Why Multiple Exposure Components Need One Offset
A claim count may depend on more than one predetermined measure of how much insurance or service was provided. For example, a fleet’s exposure may depend on both the number of covered vehicles and the fraction of a year each was insured. A model should account for both components while keeping them distinct from rating characteristics whose effects are estimated from data.
With a log link, a multiplicative exposure adjustment becomes an additive term in the linear predictor. Combining components this way gives the model the required exposure scale without estimating a separate coefficient for each known component. This tutorial is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
The Model and Its Conditions
Let \(N_i\) be the claim count for observation \(i\), and let \(\mu_i=E[N_i]\) be its conditional expected count. A log-link mean model with multiple exposure components is
Here \(\boldsymbol{x}_i\) contains modeled rating variables, and \(\boldsymbol{\beta}\) contains their coefficients. The product \(E_i\) must have a meaningful exposure interpretation in the units for which the baseline rate is defined. For instance, vehicle count multiplied by insured years gives vehicle-years. A known mileage adjustment can also be included if it has been defined as a multiplicative index relative to standard mileage and the model assumes proportionality to that index.
The product form is appropriate only when the assumed mean is proportional to each component, holding the other components and rating variables fixed. Components must be positive because their logarithms are used. A zero-exposure observation has zero expected count under this exposure model and should not be passed to a log-offset calculation; data handling for zero exposure must follow the model’s intended treatment. If a component’s effect is not known to be proportional, or needs to be estimated, it should not automatically be given an offset coefficient of 1.
Why the Components Add on the Log Scale
The justification is the logarithm’s product rule. Begin with the assumption that expected count equals combined exposure multiplied by a modeled rate:
Taking the logarithm of both sides gives
Thus the sum of the log components is not an additional estimated effect: its value is fully determined by known exposure data. Exponentiating the linear predictor recovers the original product. This also explains why the offset coefficient is fixed at 1: changing exposure by a factor \(c\), with other terms held constant, changes the modeled mean by that same factor \(c\).
Confirm that the components are positive, known in advance, and intended to multiply into exposure on the model’s scale.
Compute \(E_i=\prod_j e_{ij}\), keeping track of units and any dimensionless adjustment indices.
Use \(\log(E_i)=\sum_j\log(e_{ij})\), then add the rating-variable contribution.
Calculate \(\mu_i=E_i\exp(\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta})\) and verify that the result equals exposure times modeled rate.
Worked Calculations
Worked Example: Vehicle Count and Insured Time
A hypothetical fleet account has 8 covered vehicles, each represented by the same 0.75-year exposure period. Its modeled baseline frequency is 0.12 claims per vehicle-year, and a known rating classification has fitted relativity 1.5. Calculate the fitted expected claim count. Assume the mean is proportional to vehicle-years and that the classification relativity multiplies the baseline frequency.
Method. Multiply the known exposure components, then multiply by the fitted rate. The exposure is \(E=8(0.75)=6\) vehicle-years. The adjusted frequency is \(0.12(1.5)=0.18\) claims per vehicle-year.
The expected count is therefore
On the linear-predictor scale, the offset is \(\log(8)+\log(0.75)=\log(6)\). The rate contribution is \(\log(0.12)+\log(1.5)=\log(0.18)\), so the complete predictor is \(\log(6)+\log(0.18)=\log(1.08)\). Exponentiating gives \(1.08\), agreeing with the direct calculation. The common mistake avoided here is multiplying by the vehicle count or time a second time after already using their product as the offset.
Worked Example: Adding a Known Utilization Index
A hypothetical service fleet has 12 covered units insured for 0.5 years. A predetermined usage index is 1.2 relative to standard usage. Assume expected counts are proportional to all three exposure components. The baseline frequency is 0.08 claims per standard unit-year. Two fitted rating relativities are 1.25 and 0.9. Find the fitted expected count.
Method. First combine exposure: \(E=12(0.5)(1.2)=7.2\) standard-equivalent unit-years. Next calculate the modeled rate: \(0.08(1.25)(0.9)=0.09\) claims per standard-equivalent unit-year. Thus
The offset can also be written as \(\log(12)+\log(0.5)+\log(1.2)=\log(7.2)\). The full predictor adds \(\log(0.08)+\log(1.25)+\log(0.9)\), giving \(\log(7.2)+\log(0.09)=\log(0.648)\). Exponentiating again gives 0.648. The usage index belongs in the offset here because the problem states that it is known and exposure-proportional; an estimated usage effect would instead need a modeled coefficient.
