Why Build an Offset From a Baseline Rate?
A count model may need to reflect several known components before the data are used to estimate any remaining effects. For example, a group’s expected count may depend on its exposure, an externally supplied baseline rate, and a known adjustment for a particular rating class. If these components multiply, the natural way to include them in a Poisson GLM with a log link is as an offset: a known term in the linear predictor whose coefficient is fixed at one.
The offset establishes the reference rate structure. Estimated coefficients then describe how the modeled groups differ from that structure. This is useful when the baseline rate comes from a separate, specified source and the analysis is intended to estimate relative adjustments rather than rebuild the baseline from scratch. The resulting predictions are conditional on the supplied baseline and on the model assumptions.
When no additional known multiplier \(K_i\) is present, take \(K_i=1\). On the original scale, the model says $$ \mu_i=E_i r_{0i}K_i\exp\!\left(\beta_0+\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta}\right). $$ Thus, estimated terms multiply the known expected count \(E_i r_{0i}K_i\). A coefficient \(\beta_j\) corresponds to a multiplicative rate adjustment of \(\exp(\beta_j)\) for a one-unit increase in its predictor, holding other modeled predictors and the offset components fixed.
Constructing the Offset and Reading Its Meaning
The order of operations is conceptually simple: multiply the known components on the count scale, then take the logarithm to obtain the offset on the linear-predictor scale. Products become sums because \(\log(ab)=\log(a)+\log(b)\). The modeled terms are added to that sum, but they are not part of the offset unless their values are known and deliberately held fixed.
A supplied baseline rate may vary by group. For instance, if a benchmark table gives a different rate for each age band, each observation receives the rate corresponding to its age band. That table value remains fixed in the offset; a separate modeled age effect should be included only if the analysis is intended to estimate an additional age adjustment beyond the supplied baseline. The same reasoning applies to other known factors: the analyst must state what the baseline already accounts for, so the modeled terms represent only additional effects.
An intercept is often included. Its exponential, \(\exp(\beta_0)\), is a common estimated multiplier applied to all observations beyond their offsets. If the intercept is zero, its multiplier is one. Therefore, an externally supplied baseline is an anchor, not a guarantee that fitted predictions equal the baseline: the estimated intercept and other coefficients can shift predictions away from it.
Worked Example: Combining Exposure, a Baseline Rate, and Two Estimated Effects
A hypothetical portfolio segment has exposure of \(2.4\) policy-years. An external benchmark specifies a baseline claim rate of \(0.05\) claims per policy-year. For this example, the fitted intercept is zero, and two estimated effects have rate multipliers \(1.5\) and \(0.8\). Calculate the offset, the fitted mean count, and the total modeled rate multiplier relative to the external baseline.
Method. Multiply the known exposure and rate to obtain the offset’s count-scale basis. Multiply the modeled effects together to obtain the adjustment to that basis.
The known expected count before modeled adjustments is $$ E_i r_{0i}=2.4(0.05)=0.12. $$ Therefore, the offset is $$ o_i=\log(0.12)\approx -2.1203. $$ The product of the two fitted effect multipliers is $$ 1.5(0.8)=1.2. $$ The fitted mean count is consequently $$ \mu_i=0.12(1.2)=0.144. $$ Equivalently, the modeled terms sum on the log scale to $$ \log(1.5)+\log(0.8)=\log(1.2)\approx 0.1823. $$ Using full precision before rounding, the linear predictor is \(\log(0.12)+\log(1.2)=\log(0.144)\), so its exponential is \(0.144\), as calculated directly. The estimated effects increase the baseline rate by a factor of \(1.2\), or 20%. The rate after those effects is \(0.05(1.2)=0.06\) claims per policy-year.
Check. Multiplying the adjusted rate by exposure gives \(0.06(2.4)=0.144\), matching the fitted mean. The calculation avoids a common mistake: rounding \(\log(1.5)\) and \(\log(0.8)\) separately and then using their rounded difference as the final result. Combine the exact product first and round \(\log(1.2)\) only at the end.
