Estimable Biological Contrasts Under Confounding

Why Contrast Estimability Matters

A batch design can fail to identify every model coefficient and still identify some scientific comparisons. For example, if two biological groups are both measured in the same batch, their difference may be estimable even when another group's effect is inseparable from its batch. The relevant question is therefore not only whether the design matrix has full column rank, but whether the particular biological contrast being asked about is determined by the observed design.

This distinction is important in multi-omics analyses because biological effects are often reported as contrasts: a difference between conditions, a contrast between cell types, or a weighted combination of group effects. The earlier material on rank and connectivity checks for batch designs addresses design structure. Here, the target is a precise criterion for deciding whether a specified contrast is estimable.

Definition: Consider the fixed-design linear model \(Y=X\beta+\varepsilon\), where \(Y\) is an \(n\)-vector of responses, \(X\) is a known \(n\times p\) design matrix, and \(\beta\) is a \(p\)-vector of unknown coefficients. A linear contrast \(c^\mathsf{T}\beta\), for specified \(p\)-vector \(c\), is estimable if there is an \(n\)-vector \(a\) such that \(c^\mathsf{T}\beta=a^\mathsf{T}X\beta\) for every \(\beta\). Equivalently, \(c=X^\mathsf{T}a\), so \(c\) belongs to the row space of \(X\).

A linear contrast is a specified linear combination of model coefficients. Its weights are fixed by the scientific question; they are not chosen after seeing which coefficients happen to be estimable. For example, if a model's coefficients encode a reference-group mean and two group differences, the difference between the second and first non-reference groups is a particular combination of those coefficients.

The model's design matrix and the proposed contrast determine estimability. A probability distribution for the errors is not needed for this algebraic question. To use the model as a statistical description, the mean is assumed to be \(E(Y)=X\beta\); the usual linear-model error assumptions, such as independent errors with mean zero and a common variance, are relevant to uncertainty estimates and tests, but do not change the row-space criterion.

The Row-Space and Null-Space Criteria

Every row of \(X\) gives the coefficients of one observed linear predictor. If \(c^\mathsf{T}\) is a linear combination of those rows, then the target \(c^\mathsf{T}\beta\) is the same linear combination of observed means. In symbols, if \(c=X^\mathsf{T}a\), then

c^\mathsf{T}\beta = a^\mathsf{T}X\beta = a^\mathsf{T}E(Y).

That identity gives a direct interpretation: an estimable contrast can be recovered from the model's expected observations. If no such \(a\) exists, the observed mean structure does not determine the contrast.

The equivalent null-space criterion makes the reason for non-estimability explicit. The null space of \(X\) consists of all \(p\)-vectors \(v\) satisfying \(Xv=0\). Moving from \(\beta\) to \(\beta+v\) in such a direction leaves the expected data unchanged, because \(X(\beta+v)=X\beta\). A contrast is estimable exactly when none of these invisible changes alters its value.

c^\mathsf{T}\beta \text{ is estimable} \quad\Longleftrightarrow\quad c^\mathsf{T}v=0\text{ for every }v\text{ satisfying }Xv=0 \quad\Longleftrightarrow\quad c\in\operatorname{Row}(X).

The final equivalence is the fundamental theorem of linear algebra: the row space of \(X\) is the orthogonal complement of the null space of \(X\). Thus, a contrast fails the criterion if even one null-space direction changes it.

Practical test: To assess a contrast \(c^\mathsf{T}\beta\), check whether the equation \(X^\mathsf{T}a=c\) has a solution. Equivalently, find a basis for the null space of \(X\) and check that \(c^\mathsf{T}v=0\) for every basis vector \(v\). The rank of \(X\) alone does not answer whether a particular contrast is estimable.

Worked Examples

Worked Example: A Contrast Under Perfect Confounding

Suppose a hypothetical study measures a control group only in batch A and a treated group only in batch B. Let the model be

E(Y)=X\beta,\qquad \beta= \begin{pmatrix} \beta_0\\ \beta_T\\ \beta_B \end{pmatrix}, \qquad X= \begin{pmatrix} 1&0&0\\ 1&1&1 \end{pmatrix}.

