Exposure Offsets and Event Rates

Why Counts Need an Exposure Denominator

A count is meaningful only in relation to the opportunity for events to occur. Ten events over 100 person-years and ten events over 1,000 person-years are the same count, but they do not indicate the same event rate. In many applications, observations differ in follow-up time, population at risk, or another known measure of exposure. A count model needs to account for that difference rather than treat every observed count as if it had the same opportunity to arise.

An exposure offset incorporates a known denominator into a count model. It makes the expected count proportional to exposure while allowing explanatory variables to describe differences in the event rate. This distinction matters: a predictor can imply a higher rate, while a larger exposure can imply a higher count even when the rate is unchanged.

Definition: Let $Y_i$ be the event count for observation $i$, $E_i$ its known, positive exposure, and $x_i$ a vector of explanatory variables. A log-link count model with an exposure offset specifies $$ \log \mu_i = \log E_i + x_i^\top\beta, \qquad \mu_i = E_i\exp(x_i^\top\beta), $$ where $\mu_i=E(Y_i\mid x_i,E_i)$ is the conditional mean count. The term $\log E_i$ is the offset: its coefficient is fixed at one. The conditional event rate per unit of exposure is $\mu_i/E_i=\exp(x_i^\top\beta)$.

What the Offset Assumes

An offset is not merely a convenient way to add a denominator to a table. It imposes a substantive restriction: holding the explanatory variables fixed, doubling exposure doubles the conditional expected count. More generally, multiplying exposure by $c>0$ multiplies the expected count by $c$. This is appropriate when exposure measures the amount of opportunity for events and the conditional rate is stable over those exposure units, or when that proportionality is a reasonable model for the application.

The standard Poisson specification assumes $Y_i\mid x_i,E_i$ has a Poisson distribution with mean $\mu_i$. For estimating the conditional mean, Poisson quasi-maximum likelihood can also be used when the conditional variance is not Poisson, provided the mean specification is correct and the relevant regularity conditions hold. In that case, variance misspecification must be handled for inference, for example with sandwich standard errors. The observations must also be independent for the usual independent-observation variance calculation; dependent data require an appropriate covariance estimator or model.

The exposure must be known and strictly positive because $\log E_i$ is part of the linear predictor. A zero event count is allowed; zero exposure is different, since a rate is undefined for an observation with no opportunity for events. Exposure should be measured on a meaningful scale, such as person-years, resident-years, or thousands of resident-years. Changing units changes the numerical rate and typically the intercept, but not fitted counts when the model is transformed consistently.

Rates and Coefficients

The offset separates the conditional mean count into exposure multiplied by rate. If $x_{ij}$ is a covariate whose coefficient is $\beta_j$, increasing that covariate by one unit while holding the other covariates fixed multiplies the conditional rate by $\exp(\beta_j)$. This is a rate ratio. At the same exposure, it is also the conditional mean-count ratio. At different exposures, the count ratio includes the exposure ratio as well.

\frac{\mu_i}{\mu_k} = \frac{E_i}{E_k} \exp\!\left((x_i-x_k)^\top\beta\right), \qquad \frac{\mu_i/E_i}{\mu_k/E_k} = \exp\!\left((x_i-x_k)^\top\beta\right).

Thus, count comparisons and rate comparisons answer different questions. A count ratio describes expected event totals over the actual exposure amounts. A rate ratio compares expected events per exposure unit, conditional on the covariates. In regression, $\exp(\beta_j)$ for a one-unit covariate change does not by itself say how much the expected count changes if the compared observations have different exposures.

Worked Example: Equal Person-Time Rates, Unequal Counts

Suppose two hypothetical regions record 18 and 30 events over 900 and 1,500 person-years, respectively. Calculate the observed rates and compare the counts.