Worked Example: Summing Means Across Different Rating Groups
A hypothetical portfolio has two groups. Group A has 4 insured units for 0.5 years and a fitted rating relativity of 1.5. Group B has 6 insured units for 0.25 years and a fitted relativity of 0.8. Both groups have baseline frequency 0.10 claims per unit-year. Find the portfolio’s fitted expected count and explain whether one can use total exposure with either group’s relativity.
Method. Calculate each group’s exposure and fitted mean separately because their rating relativities differ. For A, exposure is \(4(0.5)=2\) unit-years, and the fitted frequency is \(0.10(1.5)=0.15\). Therefore \(\mu_A=2(0.15)=0.30\).
For B, exposure is \(6(0.25)=1.5\) unit-years, and the fitted frequency is \(0.10(0.8)=0.08\). Therefore \(\mu_B=1.5(0.08)=0.12\). Add the expected counts:
Total exposure is \(2+1.5=3.5\) unit-years. The exposure-weighted effective relativity is \([2(1.5)+1.5(0.8)]/3.5=(3+1.2)/3.5=1.2\). Checking, \(3.5(0.10)(1.2)=0.42\). It would be wrong to multiply all 3.5 unit-years by the baseline rate and either 1.5 or 0.8: neither relativity applies to the entire portfolio. The weighted calculation works because it reproduces the sum of the group means.
Interpretation and Common Traps
The offset accounts for known exposure scale; rating predictors account for modeled differences in rate. Those roles should remain separate. In particular, multiplying exposure components does not mean that each component gets an estimated coefficient. Fixing the coefficient at 1 encodes the assumption of direct proportionality. If the data or business definition calls for a different relationship, that assumption must be reconsidered rather than silently changed in the offset.
- Do not add raw exposure values. If the intended exposure is units multiplied by insured time, adding those values does not produce the same scale. Add their logarithms, which is equivalent to taking the log of their product.
- Do not count an adjustment twice. Once a component is included in \(E_i\), do not also multiply the fitted mean by that same component outside the model.
- Check units and scaling. An exposure product must align with the baseline rate. If time is measured in months rather than years, convert consistently or adjust the rate’s unit.
- Do not collapse unlike groups carelessly. When rating effects differ, calculate group means and sum them, or use an equivalent exposure-weighted rate. Total exposure alone does not determine the portfolio mean.
- Do not put an unknown effect in an offset. An offset is known and fixed. A factor that needs its effect estimated belongs among the model predictors, subject to the selected model specification.
A useful final check is to express the answer in two equivalent ways: combined exposure multiplied by fitted rate, and the exponential of the full linear predictor. A disagreement often signals a missing component, a duplicated adjustment, or a unit mismatch.
Check Your Understanding
For each problem, form the exposure carefully before calculating the fitted mean.
- A hypothetical account has 9 covered units insured for 0.4 years. The baseline frequency is 0.11 claims per unit-year, and its fitted classification relativity is 1.25. Find the fitted expected count. (Answer: 0.495 claims)
- A hypothetical operation has 5 locations, 0.8 insured years, and a known exposure index of 1.1. The baseline frequency is 0.06 per standard location-year, and fitted relativities are 1.4 and 0.75. Find the fitted expected count. (Answer: 0.2772 claims)
- A segment has 7 covered units for 0.5 years. Its log baseline rate is \(\log(0.10)\), and its fitted rating relativity is 2.0. State the combined log offset and calculate the fitted expected count. (Answer: offset \(\log(3.5)\); expected count \(0.70\))
- Two groups share a baseline frequency of 0.20. Group A has exposure 2 unit-years and relativity 1.1; Group B has exposure 3 unit-years and relativity 0.7. Calculate the total fitted expected count. (Answer: 0.86 claims)
- A calculation uses 4 insured units and 0.5 years, a baseline frequency of 0.20, and a fitted relativity of 1.5. A student computes exposure as \(4+0.5\), then multiplies by the adjusted frequency. Give the correct exposure and expected count, and identify the error. (Answer: exposure \(2\) unit-years; expected count \(0.60\) claims; the components should be multiplied, not added.)