Translating Scientific Questions Into Linear Restrictions

From a Scientific Claim to a Matrix Equation

A scientific question is often phrased in terms of effects: Are two slopes equal? Is an intervention effect zero? Does a weighted combination of outcomes have a specified effect? In a multivariate linear model, these questions concern entries or contrasts of the coefficient matrix. Translating them carefully means deciding which coefficients to compare and which responses to combine, while keeping other coefficients unrestricted when they are not part of the question.

Write the multivariate linear model as \(Y=XB+U\), where \(Y\) is an \(n\times p\) response matrix, \(X\) is an \(n\times q\) design matrix, and \(B\) is the \(q\times p\) coefficient matrix. Each row of \(B\) corresponds to a design-matrix column, and each column corresponds to a response. The restriction notation used here is \(C B M=\Theta\).

Definition: A linear restriction on \(B\) of the form \(C B M=\Theta\) uses a coefficient-contrast matrix \(C\) of size \(r\times q\), a response-contrast matrix \(M\) of size \(p\times s\), and a specified \(r\times s\) matrix \(\Theta\). The equation requires every entry of \(C B M\) to equal the corresponding entry of \(\Theta\). The dimensions must allow the products: \((r\times q)(q\times p)(p\times s)=r\times s\).

The order of multiplication reflects two distinct parts of the scientific question. \(C\) forms contrasts among coefficient rows, and \(M\) forms contrasts or combinations among response columns. For example, a row of \(C\) can select one predictor’s slope, while a column of \(M\) can subtract the slopes for two responses. No rank assumption is needed just to write the restriction. Whether its coefficients can be estimated from a particular design is a separate issue.

How the Products Encode the Claim

For coefficient indices \(a=1,\ldots,q\), response indices \(b=1,\ldots,p\), and contrast indices \(i=1,\ldots,r\), \(j=1,\ldots,s\), matrix multiplication gives \((CBM)_{ij}=\sum_{a=1}^{q}\sum_{b=1}^{p}C_{ia}B_{ab}M_{bj}\). Thus each entry of the restriction is a weighted sum of coefficients. A zero in \(C\) excludes a design coefficient from that coefficient contrast; a zero in \(M\) excludes a response from that response combination.

Result: The matrix equation \(CBM=\Theta\) is equivalent to the \(rs\) scalar equations \(\sum_{a=1}^{q}\sum_{b=1}^{p}C_{ia}B_{ab}M_{bj}=\Theta_{ij}\), one for each \(i,j\).

Proof. By the definition of matrix multiplication, the \((i,j)\) entry of \(CBM\) is exactly the displayed double sum. Equality of two matrices of the same dimensions holds if and only if their corresponding entries are equal. Therefore \(CBM=\Theta\) holds if and only if all \(rs\) scalar equations hold. \(\square\)

This entrywise view is useful before writing any matrices: say the intended coefficient combination aloud, then say the intended response combination. The matrices should reproduce those words when multiplied in order.

Worked Example: Equal Treatment Slopes for Two Responses

Suppose a model has three design columns—intercept, treatment indicator, and age—and two responses, \(Y_1\) and \(Y_2\). Write \[ B= \begin{pmatrix} \beta_{01}&\beta_{02}\\ \beta_{T1}&\beta_{T2}\\ \beta_{A1}&\beta_{A2} \end{pmatrix}. \] The scientific claim is that the treatment slope is equal for the two responses. The intercepts and age slopes are not part of this claim.

Select the treatment row with \(C=(0,1,0)\), and take the response difference with \(M=(1,-1)^\mathsf{T}\). Set \(\Theta=(0)\). The product is \[ CBM=(0,1,0) \begin{pmatrix} \beta_{01}&\beta_{02}\\ \beta_{T1}&\beta_{T2}\\ \beta_{A1}&\beta_{A2} \end{pmatrix} \begin{pmatrix}1\\-1\end{pmatrix} =\beta_{T1}-\beta_{T2}. \] Consequently, \(CBM=\Theta\) says exactly \(\beta_{T1}-\beta_{T2}=0\), or \(\beta_{T1}=\beta_{T2}\). The zero entries in \(C\) leave the intercept and age coefficients out of this restriction.

For a numerical check, suppose the treatment-slope row is \((1.2,1.2)\). Then \((1.2,1.2)(1,-1)^\mathsf{T}=1.2-1.2=0\), as the equality claim requires. If the row were \((1.2,0.9)\), the contrast would be \(1.2-0.9=0.3\), not zero.

Zero Effects and Nuisance Coefficients

A question about a predictor’s effect across every response usually uses an identity matrix on the response side. If the target predictor is the \(k\)th design column, let \(e_k^\mathsf{T}\) be the row vector with a 1 in position \(k\) and zeros elsewhere. Then \(e_k^\mathsf{T}B\) is the full row of coefficients for that predictor. Requiring that row to be zero is the restriction \(e_k^\mathsf{T}B I_p=0\).

This construction also shows how nuisance coefficients remain unrestricted: they do not appear in the chosen row contrast. More generally, if a change \(D\) to \(B\) represents coefficients that should not affect the scientific contrast, the contrast is unchanged exactly when \(C D M=0\). This condition follows directly by subtracting the two restricted quantities: \[ C(B+D)M-CBM=CDM. \] So nuisance coefficients may be omitted by the matrix choices only when their contribution to the contrast is zero.

Worked Example: No Intervention Effect on Any of Three Outcomes

Consider a model with design columns for an intercept, intervention, baseline severity, and site indicator, and with three responses. The intervention coefficients are \((\beta_{T1},\beta_{T2},\beta_{T3})\). The question is whether the intervention has zero effect on all three responses, with the other coefficients left unrestricted.

