Why Rating-Cell Aggregation Can Work
A frequency model is often developed from policy-level observations, but a rating dataset may contain many policies with identical predictor values. Summarizing those policies by rating cell can make estimation more compact. The key is to preserve the information the model uses: the number of claims, the total exposure, and the predictor values associated with that exposure.
Aggregation is not automatically harmless. If policies in a reported cell have different predictor values, a single cell-level predictor may not represent their separate expected frequencies. The conditions for preserving the likelihood therefore matter as much as the arithmetic.
This formulation assumes that exposures are known and treated as fixed, policy-level counts are conditionally independent given the predictors, and all policies within a cell have the same modeled rate. The model does not require the policies to have equal exposures. It does require that their predictor vectors, and hence their modeled rates, agree within the cell.
The Grouped Likelihood
For a cell with total count \(N_g\) and positive total exposure \(E_g\), the Poisson log-likelihood contribution, apart from terms that do not depend on \(\beta\), is
The first term records the observed number of claims and the cell’s predictors. The second term is the modeled expected count, since \(E[N_g]=E_g\exp(x_g^\mathsf{T}\beta)\). Across cells, add these contributions. The exposure enters as an offset \(\log E_g\) in the model for the expected count; it is not an additional rating factor to estimate.
Why the Aggregated Likelihood Is Equivalent
The equivalence follows by multiplying the individual Poisson probabilities. For policies \(j\) in cell \(g\), write \(\lambda_g=\exp(x_g^\mathsf{T}\beta)\). The part of the likelihood involving the model parameters is
The first two factors depend on the observations only through \(E_g\) and \(N_g\). The final factor depends on known exposures and observed counts, but not on \(\beta\). The grouped Poisson likelihood has the same parameter-dependent factors: \(\exp(-E_g\lambda_g)(E_g\lambda_g)^{N_g}\). Its additional factor \(E_g^{N_g}\), as well as its factorial term, is also independent of \(\beta\). Therefore both likelihoods give the same likelihood maximizer for \(\beta\).
This reasoning also shows what aggregation does not claim: the full probabilities assigned to a particular allocation of claims across policies need not equal the probability assigned to the total count. The allocation contains information about the known exposure pattern, but, under the stated model, it contains no additional information about \(\beta\).
Confirm that every observation in a cell has the same modeled predictors and that the count model is conditionally independent Poisson.
Sum claim counts to obtain \(N_g\), and sum the corresponding exposures to obtain \(E_g\).
Use \(N_g\) as the response and \(\log E_g\) as the offset, with the cell predictors in the linear predictor.
If predictors vary within a cell, or the count distribution is not the assumed Poisson model, do not presume this aggregation preserves the original likelihood.
Worked Examples
Worked Example: Estimating Two Cell Rates
A hypothetical insurer groups policies into two rating cells. Cell A has 14 claims over 700 policy-years; cell B has 15 claims over 500 policy-years. Assume independent Poisson counts, a constant frequency within each cell, and a separate rate for each cell. Estimate both rates and the rate ratio for B relative to A.
Method. This is a saturated two-cell Poisson frequency model: each cell has its own rate. For a cell with count \(N\) and exposure \(E\), its log-likelihood, omitting terms independent of the rate \(\lambda\), is \(N\log\lambda-E\lambda\). Differentiating gives \(N/\lambda-E=0\), so the maximum-likelihood estimate is \(\widehat{\lambda}=N/E\).
For cell A, \(\widehat{\lambda}_A=14/700=0.020\) claims per policy-year. Its fitted expected count is \(700(0.020)=14\), equal to the observed count.
For cell B, \(\widehat{\lambda}_B=15/500=0.030\) claims per policy-year. Its fitted expected count is \(500(0.030)=15\), also equal to the observed count.
The estimated rate ratio is \(0.030/0.020=1.5\). Thus cell B’s fitted frequency is 1.5 times cell A’s, or 50% higher. Check: multiplying each estimated rate by its own exposure recovers the respective observed total. The common mistake avoided here is dividing the total claims by the number of cells, or comparing counts without accounting for different exposures.
Worked Example: Verifying Likelihood Equivalence With Unequal Exposures
Three hypothetical policies share the same rating-cell predictors. Their exposures are 100, 250, and 150 policy-years, and their claim counts are 1, 4, and 3. Assume independent Poisson counts with a common cell rate \(\lambda\). Find the maximum-likelihood rate and verify that aggregation changes the log-likelihood only by a parameter-independent amount.
First sum the counts and exposures: \(N=1+4+3=8\), and \(E=100+250+150=500\) policy-years. The grouped count is Poisson with mean \(500\lambda\), so its log-likelihood is \(8\log(500\lambda)-500\lambda-\log(8!)\). Differentiating with respect to \(\lambda\) gives \(8/\lambda-500=0\), hence \(\widehat{\lambda}=8/500=0.016\) claims per policy-year.
For the individual data, the log-likelihood is
At \(\widehat{\lambda}=0.016\), the terms are \(-500(0.016)=-8\), \(8\log(0.016)\approx-33.0813\), and \(\log(100)+4\log(250)+3\log(150)\approx41.7229\). The factorial contribution is \(-\log(144)\approx-4.9698\). Therefore \(\ell_{\mathrm{ind}}(\widehat{\lambda})\approx-4.3282\).
For the grouped observation, \(\ell_{\mathrm{grp}}=-8+8\log(8)-\log(8!)\approx-1.9691\). The difference between these log-likelihood values is approximately \(-2.3592\), a constant determined by the observed counts and known exposures. It does not depend on \(\lambda\), so both likelihoods have the same maximizer. Check: the grouped mean at the estimate is \(500(0.016)=8\), the observed total. The common mistake avoided is assuming unequal policy exposures prevent aggregation; they do not, provided the modeled rate is the same within the cell.
