Identity Links And Additive Mean Changes

What an Identity Link Says About the Mean

A coefficient is easiest to interpret when we identify the scale on which it acts. In a generalized linear model, the linear predictor is built from the covariates, while the link connects that predictor to the conditional mean. With the identity link, the two are equal. Consequently, a change in the linear predictor is an equal-sized change in the mean—not a change that must first be transformed.

For observation \(i\), let \(\mathbf{x}_i\) be its design row and let \(\boldsymbol{\beta}\) be the coefficient vector. The linear predictor is \(\eta_i=\mathbf{x}_i^\mathsf{T}\boldsymbol{\beta}\). Writing the identity link as \(g(\mu)=\mu\), the model’s mean equation is \(\mu_i=\eta_i\). The distribution of the response still matters: it determines which values of \(\mu_i\) are admissible.

The inverse link being the identity makes the link-scale and mean-scale values numerically equal. Therefore, in a model with a single untransformed predictor \(x\), such as \(\mu(x)=\beta_0+\beta_1x\), the coefficient \(\beta_1\) is the change in conditional mean for a one-unit increase in \(x\). More generally, a change of \(d\) units changes the mean by \(\beta_1d\), provided other predictors and model terms are held fixed.

Exact Additive Changes

The additive interpretation is an exact difference, not an approximation for small changes. In a model with several predictors, changing one design component while leaving the others unchanged gives a particularly direct comparison.

Proof. Let \(\mathbf{x}\) denote the first design row and \(\mathbf{x}'\) the second. By assumption, \(x'_j=x_j+d\) and \(x'_k=x_k\) for every \(k\ne j\). Subtract the first conditional mean from the second:

This difference does not depend on the starting value \(x_j\). It does depend on what counts as changing only component \(j\): if changing a raw covariate also changes other design terms, the full design-row difference must be used instead. The claimed result is about a single design component, not automatically every possible change in the original covariates. \(\square\)

For multiple predictors, a coefficient comparison is conditional on the other design components. If a predictor is centered, rescaled, or represented by more than one design column, interpretation must follow that representation. The additive change comes from multiplying the change in the design row by the coefficient vector.

Mean Changes Must Still Be Admissible

An additive mean equation does not automatically respect the response distribution. The identity inverse link returns any real-valued linear predictor unchanged, but a response family may allow only a restricted set of means. For example, a Bernoulli response with success probability \(p\) requires \(0\leq p\leq1\), and a Poisson response with mean \(\mu\) requires \(\mu\geq0\). A Gaussian response with unrestricted real support has no corresponding positivity or upper-bound constraint on its mean.

This distinction is important: the linear predictor may be calculated for any design row, but that calculation is a valid conditional mean only if it is in the mean space allowed by the response distribution. A coefficient can have the additive interpretation and yet produce an inadmissible prediction at some covariate values.

Proof. Under the identity link, the proposed conditional mean at \(\mathbf{x}\) is exactly \(\mathbf{x}^\mathsf{T}\boldsymbol{\beta}\). By definition, a value is an admissible mean for the specified response distribution exactly when it belongs to \(\mathcal{M}\). Thus, for each \(\mathbf{x}\in\mathcal{C}\), the proposed mean is admissible if and only if \(\mathbf{x}^\mathsf{T}\boldsymbol{\beta}\in\mathcal{M}\). Requiring this for every covariate value in \(\mathcal{C}\) gives the stated condition. \(\square\)

In the special case of one predictor, \(\mu(x)=\beta_0+\beta_1x\), and a covariate range \([a,b]\), checking the two endpoints is enough when \(\mathcal{M}\) is an interval. For any \(x\in[a,b]\), write \(x=(1-t)a+tb\) for some \(t\in[0,1]\). Then \(\mu(x)=(1-t)\mu(a)+t\mu(b)\). If both endpoint means belong to an interval \(\mathcal{M}\), every such weighted average belongs to \(\mathcal{M}\) as well. This argument applies, for instance, to the Bernoulli mean interval \([0,1]\).

Checking the Scope of an Additive Interpretation

The identity link is especially transparent, but its mean-scale consequences should be checked across the covariate region relevant to the analysis. A model may produce valid means for observed design rows and invalid means for new rows. Conversely, a model that is admissible over a narrow observed range may fail under extrapolation. The check concerns the mean equation and the intended covariate domain, not just the values encountered in the fitting data.

For response families whose means are unrestricted real numbers, this particular constraint check imposes no additional bounds. For restricted mean spaces, verify the fitted or proposed means against the family’s requirements. For a single linear predictor and interval covariate range, endpoint checks can suffice as shown above; with multiple predictors, the relevant check must cover the intended design region.

Common Mistake

• Wrong reasoning: “The identity link has an inverse that accepts every real number, so every identity-link prediction is a valid mean.”
• Correction: The inverse link returns a real number, but the response distribution may permit only a subset of real numbers as means. For Bernoulli outcomes, a predicted mean must be between 0 and 1. Check the response family’s mean space over the covariate values where the model will be used.
• Wrong reasoning: “The coefficient is the additive effect of changing the original predictor, no matter how the model is written.”
• Correction: The exact additive change is the coefficient vector multiplied by the change in the design row. A coefficient for a transformed or interaction column must be interpreted according to the change in that column; changing a raw predictor may alter several columns at once.