Separating Overlapping Classification Effects

Why One-Way Class Averages Can Mislead

A one-way classification comparison asks how loss costs differ between levels of one variable, such as vehicle use or territory. But if that variable is associated with another risk characteristic, the class averages also reflect differences in the mix of that other characteristic. The observed average for a class can therefore be higher or lower even when its loss cost is lower within every comparable subgroup.

The practical question is not simply whether two marginal averages differ. It is whether that difference remains when the classes are compared among risks with similar values of the overlapping variable. Cross-tabulated experience makes this possible: it shows exposures and losses for combinations of the classification variables, so the analyst can compare like with like and, when appropriate, calculate averages using a common mix.

Definition: Let $A$ denote the classification variable being evaluated and $B$ an overlapping variable. For level $a$ of $A$ and level $b$ of $B$, let $E_{ab}$ be exposure and $L_{ab}$ be loss. The cell loss cost is $r_{ab}=L_{ab}/E_{ab}$, provided $E_{ab}>0$. The marginal loss cost for level $a$ is the total loss for that level divided by its total exposure across all levels of $B$. It reflects both the cell loss costs and the exposure mix of $B$ within level $a$.

A marginal class average is not generally the simple average of its cell loss costs. Cells with more exposure contribute more to the class result. This weighting is essential both when calculating the observed averages and when comparing classes on a standardized basis.

Why Marginal Loss Costs Are Exposure-Weighted

For a particular level $a$, total losses are the sum of losses across the levels of $B$, and total exposure is the corresponding sum of exposures. Since $L_{ab}=E_{ab}r_{ab}$, the marginal loss cost can be written as:

r_a^{\mathrm{marg}} = \frac{\sum_b L_{ab}}{\sum_b E_{ab}} = \frac{\sum_b E_{ab}r_{ab}}{\sum_b E_{ab}} = \sum_b \left(\frac{E_{ab}}{\sum_j E_{aj}}\right)r_{ab}.

The last expression shows why the class's own mix matters: each cell loss cost is weighted by that cell's share of the level's exposure. If two levels of $A$ have different distributions across $B$, their marginal averages combine different weights with the cell results. A marginal difference therefore need not represent a difference between the levels at comparable values of $B$.

One way to make a descriptive comparison less dependent on different observed mixes is to choose common weights $w_b$, where $w_b\geq 0$ and $\sum_b w_b=1$, and apply them to every level of $A$:

r_a^{\mathrm{std}}=\sum_b w_b r_{ab}.

This is a standardized loss cost for the stated common mix. The weights might represent a selected reference portfolio or another appropriate target mix. They must be identical across the classes being compared. Standardization changes the composition used in the summary; it does not change the observed cell loss costs.

A Practical Comparison

A useful sequence is to calculate cell loss costs, compare the levels of interest within each level of the overlapping variable, and then inspect the marginal averages. If a single overall comparison is useful, apply explicitly stated common weights. This sequence helps distinguish a mix-driven marginal difference from a pattern that persists within subgroups.

Check the exposure and loss basis.
Make sure the cells use a consistent definition of exposure and losses, and calculate each cell loss cost as losses divided by exposure.

Compare within levels of the overlapping variable.
For each level of $B$, compare the $A$ classes using their cell loss costs. This removes differences in the observed mix of $B$ from that particular comparison.

Calculate the marginal results.
Sum losses and exposures over $B$ separately for each $A$ class. These results describe the actual class mixes, not a common mix.

Standardize only with a stated purpose.
If an overall comparison is needed, choose common weights, apply them consistently, and report the resulting comparison as mix-adjusted rather than as the observed marginal result.

A standardized comparison is descriptive. It does not, by itself, establish that a classification variable causes losses to change, nor does it prove that a proposed rating factor is appropriate. Other characteristics, data limitations, and the intended rating use still matter. In particular, a within-stratum comparison requires experience for the classes being compared in that stratum; if one class is absent, the data do not provide that direct comparison there.

Worked Example: A Marginal Reversal Across Territories

A hypothetical insurer is reviewing two vehicle-use classes, Private and Delivery. The portfolio also contains two territories. Assume the following cell exposures and loss costs have been calculated on the same basis. Determine the marginal loss cost for each use class, compare the classes within each territory, and calculate their loss costs using a 50% / 50% territory mix.

Territory	Use class	Exposure	Cell loss cost
A	Private	10	$100
A	Delivery	90	$150
B	Private	90	$500
B	Delivery	10	$550

Method. First calculate cell losses by multiplying exposure by cell loss cost. Then calculate actual marginal averages and compare the cell results within each territory. The common mistake to avoid is interpreting the marginal use-class comparison as if both classes had the same territory mix.

