group difference can be described; the separate condition and batch coefficients cannot be identified because their indicators coincide. (2) No. Cells are not independent biological replicates of the sample-level condition assignment, and the two conditions still occur in different batches. (3) The conditional calculation is \(2.4-0.7=1.7\) units; it depends on the assumed batch contrast and is not identified by the aliased data alone. (4) The interval targets the observed group contrast under its sampling assumptions. It does not separate that contrast into condition and batch effects, so it cannot be interpreted as an interval for the separately estimated condition effect.