Design Matrices for Group Comparisons

Why the Design Matrix Matters

In a gene-wise linear model, the observed expression values for one gene are related to a set of sample-level predictors through a design matrix. For a group comparison, that matrix records which experimental group each sample belongs to. Its columns determine how the fitted model describes group expression, so the same group means can have different coefficient interpretations under different codings.

Here, \(y_g\) is the vector of expression measurements for gene \(g\), one entry per array, \(X\) is the design matrix, and \(\beta_g\) is the vector of gene-specific coefficients. A standard model is \(y_g=X\beta_g+\varepsilon_g\). The entries of \(y_g\) are typically on a log-expression scale. The coding of \(X\), not a change in the observed data, determines what each element of \(\beta_g\) means.

For the group-mean interpretation, suppose the model’s conditional mean is correctly specified, errors have mean zero given the design, and the design has full column rank. For conventional model-based inference, the usual additional assumptions are independent errors with common variance; normal errors are used for exact small-sample inference. These assumptions concern the observations and error distribution, not whether a particular indicator column is written correctly. A full-rank matrix permits unique coefficient estimates, but cannot by itself guarantee that the experimental design or error assumptions are appropriate.

Two Groups: Reference-Group Coding

For two groups, choose one group as the reference. Code it as 0 in an indicator column and code the other group as 1. With control as the reference and treated as the indicated group, the model for one gene is

The corresponding design matrix has an intercept column of ones and an indicator column \(z\). For a control sample, the fitted mean is \(\beta_{0g}\); for a treated sample, it is \(\beta_{0g}+\beta_{1g}\). Thus \(\beta_{0g}\) is the control-group mean in the fitted model, and \(\beta_{1g}\) is the treated-minus-control difference in fitted means. The coefficients have the same units as the response: for log-expression data, they are on the log-expression scale.

The indicator’s meaning follows directly by substituting its values into the model: setting \(z_i=0\) leaves the intercept, while setting \(z_i=1\) adds the indicator coefficient. Reversing the reference group changes the coefficient interpretation and signs, but not the fitted group means. This is why a coefficient should always be read together with the coding.

More Than Two Groups

For \(K\) groups, reference-group coding uses an intercept and \(K-1\) indicator columns. If group 1 is the reference, the model is

where \(z_{ik}=1\) if sample \(i\) is in group \(k\), and is 0 otherwise. Each sample belongs to exactly one group. Therefore, group 1 has all indicator values zero and fitted mean \(\beta_{0g}\). Group \(k\), for \(k\ge 2\), has fitted mean \(\beta_{0g}+\beta_{kg}\). Consequently, \(\beta_{kg}\) is the fitted mean in group \(k\) minus the fitted mean in the reference group.

A different, also valid, parameterization is cell-means coding: use one indicator column for every group and omit the intercept. Each coefficient is then the fitted mean for its group. For the preceding example, the columns are indicators for A, B, and C, and the fitted coefficients are 5.000, 5.800, and 5.000. The fitted values are identical to those from reference-group coding; only the coefficient meanings differ.

Checking a Design Matrix

A correctly constructed matrix reflects the actual sample allocation and uses a consistent column order. Before fitting the gene-wise models, check that each row corresponds to the intended array, each indicator matches that array’s group, and the matrix has full column rank. If there are \(K\) groups and all are represented, the reference-coded matrix with an intercept and \(K-1\) indicators has \(K\) independent columns. In cell-means coding, all \(K\) group columns are independent when each group is represented.

A common mistake is to include an intercept and indicators for all \(K\) groups at once. Every row then has exactly one group indicator equal to 1, so the sum of the group-indicator columns equals the intercept column. The columns are linearly dependent, and the coefficients cannot be uniquely determined. Omit one indicator when using an intercept, or use all group indicators without an intercept.

Other warning signs include an indicator column that is all zero because a group has no samples, a row whose indicators assign it to multiple groups in a mutually exclusive design, or sample labels that do not match the expression matrix’s sample order. These are not cosmetic issues: they can change the fitted values or make the model unidentifiable. Inspect group counts and a few sample-to-row mappings before fitting.

What the Coefficients Do—and Do Not—Say

For a design containing only group indicators, the fitted value for each sample is its group’s sample mean. If the design has additional columns, such as a batch indicator, the group coefficient interpretation depends on that expanded model and its coding; it need not equal the raw difference between group averages. A group comparison can also be impossible to separate from batch when group and batch are perfectly confounded. In that case, the design lacks information to distinguish their effects.

Reference-group coding is useful when the scientific description naturally names a baseline group and asks how other groups differ from it. Cell-means coding is convenient when the quantities of direct interest are the group means themselves. Neither encoding is universally preferable: choose based on clear interpretation and correct model specification. The fitted values are unchanged by a valid reparameterization, but individual coefficients are not interchangeable across codings.

These models describe gene-level expression on the response scale used in the analysis. They do not resolve whether the measured arrays are independent biological replicates, whether the error variance is appropriate, or whether the evidence across thousands of genes requires multiplicity control. Those questions require attention beyond the construction and interpretation of the group design matrix.