Common Design and Contrast Errors

Why Design and Contrast Errors Matter

A gene-wise linear model can be fitted and its output can look orderly even when the design does not separate the biological effects of interest. A coefficient may compare the wrong groups, a contrast may reverse the intended direction, or a test may answer a different question from the one motivating the analysis. These are not cosmetic issues: they change what can be estimated and what the reported result means.

For gene \(g\), let \(y_g\) be the vector of measured expression values across \(n\) arrays, \(X\) the design matrix, and \(\beta_g\) the coefficient vector. The model is \(y_g=X\beta_g+\varepsilon_g\). In the usual unweighted model, errors have mean zero and variance \(\sigma_g^2\) within gene; samples are independent conditional on the design, and normal errors are used for the usual small-sample linear-model tests. The design must represent the sample structure correctly. Moderating variance estimates, as in moderated linear-model inference, does not repair a design that confounds effects or encodes the wrong comparison.

In a full-column-rank design, \(X\) has \(p\) linearly independent columns, so the coefficients are uniquely identified by the model. For a general linear hypothesis, \(R\beta_g=r\), the rows of \(R\) specify the coefficient combinations being tested and \(r\) specifies their null values. A one-dimensional comparison is commonly written \(c^T\beta_g=0\) under \(H_0\). The meaning of that equation comes from the design coding and the contrast together, not from the contrast vector alone.

Rank Deficiency: What the Data Can and Cannot Separate

A design is rank-deficient when one or more columns can be written as linear combinations of the others. For example, including an intercept and indicator columns for every level of a factor creates an exact dependence: the indicators sum to the intercept column. This is often called the dummy-variable trap. It is a coding redundancy, not necessarily a flaw in the scientific design; dropping one indicator or using a cell-means parameterization removes that redundancy.

A more serious case occurs when distinct scientific variables are perfectly confounded in the observed samples. Then no change of coding can recover their separate effects. For example, if every control sample is processed in one batch and every treated sample in another, treatment and batch indicators are identical. The data cannot say which difference is due to treatment and which is due to batch.

The condition follows from the ambiguity in the coefficients. If \(Xv=0\), then \(X(\beta_g+v)=X\beta_g\), so \(\beta_g\) and \(\beta_g+v\) give identical fitted values. The contrast has the same value under all coefficient vectors that fit those values exactly when \(c^T(\beta_g+v)=c^T\beta_g\), or \(c^Tv=0\), for every such \(v\). By the fundamental row-space/null-space relationship, this is equivalent to \(c\) being in the row space of \(X\). Thus, checking rank alone is not enough: also check whether the particular scientific comparison is estimable.

Reference Levels and Coefficient Meaning

With treatment (reference-level) coding, one level of a factor is chosen as the reference. The intercept represents the expected expression for the reference group, and each other group coefficient represents that group minus the reference, holding any other model terms fixed. The coefficient is not automatically the effect of every group, nor is it a comparison between two non-reference groups.

For a three-group factor with control as reference, the model is \(\mu_{\text{control}}=\beta_0\), \(\mu_{\text{low}}=\beta_0+\beta_{\text{low}}\), and \(\mu_{\text{high}}=\beta_0+\beta_{\text{high}}\). Consequently, high minus low is \(\beta_{\text{high}}-\beta_{\text{low}}\), not \(\beta_{\text{high}}\). Releveling changes coefficient interpretation and numerical values, but not the fitted group means or a correctly translated scientific comparison.

Signs and Scientific Hypotheses

Before testing a contrast, write the comparison in words and fix its direction, such as “treated minus control.” Then express that direction using the model coefficients. This prevents a common error: a correct calculation reported with the opposite biological interpretation. If expression is on the log2 scale, a positive treated-minus-control contrast means higher fitted expression in treated samples; a negative value means lower fitted expression. The units are log2 expression units, and a log2 difference of \(d\) corresponds to a ratio of \(2^d\) on the original scale.

For a two-sided test of equality, reversing the sign of a contrast reverses its estimate and test statistic but leaves the absolute test statistic and two-sided p-value unchanged. For a one-sided test, the direction is part of the alternative hypothesis: a test for treated greater than control cannot be interpreted as a test for treated less than control. State the alternative in the same direction as the contrast.

Does the Fitted Hypothesis Answer the Question?

In factorial models, the distinction between a main effect and an interaction is especially important. An interaction asks whether an effect changes across levels of another factor. A treatment coefficient in a reference-coded model may instead describe the treatment effect only at the reference level of that other factor. Neither quantity should be described as a universal treatment effect unless the model and contrast support that interpretation.

For two factors with two levels each, let A be the reference level of factor A and B the reference level of factor B. Under treatment coding, a model with interaction has fitted means \(\mu_{AB}=\beta_0\), \(\mu_{A'B}=\beta_0+\beta_A\), \(\mu_{AB'}=\beta_0+\beta_B\), and \(\mu_{A'B'}=\beta_0+\beta_A+\beta_B+\beta_{AB}\). The interaction coefficient is the difference-in-differences: \((\mu_{A'B'}-\mu_{AB'})-(\mu_{A'B}-\mu_{AB})=\beta_{AB}\).

Practical Checks, Alternatives, and Limitations

A reliable design review combines structural checks with a plain-language check of each contrast. Inspect the sample annotation and factor levels before fitting; confirm that indicator columns correspond to the intended samples; and verify the column rank and contrast estimability. Software may omit redundant columns, report non-estimable coefficients, or use a generalized inverse, but these behaviors do not establish that the desired biological effect is identifiable.

Cell-means coding is a useful alternative to reference coding: each coefficient directly represents a fitted group mean, and a comparison is formed by subtracting means. It can make contrast construction more transparent, but it cannot solve confounding or create information absent from the samples. Reference coding is compact and convenient when comparisons to a chosen baseline are central. Either parameterization gives the same fitted values and scientific comparison when the model space and contrast are translated correctly.

A full-rank design is a necessary diagnostic for unique coefficients, but it is not proof that the model is scientifically appropriate. A full-rank model can still omit an important design feature or encode the wrong factor labels. Conversely, a rank-deficient parameterization may contain redundant columns even when the underlying group comparisons are estimable. Near-dependence among columns can also make estimates unstable; this is distinct from exact non-estimability and should prompt scrutiny of the sample allocation and model specification.

• Pitfall: treating every coefficient as a group comparison of interest. Correct this by writing each coefficient in terms of fitted means and forming the requested linear combination.
• Pitfall: interpreting a positive number without checking contrast order. “Positive” means higher expression for the first group only if the contrast was explicitly defined as first group minus second group.
• Pitfall: assuming variance moderation fixes confounding. It can affect uncertainty estimation, but it cannot distinguish treatment from batch when their indicators coincide.
• Pitfall: describing an interaction test as a test of an effect in one subgroup. Test the sum of the relevant main-effect and interaction coefficients for that subgroup-specific effect.

The main limitation is that design and contrast checks clarify the meaning and identifiability of the fitted model; they cannot guarantee that the model captures every relevant biological or technical feature. Any correction requiring information not represented in the observed design, such as separating perfectly confounded treatment and batch effects, requires a better-designed experiment rather than a different contrast formula.