Overdispersion and Claim-Count Standards

Why Overdispersion Changes the Standard

A claim-count standard asks how much experience is needed for the observed count to be sufficiently close to its expected value with a specified probability. Under a Poisson model, the variance equals the mean. But real or hypothetical portfolios may be modeled with a claim-count variance larger than the mean. Since greater variance means greater fluctuation around the expected count, a larger expected count is then needed to meet the same confidence and tolerance targets.

The adjustment is determined by the count’s variance-to-mean ratio. This tutorial focuses on how that ratio changes the limited-fluctuation threshold and how to convert the threshold into exposure. The result uses the normal approximation for discrete claim counts; it is a planning standard, not an exact probability calculation for every discrete count model.

Definition: Let $N$ be a claim count with mean $\mu=E[N]>0$ and variance $\operatorname{Var}(N)$. Its variance-to-mean ratio is $$ \phi=\frac{\operatorname{Var}(N)}{E[N]}=\frac{\operatorname{Var}(N)}{\mu}. $$ When $\phi>1$, the count is overdispersed relative to a Poisson count with the same mean. When $\phi=1$, it has the Poisson variance relationship; when $\phi<1$, it is underdispersed.

For a claim-count limited-fluctuation standard, let $r$ be the allowable relative deviation and let $q$ be the desired central probability. The target is that the count be within $r\mu$ of its mean. For example, $r=0.10$ means a tolerance of 10% of the expected count. The confidence probability is central: the count may be above or below its mean, and both tails are included.

Deriving the Adjusted Threshold

Suppose the count has mean $\mu$, variance $\phi\mu$, and is approximated by a normal distribution with those parameters. Standardizing the deviation from the mean gives the approximate probability of meeting the relative-tolerance requirement:

P\bigl(|N-\mu|\le r\mu\bigr) \approx 2\Phi\left(\frac{r\mu}{\sqrt{\phi\mu}}\right)-1 = 2\Phi\left(r\sqrt{\frac{\mu}{\phi}}\right)-1,

where $\Phi$ is the standard normal cumulative distribution function. Let $z_{(1+q)/2}$ be the standard normal quantile that leaves probability $q$ between its negative and positive values. To attain central probability $q$ under the approximation, require the standardized tolerance to be at least this quantile:

r\sqrt{\frac{\mu}{\phi}}\ge z_{(1+q)/2}.

Both sides are nonnegative. Squaring and rearranging therefore preserves the inequality:

\mu\ge \phi\left(\frac{z_{(1+q)/2}}{r}\right)^2.

Result: Under the normal approximation, the minimum expected claim count for central confidence $q$ and relative tolerance $r$ is $$ \mu_{\min}=\phi\left(\frac{z_{(1+q)/2}}{r}\right)^2. $$ For a Poisson count, $\phi=1$, so this reduces to the Poisson count threshold. If $\phi>1$, the threshold is multiplied by $\phi$, not by $\phi^2$.

The formula shows the separate roles of the inputs. A larger confidence level gives a larger normal quantile and raises the threshold. A tighter tolerance means a smaller $r$, which also raises the threshold. Overdispersion increases the threshold in direct proportion to $\phi$: for fixed confidence and tolerance, doubling the variance-to-mean ratio doubles the required expected count.

From Expected Claims to Exposure

A threshold in expected claims can be translated into exposure when an expected claim frequency per exposure unit is specified. Suppose a portfolio with $m$ exposure units has mean claim count $\mu=\lambda m$, where $\lambda$ is the expected count per exposure unit. If the variance-to-mean ratio is $\phi$, then $\operatorname{Var}(N)=\phi\lambda m$, and the threshold becomes:

m_{\min}=\frac{\mu_{\min}}{\lambda} =\frac{\phi}{\lambda}\left(\frac{z_{(1+q)/2}}{r}\right)^2.

This conversion assumes the specified mean frequency and variance-to-mean ratio apply over the exposure range being considered. If exposure units must be whole units, round the calculated requirement up to the next whole unit. The expected-claim threshold itself is a model-based threshold and need not be an integer.

Worked Example: Adjusting a Poisson Threshold for Overdispersion

A hypothetical portfolio’s claim count is modeled with mean $\mu$ and variance $1.8\mu$, so $\phi=1.8$. Determine the expected-claim threshold for 90% central confidence within a 10% relative tolerance, and compare it with the Poisson threshold. No interest rate or mortality assumption enters this claim-count calculation. Use the normal approximation $N\approx N(\mu,1.8\mu)$, where the second normal parameter is the variance.

Step 1: Set the confidence quantile. For central confidence $q=0.90$, the quantile is $z_{0.95}=1.645$. The tolerance is $r=0.10$.

Step 2: Calculate the Poisson baseline. If $\phi=1$, the expected-claim threshold is $$ \mu_{\min,\text{Pois}}=\left(\frac{1.645}{0.10}\right)^2 =(16.45)^2=270.6025. $$

Step 3: Adjust for overdispersion. Multiply the Poisson threshold by $\phi=1.8$: $$ \mu_{\min}=1.8(270.6025)=487.0845. $$ Thus the approximate threshold is 487.1 expected claims, to one decimal place. If expressing a conservative whole-claim count threshold, round up to 488 claims.

Check. The adjusted threshold is $1.8$ times the Poisson threshold, as required by the formula. The adjustment increases the threshold because the count has greater variance at the same mean.

Method and mistake avoided. The method is to calculate the Poisson baseline and multiply by the variance-to-mean ratio. