Avoiding Double Weighting Across Classification Levels

Why Nested Credibility Needs an Overlap Check

Classification systems often organize experience at more than one level. A portfolio may have a collective rate, a rate for a broad class, and rates for subgroups within that class. Credibility blending can be useful at each level, but the weights do not make the underlying observations independent. If a subgroup’s experience contributes to its class indication and that class indication is then used as the subgroup’s complement, the subgroup’s data can enter the calculation along two paths.

The practical question is not simply whether two credibility weights appear in a formula. It is which data each term represents. A class indication based on the full class includes the experience of every subgroup in that class. Treating that indication as if it were an independent complement for one of those subgroups can give the subgroup experience more influence than the stated subgroup weight suggests.

Definition: Let $m_{cs}$ be the observed indication for subgroup $s$ in class $c$, measured on a consistent exposure basis; let $E_{cs}$ be its exposure; and let $\mu$ be a collective indication. The exposure-weighted class experience is $m_c=\sum_s E_{cs}m_{cs}/E_c$, where $E_c=\sum_s E_{cs}$. A class indication is $C_c=Z_c m_c+(1-Z_c)\mu$. A subgroup indication using that class indication as its complement is $Q_{cs}=z_{cs}m_{cs}+(1-z_{cs})C_c$. Here $Z_c$ and $z_{cs}$ are the specified class and subgroup credibility weights.

These equations describe credibility blends, not a claim that the class and subgroup indications are independent observations. For the arithmetic to be meaningful, the indications must be on a consistent basis: for example, the same coverage, period, loss adjustments, and measure of cost per exposure. The exposure weights used to form $m_c$ must match the basis on which the subgroup indications are measured. A weight supplied for one level should not automatically be reused at another.

Nested blending is not automatically wrong. A jointly specified hierarchical model can account for the dependence between levels. The warning here concerns a common calculation error: interpreting the subgroup weight $z_{cs}$ as the total weight on subgroup experience even though $C_c$ already incorporates some of that same experience.

Trace the Experience Through the Formula

The overlap becomes visible by substituting the class formula into the subgroup formula. Define the subgroup’s exposure share in its class as $w_{cs}=E_{cs}/E_c$. Since the class mean contains the term $w_{cs}m_{cs}$, substitution gives:

\begin{aligned} Q_{cs} &=z_{cs}m_{cs}+(1-z_{cs})\left[Z_c\sum_j w_{cj}m_{cj}+(1-Z_c)\mu\right] \\ &=\left[z_{cs}+(1-z_{cs})Z_cw_{cs}\right]m_{cs} +(1-z_{cs})Z_c\sum_{j\ne s}w_{cj}m_{cj} +(1-z_{cs})(1-Z_c)\mu . \end{aligned}

This is the key diagnostic. The direct subgroup term contributes $z_{cs}$ of the subgroup indication. The class complement contributes an additional $(1-z_{cs})Z_cw_{cs}$, because the subgroup is part of the class experience. Thus, the effective coefficient on $m_{cs}$ is $z_{cs}+(1-z_{cs})Z_cw_{cs}$, not just $z_{cs}$.

Derivation: The class mean is $m_c=w_{cs}m_{cs}+\sum_{j\ne s}w_{cj}m_{cj}$, with exposure shares summing to one. Inserting this expression into $C_c=Z_cm_c+(1-Z_c)\mu$, then inserting $C_c$ into $Q_{cs}=z_{cs}m_{cs}+(1-z_{cs})C_c$, produces the expanded formula above. The extra subgroup coefficient is exactly the portion of the class complement that came from that subgroup’s own experience. It is positive when $z_{cs}<1$, $Z_c>0$, and $w_{cs}>0$.

The remaining coefficients also provide a useful check. The other subgroups collectively receive coefficient $(1-z_{cs})Z_c(1-w_{cs})$, and the collective receives $(1-z_{cs})(1-Z_c)$. Adding these to the effective subgroup coefficient gives one. The arithmetic is therefore a valid weighted combination, even when its interpretation as an independent-complement blend is not justified.

Worked Examples

Worked Example: Measuring the Subgroup’s Embedded Weight

A hypothetical class has two subgroups. Subgroup A has exposure 25 and an observed indication of $1,200 per exposure. Subgroup B has exposure 75 and an indication of $800 per exposure. The collective indication is $1,000. The class credibility weight is $Z_c=0.50$, and the subgroup A weight is $z_{cA}=0.40$. Calculate the class indication, the blended subgroup A indication, and the effective coefficient on subgroup A experience. State the method and identify the mistake the coefficient check avoids.

Method: exposure-weighted aggregation followed by coefficient tracing. Total class exposure is $25+75=100$. Subgroup A’s exposure share is $25/100=0.25$, and subgroup B’s share is $75/100=0.75$.

The class experience is:

m_c=0.25(1{,}200)+0.75(800)=300+600=900.

The class indication is:

C_c=0.50(900)+0.50(1{,}000)=450+500=950.

The subgroup A indication using the full class complement is:

Q_{cA}=0.40(1{,}200)+0.60(950)=480+570=1{,}050.

Now trace subgroup A’s coefficient. Its direct coefficient is 0.40. Its additional coefficient through the class complement is $0.60(0.50)(0.25)=0.075$. The effective coefficient is therefore $0.40+0.075=0.475$, or 47.5%. The full expansion checks the calculated indication: $0.475(1{,}200)+0.225(800)+0.300(1{,}000)=570+180+300=1{,}050$. The coefficients sum to one.

Common mistake avoided: Saying that subgroup A has exactly 40% weight because $z_{cA}=0.40$. That overlooks the additional 7.5% arriving through the class complement. The calculation is correct as a nested blend only if the overlapping contribution is intended and the model supports it; it is not a 40%-weight estimate of A with an independent complement.

Worked Example: Using a Leave-One-Subgroup-Out Complement

A hypothetical class contains target subgroup A, with exposure 20 and an indication of $1,500, and all remaining subgroups, with combined exposure 80 and an indication of $750. The collective indication is $1,000. Use a class credibility weight of 0.60 and a target subgroup weight of 0.50. Calculate the full-class blend, the subgroup indication using that blend, and an alternative indication using a leave-one-subgroup-out class complement. Explain what the comparison reveals.

Method: compare a full-class complement with a complement that excludes the target subgroup. The full class experience is:

m_c=0.20(1{,}500)+0.80(750)=300+600=900.

Its class indication is:

C_c=0.60(900)+0.40(1{,}000)=540+400=940.

Using that full-class indication, the subgroup indication is:

Q_{cA}=0.50(1{,}500)+0.50(940)=750+470=1{,}220.

The target subgroup’s effective coefficient is $0.50+(0.50)(0.60)(0.20)=0.56$. Thus, the calculation gives A 56% effective weight, not the stated direct weight of 50%.

For the alternative, first form a class complement from the experience outside A. The other-subgroup mean is $750. Blending that mean with the collective at the stated class weight gives $C_{c,-A}=0.60(750)+0.40(1{,}000)=450+400=850$. The subgroup indication using this leave-one-out complement is $0.50(1{,}500)+0.50(850)=750+425=1{,}175$.

The two results differ by $1{,}220-1{,}175=45$. The full-class complement imports part of A’s own experience; the leave-one-out complement does not. The 1,175 calculation removes this particular overlap, but it does not prove that the original weights are statistically optimal. In a practical model, the weights and complement definition must be chosen consistently.

Common mistake avoided: Using the full-class mean as if it excluded A. The class mean is 20% based on A, so the class complement necessarily carries some of A’s experience into the subgroup blend.

Worked Example: Detecting a Second Class Adjustment

A hypothetical class has three subgroups. Their exposures and observed indications are: A, 40 and $1,300; B, 60 and $900; C, 100 and $700. The collective indication is $1,000. The class weight is 0.50, and subgroup A’s weight is 0.25. Calculate A’s indication using the specified subgroup blend. Then calculate the result from an additional class blend of that already blended indication with the collective, using the class weight again. Explain why the second calculation should not be done unless a separate model explicitly calls for it.

Method: calculate once at each intended level, then audit any additional blending. Total exposure is $40+60+100=200$. The exposure shares are 0.20, 0.30, and 0.50. Therefore:

m_c=0.20(1{,}300)+0.30(900)+0.50(700)=260+270+350=880.

The class indication is $C_c=0.50(880)+0.50(1{,}000)=440+500=940$. The subgroup A indication is $Q_{cA}=0.25(1{,}300)+0.75(940)=325+705=1{,}030$. The coefficient on A is $0.25+(0.75)(0.50)(0.20)=0.325$, or 32.5%.

If someone then applies the class weight a second time, the result is $0.50(1{,}030)+0.50(1{,}000)=515+500=1{,}015$. This is not the subgroup blend specified in the problem; it is a new formula that shrinks an already class-adjusted indication toward the collective again. The second class adjustment is not an automatic consequence of nesting. It is appropriate only when a model explicitly defines that additional step and accounts for what each blend represents.

Common mistake avoided: Assuming that every rate at a lower classification level must receive another application of the class weight. Once the subgroup indication has used the class indication as its complement, applying the same class adjustment again changes the estimator.

A Practical Audit for Nested Indications

When a class and subgroup indication appear in the same calculation, use this sequence before interpreting the result:

Write the data represented by each term.
Identify the collective, class experience, and subgroup experience. Confirm that the observations are on the same basis and period.

Check whether the class experience includes the target subgroup.
If it does, calculate the subgroup exposure share used in the class mean.

Substitute before interpreting the weights.
Expand the nested formula and calculate the effective coefficient on the target subgroup’s experience.

Choose a complement consistent with the intended model.
If the subgroup experience should not re-enter through the complement, consider a leave-one-subgroup-out complement. Do not assume that the original weights remain optimal after changing the data used in the complement.

Check for an extra blend.
Do not apply a class adjustment again merely because the subgroup belongs to that class. Verify that every stated weighting step has a distinct intended role.

The exposure share matters: for fixed weights, a subgroup that accounts for more of the class mean receives a larger indirect contribution through the class complement. The class credibility weight matters too: if $Z_c=0$, the class complement is just the collective and carries no class experience; if $Z_c=1$, it is the full class experience. In the latter case, the effective subgroup coefficient becomes $z_{cs}+(1-z_{cs})w_{cs}$.

A leave-one-subgroup-out complement is a useful way to avoid direct overlap when the complement is meant to represent experience outside the target subgroup. But it is not a universal replacement for a coherent credibility model. If weights were developed for the full-class data, changing the data behind the complement can alter their interpretation. A sound indication therefore states both the weight and the experience to which that weight is applied.

Common Traps and Key Takeaway

Using an unweighted average of subgroup indications. When subgroup exposures differ, the class experience is generally exposure-weighted, not the simple average across subgroups.
Reading the direct subgroup weight as its total weight. If the complement includes the subgroup, expand the formula to find its effective coefficient.
Applying a class weight twice by habit. A class-adjusted subgroup indication does not automatically need another class blend.
Assuming overlap always makes the arithmetic invalid. Overlap is a modeling and interpretation issue; a jointly specified model can account for dependence. The concern is treating overlapping terms as independent or claiming a direct weight that the expanded formula does not support.

Key takeaway: In nested credibility, trace each observation through every blended term. When a subgroup’s experience is already inside the class complement, its effective weight is $z_{cs}+(1-z_{cs})Z_cw_{cs}$. Use a complement and weighting structure that match the intended model, and do not add another class adjustment without a defined reason.

Check Your Understanding

For each problem, track which experience enters through the complement as well as through the direct subgroup term.

A target subgroup has 30% of class exposure. The class weight is 0.40 and its subgroup weight is 0.50. Find the effective coefficient on the target subgroup’s experience when the full-class indication is its complement. (Answer: 0.56, or 56%)
A target subgroup has 20% of class exposure and an indication of $1,400. The remaining class experience is $800, the collective indication is $1,000, the class weight is 0.50, and the subgroup weight is 0.40. Calculate the full-class experience, class indication, subgroup indication, and effective coefficient on target experience. (Answer: class experience $920; class indication $960; subgroup indication $1,136; effective target coefficient 0.46, or 46%)
Subgroup A has 30% of class exposure and an indication of $1,200; the remaining subgroups have an indication of $800. The collective is $900, the class weight is 0.50, and A’s subgroup weight is 0.50. Calculate A’s indication first with the full-class complement and then with a leave-one-out complement formed from the remaining experience. (Answer: full-class indication $910 and subgroup indication $1,055; leave-one-out complement $850 and subgroup indication $1,025)
A target subgroup has exposure share 0.25, class weight 0.60, and subgroup weight 0.30. Calculate its effective coefficient when its experience is included in the class complement. (Answer: $0.30+(0.70)(0.60)(0.25)=0.405$, or 40.5%)
A collective indication is $1,000. A class indication already calculated for a target subgroup is $1,040; the target subgroup indication is $1,300, its weight is 0.40, and the class weight is 0.50. Calculate the specified subgroup blend, then calculate the result if someone incorrectly applies another class blend toward the collective. (Answer: specified subgroup blend $1,144; additional class blend $1,072)

What You'll Learn