What a GLM Linear Predictor Represents

What the Linear Predictor Represents

A generalized linear model uses covariates to describe how the response distribution changes across observations. The linear predictor is the model’s covariate-based summary for an observation: it adds the observation’s covariate values, each weighted by its coefficient. The coefficients describe how the model combines covariates; the covariate values determine the resulting predictor for a particular observation.

This distinction matters because a fitted model usually has one shared set of coefficients but a different collection of covariate values for each observation. Consequently, the coefficients can stay the same while the linear predictor changes from one observation to another. The predictor is commonly denoted by \(\eta_i\), where \(i\) identifies the observation.

The word “linear” refers to the way the coefficients enter: each coefficient is multiplied by a covariate value, and these products are added. The predictor is not necessarily a linear function of the response, and its numerical value is not necessarily a predicted response. In a GLM, the predictor is typically on the scale specified by the link function. In the usual notation, the model relates it to the mean \(\mu_i=E[Y_i]\) through \(\eta_i=g(\mu_i)\), where \(g\) is the link function. Thus, interpreting \(\eta_i\) requires keeping its scale distinct from the response scale.

Coefficients and Observation-Specific Values

A coefficient \(\beta_j\) is a model parameter attached to covariate \(j\). It is shared: the same \(\beta_j\) is used when calculating the predictor for every observation in the model. By contrast, \(x_{ij}\) records the value of that covariate for a particular observation. For example, patients can have different ages \(x_{i1}\), but the age coefficient \(\beta_1\) is one common model coefficient.

The contribution of covariate \(j\) to observation \(i\)’s predictor is the product \(x_{ij}\beta_j\). A large covariate value does not by itself guarantee a large contribution: the coefficient’s size and sign also matter. A positive coefficient makes the contribution increase as the covariate increases, while a negative coefficient makes it decrease, with other quantities held fixed.

In an application, the coefficients may be unknown and estimated from data. Once estimates \(\hat{\beta}_j\) are available, the fitted linear predictor is calculated by substituting those estimates: \(\hat{\eta}_i=\sum_{j=1}^{p}x_{ij}\hat{\beta}_j\). This calculation does not change the roles of the two ingredients: the covariate values belong to the observation, and the coefficient estimates are shared by observations in the fitted model.

How Predictors Differ Across Observations

Shared coefficients do not imply identical predictors. If two observations have different covariate values, their weighted sums may differ. Subtracting their predictor expressions gives an exact accounting of the difference: each coefficient weights the difference in its covariate.

To verify the result, write the predictor for each observation using the definition, then subtract and group the terms by covariate:

If \(x_{ij}=x_{kj}\) for every \(j\), every difference \(x_{ij}-x_{kj}\) is zero, so the expression equals zero and \(\eta_i=\eta_k\). This is an algebraic statement about the model predictors. It does not assert that the two observed responses must match. □

Changing One Covariate

The weighted-sum definition also gives a direct way to understand a change in one covariate. Hold the coefficients and every other covariate fixed, and change covariate \(r\) from \(x_{ir}\) to \(x_{ir}+\Delta\). Only one term in the predictor changes.

This result describes change on the predictor scale. It says that increasing a covariate by one unit changes the predictor by its coefficient, provided the other covariates and coefficients stay fixed. If the coefficient is negative, a positive increase in the covariate decreases the predictor. If the coefficient is zero, changing that covariate alone does not change the predictor. Any corresponding change in the mean or response probability depends on how the link function maps predictors to means.

A Categorical-Covariate Calculation

Covariates need not be measured on a continuous numerical scale. A two-category predictor can be represented by an indicator with values 0 and 1. In the linear predictor, that indicator’s contribution is zero for the category coded 0 and its coefficient for the category coded 1. The meaning of this coefficient depends on the model and coding; the arithmetic is still the same weighted sum.

Common Mistakes and Why the Predictor Matters

• Treating coefficients as observation-specific. The wrong reasoning is that each patient gets a different age coefficient because each patient has a different age. The correction is that the age values vary by patient, while the model’s age coefficient is shared.
• Calling a covariate value the predictor. A value such as \(x_{i1}=6\) is only one input. The predictor is the sum of all covariate contributions, such as \(x_{i1}\beta_1+x_{i2}\beta_2\).
• Ignoring coefficient signs. A positive covariate value multiplied by a negative coefficient contributes a negative quantity. Calculate each product with its sign before adding.
• Interpreting the predictor as a response without checking the scale. The predictor is on the link scale in a GLM. A probability, expected count, or other mean is obtained through the model’s link relationship, not by assuming \(\eta_i\) itself is that response-scale quantity.

The linear predictor matters because it is the part of a GLM that directly combines covariate information with model coefficients. It gives a compact, observation-specific value that can be compared across covariate profiles and then connected to the modeled mean through the link. Keeping the roles of values and coefficients clear makes both calculations and interpretations more reliable.