MerQuR+Statistics Glossary

Statistics Glossary → Offset (exposure term)

Offset (exposure term)concept

A term added to the linear predictor of a count model with its coefficient fixed at 1 (usually the logarithm of exposure), to account for observations being counted over different lengths of time, areas or populations.

The number of infections a hospital sees in a year depends on how many patient-days of care it provides; a large hospital seeing more infections does not mean it is riskier. Poisson and negative binomial regression use a log link, so log(expected count) = log(exposure) + β₀ + β₁x... The coefficient of log(exposure) is not estimated but fixed at 1, so the model in effect explains the rate per unit of exposure (log(count / exposure)) rather than the raw count, and the coefficients are read as rate ratios.

An offset is not a predictor: if exposure were entered as an ordinary predictor, its coefficient would be estimated freely, relaxing the assumption that counts rise in proportion to exposure. That assumption is sometimes worth testing, but when a rate model is wanted, the offset is the right tool. The offset must be supplied on the log scale; if exposure is already stored as a logarithm, it should not be logged again. Because the logarithm of zero is undefined, observations with zero exposure must be left out.

An everyday example

Traffic accidents in two districts are compared: district A has 40 a year and district B has 20. But vehicles travel 8 million vehicle-km a year on A's roads and only 2 million on B's. With log(vehicle-km) as an offset, the model compares accidents per million vehicle-km: 5 in A and 10 in B. So B is more dangerous despite having fewer accidents.

In MerQur

It appears as the 'Offset (optional)' and 'Offset already on log scale' options in Poisson / Negative Binomial Regression, 'Offset (log exposure)' in Generalized Estimating Equations and 'Offset (linear-predictor scale)' in Generalized Linear Mixed Models.

Analyses where this concept appears

Poisson / Negative Binomial Regression (Count Regression)Generalized Estimating Equations (GEE)Generalized Linear Mixed Models (GLMM)

Related concepts

Incidence rate ratio (IRR)Distribution family (GLM)