The columns represent an intercept, treatment status, and batch B status. The two rows represent the control observations in batch A and the treated observations in batch B. The treatment and batch columns are identical, so \(X\) has rank 2 rather than 3.

The requested biological contrast is the treatment effect \(\beta_T\), with \(c=(0,1,0)^\mathsf{T}\). The row space is spanned by \((1,0,0)\) and \((1,1,1)\). No combination of these rows equals \((0,1,0)\): a combination with a nonzero treatment coordinate also has the same nonzero batch coordinate. Therefore, the treatment effect is not estimable.

The null-space check reaches the same conclusion. Take \(v=(0,1,-1)^\mathsf{T}\). Then \(Xv=(0,0)^\mathsf{T}\), but \(c^\mathsf{T}v=1\). The model cannot distinguish an increase of one unit in the treatment coefficient from a decrease of one unit in the batch coefficient.

For hypothetical observed group means of 5.2 and 8.1 expression units, the difference \(8.1-5.2=2.9\) expression units is estimable as the combined treatment-plus-batch difference, \(\beta_T+\beta_B\). It is not an estimate of \(\beta_T\) alone. Assigning the entire difference to treatment would require additional information or assumptions not supplied by this design.

Worked Example: A Biological Difference That Is Estimable

Consider three biological groups, A, B, and C, and two batches. In a hypothetical design, A and B are measured in batch 1, while C is measured in batch 2. Use A as the reference group and batch 1 as the reference batch. The model coefficients are \(\beta_0,\beta_B,\beta_C,\gamma\), where \(\gamma\) is the batch 2 effect. The observed group-by-batch design rows are

X= \begin{pmatrix} 1&0&0&0\\ 1&1&0&0\\ 1&0&1&1 \end{pmatrix}.

The rows correspond to A in batch 1, B in batch 1, and C in batch 2. The column for group C and the column for batch 2 are identical, so the model is rank deficient. Nevertheless, the contrast between B and A is \(c^\mathsf{T}\beta=\beta_B\), with \(c=(0,1,0,0)^\mathsf{T}\). This contrast is the second design row minus the first:

$$ (1,1,0,0)-(1,0,0,0)=(0,1,0,0). $$

Hence it is in the row space and is estimable. For a hypothetical set of two replicate values in each observed group-batch cell, suppose A has values 7 and 9, and B has values 11 and 13, all in normalized log-expression units. The sample means are \(\bar{x}_A=(7+9)/2=8\) and \(\bar{x}_B=(11+13)/2=12\). The contrast estimate is

\bar{x}_B-\bar{x}_A=12-8=4\text{ normalized log-expression units}.

The calculation compares groups measured in the same batch, so the shared batch shift cancels. In contrast, the C-versus-A biological effect has coefficient vector \((0,0,1,0)^\mathsf{T}\), which is not in the row space. The null vector \(v=(0,0,1,-1)^\mathsf{T}\) satisfies \(Xv=0\), while \((0,0,1,0)v=1\). Thus the C effect cannot be separated from the batch 2 effect.

Worked Example: Checking a Combined Contrast

Now suppose four biological groups are measured in a hypothetical design: A and B in batch 1, and C and D in batch 2. With A as reference and batch 1 as reference, let the coefficient vector be \(\beta=(\beta_0,\beta_B,\beta_C,\beta_D,\gamma)^\mathsf{T}\). The design rows for A, B, C, and D are respectively

(1,0,0,0,0),\quad (1,1,0,0,0),\quad (1,0,1,0,1),\quad (1,0,0,1,1).

The batch column equals the sum of the group C and group D columns, so the matrix is rank deficient. Consider the contrast \((B-A)+(D-C)\), whose coefficient vector is \(c=(0,1,-1,1,0)^\mathsf{T}\). It can be expressed as the difference of the B and A rows, plus the difference of the D and C rows:

\{(1,1,0,0,0)-(1,0,0,0,0)\} +\{(1,0,0,1,1)-(1,0,1,0,1)\} =(0,1,-1,1,0).

Therefore the combined contrast is estimable. If hypothetical group means are A = 5, B = 8, C = 10, and D = 13 normalized log-expression units, its estimate is

(8-5)+(13-10)=3+3=6\text{ normalized log-expression units}.

