The Log Link And Positive Means

From a Linear Predictor to a Positive Mean

In a generalized linear model, the linear predictor is built from model terms, while the response of interest is often a conditional mean. A link function connects these two quantities. For ratemaking targets such as expected claim counts, the fitted mean must not be negative. The log link is useful because its inverse maps every finite real-valued predictor to a strictly positive mean.

This tutorial focuses on that mapping and its consequences. It does not require a particular response distribution: the log-link relationship concerns the mean, although a GLM also specifies a response distribution and its assumptions. The examples use hypothetical claim-count models with a Poisson response so that the distribution is explicit; the arithmetic of the link itself does not depend on the Poisson choice.

The predictor \(\eta\) may be any finite real number, including zero or a negative value. The fitted mean is not the predictor itself: the inverse link must be applied. If \(\eta=0\), then \(\mu=\exp(0)=1\); if \(\eta<0\), then \(0<\mu<1\). A negative predictor therefore does not imply a negative expected claim count.

Why the Exponential Guarantees Positivity

The log link is defined by \(\eta=\ln(\mu)\). Because the natural logarithm and exponential function are inverses, exponentiating both sides recovers \(\mu\). This is also the reason the inverse link has the required positivity property: the exponential of every finite real number is strictly positive.

This argument has an important boundary condition. The log link requires a strictly positive mean, because \(\ln(0)\) is undefined. A finite predictor cannot produce a fitted mean of exactly zero; as \(\eta\) decreases without bound, \(\exp(\eta)\) approaches zero but remains positive. Thus the link ensures positivity, not nonnegativity including zero.

In a model, \(\eta\) is usually a sum of terms. For example, a hypothetical model might have an intercept and one or more predictor contributions. Once those terms produce \(\eta\), the fitted mean is obtained by exponentiating the total—not by exponentiating only one term or by adding the terms to a mean on its original scale.

Additive Predictor Changes and Multiplicative Mean Changes

The exponential function converts addition in the predictor into multiplication on the mean scale. Suppose two fitted predictors differ by an amount \(\Delta\). Write the second predictor as \(\eta_2=\eta_1+\Delta\). Their fitted means are \(\mu_1=\exp(\eta_1)\) and \(\mu_2=\exp(\eta_2)\). Dividing the second mean by the first gives:

This derivation shows that the multiplicative change in the fitted mean depends only on the predictor difference \(\Delta\), not on the starting predictor \(\eta_1\). The corresponding relative change is \(\exp(\Delta)-1\). If \(\Delta>0\), the mean increases; if \(\Delta<0\), it decreases; and if \(\Delta=0\), the fitted means are equal. The result describes a difference between predictors and does not, by itself, require assigning a particular interpretation to any individual model term.

Negative Predictors, Scale, and Model Boundaries

A negative predictor is entirely compatible with a positive fitted mean. It often corresponds to a mean below one when the modeled response is a count per stated unit, but it never makes the mean negative. The predictor and mean have different roles and may also be expressed on different scales: \(\eta\) is a logarithmic quantity, while \(\mu\) is in the response’s measurement units.

The unit used for the response must be consistent. If a count mean is stated per policy-year, the model output is interpreted per policy-year. Changing the response’s unit changes the numerical mean and requires a corresponding change in the model’s predictor representation. The log-link equation does not make units irrelevant; it specifies how the numerical mean in the chosen unit relates to \(\eta\).

The positivity guarantee is a property of the inverse link, not a guarantee that the model is otherwise appropriate. A log-link GLM still requires a suitable response distribution, a defensible predictor structure, and appropriate estimation and validation. In a Poisson model, for example, the conditional distribution is Poisson with mean \(\mu\) and variance \(\mu\). If that distributional assumption is unsuitable for the data, the fact that \(\exp(\eta)\) is positive does not repair the mismatch.

Common Exam Trap: Using the Link Instead of Its Inverse

The most direct error is to compute the log of the predictor or to use the predictor as the fitted mean. With the log link, the model statement is \(\ln(\mu)=\eta\), not \(\mu=\ln(\eta)\) and not \(\mu=\eta\). The correct operation is to exponentiate \(\eta\). A quick reasonableness check is that a finite predictor must produce a positive fitted mean.

This distinction matters in ratemaking because fitted means are used as estimates of expected response on the original scale, while the model is constructed on the predictor scale. The log link supports positive fitted means and gives a precise relationship between predictor differences and mean ratios. These are core properties aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus; they do not replace the need to assess whether the chosen response family and model structure suit the intended application.