Worked Example: Estimating a Log-Linear Trend From Cell Totals
A hypothetical portfolio has three cells indexed by a numerical rating variable \(x\). The cells have \((x,E,N)\) values \((0,100,2)\), \((1,200,8)\), and \((2,100,8)\), where exposure is in policy-years. Assume independent Poisson cell counts with mean \(E_g\exp(\alpha+\beta x_g)\). Estimate \(\alpha\) and \(\beta\).
The score equations for this model are the sums of predictor values multiplied by observed-minus-fitted counts. For an intercept and slope they are
The observed total count is \(2+8+8=18\). The observed count weighted by \(x\) is \(0(2)+1(8)+2(8)=24\). Consider fitted rates \(0.02,\ 0.04,\ 0.08\), which double from one \(x\)-level to the next. The corresponding fitted means are \(100(0.02)=2\), \(200(0.04)=8\), and \(100(0.08)=8\).
The fitted total is \(2+8+8=18\), matching the observed total. The fitted count weighted by \(x\) is \(0(2)+1(8)+2(8)=24\), also matching. Thus both score equations are satisfied. The model is identifiable because the cells have varying \(x\), and the Poisson log-likelihood is concave in \((\alpha,\beta)\); these score solutions give the maximum.
Since the rate at \(x=0\) is \(\exp(\alpha)=0.02\), \(\widehat{\alpha}=\log(0.02)\approx-3.9120\). The rate doubles for each one-unit increase in \(x\), so \(\exp(\widehat{\beta})=2\) and \(\widehat{\beta}=\log 2\approx0.6931\). Check: the fitted rates are \(0.02\), \(0.02(2)=0.04\), and \(0.02(2^2)=0.08\), reproducing all three observed cell rates. The common mistake avoided is fitting a straight line to the raw counts; the model relates the log rate to \(x\), while exposure determines the expected count.
When Aggregation Does Not Preserve the Model
The grouped likelihood result depends on the grouping matching the model. Suppose a cell combines policies with different values of a rating factor. Their rates are then \(\exp(x_j^\mathsf{T}\beta)\), not one common \(\lambda_g\). The expected total is a sum of different policy-level means, and it generally cannot be represented by \(E_g\exp(x_g^\mathsf{T}\beta)\) for one ordinary cell predictor \(x_g\). Keeping only the total count and exposure can therefore lose information needed to estimate the model.
Similarly, a zero-claim cell is valid information: it contributes \(-E_g\exp(x_g^\mathsf{T}\beta)\) to the log-likelihood up to constants. A cell with zero exposure, however, cannot have positive exposure-based information about a rate; a positive count paired with zero exposure contradicts this model. Do not manufacture a rate by dividing by zero.
| Grouping situation | What to retain or check |
|---|---|
| Same predictors, unequal policy exposures, independent Poisson counts | Sum counts and exposures; the coefficient likelihood is preserved up to a constant. |
| Same predictors, zero claims, positive exposure | Retain the cell; a zero count contributes valid likelihood information. |
| Different predictors combined into one cell | Retain finer groupings or use a justified model for the predictor distribution; totals alone need not preserve the likelihood. |
| Non-Poisson or dependent counts | Reassess the distribution and aggregation property rather than assuming the Poisson result applies. |
Offsets, Weights, and an Exam Trap
For grouped frequency estimation, the direct specification is a count response \(N_g\), a log link, and offset \(\log E_g\). This makes the expected count \(E_g\exp(x_g^\mathsf{T}\beta)\). Exposure is not normally entered as an estimated coefficient, because doubling exposure should double expected claims while leaving the underlying frequency rate unchanged.
A frequent error is to use the observed rate \(N_g/E_g\) as the response in an ordinary unweighted Poisson model. For example, in the first worked example, the rates are \(0.020\) and \(0.030\); treating each as one equally weighted observation ignores that the first rate is based on 700 policy-years and the second on 500. The correction is to model counts with exposure offsets, or to use a rate-response formulation only when the software’s weighting convention is known to reproduce the intended Poisson likelihood. Do not apply exposure both as an offset and as a frequency weight without a specific likelihood justification.
Check Your Understanding
For each problem, use the stated Poisson assumptions and determine what the cell totals imply.
- A hypothetical cell has 9 claims over 300 policy-years. Under a constant-rate Poisson model, find the maximum-likelihood frequency estimate and its fitted expected count. (Answer: \(0.030\) claims per policy-year; fitted expected count \(=9\).)
- Two hypothetical cells have 12 claims over 400 policy-years and 18 claims over 600 policy-years. Find their separate saturated rate estimates and the second-to-first rate ratio. (Answer: \(0.030\) and \(0.030\) claims per policy-year; ratio \(=1.000\).)
- Four policies in one rating cell have exposures 40, 60, 100, and 200 policy-years, with counts 0, 1, 2, and 3. Find the grouped count, grouped exposure, and maximum-likelihood rate. (Answer: \(N=6\), \(E=400\) policy-years, and \(\widehat{\lambda}=0.015\) claims per policy-year.)
- In a log-linear model with rate \(\exp(\alpha+\beta x)\), cell \(x=0\) has exposure 200 and 4 claims; cell \(x=1\) has exposure 100 and 4 claims. If each cell has its own rate estimate, find the rate ratio for \(x=1\) relative to \(x=0\). (Answer: \((4/100)/(4/200)=2.000\).)
- A cell combines policies with different rating-factor values, and only its total count and total exposure are available. Can those totals generally be treated as an exact aggregation for a GLM that includes the varying factor? (Answer: No; the policies have different modeled rates, so the totals alone generally do not preserve the policy-level coefficient likelihood.)