By comparison, C minus A is not estimable: its coefficient vector is \((0,0,1,0,0)^\mathsf{T}\). The vector \(v=(0,0,1,1,-1)^\mathsf{T}\) satisfies \(Xv=0\), but the C-minus-A contrast has inner product 1 with \(v\). This design therefore identifies the stated within-batch differences and their sum, but not an arbitrary comparison across batches.

How to Use the Criterion in Practice

For a specified target, a reliable analysis proceeds from the model's actual design matrix and the exact coefficient coding used to define the contrast. The following steps keep the scientific question separate from any software convention for handling rank deficiency.

Write the target as a coefficient contrast.
Specify the coefficient vector \(c\) for the biological quantity of interest, including its signs and weights.

Form the design matrix.
Use the rows corresponding to the observations actually included in the model and the same coding used to define \(\beta\).

Check the row-space condition.
Solve \(X^\mathsf{T}a=c\), or check whether \(c\) is orthogonal to a basis of the null space of \(X\).

Interpret the result in context.
If the condition holds, explain which observed means or combinations identify the contrast. If it fails, identify the null direction that changes the contrast without changing fitted means.

For a numerical computation, rank and row-space checks use a tolerance, especially when columns are nearly linearly dependent. Exact perfect confounding is an algebraic property; a numerical tolerance can also classify a very small singular value as effectively zero. The tolerance, scaling, and contrast should therefore be inspected rather than treated as automatic proof. This check is about identifiability, not a diagnostic of whether the linear model's mean assumptions fit the data.

If a contrast is estimable, any coefficient solution that reproduces the fitted means gives the same value for that contrast. A generalized inverse may return one particular coefficient vector even when many coefficient vectors produce the same fitted means. Its displayed individual coefficients do not by themselves show that those coefficients are uniquely identified or that a requested contrast is estimable. The contrast must be checked directly.

Common Pitfalls and Limits

Using full rank as the only criterion. A rank-deficient model can still estimate contrasts, as the within-batch comparisons above show. Check the proposed contrast, not just the coefficient vector as a whole.
Treating a printed coefficient as identified. Software can choose one solution among multiple coefficient vectors. A reported number does not establish that the corresponding scientific quantity is estimable.
Confusing a combined observed difference with a biological effect. Under perfect confounding, a treatment-plus-batch difference can be estimable even when neither component is separately estimable.
Changing coding without transforming the contrast. Reference levels and contrast coding change the coefficient coordinates. The scientific target remains the same, but its vector \(c\) must be rewritten in the new coordinates.
Interpreting estimability as causal identification or precision. Estimability says the model's expected observations determine a linear contrast. It does not establish that the model is scientifically adequate, that the contrast is causal, or that its estimate is precise.

A useful alternative to testing a contrast directly is to re-express the model in terms of observed cell means and identifiable combinations. This can make the estimable quantities easier to explain. It does not recover a biological effect that the design cannot distinguish from batch. Likewise, imposing a constraint such as setting one batch coefficient to zero selects a parameter representation; it does not supply the missing information needed to identify a confounded biological effect.

Check Your Understanding

Answer each question using the linear-model definition and criteria above.

In a design with \(X\) of rank less than \(p\), must every linear contrast be non-estimable? Explain.
If \(Xv=0\) and \(c^\mathsf{T}v\ne 0\), what does this imply about \(c^\mathsf{T}\beta\)?
In the three-group example, why is B minus A estimable even though the model matrix is rank deficient?
Suppose a model uses a different reference group. Does that necessarily change whether the same scientific contrast is estimable?

Solutions. (1) No. A rank-deficient matrix has some non-estimable coefficient combinations, but a particular contrast can still lie in its row space. (2) The contrast is not estimable: \(\beta\) and \(\beta+v\) give the same expected observations, but their contrast values differ by \(c^\mathsf{T}v\ne0\). (3) The contrast vector is the B row minus the A row, so it belongs to the row space; the common batch shift cancels. (4) No. A change of reference changes coefficient coordinates, so the contrast vector must be transformed as well. Estimability of the same scientific quantity is unchanged under an equivalent reparameterization.