Interpreting Coefficients in Offset Models

What an Offset-Model Coefficient Describes

In a count model with exposure, the observed count depends both on the underlying event rate and on how much opportunity there was for events to occur. A policy observed for two years, for example, may be expected to generate more claims than a similar policy observed for six months, even when their modeled annual claim rates are equal. A fixed exposure offset accounts for this difference in observation time.

The key interpretive point is that a coefficient in a log-link offset model describes a change in the modeled rate, conditional on the other predictors. It does not, by itself, describe a fixed additive change in the count. The exposure determines how a rate translates into an expected count.

Definition: Suppose $Y_i$ is a count for observation $i$, $e_i>0$ is its known exposure measured in a stated unit, and the conditional mean follows a Poisson log-link model: $$ Y_i \mid \boldsymbol{x}_i \sim \operatorname{Poisson}(\mu_i), \qquad \log(\mu_i)=\log(e_i)+\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta}. $$ The term $\log(e_i)$ is a fixed offset: its coefficient is constrained to be 1. The modeled rate per exposure unit is $$ \lambda_i=\frac{\mu_i}{e_i} =\exp\!\left(\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta}\right). $$ This interpretation assumes exposure is measured consistently and is known, the conditional mean is correctly specified, and the log link is used. The Poisson model additionally assumes the conditional variance equals the conditional mean; that distributional assumption is not needed for the algebraic rate-ratio interpretation if the same log-mean form is used with another suitable count model.

The intercept and slopes therefore operate on the log rate. A one-unit increase in a predictor with coefficient $\beta_j$, holding the other predictors fixed, changes the log rate by $\beta_j$. On the rate scale, the corresponding multiplicative change is $\exp(\beta_j)$. For a categorical predictor, this is the rate ratio for that category compared with its reference category, subject to the model's other predictors.

Why Exponentiating Gives a Rate Ratio

The rate-ratio result follows directly by comparing two model profiles. Suppose profiles A and B differ by one unit in predictor $x_j$, while all other predictors and the exposure unit are held fixed. Their modeled log rates differ by $\beta_j$. Exponentiating both log rates and dividing cancels the common terms:

\frac{\lambda_B}{\lambda_A} = \frac{\exp(\boldsymbol{x}_B^\mathsf{T}\boldsymbol{\beta})} {\exp(\boldsymbol{x}_A^\mathsf{T}\boldsymbol{\beta})} = \exp\!\left(\boldsymbol{x}_B^\mathsf{T}\boldsymbol{\beta} -\boldsymbol{x}_A^\mathsf{T}\boldsymbol{\beta}\right) = \exp(\beta_j).

This is a ratio of rates, not a difference in expected counts. In general, $\mu_i=e_i\lambda_i$, so the expected-count ratio for two observations is $\mu_B/\mu_A=(e_B/e_A)(\lambda_B/\lambda_A)$. The rate ratio equals the expected-count ratio only when the exposures are equal.

Interpretation rule: For a one-unit increase in a predictor with coefficient $\beta$, the modeled rate is multiplied by $\exp(\beta)$, holding other predictors fixed. The percentage change in rate is $100\{\exp(\beta)-1\}\%$. If $\beta<0$, this percentage is negative, indicating a rate decrease.

Worked Examples

Worked Example: Comparing Two Risk Categories

A hypothetical insurer models claim counts using a Poisson log-link model. Exposure is measured in policy-years. The fitted model is $\log(\mu_i)=\log(e_i)-3+(\log 2)H_i+(\log 1.5)R_i$, where $H_i=1$ for a higher-risk category and $R_i=1$ for a rural territory; both indicators are 0 for their reference categories. Interpret the coefficients and calculate the expected count for a higher-risk rural policy observed for 2.4 policy-years. Round final rates and counts to four decimals.

Method. First calculate the reference rate by setting both indicators to 0. Then apply the exponentiated coefficients as rate multipliers. Finally multiply the resulting rate by exposure to obtain the expected count.