Known Multipliers and Estimated Effects
A known multiplicative factor belongs in the offset only when its value is supplied and treated as fixed for the analysis. A factor whose effect is to be estimated belongs among the predictors. This distinction is about the role of the quantity in the model, not its subject matter: a geographic classification might have a known external adjustment in one analysis and an estimated coefficient in another.
| Component | How it enters | What it means |
|---|---|---|
| Exposure \(E_i\) | \(\log(E_i)\) in the offset | Scales the expected count with the amount of exposure |
| External baseline rate \(r_{0i}\) | \(\log(r_{0i})\) in the offset | Supplies the reference rate per exposure unit |
| Known multiplier \(K_i\) | \(\log(K_i)\) in the offset | Applies a specified adjustment fixed for the analysis |
| Estimated predictor effect | Predictor multiplied by an estimated coefficient | Estimates an additional association with the rate |
Worked Example: A Group-Specific Baseline and a Known Adjustment
A hypothetical risk group has \(40\) policy-years of exposure. Its external baseline is \(0.025\) claims per policy-year, and a supplied classification table gives a known multiplier of \(1.2\). The fitted intercept is zero. Two estimated predictor effects have multipliers \(1.4\) and \(0.9\). Find the offset, predicted mean count, and overall multiplier relative to the group’s external baseline.
Step 1: Build the known count-scale basis. The exposure and baseline rate give \(40(0.025)=1.00\) expected claims before the known classification adjustment. Including that adjustment gives $$ E_i r_{0i}K_i=40(0.025)(1.2)=1.20. $$
Step 2: Take the logarithm for the offset. $$ o_i=\log(1.20)\approx 0.1823. $$
Step 3: Combine the estimated effects. Their multiplier is \(1.4(0.9)=1.26\). Thus the fitted mean is $$ \mu_i=1.20(1.26)=1.512. $$ The estimated effects alone multiply the known adjusted count by \(1.26\). Relative to the external baseline before the known classification adjustment, the overall multiplier is \(1.2(1.26)=1.512\); the fitted rate is \(1.512/40=0.0378\) claims per policy-year.
Check. The baseline rate multiplied by all factors gives \(0.025(1.2)(1.4)(0.9)=0.0378\), and \(40(0.0378)=1.512\). The result agrees on both the rate and count scales. The known classification adjustment is part of the offset; it must not also be counted as an estimated effect in this calculation.
Why the Log Offset Works
The offset form follows directly from the multiplicative mean model. Suppose the expected count is the product of exposure, a baseline rate, a known multiplier, and an estimated positive rate adjustment \(M_i\): $$ \mu_i=E_i r_{0i}K_i M_i. $$ Taking logarithms of both sides gives $$ \log(\mu_i)=\log(E_i r_{0i}K_i)+\log(M_i). $$ If the estimated adjustment has log-linear form \(M_i=\exp(\beta_0+\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta})\), then \(\log(M_i)=\beta_0+\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta}\). Substitution yields the stated GLM equation. This is why the known product appears as an additive offset and why its coefficient is one: it represents the specified multiplicative contribution without estimating a new coefficient for it.
Under the Poisson assumption, the conditional likelihood for an observed count \(n_i\) is $$ P(N_i=n_i)=\frac{\exp(-\mu_i)\mu_i^{n_i}}{n_i!}. $$ The offset determines part of \(\mu_i\), while the regression coefficients determine the remaining modeled adjustment. Estimation varies the coefficients, not the offset. This interpretation relies on the stated mean structure and the assumption that the supplied components are fixed. If the baseline is not genuinely known, treating it as fixed can give an overly certain and potentially misspecified analysis; the model should reflect how that baseline is obtained.
Worked Example: Predicting Two Groups With Different Offsets
A hypothetical model uses a baseline rate of \(0.04\) claims per policy-year. Both groups share a fitted predictor effect with multiplier \(\exp(0.2)\); the second group also has a modeled category effect with multiplier \(1.5\). The fitted intercept is zero. Group A has \(10\) policy-years of exposure and a known multiplier of \(1.25\); Group B has \(25\) policy-years and a known multiplier of \(0.8\). Calculate each offset and mean count, then interpret the category effect.
