Why Exposure Belongs in a Positive-Amount Model
A positive amount observed over a longer exposure period will often have a larger expected total, even when the underlying rate is unchanged. For example, a hypothetical policy with three years of coverage might accumulate a larger positive covered amount than an otherwise similar policy with one year of coverage. A regression model should distinguish that difference in exposure from differences in risk characteristics.
For a positive continuous response, a Gamma regression with a log link can model the expected total as exposure multiplied by a covariate-dependent mean amount per exposure unit. The exposure term is included as an offset: its coefficient is fixed at one rather than estimated. This expresses proportionality to exposure while allowing the regression coefficients to describe how the amount per unit changes with predictors.
The response \(Y_i\) is the positive total amount for the observation, not a count and not necessarily an amount per claim. The model assumes that the conditional response is Gamma distributed, with the usual Gamma GLM mean-variance relationship $$ \operatorname{Var}(Y_i\mid \mathbf{x}_i,e_i)=\phi\mu_i^2, $$ where the dispersion parameter \(\phi>0\). A Gamma distribution has positive support, so this model is not appropriate for observations that can be zero or negative without an additional modeling strategy. Conditional independence is commonly assumed across observations when fitting the GLM; the mean formula itself states the conditional mean, not a claim that totals from different observations have any particular aggregate distribution.
The exposure must be positive and measured consistently. It might be policy-years, covered months, or another meaningful unit. The regression coefficients depend on the chosen unit through the intercept, but fitted expected totals do not change when the unit conversion is carried out correctly.
Why the Offset Has Coefficient One
Suppose the expected amount is proportional to exposure, with a covariate-dependent rate \(r_i\) per exposure unit. The proportionality assumption says \(\mu_i=e_ir_i\). If a log link is used for the mean and the rate follows a linear predictor, then taking logarithms gives the offset form directly.
Write \(\mu_i=e_ir_i\), where \(e_i\) is exposure and \(r_i\) is expected amount per exposure unit.
Let \(\log(r_i)=\beta_0+\sum_{j=1}^p\beta_jx_{ij}\), so \(r_i=\exp(\beta_0+\sum_{j=1}^p\beta_jx_{ij})\).
Because \(\mu_i=e_ir_i\), \(\log(\mu_i)=\log(e_i)+\log(r_i)=\log(e_i)+\beta_0+\sum_{j=1}^p\beta_jx_{ij}\). Thus the coefficient of \(\log(e_i)\) is exactly one.
This is the justification for using an offset rather than treating log exposure as an ordinary predictor. If its coefficient were estimated freely, the fitted mean would be proportional to \(e_i^\gamma\) for an estimated \(\gamma\), not necessarily to \(e_i\). That may be a different scientific model, but it does not impose the proportionality assumption described here.
Reading the Fitted Rate and Total
The linear predictor excluding the offset describes the log of the expected amount per exposure unit. Exponentiating it gives the rate on the original scale. A one-unit increase in a quantitative predictor \(x_j\), holding the other predictors fixed, multiplies this rate by \(\exp(\beta_j)\). It also multiplies the expected total by the same factor when exposure is unchanged.
| Quantity | Expression | Interpretation |
|---|---|---|
| Rate per exposure unit | \(\exp(\beta_0+\sum_j\beta_jx_{ij})\) | Expected amount for one exposure unit at the given predictors |
| Expected total amount | \(e_i\exp(\beta_0+\sum_j\beta_jx_{ij})\) | Expected amount over the observation’s actual exposure |
| Rate ratio for a one-unit increase in \(x_j\) | \(\exp(\beta_j)\) | Multiplicative change in rate, holding other predictors fixed |
In a model with a categorical indicator, its coefficient has the same multiplicative interpretation: changing the indicator from zero to one multiplies the fitted rate by the exponential of its coefficient. The offset itself is not a predictor effect to interpret in this way; it is the known exposure adjustment imposed by the model.
