Parameterizing Linear Contrast Families

From Mean Vectors to Contrast Parameters

In multivariate inference, a scientific question often concerns a small collection of comparisons among the components of a mean vector rather than every component separately. For example, an investigator may want to compare four treatment means in a sequence, or compare the average response in one group of conditions with the average response in another. A matrix gives a precise way to specify all such comparisons together.

Let \(\mu=(\mu_1,\ldots,\mu_p)^T\) be a \(p\)-dimensional mean vector. A row vector \(c^T=(c_1,\ldots,c_p)\) defines the linear functional \(c^T\mu=\sum_{k=1}^p c_k\mu_k\). In the classical treatment-means setting, a contrast is such a linear functional whose coefficients sum to zero. This condition makes the comparison unchanged if the same constant is added to every mean. More generally, a row that does not sum to zero still defines a linear functional, but is not a classical contrast under this convention.

Writing the questions as rows fixes their coefficients and order. If the \(j\)th row is \(c_j^T\), then the \(j\)th component of \(\eta\) is \(\eta_j=c_j^T\mu\). The components of \(\eta\) are the parameters being targeted; the number \(q\) of listed rows need not equal the number of independent target parameters.

Encoding a Family of Comparisons

Consider four treatment means. The adjacent differences can be recorded as three rows, each with a positive coefficient for the first mean in the comparison and a negative coefficient for the second:

Each row sums to zero, so each is a classical contrast. The matrix records both the questions and their signs: reversing one row reverses the corresponding contrast parameter. Stating a convention for row order and signs prevents ambiguity when reporting results.

A contrast matrix also makes it possible to tell whether some listed questions repeat information already contained in others. This matters when a family is described with more rows than it has independent directions.

Rank, Dimension, and Unmeasured Directions

The dimension statement follows from a basic fact about linear maps. The matrix \(C\) maps each possible mean vector in \(\mathbb{R}^p\) to a point \(C\mu\) in \(\mathbb{R}^q\). Its image, also called its column space, is the set of all such output vectors. The rank of \(C\) is the dimension of that image.

Proof. The set of possible outputs is the image of the linear map \(\mu\mapsto C\mu\). It is a subspace: if \(C\mu=u\) and \(C\tilde\mu=v\), then \(u+v=C(\mu+\tilde\mu)\), and for any scalar \(a\), \(au=C(a\mu)\). By the definition of matrix rank, the dimension of this image is \(\operatorname{rank}(C)\).

For the second claim, \(C\mu=C\tilde\mu\) holds exactly when subtracting the two sides gives \(C(\mu-\tilde\mu)=0\). This says precisely that \(\mu-\tilde\mu\) is in the null space of \(C\). Thus directions in the null space change the mean vector without changing the target parameters. By the rank-nullity theorem, that null space has dimension \(p-\operatorname{rank}(C)\). \(\square\)

When \(C\) has dependent rows, its target vectors satisfy linear constraints and occupy a proper subspace of \(\mathbb{R}^q\). When the rows are independent, the image has dimension \(q\), so there are no such constraints among the \(q\) coordinates. This is a statement about the parameterization, not a claim that the observed estimates or their uncertainty have any particular distribution.

Choosing Coordinates for the Same Contrast Family

A family of contrasts can be represented using different rows that span the same row space. Changing the basis changes the numerical coordinates used to describe the family, but does not add or remove its independent directions. For an invertible \(r\times r\) matrix \(A\), replacing a full-row-rank matrix \(C\) by \(C^*=AC\) gives \(\eta^*=C^*\mu=A(C\mu)=A\eta\). The parameters have been re-expressed by an invertible transformation.

Proof. Every row of \(AC\) is a linear combination of rows of \(C\), so the row space of \(AC\) is contained in the row space of \(C\). Since \(A^{-1}\) exists, \(C=A^{-1}(AC)\), so every row of \(C\) is also a linear combination of rows of \(AC\). The row spaces are therefore equal, and their dimensions, hence their ranks, are equal. Finally, \((AC)\mu=A(C\mu)\); invertibility of \(A\) makes this a one-to-one change of coordinates. \(\square\)

Common Pitfall and Practical Use

Common mistake: Counting the rows and declaring that a family of \(q\) listed contrasts has \(q\) independent parameters. The three pairwise differences among three means are a counterexample: there are three rows, but the third contrast is the sum of the first two, so the rank is only 2. The correction is to check linear dependence, either by row reduction or by expressing rows as combinations, before interpreting the number of target dimensions.

Another useful check is whether a row really is a classical contrast. For example, \((1,1,0)\mu=\mu_1+\mu_2\) is a linear functional but its coefficients sum to 2, not 0. Adding a constant \(a\) to every mean changes this functional by \(2a\), so it is not a treatment contrast under the zero-sum convention. The row itself makes this distinction visible.

Parameterizing the target before inference clarifies exactly which mean features a later analysis concerns. It identifies the requested quantities, shows whether some are algebraically redundant, and records what changes in the mean vector cannot be detected by those quantities. The matrix does not by itself supply a confidence procedure or a test; it gives the parameter space that such procedures must address.