Private losses. In Territory A, losses are $10\times \$100=\$1{,}000$. In Territory B, losses are $90\times \$500=\$45{,}000$. Total Private losses are $\$1{,}000+\$45{,}000=\$46{,}000$, and total Private exposure is $10+90=100$. Its marginal loss cost is $\$46{,}000/100=\$460$.

Delivery losses. In Territory A, losses are $90\times \$150=\$13{,}500$. In Territory B, losses are $10\times \$550=\$5{,}500$. Total Delivery losses are $\$13{,}500+\$5{,}500=\$19{,}000$, and total Delivery exposure is $90+10=100$. Its marginal loss cost is $\$19{,}000/100=\$190$.

The marginal comparison makes Delivery look less costly: $\$190$ versus $\$460$. But within Territory A, Delivery is $\$150-\$100=\$50$ higher than Private; within Territory B, it is $\$550-\$500=\$50$ higher. Delivery has most of its exposure in the lower-cost Territory A, while Private has most in Territory B.

With equal territory weights, standardized Private loss cost is $0.5(\$100)+0.5(\$500)=\$300$. Standardized Delivery loss cost is $0.5(\$150)+0.5(\$550)=\$350$. On that common mix, Delivery is $\$50$ higher, or $\$350/\$300=1.1667$ times the Private loss cost. The check is that the standardized difference, $\$50$, agrees with the within-territory differences because those differences are equal in both territories. The marginal reversal is explained by territory mix, not by lower Delivery loss costs within either territory.

Choosing and Interpreting Common Weights

The common weights answer a specific question: what would each class's average loss cost be if the classes were compared using this same distribution of the overlapping variable? Different reasonable reference mixes can produce different standardized averages when cell comparisons vary across strata. Therefore, state the weights and avoid presenting a standardized number as though it were the unique mix-free answer.

If the within-stratum differences point in the same direction and have similar sizes, a common-mix summary can be a useful concise comparison. If those differences vary substantially, a single standardized result can hide important variation. The cells should remain visible in the analysis, and any summary should be interpreted alongside them. This is a diagnostic use of cross-tabulated experience, not a claim that all combinations have the same effect.

Worked Example: Standardizing Equipment-Age Experience

A hypothetical insurer reviews loss costs for New and Older equipment in Coastal and Inland areas. The observed exposures and cell loss costs are shown below. Calculate each actual marginal loss cost, compare the age classes within each area, and standardize to equal Coastal and Inland weights. State why the marginal ratio does not equal the standardized ratio.

Area	Equipment age	Exposure	Cell loss cost
Coastal	New	30	$600
Inland	New	70	$300
Coastal	Older	70	$800
Inland	Older	30	$500

Method. Use exposure-weighted cell results for the observed marginal averages, then apply the same 50% / 50% area weights to both age classes. A common mistake is to compare the marginal ratio with a within-area ratio without noticing that they use different exposure mixes.

For New equipment, Coastal losses are $30\times \$600=\$18{,}000$, and Inland losses are $70\times \$300=\$21{,}000$. Total losses are $\$39{,}000$ over 100 exposures, giving a marginal loss cost of $\$390$.

For Older equipment, Coastal losses are $70\times \$800=\$56{,}000$, and Inland losses are $30\times \$500=\$15{,}000$. Total losses are $\$71{,}000$ over 100 exposures, giving a marginal loss cost of $\$710$. The marginal ratio of Older to New is $\$710/\$390=1.8205$.

Within Coastal, Older is $\$800/\$600=1.3333$ times New. Within Inland, Older is $\$500/\$300=1.6667$ times New. Thus both within-area comparisons show higher loss costs for Older equipment, but the relative differences are not equal.

For the equal-area mix, New has standardized loss cost $0.5(\$600)+0.5(\$300)=\$450$. Older has $0.5(\$800)+0.5(\$500)=\$650$. The standardized ratio is $\$650/\$450=1.4444$. This differs from 1.8205 because the observed marginal results weight areas according to each age class's own exposure distribution, while the standardized results assign both age classes the same area weights. As a check, New's observed mix is 30% Coastal and Older's is 70% Coastal, exactly as the exposure counts show.

When the Data Do Not Separate the Effects

Cross-tabulation helps distinguish overlapping effects only where the data support comparisons. If every risk in one level of $A$ is in one level of $B$, while every risk in another level of $A$ is in a different level of $B$, then class and subgroup differences cannot be disentangled by direct within-stratum comparisons. The marginal results still describe the observed groups, but they do not isolate the contribution associated with either variable.

This limitation is different from ordinary exposure weighting. A large exposure imbalance can make marginal averages misleading while still leaving observations in each relevant subgroup for comparison. A lack of exposure for a needed cell means that the direct comparison is unavailable. Do not fill that gap by treating an unobserved cell loss cost as known.