Group A. Its known count-scale basis is \(10(0.04)(1.25)=0.50\), so its offset is \(\log(0.50)\approx -0.6931\). Its modeled multiplier is \(\exp(0.2)\approx 1.2214\). Therefore, $$ \mu_A=0.50\exp(0.2)\approx 0.50(1.2214)=0.6107. $$
Group B. Its known count-scale basis is \(25(0.04)(0.8)=0.80\), so its offset is \(\log(0.80)\approx -0.2231\). Its modeled multiplier is \(\exp(0.2)(1.5)\approx 1.2214(1.5)=1.8321\). Thus, $$ \mu_B=0.80\exp(0.2)(1.5)\approx 0.80(1.8321)=1.4657. $$
Interpretation and check. Holding the other modeled effect and the offset components fixed, the category effect multiplies the rate by \(1.5\), a 50% increase. It does not imply that Group B’s expected count is 50% greater than Group A’s: the groups have different exposures and known multipliers. Directly, the Group B rate is \(1.4657/25\approx 0.0586\); Group A’s is \(0.6107/10\approx 0.0611\). The group-level rates differ because their known multipliers differ, even though the modeled category effect increases Group B’s rate relative to its own offset baseline.
Common Specification Traps
Offsets are easy to write down but easy to misinterpret. The most important check is to reconstruct the intended mean on the original scale: multiply exposure, baseline, known factors, and estimated multipliers exactly once. This makes omissions and duplications visible.
- Leaving out the baseline rate. In the first worked example, using exposure alone as the offset basis and then applying only the two estimated effects gives \(2.4(1.5)(0.8)=2.88\) expected claims, rather than \(0.144\). That calculation has omitted the rate of \(0.05\) claims per policy-year.
- Applying a known component twice. If the offset already contains \(E_i r_{0i}K_i\), multiplying a prediction by \(r_{0i}\) again duplicates the baseline. In the first example, doing so would give \(0.144(0.05)=0.0072\), not \(0.144\).
- Confusing a known adjustment with an estimated effect. A fixed external multiplier belongs in the offset if the model treats it as known. Adding the same factor again among modeled effects changes the intended mean.
- Interpreting a coefficient as an absolute rate. A coefficient’s exponential is a relative multiplier, conditional on the included modeled terms. It is not itself the claim rate; the offset and other terms also determine the predicted rate and count.
Check Your Understanding
For each problem, assume a Poisson log-link model and a fitted intercept of zero unless stated otherwise. Use the supplied external rates and known factors as fixed components.
- A group has \(12\) exposure units, a baseline rate of \(0.03\) per unit, a known multiplier of \(0.5\), and an estimated effect multiplier of \(1.25\). Find the offset and fitted mean count. (Answer: offset \(\log(0.18)\approx -1.7148\); mean \(0.225\).)
- A segment has \(50\) policy-years, an external baseline rate of \(0.04\) per policy-year, and a known factor of \(1.1\). Its estimated log-scale effects are \(0.2\) and \(-0.1\). Find the offset and fitted mean count. (Answer: offset \(\log(2.2)\approx 0.7885\); mean \(2.2\exp(0.1)\approx 2.4314\).)
- Two groups share a modeled multiplier of \(1.2\). Group A has exposure \(20\), baseline rate \(0.02\), and known factor \(1.5\); Group B has exposure \(30\), the same baseline rate, and known factor \(0.8\). Find both fitted means. (Answer: Group A \(20(0.02)(1.5)(1.2)=0.72\); Group B \(30(0.02)(0.8)(1.2)=0.576\).)
- A model’s offset is \(\log(E_i r_{0i}K_i)\), and a predictor has estimated coefficient \(\log(0.7)\). Interpret a one-unit increase in that predictor, holding the other terms fixed. (Answer: it multiplies the modeled rate and mean count by \(0.7\), a 30% decrease; it does not replace or modify the offset.)
- A group has exposure \(8\), baseline rate \(0.05\) per unit, known factor \(1.25\), and estimated effect multipliers \(0.8\) and \(1.5\). Find the fitted mean and identify the result of mistakenly applying the baseline rate twice. (Answer: correct mean \(8(0.05)(1.25)(0.8)(1.5)=0.60\); applying the baseline twice gives \(0.03\).)