Segmentation Trade-Offs In Credibility

Why Segmentation Is a Trade-Off

A rating plan that separates risks into finer groups can reflect genuine differences between them. But each finer estimate is based on fewer observations, so random fluctuation has greater influence. Combining groups provides more data and can stabilize an estimate, but the combined estimate may not represent any one subgroup well. The central question is therefore not simply whether more detail is desirable: it is whether the reduction in sampling variation from pooling is worth the resulting mismatch for the subgroup being estimated.

A useful way to make that comparison is mean squared error (MSE), measured relative to a clearly specified target. MSE includes both sampling variation and systematic error. In this tutorial, a hypothetical portfolio has two groups, and the target is the underlying mean for one particular group. The calculations show why either fine or broad segmentation can be preferable, depending on how different the groups are and how much data each contains.

Definition: A fine estimate for a group uses only observations from that group. A broad estimate pools observations across multiple groups. For an estimate $\hat{\theta}$ of a target $\theta$, its mean squared error is $$ \operatorname{MSE}(\hat{\theta})=E[(\hat{\theta}-\theta)^2] =\operatorname{Var}(\hat{\theta})+\big(E[\hat{\theta}]-\theta\big)^2. $$ The second term is squared bias. MSE compares estimates for the same target; a pooled portfolio average is not automatically an appropriate estimate of a subgroup mean.

A Two-Group Model

Suppose observations in group 1 are independent and identically distributed with mean $\mu_1$ and variance $\sigma^2$, while observations in group 2 are independent and identically distributed with mean $\mu_2$ and the same variance $\sigma^2$. Assume observations are independent across the groups as well. Let there be $n_1$ observations in group 1 and $n_2$ in group 2. The target is $\mu_1$, and the observations are on the same defined exposure and measurement basis.

The fine estimate of the target is the group 1 sample mean. The broad estimate is the sample mean after pooling both groups. Under the model, the fine estimate is unbiased for $\mu_1$, while the pooled estimate generally is not. Its expectation is a weighted average of the two group means, with weights determined by the numbers of observations.

\bar{X}_1=\frac{1}{n_1}\sum_{j=1}^{n_1}X_{1j}, \qquad \bar{X}_{\text{pool}}= \frac{n_1\bar{X}_1+n_2\bar{X}_2}{n_1+n_2}.

Because independent observations with variance $\sigma^2$ have a sample mean variance equal to $\sigma^2$ divided by the sample size, the fine estimate has variance $\sigma^2/n_1$. The pooled estimate has variance $\sigma^2/(n_1+n_2)$. Pooling therefore reduces variance in this model. But for the group 1 target, the pooled estimate’s bias is its expected value minus $\mu_1$:

E[\bar{X}_{\text{pool}}]-\mu_1 = \frac{n_1\mu_1+n_2\mu_2}{n_1+n_2}-\mu_1 = \frac{n_2}{n_1+n_2}(\mu_2-\mu_1).

The bias is larger when the other group is more different, and its weight is larger when group 2 contributes more observations to the pool. These results give the following MSEs for estimating $\mu_1$.

\operatorname{MSE}(\bar{X}_1)=\frac{\sigma^2}{n_1}, \qquad \operatorname{MSE}(\bar{X}_{\text{pool}}) = \frac{\sigma^2}{n_1+n_2} + \left[ \frac{n_2}{n_1+n_2}(\mu_2-\mu_1) \right]^2.

Why the MSE Formula Holds

The identity separating MSE into variance and squared bias follows by adding and subtracting the estimator’s expectation. Let $m=E[\hat{\theta}]$. Then $\hat{\theta}-\theta=(\hat{\theta}-m)+(m-\theta)$. Squaring and taking expectations gives a variance term, a squared-bias term, and a cross term. The cross term is zero because $E[\hat{\theta}-m]=0$.

\begin{aligned} E[(\hat{\theta}-\theta)^2] &=E[((\hat{\theta}-m)+(m-\theta))^2]\\ &=E[(\hat{\theta}-m)^2] +2(m-\theta)E[\hat{\theta}-m] +(m-\theta)^2\\ &=\operatorname{Var}(\hat{\theta})+(m-\theta)^2. \end{aligned}

This explains why comparing variances alone is insufficient: an estimate with lower variance can still have higher MSE if its bias is large. It also explains why the target must be stated. If the target were the overall mean of the two groups rather than $\mu_1$, the pooled estimate could be appropriate and its bias calculation would be different.