For Region A, the rate is $$ \frac{18}{900}=0.0200 \text{ events per person-year}. $$ For Region B, it is $$ \frac{30}{1500}=0.0200 \text{ events per person-year}. $$ The observed rates are equal. Yet Region B records more events because it contributes more person-time. The count ratio is $30/18=5/3\approx1.667$, and the exposure ratio is $1500/900=5/3\approx1.667$. These are equal ratios, not a claim that either unrounded ratio is exactly 1.667.

A common-rate model for these two observations has pooled rate $$ \frac{18+30}{900+1500} = \frac{48}{2400} = 0.0200 \text{ events per person-year}. $$ It gives expected counts $900(0.0200)=18$ and $1500(0.0200)=30$. In a model with a Region B indicator, the observed rate ratio is $0.0200/0.0200=1$, so the corresponding log rate-ratio coefficient is $\log(1)=0$. The larger count in Region B is accounted for by its larger exposure, not by a higher rate.

Worked Interpretation of a Rate Effect

Consider a model with a treatment indicator $T_i$, equal to one for treated observations and zero otherwise. Suppose exposure is measured in thousands of person-years and the fitted model is $$ \log\mu_i=\log E_i+\log(4)+\log(1.25)T_i. $$ The baseline conditional rate is 4 events per 1,000 person-years. The treatment coefficient is $\log(1.25)\approx0.2231$, so treatment status is associated with a 1.25 rate ratio, conditional on the model's specification. The model's rate comparison does not require the two observations to have the same exposure; the model compares their rates at the stated covariate values.

Worked Example: Different Exposures and a Treatment Rate Ratio

Use the model above to compare a hypothetical control observation with exposure $E=2$ thousand person-years and a treated observation with exposure $E=5$ thousand person-years. Compute their expected counts and distinguish the count ratio from the rate ratio.

For the control observation, $T=0$, so its rate is $$ \exp\{\log(4)+\log(1.25)(0)\}=4 \text{ events per 1,000 person-years}. $$ Its expected count is $\mu=2(4)=8$ events. For the treated observation, the rate is $4(1.25)=5$ events per 1,000 person-years, and its expected count is $\mu=5(5)=25$ events.

The rate ratio is $5/4=1.25$, matching the exponentiated treatment coefficient. The expected-count ratio is instead $25/8=3.125$. It decomposes into the exposure ratio times the rate ratio: $$ \frac{25}{8} = \frac{5}{2}\times1.25 = 3.125. $$ At the same exposure of 2 thousand person-years, the expected counts would be 8 for control and $2(5)=10$ for treated, a count ratio of 1.25. The 3.125 ratio compares observations that differ in both treatment status and exposure.

Population Denominators as Exposure

Exposure can represent population at risk over a specified time, not only individually accumulated person-time. For example, in a study where each observation is a district-year, the district's population may be used as exposure when the outcome is a count of events over one year and the rate is defined per resident-year. If observation periods differ in duration, population alone may not capture opportunity; a population-time measure is then more appropriate. If the population at risk changes during follow-up, the exposure should reflect that change as closely as the study design requires.

To use population values conveniently, analysts often choose a scale such as 10,000 residents. With an exposure $E_i$ measured in units of 10,000 resident-years, the rate $\mu_i/E_i$ is events per 10,000 resident-years. The scale is a choice of units; it does not alter the model's proportionality restriction.

Worked Example: District Counts With Population Exposure

Two hypothetical districts each contribute one year of observation. District A has 12,000 residents and records 24 events; District B has 20,000 residents and records 30 events. Express exposure in units of 10,000 resident-years, calculate the rates, and compare the rate and count ratios.

The exposures are $E_A=12{,}000/10{,}000=1.2$ and $E_B=20{,}000/10{,}000=2.0$ units of 10,000 resident-years. The rates are $$ \frac{24}{1.2}=20 \quad\text{and}\quad \frac{30}{2.0}=15 $$ events per 10,000 resident-years. The observed rate ratio for B relative to A is $15/20=0.750$: District B's rate is 25% lower in this comparison. Its count ratio relative to A is $30/24=1.25$, because its larger population exposure more than offsets its lower rate.