Reference rate. With $H_i=R_i=0$, the log rate is $-3$, so $\lambda_0=\exp(-3)=0.049787\ldots$, or $0.0498$ claims per policy-year.

Higher-risk effect. The coefficient is $\log 2$, so its rate ratio is $\exp(\log 2)=2$. Holding territory fixed, the higher-risk category has twice the modeled rate of the reference risk category, a 100% increase.

Rural effect. The coefficient is $\log 1.5$, so its rate ratio is $\exp(\log 1.5)=1.5$. Holding risk category fixed, rural territory has 1.5 times the modeled rate of the reference territory, a 50% increase.

Combined profile. The two factors multiply, giving rate $0.049787\ldots\times2\times1.5=0.149361\ldots$, or $0.1494$ claims per policy-year. For exposure $e_i=2.4$, the expected count is $2.4\times0.149361\ldots=0.358467\ldots$, or $0.3585$ claims.

Check and common mistake. The combined rate is three times the reference rate, and the expected count is $2.4$ times that combined rate. The common error is to treat $\log 2$ as a 69.3% rate increase because its numerical value is about 0.693. The correct increase is $100\{\exp(\log 2)-1\}=100\%$.

Worked Example: A Continuous Predictor and a Rate Prediction

A hypothetical model for service calls uses exposure in machine-years and has $\log(\mu_i)=\log(e_i)+\log(0.04)+(\log 1.2)z_i+(\log 0.75)I_i$, where $z_i=(\text{age in years}-50)/10$, and $I_i=1$ if a machine received a preventive inspection, otherwise 0. Interpret the age effect and find the expected calls for a 60-year-old inspected machine with 1.5 machine-years of exposure. Round the final rate and expected count to four decimals.

Age interpretation. Increasing age by 10 years increases $z_i$ by 1. Thus the rate ratio for a 10-year increase, holding inspection status fixed, is $\exp(\log 1.2)=1.2$: a 20% rate increase. For a 20-year increase, the multiplier is $1.2^2=1.44$, a 44% increase.

Inspection interpretation. At the same age, inspection multiplies the modeled rate by $\exp(\log 0.75)=0.75$, or a 25% decrease.

Prediction. For age 60, $z_i=(60-50)/10=1$. For an inspected machine, $I_i=1$. Therefore the rate is $0.04\times1.2\times0.75=0.0360$ calls per machine-year. Multiplying by exposure gives $1.5\times0.0360=0.0540$ expected calls.

Check and common mistake. The intercept corresponds to age 50 and no inspection, because $z_i=0$ and $I_i=0$ there. Treating the coefficient of $z_i$ as the effect of one year would be wrong: $z_i$ increases by one for a 10-year, not a one-year, age increase.

Worked Example: Interpreting an Interaction

A hypothetical model estimates claim rates per vehicle-year using two binary predictors: $A=1$ indicates a selected safety feature, and $B=1$ indicates a selected use category. The model is $\log(\mu_i)=\log(e_i)+\log(0.03)+(\log 2)A_i+(\log 1.5)B_i+\{\log(2/3)\}A_iB_i$. Interpret the effect of the safety feature at each value of $B$, and calculate the expected count for a policy with $A=B=1$ and exposure of 4 vehicle-years. Round final results to four decimals.

Effect of $A$ when $B=0$. Changing $A$ from 0 to 1 changes the log rate by $\log 2$. The rate ratio is $\exp(\log 2)=2$.

Effect of $A$ when $B=1$. The change in log rate now includes both the $A$ coefficient and the interaction coefficient: $\log 2+\log(2/3)=\log(4/3)$. The rate ratio is therefore $4/3=1.3333\ldots$, not 2.

Joint-profile prediction. With both indicators equal to 1, the rate multiplier relative to $A=B=0$ is $2\times1.5\times(2/3)=2$. The rate is $0.03\times2=0.0600$ claims per vehicle-year. At exposure 4, the expected count is $4\times0.0600=0.2400$ claims.

