Auditing Offsets and Denominators in Rate Comparisons

Why a Rate Comparison Needs an Audit

A group with more claims does not necessarily have a higher claim rate: it may simply have more policies or more time at risk. A meaningful comparison starts by checking that each count has the correct denominator, then uses that denominator explicitly in the model. This tutorial focuses on that audit for Poisson count models, aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Let $Y_i$ be a count observed for unit $i$, and let $E_i>0$ be its exposure, measured in the denominator units appropriate to the question. Examples include policy-years, vehicle-years, or insured months. A rate is a count per exposure unit; it is not the count itself. In a hypothetical insurance example, claim counts should be related to time insured, rather than compared without regard to how long each policy was observed.

Definition: In a Poisson rate model with a log link, the conditional count $Y_i$, given its covariates and exposure, has mean $\mu_i$ and is modeled as $$ Y_i\mid \mathbf{x}_i,E_i\sim \operatorname{Poisson}(\mu_i), \qquad \log(\mu_i)=\log(E_i)+\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}. $$ The term $\log(E_i)$ is an offset: its coefficient is fixed at 1. Equivalently, the modeled rate is $$ \lambda_i=\frac{\mu_i}{E_i} =\exp(\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}). $$

This formulation assumes that, conditional on the modeled covariates, the expected count is proportional to exposure. The count is the Poisson response; the exposure enters through the offset. The Poisson model also assumes conditional independence across observational units and equality of the conditional mean and variance. If the exposure does not represent the time or population actually at risk, an offset cannot repair that denominator error. If expected counts are not proportional to exposure, fixing the exposure coefficient at 1 may also be inappropriate.

Why the Offset Has Coefficient One

The offset is not an extra risk factor whose effect is estimated. It enforces the meaning of exposure. If the modeled rate for a unit is $\lambda_i$, then the expected count accumulated over $E_i$ exposure units is $E_i\lambda_i$. Taking logarithms gives the model form directly:

\mu_i=E_i\lambda_i \quad\Longrightarrow\quad \log(\mu_i)=\log(E_i)+\log(\lambda_i).

If the log rate is modeled as $\beta_0+\sum_{j=1}^p\beta_jx_{ij}$, the last term is the covariate part of the model. The coefficient of $\log(E_i)$ must therefore be 1 to preserve $\mu_i=E_i\lambda_i$. For example, doubling exposure while keeping the risk profile fixed doubles the expected count, but leaves the expected rate unchanged.

Under a model with one rate per group and no additional covariates, the Poisson likelihood also shows why the exposure denominator appears. For group $g$, suppose its counts are independent Poisson variables with means $E_i\lambda_g$. The log-likelihood terms involving $\lambda_g$ are

\ell(\lambda_g) =\sum_{i\in g}\left\{Y_i\log(E_i\lambda_g)-E_i\lambda_g-\log(Y_i!)\right\} =\text{constant}+\left(\sum_{i\in g}Y_i\right)\log(\lambda_g) -\lambda_g\sum_{i\in g}E_i.

Differentiating with respect to $\lambda_g$ and setting the result to zero gives $\sum_{i\in g}Y_i/\lambda_g-\sum_{i\in g}E_i=0$. Thus the maximum-likelihood estimate is total count divided by total exposure. This is the rate, not the group’s total count.

Worked Example: More Counts but a Lower Rate

A hypothetical insurer compares two underwriting groups. Group A has 45 claims over 1,500 policy-years; Group B has 32 claims over 800 policy-years. With no other covariates in the comparison, calculate each estimated rate, the Group B to Group A rate ratio, and the raw-count ratio. State which comparison answers the rate question.

Method. Estimate each group’s rate as its total count divided by its total policy-years. The common mistake is to use the ratio of counts as though it were a rate ratio.

For Group A, $\widehat{\lambda}_A=45/1500=0.03$ claims per policy-year, or 30 claims per 1,000 policy-years.

For Group B, $\widehat{\lambda}_B=32/800=0.04$ claims per policy-year, or 40 claims per 1,000 policy-years.

