Exposure Time Offsets in Count PK/PD Models

Why Count Models Need an Exposure-Time Offset

A count is meaningful only in relation to the time over which events could have been observed. Twelve episodes recorded during 20 hours of monitoring represent a different event rate from twelve episodes recorded during 200 hours. If observation durations vary across participants or periods, modeling the counts as though they had equal opportunities to accrue can distort the exposure-response relationship.

In a Poisson PK/PD model, an exposure-time offset accounts for these unequal observation opportunities. It expresses the expected count as the product of the observation duration and an event rate that may depend on drug concentration or other predictors. The central distinction is that a coefficient describes a contrast in rates, whereas expected counts also depend on how long observation continues.

Definition: Let $Y_i$ be the nonnegative integer count observed for record $i$, and let $T_i>0$ be its time at risk or monitored exposure duration, expressed in a consistent unit. A Poisson model with a log-exposure-time offset is $$ Y_i\mid T_i,\mathbf{x}_i \sim \operatorname{Poisson}(\mu_i), \qquad \log(\mu_i)=\log(T_i)+\beta_0+\boldsymbol{\beta}^{\mathsf T}\mathbf{x}_i. $$ The offset is $\log(T_i)$, whose coefficient is fixed at one rather than estimated. The corresponding event rate per unit time is $\lambda_i=\mu_i/T_i=\exp(\beta_0+\boldsymbol{\beta}^{\mathsf T}\mathbf{x}_i)$.

Here, $T_i$ is the numerical duration in a chosen unit, such as hours, and $\mu_i=E(Y_i\mid T_i,\mathbf{x}_i)$ is the conditional expected count over that duration. The rate $\lambda_i$ has units of events per chosen time unit. For example, if $T_i$ is in hours, $\lambda_i$ is in events per hour. The predictor $\mathbf{x}_i$ may include concentration, treatment, or other covariates; this tutorial concerns the time offset, not the form of the concentration-response function.

The model assumes that, conditional on the predictors and duration, the count follows a Poisson distribution and the expected count grows in direct proportion to duration. In particular, the same rate applies across the interval represented by a record. Records are treated as conditionally independent for the likelihood unless a dependence structure is specified. The offset does not by itself adjust for differences in event definitions, incomplete monitoring, or changes in the event rate during an interval.

Why the Offset Has a Fixed Coefficient

The fixed coefficient follows from the intended rate model. If the expected count over $T_i$ time units is the rate multiplied by time, then $\mu_i=T_i\lambda_i$. Taking logarithms gives $\log(\mu_i)=\log(T_i)+\log(\lambda_i)$. Modeling the log rate as $\beta_0+\boldsymbol{\beta}^{\mathsf T}\mathbf{x}_i$ therefore yields the offset form. Estimating a coefficient for $\log(T_i)$ would instead allow the expected count to grow as a power of duration, which is a different model and does not enforce proportional accumulation over time.

The likelihood makes the role of the offset explicit. For one observation, the Poisson probability is $P(Y_i=y_i)=\exp(-\mu_i)\mu_i^{y_i}/y_i!$. Substituting $\mu_i=T_i\exp(\eta_i)$, where $\eta_i=\beta_0+\boldsymbol{\beta}^{\mathsf T}\mathbf{x}_i$, gives the log-likelihood contribution

\ell_i = y_i\{\log(T_i)+\eta_i\} -T_i\exp(\eta_i)-\log(y_i!).

Thus duration contributes both to the log mean and to the expected count in the likelihood. It is not an additional estimated effect: its coefficient remains one. For a model with an intercept and one binary predictor $x_i$, differentiating the summed log-likelihood with respect to the intercept gives

\frac{\partial \ell}{\partial \beta_0} =\sum_i\{y_i-T_i\exp(\beta_0+\beta_1x_i)\}.

At the maximum likelihood estimate, this score is zero. The fitted total expected count therefore equals the observed total count, with each record's duration included in its expected count. This score equation is also a useful check on what the offset is doing: longer observation periods contribute more expected events at the same rate.

