Contrasts for Gene-Level Questions

From a Biological Question to a Contrast

A fitted gene-wise linear model describes expression through model coefficients, but a biological question is usually phrased in terms of group means. It may ask whether expression is higher in one condition than another, or whether one group differs from the average of several others. A contrast turns that question into a specified linear combination of coefficients or means. Writing the combination explicitly makes clear what is being estimated, how its units should be interpreted, and whether the study design can identify it.

Suppose that for gene \(g\), the observed expression values are collected in the vector \(y_g\), and the design matrix is \(X\). The gene-wise linear model is \(y_g=X\beta_g+\varepsilon_g\). Here \(\beta_g\) is the coefficient vector for gene \(g\). For the usual least-squares expression model, the design is treated as fixed, the model mean is correctly specified, and errors have mean zero and a common variance within that gene: \(\operatorname{E}(\varepsilon_g)=0\) and \(\operatorname{Var}(\varepsilon_g)=\sigma_g^2 I\). Independence is a standard working assumption for the errors. The estimate \(\hat{\beta}_g\) is obtained from the fitted model; if \(X\) has full column rank, it is \((X'X)^{-1}X'y_g\).

The distinction between a coefficient and a contrast matters. Under reference-group coding, a coefficient may already represent a particular difference of means. Under other codings, the same difference may require combining several coefficients. The model coding is a way to represent the fitted means; the question being asked determines the contrast.

Writing Contrasts for Group Means

First consider a model that directly represents the means of three groups as \(\beta_{g,1}=\mu_{g,1}\), \(\beta_{g,2}=\mu_{g,2}\), and \(\beta_{g,3}=\mu_{g,3}\). The expression mean in group 2 minus the mean in group 1 is \(c'\beta_g\) for \(c=(-1,1,0)'\). The order of subtraction is deliberate: a positive estimate means higher fitted expression in group 2 than group 1.

More generally, a contrast among \(K\) group means has weights \(a_1,\ldots,a_K\) with \(\sum_{k=1}^K a_k=0\). For example, \(\mu_{g,1}-(\mu_{g,2}+\mu_{g,3}+\mu_{g,4})/3\) has weights \(1,-1/3,-1/3,-1/3\). Their sum is zero, and the contrast compares group 1 with the equally weighted average of the other three means. These weights are not determined by the sample sizes: they encode the scientific comparison.

For a general design matrix, write the intended quantity in the model's coefficient coordinates. With reference-group coding for three groups, \(\beta_{g,0}\) may be the reference mean, while \(\beta_{g,1}\) and \(\beta_{g,2}\) are differences from that mean. Then the difference between group 2 and group 3 is \(\beta_{g,1}-\beta_{g,2}\), not either coefficient alone. Its contrast vector is \((0,1,-1)'\) in the coefficient order \((\beta_{g,0},\beta_{g,1},\beta_{g,2})\).

Estimability and Standard Errors

A requested contrast must be supported by the design. If \(X\) has full column rank, every coefficient combination \(c'\beta_g\) is estimable. If the design is rank deficient, some coefficient vectors are not uniquely determined by the fitted mean \(X\beta_g\). In that case, \(c'\beta_g\) is estimable exactly when \(c\) lies in the row space of \(X\). Equivalently, every vector \(v\) satisfying \(Xv=0\) must also satisfy \(c'v=0\). This condition ensures that different coefficient vectors producing the same fitted means also produce the same requested contrast.

When \(X\) has full column rank and errors have covariance \(\sigma_g^2 I\), the least-squares coefficient covariance is \(\sigma_g^2(X'X)^{-1}\). Replacing the unknown variance by the fitted gene-wise residual variance gives the estimated standard error of the contrast:

This expression follows directly from the variance of a linear transformation. Since \(c'\hat{\beta}_g\) is a scalar linear combination of the coefficient estimates, \(\operatorname{Var}(c'\hat{\beta}_g)=\operatorname{Var}(c'\hat{\beta}_g)\) can be expanded as \(c'\operatorname{Var}(\hat{\beta}_g)c\). Substituting \(\operatorname{Var}(\hat{\beta}_g)=\sigma_g^2(X'X)^{-1}\) gives \(\sigma_g^2c'(X'X)^{-1}c\). Taking the square root after replacing \(\sigma_g^2\) by \(s_g^2\) yields the stated standard error.

For a cell-means model with independent groups, the same result simplifies. If group \(k\) has \(n_k\) observations and the contrast weights are \(a_k\), then \(\widehat{\operatorname{Var}}(\sum_k a_k\hat{\mu}_{g,k})=s_g^2\sum_k a_k^2/n_k\). The squared weights appear because variances scale by the square of a multiplier. This is why the standard error for a difference depends on both group sizes and the gene-wise residual variance.

Practical Checks and Common Pitfalls

Before interpreting a gene-wise contrast, verify that the vector's coefficient order matches the fitted design, that the signs reflect the intended subtraction, and that the weights define the desired comparison. Then check estimability and calculate the standard error from the same fitted model and contrast. If the estimated variance is unusually large, inspect the group sample sizes and the design matrix: sparse groups, near-confounding, or redundant columns can make a contrast imprecise or nonestimable.

A contrast is not automatically a causal effect. It is a comparison of modeled expression means; causal interpretation requires a design and assumptions that justify it. Likewise, a small standard error does not establish that the model is correctly specified. Plot residuals and fitted values where appropriate, and examine whether the design has the intended observations in each group. These checks can reveal outliers, variance patterns, or confounding that undermine the working model.

A common mistake is to use the sample-size-weighted average of comparison groups when the scientific question asks for their equal-weighted average. Another is to treat a reference-coded coefficient as the only possible group comparison. The correction is to write the estimand in words, translate those words into explicit weights, and then express those weights in the model's coefficient coordinates. Pairwise comparisons are useful when each pair is of interest, but they do not replace a specified multi-group contrast. An omnibus test addresses whether a collection of effects is jointly zero; it does not, by itself, report the particular direction and magnitude captured by one contrast.

Multiplying every contrast weight by a nonzero constant also multiplies the estimate and standard error by that constant's absolute value. The underlying comparison is the same up to scale, but its reported magnitude and units change. For interpretability, use weights that directly match the question, such as a difference in means or a difference from an average.