Exposure Offsets in Count Models

Why Exposure Belongs in a Count Model

A count depends not only on the underlying event rate but also on how long, or how much, an individual was observed. Six events during six person-years of follow-up and six events during one person-year are not evidence of the same event rate. A count model that ignores this difference can confuse more opportunity to observe events with a higher rate of events.

A fixed exposure offset adjusts the expected count for known observation opportunity. It does not model exposure as an ordinary predictor with an estimated effect. Instead, it assumes that expected counts increase in direct proportion to exposure, conditional on the model's covariates and random effects. The resulting regression describes rates; multiplying a rate by a particular exposure gives an expected count for that exposure.

The numerical value of \(T_{ij}\) must be expressed in a stated unit. If exposure is measured in person-years, for example, the model's rate is per person-year. Changing the exposure unit changes the numerical rate and intercept, but not fitted expected counts when the offset and rate units are changed consistently.

Model Specification and Interpretation

Write \(\eta_{ij}=\mathbf{x}_{ij}^{\mathsf T}\boldsymbol{\beta}+\mathbf{z}_{ij}^{\mathsf T}\mathbf{b}_i\), where the fixed effects are \(\boldsymbol{\beta}\) and \(\mathbf{b}_i\) denotes the individual-level random effects in the mixed model. For a Poisson model, the conditional mean and distribution are

Equivalently, \(\mu_{ij}=T_{ij}\exp(\eta_{ij})\). Thus, \(\exp(\eta_{ij})\) is the conditional event rate per unit of exposure and \(\mu_{ij}\) is the conditional expected count in the actual interval. A fixed-effect coefficient \(\beta_k\) describes a multiplicative change \(\exp(\beta_k)\) in the rate for a one-unit increase in its predictor, holding other predictors and the random effects fixed. It is not, in general, a fixed multiplicative change in the expected count across observations with different exposures.

For an overdispersed count outcome, a negative-binomial model can use the same mean structure. In one common parameterization, conditional on the random effects,

where \(\alpha>0\) is a dispersion parameter. The offset adjusts the mean for exposure in both model families; it does not by itself account for extra-Poisson variation. The distributional assumptions and variance structure remain different.

Why the Offset Coefficient Is One

The fixed coefficient follows from the assumed proportionality between exposure and expected count. If a rate is constant over an observation interval, doubling the exposure doubles the expected count. For a rate \(r_{ij}=\exp(\eta_{ij})\), the expected count is therefore \(T_{ij}r_{ij}\). Taking logarithms gives the offset form.

This also follows from a constant-rate Poisson process: counts from disjoint pieces of exposure add, and their expected counts add. If the rate is \(r\) per person-year, then an interval of \(T\) person-years has expected count \(rT\). The offset is not an arbitrary convention; it encodes this particular exposure relationship.

For example, the conditional Poisson log-likelihood contribution for one interval, apart from terms independent of the model parameters, is

The known exposure enters both the log-mean and the expected-count term. Fixing its coefficient at one preserves the assumed mean \(\mu_{ij}=T_{ij}\exp(\eta_{ij})\); estimating a coefficient on \(\log(T_{ij})\) instead would allow a different, data-estimated relationship between exposure and expected count.

Worked Rate Comparison With Unequal Exposure

With a log link and an exposure offset, a coefficient for a group indicator compares rates while accounting for exposure. This is the central reason to use an offset rather than compare raw counts when observation times differ.

Negative-Binomial Means Still Use the Offset

An offset is about the mean-exposure relationship, not the choice between Poisson and negative-binomial variation. In either model, if the mean is proportional to exposure, the log mean includes \(\log(T_{ij})\) with coefficient one. The negative-binomial model allows conditional variance above the Poisson variance while retaining this mean specification.

Assumptions, Diagnostics, and Alternatives

Using \(\log(T_{ij})\) as a fixed offset is appropriate only when the exposure is known and the conditional expected count is proportional to it. In practice, check that the exposure measures actual opportunity for events and is positive for every modeled interval. A person-year at risk is not interchangeable with calendar time if individuals contribute different amounts of time at risk.

The proportionality assumption can fail when the rate changes systematically during an interval, when risk depends on time since entry, or when an event changes subsequent opportunity to experience further events. Covariates or a more suitable model may then be needed. A single offset cannot repair an incorrectly defined risk period, omitted time patterns, dependence, or a misspecified count distribution.

• Check the exposure records. Confirm units, interval boundaries, and that time not at risk is excluded. An incorrect exposure produces an incorrect rate even if the model is otherwise fitted correctly.
• Inspect fitted counts and rates by exposure. Systematic residual patterns across exposure levels can indicate that proportional scaling is inadequate or that relevant predictors are missing. Use count-model diagnostics and predictive checks rather than judging the offset assumption from one coefficient alone.
• Assess the count distribution. If variation exceeds the Poisson model's conditional mean-variance relationship, a negative-binomial model may be a reasonable alternative. An offset does not remove overdispersion; conversely, extra dispersion does not establish that the exposure was measured correctly.
• Check for zero exposure and structural design issues. Since \(\log(0)\) is undefined, an interval with no exposure is not handled by this offset formula. Such an interval should not be treated as an ordinary at-risk count interval with a fabricated small exposure.

A useful alternative is to model \(\log(T_{ij})\) as a predictor with an estimated coefficient. That asks whether expected counts scale as \(T_{ij}^{\gamma}\) for an estimated \(\gamma\), rather than imposing proportionality (\(\gamma=1\)). It is not the same scientific model, and an estimated value near one does not automatically validate the exposure definition or other assumptions. Modeling a rate by dividing each observed count by exposure is also not generally equivalent to fitting the count likelihood: it can give inappropriate weight to intervals with different exposure and does not directly represent the discrete count distribution.

The offset fixes the exposure effect on the conditional mean, but it does not determine the random-effects structure or turn conditional effects into population-averaged effects. Predictions that include random effects and predictions averaged over their distribution answer different questions. Also, the offset is not a substitute for a time-to-event model when event timing, censoring, or event-specific risk processes are central to the question.

Common Pitfall

For interpretation, state both the rate unit and the comparison. For example, \(\exp(\beta_1)=1.50\) means a 1.50-fold conditional event rate, or a 50% higher conditional rate, for a one-unit increase in the predictor, holding other modeled terms fixed. It does not say that every individual will experience 50% more events, nor that the expected counts are 50% higher when exposures differ.