Check and common mistake. The interaction does not describe the effect of $A$ alone. It modifies the $A$ effect when $B=1$. Reporting the main-effect rate ratio of 2 as the effect of $A$ for both groups would ignore this modification.

Interpreting Coefficients in Practice

A coefficient's meaning depends on the predictor's coding and units. For a continuous predictor measured in units of one year, $\exp(\beta)$ is the rate ratio for a one-year increase. If the predictor is instead coded in decades, the same coefficient describes a decade increase. Always identify the predictor scale before stating the effect. For a factor variable, identify the reference category; its coefficient compares the coded category with that reference while holding other predictors fixed.

For a change of $c$ units in a continuous predictor with no interaction involving that predictor, the rate ratio is $\exp(c\beta)$. It is usually clearer to report both the scale of the change and the multiplier, such as “a 10-year increase multiplies the modeled rate by 1.2,” rather than to state only the coefficient.

When an interaction is present, a main-effect coefficient is conditional on the interacting predictor being at its reference or zero value. To compare profiles, calculate the difference in their linear predictors and exponentiate that difference. This procedure avoids treating a conditional coefficient as though it were a universal effect.

The fixed offset does not change the coefficient-to-rate-ratio rule. It ensures that the expected count scales with known exposure while the predictors describe the rate. If two observations have different exposures, compare their modeled rates to assess their relative intensity; compare expected counts only when the exposure difference is also part of the question.

Common Interpretation Traps

Reading a log-scale coefficient as a percentage. A coefficient of 0.2 is not exactly a 20% increase. Its rate ratio is $\exp(0.2)$, and its percentage rate change is $100\{\exp(0.2)-1\}\%$.
Confusing rate and count. A rate ratio of 1.5 means a 50% higher modeled rate. If exposure differs, the expected-count ratio also includes the exposure ratio.
Ignoring the predictor's unit. A coefficient for a 10-unit increase cannot be described as the effect of one unit without rescaling it.
Ignoring the reference category. A factor coefficient describes a contrast with its model reference category, not with every other category.
Ignoring an interaction. With an interaction, the effect of one predictor depends on the interacting predictor's value; add the relevant coefficients on the log scale before exponentiating.

Key takeaway: In a log-link model with a fixed exposure offset, exponentiating a coefficient gives a multiplicative effect on the modeled rate for the predictor contrast it represents. Multiply the rate by exposure only when the question asks for an expected count.

Check Your Understanding

For each problem, interpret the requested rate effect and keep rate comparisons distinct from expected-count comparisons.

A log-link offset model has coefficient $\log(1.25)$ for a binary indicator. What is the rate ratio for indicator 1 versus 0, and what is the percentage rate change? (Answer: rate ratio $1.25$; 25% increase.)
A continuous predictor has coefficient $\log(0.8)$ per unit. Find the rate ratio and percentage rate change for a three-unit increase. (Answer: rate ratio $0.5120$; 48.8% decrease.)
The reference rate is $0.025$ claims per exposure unit. A category coefficient is $\log(2.4)$. For exposure $1.75$, calculate the category's modeled rate and expected count. (Answer: rate $0.0600$ per exposure unit; expected count $0.1050$.)
A model contains binary predictors $A$ and $B$, with coefficients $\log(1.8)$ for $A$ and $\log(5/6)$ for $AB$. What is the rate ratio for changing $A$ from 0 to 1 when $B=1$? (Answer: $1.8\times(5/6)=1.5000$, a 50% increase.)
A model has coefficients $\log(1.1)$ for $x$ and $\log(0.7)$ for $z$. Profile A has $x=2,z=0$, and exposure $0.8$; profile B has $x=1,z=1$, and exposure $2.4$. Find the rate ratio B-to-A and the expected-count ratio B-to-A. (Answer: rate ratio $0.7/1.1=0.6364$; expected-count ratio $3(0.6364)=1.9091$.)

What You'll Learn