Worked Example: Estimating a Rate Ratio From Unequal Durations

Suppose a hypothetical study records an adverse-event count for two concentration groups. In the lower-concentration group, 18 events occur over 120 monitored hours; in the higher-concentration group, 30 occur over 100 monitored hours. Assume records are independent Poisson counts, durations are measured correctly, and the rate is constant within each group during monitoring. Let $x=0$ denote the lower group and $x=1$ the higher group.

With a separate rate for each group, the maximum likelihood rate estimate is the total count divided by total exposure time in that group. The estimated lower-group rate is

\widehat{\lambda}_0=\frac{18}{120\ \text{h}} =0.1500\ \text{events/h}.

For the higher-concentration group,

\widehat{\lambda}_1=\frac{30}{100\ \text{h}} =0.3000\ \text{events/h}.

Under $\log(\mu_i)=\log(T_i)+\beta_0+\beta_1x_i$, the fitted intercept is the log rate in the reference group and the slope is the log rate ratio. Substitution gives

\widehat{\beta}_0=\log(0.1500)=-1.897,\qquad \widehat{\beta}_1=\log\left(\frac{0.3000}{0.1500}\right) =\log(2)=0.6931.

The fitted rate ratio is $\exp(\widehat{\beta}_1)=2.00$. In this hypothetical comparison, the higher-concentration group has an estimated event rate twice the lower-group rate. This is a rate contrast, not a claim that the higher group must have twice as many observed events: the groups were monitored for different total durations. The result is an association in these data; a causal interpretation would require an appropriate study design and assumptions beyond the offset model.

Rate Ratios and Expected-Count Contrasts

For a one-unit increase in a predictor $x$, holding other modeled predictors fixed, the model's rate ratio is $\exp(\beta_x)$. This ratio does not depend on duration. But for a particular duration $T$, the expected count is $T\lambda$, so an expected-count difference depends on $T$. For two predictor settings with rates $\lambda_a$ and $\lambda_b$, the rate ratio is $\lambda_b/\lambda_a$, while the expected-count difference over the same duration is $T(\lambda_b-\lambda_a)$.

When comparing different durations, calculate each expected count using its own duration. A rate ratio can remain constant while the difference in expected counts changes with the observation window. This distinction is important when translating a model coefficient into a quantity relevant to monitoring schedules or clinical follow-up.

Worked Example: Translating a Rate Ratio Into Counts

Consider a hypothetical fitted model with a reference rate of $0.1200$ events per hour and a rate ratio of $1.50$ for a higher concentration. Assume both rates apply throughout each specified observation period. The higher-concentration rate is

\lambda_{\text{high}}=0.1200\times1.50 =0.1800\ \text{events/h}.

Over four hours, the expected counts are $4(0.1200)=0.4800$ events at the reference concentration and $4(0.1800)=0.7200$ events at the higher concentration. The expected-count difference is $0.7200-0.4800=0.2400$ events over four hours. Over ten hours, the corresponding expected counts are $1.200$ and $1.800$ events, a difference of $0.6000$ events. The rate ratio remains $1.50$ in both comparisons, while the expected-count difference changes with duration.

An expected count need not be an integer: it is the mean of the count distribution, not a prediction that a particular record will contain a fractional event. The model predicts $0.7200$ expected events over four hours at the higher concentration; an observed count remains a whole number.

Why Ignoring Duration Can Reverse a Comparison

A count-only comparison implicitly treats observation opportunities as equal, unless duration is otherwise handled. That can be a serious error when groups differ in follow-up time. The next example shows how total counts can suggest the opposite direction from rates.

Worked Example: Counts Suggest One Direction, Rates Another

In a hypothetical count dataset, a reference group has 20 events over 50 hours of observation, while a second group has 24 events over 100 hours. Assume independent Poisson counts and a constant rate within each group. Looking only at total counts gives $24/20=1.20$, which might be misread as a 20% higher event rate in the second group. The exposure-adjusted rate estimates are instead

\widehat{\lambda}_{\text{ref}}=\frac{20}{50\ \text{h}} =0.4000\ \text{events/h},\qquad \widehat{\lambda}_{\text{second}}=\frac{24}{100\ \text{h}} =0.2400\ \text{events/h}.