Choose \(C=(0,1,0,0)\), \(M=I_3\), and \(\Theta=(0,0,0)\). The dimensions are \(1\times4\), \(3\times3\), and \(1\times3\), respectively, and \[ CBM=(0,1,0,0)B I_3=(\beta_{T1},\beta_{T2},\beta_{T3}). \] Thus the matrix equation is equivalent to the three scalar restrictions \(\beta_{T1}=0\), \(\beta_{T2}=0\), and \(\beta_{T3}=0\). The intercept, baseline-severity, and site coefficients do not enter because their positions in \(C\) are zero.

For a numerical check, if the intervention row of a candidate coefficient matrix is \((0.4,-0.1,0.7)\), then \(CBM=(0.4,-0.1,0.7)\), which is not the zero row. The restriction being stated is still clear: it requires each of those three coefficients to equal zero.

Specified Linear Combinations Across Responses

A response contrast need not compare two outcomes with equal weights. It may instead define a composite outcome, such as \(2Y_1-Y_2+Y_3\). If the question concerns the effect of a predictor on that composite, the coefficient contrast selects the predictor and the response contrast supplies the weights. If the predictor effect depends on a covariate through an interaction, the coefficient contrast must include the relevant interaction coefficient as well.

Worked Example: A Treatment Effect at a Chosen Age for a Weighted Outcome

Suppose the design columns are, in order, an intercept, treatment, treatment multiplied by age, and age. There are three responses. For response \(j\), the treatment effect at age 50 is \(\beta_{Tj}+50\beta_{TA,j}\), where \(\beta_{TA,j}\) is the treatment-by-age coefficient. The scientific question is whether the weighted response \(2Y_1-Y_2+Y_3\) has a treatment effect of 1 unit at age 50.

The target coefficient contrast and response weights are \(C=(0,1,50,0)\) and \(M=(2,-1,1)^\mathsf{T}\). The first and fourth entries of \(C\) exclude the intercept and age main effect; the treatment and treatment-by-age entries produce the effect at age 50. The requested restriction is \[ CBM=1. \]

For an explicit coefficient matrix, take \[ B= \begin{pmatrix} 10&8&5\\ 1.5&2&-1\\ 0.01&0&0\\ 0.02&0.01&0.03 \end{pmatrix}. \] Multiplying the coefficient contrast gives \[ CB=(0,1,50,0)B=(1.5+50(0.01),\,2+50(0),\,-1+50(0))=(2,2,-1). \] Then \[ CBM=(2,2,-1) \begin{pmatrix}2\\-1\\1\end{pmatrix} =2(2)-2+(-1)=1. \] The candidate coefficients therefore satisfy the specified restriction. This calculation also checks that the treatment-by-age term was included: selecting only the treatment row would have produced a different question.

A Reliable Translation Procedure

The same sequence of decisions works for a single equality or a matrix of restrictions. First identify the scientific target in words; then map each part of that statement to its coefficient and response contrast. The following procedure helps keep the dimensions and interpretation aligned.

Name the target coefficients.
Use the model’s design-column definitions to identify the coefficient rows that contribute to the claimed effect, including interactions when the effect is conditional on a covariate value.

Write the coefficient contrast.
Put the required weights in \(C\). A selector has a single 1; a difference has weights such as \(1,-1\); a conditional effect can combine several rows.

Write the response contrast.
Put response weights in \(M\). Use \(I_p\) when the claim concerns all responses separately, and a contrast vector when it concerns a specific combination.

Set the target value and check the product.
Choose \(\Theta\) to state the required value, verify dimensions, and expand \(CBM\) to confirm it says exactly the scientific claim.

A matrix equation can state several related scalar restrictions at once. For instance, using several rows of \(C\) can form several coefficient contrasts, while using \(I_p\) in \(M\) retains separate restrictions for all responses. In every case, the entrywise expansion is the final check on what is being asserted.

Common Mistakes and Why the Translation Matters

Reversing the contrast. If the question is whether response 1 and response 2 have equal slopes, \(M=(1,-1)^\mathsf{T}\) gives \(\beta_1-\beta_2\). Reversing the signs gives the negative contrast; equality to zero is unchanged, but a nonzero specified value must have its sign reversed too.
Leaving out an interaction. In a model with treatment-by-age interaction, the treatment effect at age 50 includes the interaction coefficient multiplied by 50. Selecting only the treatment row instead describes the effect at the reference age encoded by the model.
Restricting nuisance coefficients by accident. A full matrix equality such as \(B=0\) includes every coefficient. If the claim concerns only the intervention row, select that row rather than setting the whole coefficient matrix to zero.
Using the wrong response operation. \(M=I_p\) keeps the responses separate; a contrast vector combines them. These state different scientific claims even when they use the same \(C\).
Stopping at plausible dimensions. Compatible dimensions do not guarantee a correct translation. Expand the product and check each coefficient and response weight against the words of the question.

A well-translated restriction makes the scientific target explicit before any test is calculated. It also reveals which coefficients are being compared, which responses are combined, and which model terms are deliberately left outside the claim. The matrix equation is concise, but its meaning should always be recoverable from the corresponding scalar contrast.

Check Your Understanding

For each question, identify the coefficient contrast, response contrast, and target matrix.

A model has an intercept, a drug indicator, and a severity score, with two responses. Write a restriction stating that the drug slopes are equal across responses, leaving the other coefficients unrestricted.
With four responses, write a restriction stating that the coefficient of a specified predictor is zero for every response.
A treatment effect varies linearly with age through a treatment-by-age interaction. Which coefficient rows contribute to the treatment effect at age 60?
Explain the difference between using \(M=I_3\) and using \(M=(1,1,-1)^\mathsf{T}\) when there are three responses.
If \(C\) selects the treatment row and \(M\) contrasts two responses, what scalar coefficient difference does \(CBM\) represent?