The rate ratio is $\widehat{\lambda}_B/\widehat{\lambda}_A=0.04/0.03=4/3\approx1.3333$. Thus Group B’s estimated rate is about 1.3333 times Group A’s rate. In a two-group Poisson log-rate model, the Group B coefficient relative to Group A is $\log(4/3)\approx0.2877$.

The raw-count ratio is $32/45\approx0.7111$, which suggests fewer claims in Group B only because it ignores the smaller exposure. As a check, if Group B had Group A’s 1,500 policy-years at its estimated rate, its expected count would be $1,500(0.04)=60$, which exceeds 45. The rate comparison correctly identifies Group B’s higher estimated rate. Rates and the coefficient are rounded to four decimal places where shown.

Reading and Auditing the Offset

The offset is $\log(E_i)$, not $E_i$, and it enters with a plus sign in the log-mean equation. A quick audit should confirm the following before interpreting any fitted coefficient:

Response: Is the modeled response a nonnegative count, such as claims, events, or incidents?
Denominator: Does $E_i$ measure the population or time that could generate that count?
Scale: Are all exposures expressed in the same unit?
Model form: Is $\log(E_i)$ included as an offset with its coefficient fixed at 1?
Comparison: Is the reported contrast a rate ratio at the same values of other modeled covariates?

For a binary group indicator $G_i$, a model such as $\log(\mu_i)=\log(E_i)+\beta_0+\beta_1G_i$ implies that the rate ratio for $G_i=1$ versus $G_i=0$ is $\exp(\beta_1)$. Exposure cancels in this ratio because both rates are defined per the same exposure unit. With additional covariates, that interpretation holds while those covariates are held fixed.

Worked Example: Changing the Exposure Unit

Two hypothetical portfolios have 18 claims over 900 policy-years and 24 claims over 600 policy-years, respectively. Re-express exposure in thousands of policy-years. Calculate both rates using each exposure scale and verify the rate ratio is unchanged.

Method. Convert the denominator and recalculate count per denominator unit. The trap is to divide the counts by exposure in thousands but still label the result as a rate per policy-year.

On the policy-year scale, the rates are $18/900=0.02$ and $24/600=0.04$ claims per policy-year. The second-to-first rate ratio is $0.04/0.02=2$.

On the thousand-policy-year scale, the exposures are $900/1000=0.9$ and $600/1000=0.6$. The rates are $18/0.9=20$ and $24/0.6=40$ claims per 1,000 policy-years. Their ratio is $40/20=2$, as before.

The numerical rate changes by a factor of 1,000 because its unit changes; the rate ratio does not change. In the model, if $E_i^*=E_i/1000$, then the rate per thousand exposure units is $1000\lambda_i$. Its log is $\log(1000)+\log(\lambda_i)$, so changing the exposure unit shifts the intercept by $\log(1000)$, while leaving covariate coefficients and rate ratios unchanged. The calculations check exactly.

Adjusted Rate Ratios and Model Predictions

A model can compare groups while accounting for covariates that differ between them. In a log-link model, exponentiating a coefficient gives a multiplicative change in the modeled rate for a one-unit increase in that predictor, with other predictors fixed. For a binary indicator, exponentiating its coefficient gives the rate ratio for the indicated group against the reference group. The exposure affects the expected count, but it is not itself part of this adjusted rate ratio.

Worked Example: An Adjusted Rate Ratio and Expected Counts

Suppose a hypothetical Poisson model for event counts has exposure $E$ measured in policy-years, a treatment indicator $T$ equal to 0 or 1, and a risk indicator $z$ equal to 0 or 1: $\log(\mu)=\log(E)+\log(0.04)+\log(1.5)T+\log(2)z$. Calculate expected counts for $E=300$ at $z=1$, first for $T=0$ and then for $T=1$. Interpret the treatment rate ratio.

Method. Exponentiate the non-offset terms to obtain the rate, then multiply by exposure to obtain the expected count. Do not interpret the treatment coefficient as an additive increase in claims.

For $T=0$ and $z=1$, the modeled rate is $0.04(1.5)^0(2)^1=0.08$ events per policy-year. The expected count is $\mu=300(0.08)=24$.