Equal-Size Groups: A Simple Decision Rule

When both groups have $n$ observations, the pooled estimate gives each group equal weight. For the group 1 target, the broad estimate’s bias is $(\mu_2-\mu_1)/2$. The broad estimate has lower MSE exactly when its variance reduction is greater than its squared bias.

\begin{aligned} \operatorname{MSE}(\bar{X}_{\text{pool}}) &<\operatorname{MSE}(\bar{X}_1)\\ \frac{\sigma^2}{2n}+\frac{(\mu_2-\mu_1)^2}{4} &<\frac{\sigma^2}{n}\\ (\mu_2-\mu_1)^2&<\frac{2\sigma^2}{n}. \end{aligned}

Thus, under these assumptions, pooling is favored when the group means are sufficiently close relative to the observation variance and group size. As $n$ grows, the variance of the fine estimate falls. A fixed difference between the group means then becomes more consequential relative to the variance saved by pooling.

Worked Example: A Large Difference Between Equal-Size Groups

A hypothetical insurer is estimating the mean claim cost for group 1. Each group has $n=5$ independent observations, and the individual-observation variance in both groups is $\sigma^2=900$ dollars squared. Assume the group means are $\mu_1=\$40$ and $\mu_2=\$70$. Compare the fine and broad estimates for group 1 using MSE.

Method. Use the common-variance, equal-size two-group model. The target is $\mu_1=\$40$; the variance input $900$ is the variance of each individual observation, not the variance of a group estimate.

Fine estimate. Its expectation is $\$40$, so it has zero bias. Its variance and MSE are $$ \frac{900}{5}=180, \qquad \operatorname{MSE}(\bar{X}_1)=180+0^2=180. $$

Broad estimate. Its variance is $$ \frac{900}{5+5}=90. $$ Its expected value is $(\$40+\$70)/2=\$55$, so its bias for the group 1 target is $\$55-\$40=\$15$. Therefore, $$ \operatorname{MSE}(\bar{X}_{\text{pool}})=90+15^2=90+225=315. $$

Check and conclusion. The broad estimate has lower variance, $90$ rather than $180$, but its squared bias is $225$. Its MSE, $315$, exceeds the fine estimate’s MSE of $180$, so fine segmentation is preferable for this target under the stated model. This avoids the common mistake of comparing $90$ with $180$ and choosing the broad estimate without including its bias.

Worked Example: Similar Equal-Size Groups

A hypothetical service-contract portfolio has two equally sized segments, each with $n=5$ independent observations. The variance of each individual observation is $\sigma^2=900$, and the assumed means are $\mu_1=\$58$ and $\mu_2=\$62$. For the group 1 target, calculate both MSEs and select the lower-MSE estimate.

Method. Apply the same equal-size model, taking the group 1 mean of $\$58$ as the target.

Fine estimate. It is unbiased, and its variance is $$ \operatorname{Var}(\bar{X}_1)=\frac{900}{5}=180. $$ Thus its MSE is $180+0^2=180$.

Broad estimate. Its variance is $900/(5+5)=90$. Its expected value is $(\$58+\$62)/2=\$60$, so its bias for group 1 is $\$60-\$58=\$2$. The squared bias is $2^2=4$, and $$ \operatorname{MSE}(\bar{X}_{\text{pool}})=90+4=94. $$

Check and conclusion. The broad estimate’s MSE, $94$, is less than the fine estimate’s MSE, $180$. The broad estimate is preferable for estimating group 1 under this model. The small difference between the group means makes the bias cost of pooling modest, while pooling halves the variance relative to the fine estimate.

Unequal Group Sizes: Pooling Does Not Mean Taking the Midpoint

If group sizes differ, the pooled estimate is not the simple average of the two group means. Each observation has equal weight in the pooled sample mean, so the larger group receives more weight. For a group 1 target, this can make the pooled estimate’s bias larger or smaller than a midpoint-based calculation suggests. The formulas above incorporate the unequal weights directly.

Worked Example: Unequal Groups and a Weighted Pooled Estimate

A hypothetical insurer studies two types of small commercial policies. There are $n_1=4$ independent observations in group 1 and $n_2=12$ in group 2. The individual-observation variance is $\sigma^2=400$ in each group, and the assumed means are $\mu_1=\$50$ and $\mu_2=\$60$. Compare the estimates for the group 1 target.

Method. Use the unequal-size formulas. Group 2 supplies $12$ of the $16$ observations, so its mean receives weight $12/16$ in the pooled estimate.

