Why estimability is a question about the contrast
When batch and biological condition are aliased, the design matrix may not identify every model coefficient separately. That does not mean every comparison is lost. A particular linear combination of coefficients can still be uniquely determined by the observed design. The relevant question is therefore not simply “Is the model rank deficient?” but “Is the contrast I want estimable under this design?”
Consider a gene-wise linear mean model for a response such as normalized expression on a chosen analysis scale. Write the expected response vector as \(\mathbf{X}\boldsymbol{\beta}\), where \(\mathbf{X}\) is the design matrix and \(\boldsymbol{\beta}\) contains the model coefficients. A contrast is a specified linear combination \(c^\mathsf{T}\boldsymbol{\beta}\). Its estimability is a property of the design and the contrast, not of the observed expression values.
This definition assumes a specified design matrix and the stated linear mean model. It does not require a particular error distribution to decide estimability. A probability model for the errors, such as independent errors with common variance, may be needed for standard errors and tests, but it cannot turn a non-estimable contrast into an identifiable one.
From aliasing to constraints on contrast weights
A dependency among design columns creates a constraint on which coefficient combinations are determined. For example, if the batch indicator column equals the condition indicator column, the data reveal only the sum of their coefficients. They cannot separate the two contributions. A contrast is estimable only if it respects that ambiguity: changing coefficients along an aliased direction must leave the contrast unchanged.
The row-space test gives a direct calculation. If the rows of \(\mathbf{X}\) are \(x_1^\mathsf{T},\ldots,x_n^\mathsf{T}\), seek numbers \(a_1,\ldots,a_n\) satisfying
If a solution exists, then \(c^\mathsf{T}\boldsymbol{\beta}=\sum_i a_i(\mathbf{X}\boldsymbol{\beta})_i\), so the contrast is a combination of fitted means represented by the observed design rows. If no solution exists, the contrast asks for a coefficient combination that the observed rows cannot distinguish.
This also explains why the test is about the contrast rather than the coefficient vector as a whole. Rank deficiency prevents unique recovery of all coefficients, while some contrasts can still be represented by observed rows and hence have unique values.
Worked Example: A Perfectly Confounded Two-Group Design
Suppose a hypothetical experiment measures normalized expression for one gene in two groups. All condition A cells were processed in batch 1, and all condition B cells were processed in batch 2. Use an intercept, a batch-2 indicator, and a condition-B indicator, so that \(\boldsymbol{\beta}=(\beta_0,\gamma,\tau)^\mathsf{T}\). The design rows for the two observed groups are \((1,0,0)\) and \((1,1,1)\).
The row combinations have the form \(a(1,0,0)+b(1,1,1)=(a+b,b,b)\). Thus a contrast \(c^\mathsf{T}\boldsymbol{\beta}\) is estimable exactly when its batch and condition weights are equal: \(c=(c_0,k,k)\).
The pure condition contrast \(\tau\) has \(c=(0,0,1)\). Its batch and condition weights differ, so it is not estimable. In contrast, \(\gamma+\tau\), with \(c=(0,1,1)\), is estimable because it equals the observed difference between the two group means.
For concrete hypothetical group means, suppose condition A in batch 1 has mean expression \(2.0\) units and condition B in batch 2 has mean \(3.4\) units. The observed difference is \(3.4-2.0=1.4\) units, which estimates \(\gamma+\tau\). It does not determine \(\tau\) alone: both \((\gamma,\tau)=(0,1.4)\) and \((0.9,0.5)\) produce the same observed difference. The batch-adjusted condition effect cannot be separated from the batch effect in this design.
Contrasts that survive partial aliasing
Rank deficiency does not necessarily destroy every biological comparison. In a design with multiple conditions, two conditions measured in the same batch can have a difference that cancels a shared batch contribution. The row-space calculation identifies such comparisons without claiming that all condition effects are separately available.
Worked Example: A Within-Batch Condition Contrast
Consider a hypothetical study with three conditions, A, B, and C. Condition A is observed only in batch 1; conditions B and C are observed only in batch 2. Use the model \(\mu=\beta_0+\gamma I(\text{batch 2})+\tau_B I(\text{condition B})+\tau_C I(\text{condition C})\), with A as the reference condition. The rows for the observed groups are:
The contrast between condition B and condition C is \(\tau_B-\tau_C\), with \(c=(0,0,1,-1)\). It is the difference between two design rows:
Therefore \(\tau_B-\tau_C\) is estimable: the shared intercept and batch term cancel. Suppose hypothetical group means are \(1.2\), \(2.0\), and \(1.4\) expression units for A, B, and C, respectively. The estimable within-batch contrast is \(2.0-1.4=0.6\) expression units.