A two-district log-rate model with an indicator for B would represent the observed rate ratio by a coefficient $\log(0.750)\approx-0.2877$. The offset then accounts for population exposure: the fitted means at these observed rates are $1.2(20)=24$ and $2.0(15)=30$. These descriptive rates alone do not establish that district membership causes the rate difference.

Why the Offset Coefficient Is Fixed at One

The fixed coefficient follows directly from the intended rate model. Suppose the conditional rate is $\lambda_i=\exp(x_i^\top\beta)$. For exposure $E_i$, the conditional expected count is $E_i\lambda_i$. Taking logarithms produces an additive term $\log E_i$ with coefficient one. The offset therefore encodes a known proportional relationship; the data estimate the rate parameters $\beta$, not a separate exposure elasticity.

\begin{aligned} \mu_i &= E_i\lambda_i = E_i\exp(x_i^\top\beta),\\ \log\mu_i &= \log E_i+x_i^\top\beta. \end{aligned}

For the Poisson likelihood, the contribution from observation $i$, apart from the factorial term, is $Y_i\log\mu_i-\mu_i$. Substituting the offset model gives $$ Y_i\{\log E_i+x_i^\top\beta\}-E_i\exp(x_i^\top\beta). $$ Because $E_i$ is known, this expression is optimized over $\beta$; there is no free exposure coefficient to estimate. The derivation also shows the modeling commitment: if the relationship between exposure and expected count is not proportional, the fixed offset may be inappropriate.

Define the rate scale.
State what one exposure unit means, such as one person-year or 10,000 resident-years, and verify that each $E_i$ is positive.

Specify the conditional rate.
Write $\lambda_i=\exp(x_i^\top\beta)$, with the covariates and their units clearly identified.

Add the offset.
Use $\log\mu_i=\log E_i+x_i^\top\beta$, fixing the coefficient of $\log E_i$ at one.

Interpret comparisons on the right scale.
Exponentiate covariate contrasts for rate ratios; include $E_i/E_k$ as well when comparing expected counts at unequal exposures.

Common Mistakes and Practical Checks

Comparing raw counts as if exposure were equal. A larger count may reflect more people or longer follow-up rather than a higher event rate. Calculate or model the rate using the exposure denominator.
Calling a coefficient a count effect without stating exposure. An exponentiated coefficient is a rate ratio. It is also a mean-count ratio only for a comparison at equal exposure.
Estimating the offset coefficient by default. A free coefficient on $\log E_i$ would allow the expected count to scale as $E_i^\gamma$, rather than imposing proportionality. That is a different model, not the standard exposure-offset specification.
Using a denominator that does not represent opportunity. A population total may be inadequate when follow-up duration differs substantially or population at risk changes. Choose exposure to match the time and population at risk for the events.
Ignoring the rate units. A rate per person-year is numerically different from a rate per 1,000 person-years. State the scale, and transform exposure and the rate interpretation consistently.

Key takeaway: A log-exposure offset expresses the conditional mean count as exposure times a modeled rate. Covariate coefficients describe rate ratios; expected-count comparisons also depend on the ratio of exposures.

Check Your Understanding

Answer each question using the distinction between exposure, rate, and count.

A group records 16 events over 800 person-years, while another records 24 events over 1,200 person-years. What are the two rates, and how do the count and exposure ratios compare?
In an offset model, a covariate coefficient is $\log(0.80)$. State its rate-ratio interpretation for a one-unit increase, holding other covariates fixed.
Two observations have the same modeled rate, but the second has three times the exposure. What is the ratio of their expected counts?
Why is the coefficient of $\log E_i$ fixed at one in the standard exposure-offset model?
A population count covers six months in one district and twelve months in another. Why might population size alone be an incomplete exposure measure?