Fine estimate. It is unbiased for $\$50$, and $$ \operatorname{MSE}(\bar{X}_1)=\frac{400}{4}=100. $$

Broad estimate variance. The total number of observations is $4+12=16$, giving $$ \operatorname{Var}(\bar{X}_{\text{pool}})=\frac{400}{16}=25. $$

Broad estimate bias. Its expected value is $$ \frac{4(\$50)+12(\$60)}{16} =\frac{\$200+\$720}{16} =\$57.50. $$ The bias for group 1 is $\$57.50-\$50=\$7.50$, and the squared bias is $7.50^2=56.25$. Therefore, $$ \operatorname{MSE}(\bar{X}_{\text{pool}})=25+56.25=81.25. $$

Check and conclusion. The broad MSE of $81.25$ is below the fine MSE of $100$, so pooling is preferable for this group 1 target under the model. The pooled expectation, $\$57.50$, is closer to the larger group’s mean of $\$60$ than to $\$50$; treating it as the midpoint $\$55$ would give the wrong bias and MSE.

Assumptions and Practical Interpretation

These comparisons isolate a particular trade-off; they are not a universal rule that one segmentation level is always best. The formulas require independent observations, a common within-group variance, and a clearly specified group mean as the target. They also presume the group means and variance used for the comparison are known for the theoretical calculation. In practice, they may need to be estimated, so the calculated MSE comparison is only as credible as those assumptions and inputs.

If variances differ between groups, the pooled mean’s variance is not generally $\sigma^2/(n_1+n_2)$. If observations are dependent, the variance of a mean must account for their covariance. If observations do not share a comparable exposure basis or claim measure, averaging them may not estimate the intended quantity. In those situations, the simple formulas here do not apply without modification.

The decision also depends on the purpose. If the goal is to estimate a subgroup-specific mean, pooling can introduce bias. If the goal is instead an overall portfolio mean, a weighted pooled estimate may be exactly the desired quantity. A comparison must therefore name both the estimand (the quantity being estimated) and the candidate estimates before calculating MSE.

Common Mistakes

Ignoring bias. A smaller variance does not guarantee smaller MSE. For example, in the first worked example, the broad variance is $90$, but its squared bias is $225$, making its MSE $315$.
Using a midpoint for unequal groups. The pooled mean weights each observation, not each group equally. With four observations in one group and twelve in the other, the weights are $4/16$ and $12/16$, not $1/2$ and $1/2$.
Confusing observation variance with estimate variance. If $\sigma^2=900$ is the variance of an individual observation, a mean of five observations has variance $900/5=180$, not $900$.
Comparing estimates against different targets. MSE for a subgroup mean cannot be compared directly with MSE for an overall mean. State the target first.
Assuming fine segmentation always wins. More groups can expose structural differences, but with limited observations the resulting estimates can be less stable. The MSE comparison quantifies both effects under its assumptions.

Key takeaway: Fine segmentation reduces pooling bias for a subgroup target but uses fewer observations. Broad segmentation can lower sampling variance while pulling the estimate toward other groups. Compare variance plus squared bias for the same target, and use weights that match the actual pooled observations.

Check Your Understanding

For each problem, compare estimates only for the stated target and use the common-variance model unless otherwise stated.

Two hypothetical groups each have $n=8$ independent observations with individual-observation variance $\sigma^2=640$. Their means are $\mu_1=\$90$ and $\mu_2=\$98$. For the group 1 target, calculate fine and broad MSE. (Answer: Fine MSE $=80$; broad MSE $=40+16=56$, so broad is lower.)
Two equally sized hypothetical groups each have $n=4$ observations with individual-observation variance $\sigma^2=1{,}000$. Their means are $\mu_1=\$30$ and $\mu_2=\$50$. For group 1, calculate both MSEs. (Answer: Fine MSE $=250$; broad MSE $=125+100=225$, so broad is lower.)
A hypothetical portfolio has $n_1=6$, $n_2=3$, individual-observation variance $\sigma^2=360$, and means $\mu_1=\$44$, $\mu_2=\$50$. For the group 1 target, calculate the broad estimate’s expected value and compare MSEs. (Answer: Broad expectation $=\$46$; fine MSE $=60$; broad MSE $=40+2^2=44$, so broad is lower.)
Two equal-size groups each have $n=10$ observations and individual-observation variance $\sigma^2=500$. Their means are $\mu_1=\$100$ and $\mu_2=\$108$. For group 1, calculate both MSEs and use the equal-size decision condition to check the result. (Answer: Fine MSE $=50$; broad MSE $=25+16=41$, so broad is lower; the condition holds because $8^2=64<2(500)/10=100$.)
A hypothetical portfolio has $n_1=5$, $n_2=15$, individual-observation variance $\sigma^2=800$, and means $\mu_1=\$70$, $\mu_2=\$74$. For the group 1 target, calculate the fine and broad MSEs using the actual pooled weights. (Answer: Fine MSE $=160$; broad expectation $=\$73$, broad MSE $=40+3^2=49$, so broad is lower.)

What You'll Learn