Worked Example: Fitting a Mean at a New Exposure
A hypothetical insurer models a strictly positive aggregate covered amount \(Y\), in dollars, for each policy over its observed policy-years. Suppose the fitted Gamma log-link model is \(\log(\mu)=\log(e)+\log(80)+\log(1.5)z+\log(0.8)w\), where \(e\) is exposure in policy-years and \(z,w\) are binary risk indicators. Take the Gamma dispersion to be \(\phi=0.25\). For a policy with \(e=2.5\), \(z=1\), and \(w=0\), calculate the fitted rate, expected total, and conditional standard deviation.
Method. Evaluate the predictor excluding the offset to obtain the rate, multiply that rate by exposure to obtain the total mean, and then apply the Gamma variance relationship.
Rate. With \(z=1\) and \(w=0\), \(r=\exp(\log(80)+\log(1.5)+\log(0.8)\cdot0)=80(1.5)=\$120\) per policy-year.
Expected total. The exposure offset contributes a factor \(e=2.5\), so \(\mu=2.5(\$120\text{ per policy-year})=\$300\).
Variance and standard deviation. The conditional variance is \(\operatorname{Var}(Y\mid e,z,w)=0.25(300)^2=0.25(90{,}000)=22{,}500\) dollars squared. Therefore the conditional standard deviation is \(\sqrt{22{,}500}=\$150\).
Check. The distribution is Gamma with mean \(\$300\), dispersion \(\phi=0.25\), shape \(1/\phi=4\), and scale \(\phi\mu=\$75\). Its mean is \(4(75)=\$300\), and its variance is \(4(75)^2=\$22{,}500\), agreeing with the GLM calculation. A common mistake is to report the rate, \(\$120\), as the expected total; the total must also reflect the 2.5 policy-years.
Exposure Units Must Be Consistent
Changing from months to years changes the numerical rate per unit. It does not change the expected total for a fixed period. In the log-mean equation, changing units also changes the numerical offset and requires a corresponding intercept shift. A reliable check is to compute the rate and exposure in one unit, then repeat the calculation in the other unit and confirm that the total agrees.
Worked Example: Converting a Monthly Rate to Years
A hypothetical model has a baseline expected amount of \(\$24\) per covered month. A group characteristic multiplies that rate by \(1.25\). A policy has 15 covered months. Calculate its expected total, then verify the result using years.
Method. Apply the predictor multiplier once to the baseline rate, and multiply the resulting group rate by the exposure expressed in the same unit.
Using months. The group rate is \(\$24\times1.25=\$30\) per month. Therefore the expected total is \(15\times\$30=\$450\).
Using years. The baseline annual rate is \(\$24\times12=\$288\) per year. Applying the group multiplier gives a group rate of \(\$288\times1.25=\$360\) per year. The exposure is \(15/12=1.25\) years, so the expected total is \(1.25\times\$360=\$450\).
Check. Both unit systems give an expected total of \(\$450\). The common trap is to treat \(\$360\) as a baseline rate and multiply it by \(1.25\) again as a predictor effect. It is already the group’s annual rate; multiplying it by 1.25 years gives the total.
Comparing Different Exposures and Predictors
When comparing fitted totals, separate the exposure ratio from the rate ratio. If two observations have exposures \(e_A,e_B\) and fitted rates \(r_A,r_B\), their expected-total ratio is $$ \frac{\mu_B}{\mu_A}=\frac{e_B}{e_A}\frac{r_B}{r_A}. $$ The rate ratio equals one only when the covariate contributions to the fitted rates are the same. Similar-looking groups need not have the same predictors, so the model must be evaluated separately for each group.
Worked Example: Comparing Two Groups Correctly
A hypothetical Gamma model for positive aggregate amounts has fitted rate \(r=100\exp(\log(1.4)z+\log(0.75)w)\) dollars per policy-year, with exposure offset \(\log(e)\). Group A has \(e_A=2.0\), \(z_A=0\), \(w_A=0\); Group B has \(e_B=3.2\), \(z_B=1\), \(w_B=0\). Find both expected totals, their ratio, and the combined expected total.
Group A rate and total. Its rate is \(r_A=100(1.4)^0(0.75)^0=\$100\) per policy-year. Its expected total is \(\mu_A=2.0(\$100)=\$200\).