Worked Example: A Difference Hidden by Portfolio Mix

A hypothetical insurer compares Standard and Intensive equipment use across Flat and Hilly terrain. Calculate marginal and equally standardized loss costs, then explain what each comparison says.

Terrain	Use class	Exposure	Cell loss cost
Flat	Standard	40	$500
Hilly	Standard	60	$1,000
Flat	Intensive	80	$700
Hilly	Intensive	20	$1,200

Method. Calculate the observed exposure-weighted averages and then compare the classes within terrain. The common mistake is to conclude that there is no use-class difference because the marginal averages happen to match.

Standard losses are $40\times \$500=\$20{,}000$ on Flat terrain and $60\times \$1{,}000=\$60{,}000$ on Hilly terrain. Total Standard losses are $\$80{,}000$ over 100 exposures, so the marginal loss cost is $\$800$.

Intensive losses are $80\times \$700=\$56{,}000$ on Flat terrain and $20\times \$1{,}200=\$24{,}000$ on Hilly terrain. Total Intensive losses are also $\$80{,}000$ over 100 exposures, so its marginal loss cost is also $\$800$. The marginal comparison indicates no difference.

Within Flat terrain, Intensive is $\$700-\$500=\$200$ higher. Within Hilly terrain, Intensive is $\$1{,}200-\$1{,}000=\$200$ higher. With equal terrain weights, Standard is $0.5(\$500)+0.5(\$1{,}000)=\$750$, while Intensive is $0.5(\$700)+0.5(\$1{,}200)=\$950$. The standardized difference is $\$200$, and the ratio is $\$950/\$750=1.2667$. The equal marginal averages arise because Standard has more Hilly exposure and Intensive has more Flat exposure. The final check is that both actual totals are $\$80{,}000$, while their common-mix results differ because they use equal terrain weights.

Common Mistakes and Exam Traps

Comparing marginal averages as though mix were equal. A marginal loss cost describes both the cell results and the class's own exposure distribution. Check the exposure mix before interpreting a difference as a class effect.
Taking an unweighted average of cell loss costs. The actual marginal average is $\sum_b E_{ab}r_{ab}/\sum_b E_{ab}$, not the arithmetic mean of $r_{ab}$ unless the cell exposures are equal.
Changing the weights between classes. A valid standardized comparison uses the same $w_b$ for each class. Otherwise, mix differences remain in the comparison.
Calling a standardized result causal or uniquely correct. It is a comparison for the selected mix. The chosen weights and observed cells should be reported, and the result should not be interpreted beyond what the data support.
Ignoring a missing comparison cell. If a class has no exposure in a subgroup, its within-subgroup loss cost cannot be calculated from that experience. A marginal average does not supply the missing cell result.

Key takeaway: Cross-tabulated experience exposes the role of portfolio composition. Use within-stratum comparisons to see how class results differ at comparable levels of another variable, and use common weights when a single mix-adjusted summary is needed. Keep marginal and standardized results distinct: they answer different questions.

Check Your Understanding

For each problem, distinguish the observed mix from any stated common mix.

A class has 40 exposures with cell loss cost $\$250$ and 60 exposures with cell loss cost $\$400$. Calculate its marginal loss cost. (Answer: $\$340$)
For two regions, class P has loss costs of $\$200$ Urban and $\$500$ Rural; class Q has $\$300$ Urban and $\$600$ Rural. Using common weights of 25% Urban and 75% Rural, calculate the standardized loss costs and their difference. (Answer: P $\$425$, Q $\$525$, Q is $\$100$ higher)
Class X has 20 exposures and $\$2{,}000$ of losses in a low-cost stratum, and 80 exposures and $\$16{,}000$ of losses in a high-cost stratum. Class Y has 80 exposures and $\$9{,}600$ of losses in the low-cost stratum, and 20 exposures and $\$4{,}400$ of losses in the high-cost stratum. Calculate the marginal loss cost for each class and the equally standardized loss cost for each. (Answer: Marginal X $\$180$, Y $\$140$; standardized X $\$150$, Y $\$170$)
Class P has loss costs of $\$200$ Urban and $\$400$ Rural, with 20 Urban and 80 Rural exposures. Class Q has loss costs of $\$300$ Urban and $\$500$ Rural, with 80 Urban and 20 Rural exposures. Calculate the marginal loss costs and the standardized difference using equal region weights. (Answer: Marginal P $\$360$, Q $\$340$; standardized P $\$300$, Q $\$400$, so Q is $\$100$ higher)
Suppose all risks in class A are in Territory 1 and all risks in class B are in Territory 2. Can the observed cross-tabulated experience directly separate a class difference from a territory difference? State why. (Answer: No; there is no within-territory comparison between A and B, so the observed difference cannot be separated directly into class and territory contributions.)

What You'll Learn