Building a Marginal Linear Predictor

From a Scientific Question to a Mean Formula

A marginal model begins with a question about the average response at specified covariate values. The model formula should make clear which covariates explain that mean, how they enter, and on what scale their contributions are additive. This is a matter of specifying the mean structure; it is separate from specifying how responses within the same cluster are associated.

Write \(Y_{ij}\) for the response from cluster \(i\) at measurement \(j\), where \(i=1,\ldots,m\) and \(j=1,\ldots,n_i\). Let \(X_i\) denote the covariates recorded for cluster \(i\), and let \(x_{ij}\) be the covariate values relevant to \(Y_{ij}\). The marginal mean for that response is

This conditions on the observed covariates, not on the other responses in the cluster. In a marginal model, the mean is not defined by conditioning on a latent subject effect. The link function \(g\) maps the response mean to a scale on which a linear predictor is specified.

The word “linear” refers to linearity in the coefficients, not necessarily to the response mean itself. For example, a log link gives a mean that is an exponential function of the predictor. The predictor can include transformed covariates, such as age squared, while remaining linear in its coefficients.

Building the Predictor Deliberately

Start by stating the scientific question in terms of a mean. Then decide which covariates represent the comparison or trend of interest. For a basic additive model, each covariate contributes its own term, and the predictor can be written as an intercept plus those terms. The intercept is the predictor value when every numeric covariate is zero and every categorical covariate is at its reference category. Whether that point is scientifically useful depends on the coding.

For a categorical covariate, choose a reference category and represent the remaining categories with indicator variables. For a numeric covariate, state its units and consider centering it at a meaningful value. These choices do not change the underlying fitted means when the model is reparameterized appropriately, but they do change what the intercept describes.

This process prevents a common ambiguity: a coefficient has meaning only in relation to the covariate coding and the other terms in the predictor. In particular, an intercept is not automatically the overall sample mean, and a slope is not automatically a constant difference in response means.

What Additivity Implies

In an additive predictor, changing one covariate while holding the others fixed changes the predictor by a quantity determined by that covariate’s term. This is the basic algebra behind coefficient interpretations. The result concerns the link scale; the corresponding change in the response mean depends on the inverse link.

Proof. Write \(\eta\) for the original predictor. After changing \(z_{k,ij}\) to \(z_{k,ij}+\Delta\), the new predictor is

Since \(g(\mu_{ij})=\eta\), the link-scale change is \(\eta_{\mathrm{new}}-\eta=\beta_k\Delta\). Applying the inverse link gives the stated mean. \(\square\)

For a factor indicator, the same reasoning gives the coefficient as the change in the predictor from the reference category to that category, holding other terms fixed. Under a log link, a one-unit increase in a numeric covariate multiplies the mean by \(e^{\beta_k}\), because \(g^{-1}(\eta)=e^\eta\). Under an identity link, it changes the mean by \(\beta_k\). These response-scale translations depend on the link, not merely on the word “slope.”

Centering and the Meaning of the Intercept

Centering a numeric covariate replaces its zero point with a chosen reference value. This often makes the intercept interpretable without changing the model’s fitted means or slope. The algebra is exact, provided the same additive terms are retained.

Proof. Since \(x=x_c+c\), substitution gives

Setting \(\beta_{0,c}=\beta_0+\beta_xc\) yields the centered expression. Both expressions equal the same \(\eta\) for each value of \(x\), so applying the same inverse link produces the same fitted mean. The coefficient of \(x_c\) remains \(\beta_x\), while the new intercept is the predictor at \(x_c=0\), equivalently \(x=c\). \(\square\)

In a model with several additive terms, the same substitution applies to one numeric covariate while the other terms remain unchanged. With categorical indicators, the intercept also depends on which category is assigned all-zero indicators. It is therefore good practice to describe the intercept using the reference values explicitly, rather than simply calling it the “baseline.”

Common Mistakes and a Practical Check

• Calling the predictor the mean. The predictor is \(\eta=g(\mu)\); obtain the response mean by applying the inverse link.
• Interpreting an intercept at an impossible value. If age is measured from zero, an intercept may refer to age zero. Centering at a meaningful age gives it a more useful reference point.
• Reading a link-scale coefficient as an outcome-unit difference. That is valid for the identity link, but not in general. Translate with \(g^{-1}\) when a response-scale interpretation is needed.
• Leaving factor coding implicit. State the reference category and define each indicator, or the intercept and factor coefficients are ambiguous.
• Assuming separate coefficient estimates are identifiable from any design. If predictor columns are redundant, different coefficient vectors can produce the same linear predictor. Check that the design contains enough distinct covariate patterns to distinguish the terms being interpreted.

A useful final check is to select a particular covariate pattern, substitute its values into the predictor, and verify that the resulting intercept and terms represent the comparison intended by the scientific question. This calculation often reveals a misplaced reference value or an indicator coded in the wrong direction before any fitting is attempted.