For $T=1$ and $z=1$, the rate is $0.04(1.5)^1(2)^1=0.12$ events per policy-year. The expected count is $\mu=300(0.12)=36$.

At the same exposure and risk-indicator value, the treatment rate ratio is $0.12/0.08=1.5$. Equivalently, it is $\exp(\log(1.5))=1.5$. The expected counts also have ratio $36/24=1.5$ here because exposure is the same. If the exposure were different, the count ratio would also reflect that exposure difference. The computed rates and counts are consistent with the model.

Common Modeling Errors

Two tempting alternatives change the question being answered. First, comparing counts without exposure compares total volume, not rates. Second, putting exposure into the model as an ordinary predictor estimates its coefficient rather than enforcing the proportionality that defines an exposure-adjusted count model.

Common mistake: Suppose the fitted equation is written as $\log(\mu_i)=\alpha+\gamma\log(E_i)+\boldsymbol{\beta}^{\mathsf T}\mathbf{x}_i$. Treating $\log(E_i)$ as an ordinary predictor allows the data to estimate $\gamma$. Unless $\gamma=1$, doubling exposure multiplies the expected count by $2^\gamma$, not by 2, holding other predictors fixed. For an exposure-adjusted rate model, the intended specification is $\log(\mu_i)=\log(E_i)+\alpha+\boldsymbol{\beta}^{\mathsf T}\mathbf{x}_i$, with the offset coefficient fixed at 1.

There are settings in which a more general relationship between exposure and expected count is scientifically appropriate, but that is a different model and needs justification. It should not arise accidentally because exposure was entered as an ordinary predictor. Similarly, using a rate as the response while ignoring its denominator can give observations with very different exposure amounts inappropriate influence. The count-plus-offset formulation retains the count response and explicitly accounts for exposure.

A denominator audit is especially important when comparing groups with different observation windows, eligibility rules, or amounts of time at risk. A numerator may count events only among eligible policies, while the denominator accidentally includes time before coverage began. The resulting model may be computed correctly but answer the wrong question. Always match the numerator and denominator to the same population and time period.

Define the event count.
Specify exactly what event is counted and the unit represented by each observation.

Verify exposure.
Confirm the denominator measures the corresponding time or population at risk and uses a consistent unit.

Check the model equation.
For the standard Poisson rate model, include $\log(E_i)$ as an offset, not as an ordinary estimated predictor.

Interpret the contrast.
Exponentiate a log-rate coefficient for a rate ratio, and distinguish that ratio from a ratio of total counts.

Key takeaway: An exposure-adjusted Poisson model describes counts through $\mu_i=E_i\lambda_i$. The offset $\log(E_i)$ enforces proportionality between expected count and exposure; exponentiated covariate contrasts describe rate ratios, not raw-count differences.

Check Your Understanding

For each case, identify the relevant rate or model quantity before comparing counts.

A hypothetical group has 21 events over 700 member-years. What is its estimated rate per 1,000 member-years? (Answer: 30 events per 1,000 member-years.)
Portfolio A has 15 claims over 500 policy-years; Portfolio B has 18 claims over 450 policy-years. What is the B-to-A rate ratio, and which portfolio has the higher estimated rate? (Answer: $0.04/0.03=1.3333$; Portfolio B has the higher estimated rate.)
A model is $\log(\mu)=\log(E)-2+0.4x$. For $x=0$, calculate the expected count at exposure $E=100$. (Answer: $100e^{-2}\approx13.5335$ events.)
In a Poisson log-rate model, a binary predictor has coefficient $\log(0.8)$. What is the rate ratio for predictor value 1 versus 0, holding other covariates fixed? (Answer: $0.8$, or a 20% lower modeled rate.)
Exposures are recorded in policy-months, but a model’s intercept and rates are to be reported per policy-year. What conversion should be made to the exposure, and does that conversion change covariate rate ratios? (Answer: Convert policy-months to policy-years by dividing by 12; no, the covariate rate ratios are unchanged.)

What You'll Learn