The fitted rate ratio for the second group relative to the reference is $0.2400/0.4000=0.6000$, or a 40% lower estimated rate. In a model with a binary group indicator, the corresponding coefficient is $\log(0.6000)=-0.5108$. The second group has more total events because it contributes twice as many observation hours, despite having a lower estimated rate.

The comparison is justified only if the hours represent comparable time at risk and the stated Poisson assumptions are reasonable. If the second group's monitoring time includes periods when events could not occur or could not be detected, simply dividing by that duration would not repair the data problem.

Checking the Offset and Choosing an Approach

Before fitting or interpreting an offset model, verify the duration variable and the model's time scale. The following checks address the offset itself; they do not replace broader assessment of Poisson-model fit.

Use actual time at risk. Confirm that $T_i$ measures time during which the event could be observed and counted, rather than calendar time that includes periods outside monitoring or eligibility.
Use consistent units. All durations in a fit should use the same unit. Changing from hours to minutes changes the numerical intercept and rate units, but does not change fitted rate ratios when the conversion is applied consistently.
Check the assumed proportionality. The offset assumes doubling duration doubles the expected count at fixed modeled predictors: $\mu(2T,\mathbf{x})=2\mu(T,\mathbf{x})$. If that assumption is implausible, the simple offset model may not describe the data.
Do not use zero duration as an ordinary row. Since $\log(0)$ is undefined, a zero-duration record cannot be entered as a standard offset observation. It provides no observed time at risk; determine whether it should be excluded or whether its exposure was recorded incorrectly.
Inspect event and duration records together. Check for implausible counts, duplicated intervals, inconsistent monitoring windows, and systematic differences in observation procedures across predictor groups.

A model without an offset can be reasonable when every record has the same observation duration; then duration is a common factor in expected counts. With unequal durations, omitting the offset generally confounds differences in time observed with differences in rates. An alternative is to analyze rates $Y_i/T_i$ directly, but that does not generally preserve the Poisson count likelihood and can give observations with very different precision inappropriate influence. The offset retains the count likelihood while accounting for each record's exposure time.

The offset is also not a general remedy for misspecification. If event intensity changes systematically within a monitoring interval, a single constant rate may be inadequate. If concentration varies substantially during an interval, assigning one concentration value to the entire record may fail to represent the relevant exposure. The time offset addresses duration only; it does not model such time variation, within-subject dependence, or departures from the Poisson variance assumption.

Common Mistake

Common mistake: “The second group had 24 events and the first had 20, so the second group's event rate was 20% higher.” This conclusion uses counts without accounting for the observation time. In the worked example, the rates are $24/100=0.2400$ and $20/50=0.4000$ events per hour, so the estimated rate ratio is $0.6000$, not $1.20$. Compare rates when the question concerns event intensity; compare expected counts only after specifying the observation durations.

Another frequent error is to estimate a duration coefficient when the intended model requires proportional accumulation. The offset sets that coefficient to one by construction. If the data or scientific setting do not support proportional growth with time, that is a reason to reconsider the model, not to describe the ordinary offset coefficient as freely estimated.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A hypothetical group records 15 events over 75 monitored hours. What is its estimated rate in events per hour?
In a log-link model, a concentration coefficient is $0.4055$. What is the rate ratio for a one-unit increase in concentration, approximately?
At one predictor setting, the rate is $0.0800$ events per hour; at another, it is $0.1200$ events per hour. Find the rate ratio and the expected-count difference over five hours.
Why is a zero-duration record not entered as an ordinary observation with a log-time offset?

Solutions. (1) The estimated rate is $15/75=0.2000$ events per hour. (2) The rate ratio is $\exp(0.4055)\approx1.500$; the model associates a one-unit increase with a rate 1.50 times as large, holding other modeled predictors fixed. (3) The rate ratio is $0.1200/0.0800=1.500$. Over five hours, the expected counts are $0.6000$ and $0.4000$ events, so the difference is $0.2000$ expected events over five hours. (4) Its offset would require $\log(0)$, which is undefined, and zero duration supplies no time at risk for estimating an event rate.