Group B rate and total. Its rate is \(r_B=100(1.4)^1(0.75)^0=\$140\) per policy-year. Its expected total is \(\mu_B=3.2(\$140)=\$448\).
Ratio and combined expectation. The total ratio is \(448/200=2.24\). Independently, Group B has \(3.2/2.0=1.6\) times the exposure and \(140/100=1.4\) times the rate, giving \(1.6(1.4)=2.24\). By additivity of expectation, the combined expected total is \(\$200+\$448=\$648\); independence is not needed for this expectation.
Check. Group A and Group B have the same \(w\) value but different \(z\) values, so their fitted rates differ. Their expected-total ratio agrees with the product of exposure and rate ratios. The trap is to compare exposure alone and conclude that the ratio is 1.6, ignoring the different risk indicators.
Assumptions, Interpretation, and Common Traps
The offset model is appropriate when a positive response can reasonably be treated as a total whose conditional mean scales in direct proportion to exposure, after accounting for modeled predictors. It is an assumption about the mean structure. A Gamma distribution additionally represents positive continuous outcomes and supplies a particular variance relationship, \(\phi\mu_i^2\). The choice of that response distribution and its assumptions should be justified for the application; exposure alone does not establish that a Gamma model is suitable.
For a model using the usual Gamma GLM likelihood, observations are positive, the conditional mean follows the specified log-link equation, and conditional variance follows the Gamma mean-variance relationship. The dispersion is positive; a common dispersion parameter is used in the standard model. The offset is known rather than estimated. If exposure is measured with error, is zero, or does not have a proportional relationship with the expected amount, the stated formulation does not apply as written. In particular, \(\log(0)\) is undefined, and a zero response is outside the ordinary Gamma response support.
- Do not confuse rate and total. The exponentiated linear predictor excluding \(\log(e_i)\) is amount per exposure unit; multiply by \(e_i\) for the expected total.
- Do not estimate the offset coefficient in this model. It is fixed at one to impose direct proportionality of mean total to exposure.
- Do not mix units. A rate per month must be paired with exposure in months; a rate per year must be paired with exposure in years.
- Apply each predictor effect once. A rate computed after exponentiating the predictor effects already includes those effects.
- Do not infer a common rate from similar labels. Calculate each observation’s predictor contribution before comparing fitted rates.
This distinction is useful in actuarial work because a fitted mean per policy-year can be communicated as a rate, while a policy-level expected amount can be calculated for its actual coverage duration. The offset makes those two quantities consistent by construction. The treatment of a positive Gamma response and the interpretation of its exposure-adjusted mean are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.
Check Your Understanding
For each hypothetical model, calculate the requested rate, expected amount, or variability. Keep the exposure unit consistent with the stated rate.
- A positive-amount Gamma model has baseline rate \(\$160\) per exposure unit and a binary predictor multiplier of \(1.2\). For an observation with exposure \(0.75\) units and predictor equal to one, find the fitted rate and expected total. (Answer: \(\$192\) per unit; \(\$144\))
- A model has baseline rate \(\$18\) per month, a group multiplier of \(1.4\), and exposure of 10 months. Find the expected total and verify it using annual units. (Answer: \(\$252\); annual group rate \(\$302.40\) and exposure \(10/12\) years)
- A fitted model is \(\log(\mu)=\log(e)+\log(90)+0.2x\), with exposure in years. For \(e=3\) and \(x=2\), find the fitted rate and expected total, rounding amounts to the nearest cent. (Answer: rate \(\$134.26\) per year; expected total \(\$402.79\))
- A model’s rate is \(75(1.5)^z\) dollars per policy-year. Group A has exposure 2.4 and \(z=0\); Group B has exposure 3.0 and \(z=1\). Find both expected totals and the ratio of Group B’s total to Group A’s total. (Answer: \(\$180\) and \(\$337.50\); ratio \(1.875\))
- A Gamma model has exposure \(e=2\), baseline rate \(\$110\) per unit, predictor multiplier \(0.8\), and dispersion \(\phi=0.16\). Find the fitted mean, variance, and standard deviation. (Answer: mean \(\$176\); variance \(4{,}956.16\) dollars squared; standard deviation \(\$70.40\))