By contrast, \(\tau_B\) alone is not estimable. Its vector \((0,0,1,0)\) cannot be written as a combination of these rows: the only row with a nonzero \(\tau_B\) weight is \(x_B\), which also contributes a nonzero batch weight. The data determine the B-versus-C difference, but do not separate each condition’s coefficient from the batch contribution and reference.
A general justification and a practical test
The row-space criterion follows directly from what the design can distinguish. If \(c^\mathsf{T}=a^\mathsf{T}\mathbf{X}\), then for any coefficient vector,
Consequently, any two coefficient vectors that yield the same mean vector \(\mathbf{X}\boldsymbol{\beta}\) also yield the same value of this contrast. Conversely, if \(c^\mathsf{T}\) is not in the row space of \(\mathbf{X}\), finite-dimensional linear algebra gives a coefficient change \(d\) with \(\mathbf{X}d=0\) but \(c^\mathsf{T}d\ne 0\). Then \(\boldsymbol{\beta}\) and \(\boldsymbol{\beta}+d\) give the same mean vector, while their contrast values differ by \(c^\mathsf{T}d\). The contrast is not uniquely determined.
In practice, first write the coefficient vector and the exact contrast of interest under the chosen coding. Then inspect the design dependencies or solve \(a^\mathsf{T}\mathbf{X}=c^\mathsf{T}\). For a small design, this can be done by hand as in the examples. For a larger design, rank-revealing QR or singular-value methods can expose dependencies, after which the contrast can be checked against the row space. Numerical rank decisions should use a tolerance appropriate to the scale and precision of the design.
A change in reference category or other full-rank reparameterization changes the coefficient coordinates and the contrast vector, but not whether the underlying scientific comparison is estimable. The question must be translated correctly into the new parameterization before applying the test.
Worked Example: Testing a Contrast from the Observed Group Means
Return to the three-condition design above. Consider the mean expression difference between the average of conditions B and C and condition A. In coefficient form, this is \(\gamma+(\tau_B+\tau_C)/2\), with contrast vector \(c=(0,1,1/2,1/2)\).
The contrast is the corresponding combination of design rows:
Thus it is estimable. With hypothetical group means \(A=1.2\), \(B=2.0\), and \(C=1.4\) expression units, its value is \((2.0+1.4)/2-1.2=1.7-1.2=0.5\) expression units. This is the observed contrast of those groups; it is not the pure average condition effect \((\tau_B+\tau_C)/2\), which omits the batch contribution and is not estimable here.
Diagnostics, alternatives, and limitations
A rank-deficient fit may still return a coefficient vector, but software must use a convention such as a generalized inverse, omit a coefficient, or impose a constraint to produce that output. The resulting individual coefficient values can depend on that convention. Before interpreting a reported condition coefficient, check whether the scientific contrast is estimable in the original design. A finite-looking standard error or a printed coefficient does not, by itself, establish estimability.
The main alternatives answer different questions. Reparameterizing the model can make identifiable combinations explicit, such as \(\gamma+\tau\) in the perfectly confounded example, but cannot create information about the separate effects. Dropping a batch term may produce a condition coefficient, but under confounding it attributes the combined difference to condition and changes the model assumption; it does not identify a batch-adjusted condition effect. Restricting analysis to a subset with relevant within-batch comparisons can support some contrasts, but changes the target and may reduce precision.
Estimability is also distinct from precision. A contrast can be estimable but have a large standard error if the relevant comparison is weakly supported or noisy. Conversely, collecting more observations in exactly the same aliased groups can improve precision for the observed group means without separating the batch and condition effects. The design must contain the comparisons needed for the target contrast.
Check Your Understanding
For each question, determine what the design supports. Solutions are included for checking.
- In the two-group perfectly confounded design, is \(2\gamma+2\tau\) estimable? What observed quantity does it represent?
- In the three-condition design, is \(\tau_B-\tau_C\) estimable? Give a row-combination justification.
- In that same design, can \(\tau_B\) be estimated separately from batch? Explain using its contrast vector.
- If a generalized inverse returns a numerical value for a non-estimable contrast, does that resolve the identifiability problem?
- Would adding many more cells to the existing perfectly confounded groups make the separate batch and condition effects identifiable?
Solutions. (1) Yes. Its vector is \((0,2,2)\), twice the estimable vector \((0,1,1)\), and it represents twice the observed difference between condition B in batch 2 and condition A in batch 1. (2) Yes: \(x_B^\mathsf{T}-x_C^\mathsf{T}=(0,0,1,-1)\), so the contrast is a difference of observed group means. (3) No. Its vector \((0,0,1,0)\) cannot be formed from the design rows without also introducing a batch contribution; the B coefficient is not separated from the other model terms. (4) No. The reported value depends on the convention used to select among coefficient vectors that fit the same mean, rather than being uniquely determined by the design. (5) No. More cells in the same groups can improve precision for observed group means, but the batch and condition indicators remain aliased, so their separate effects are still not identifiable.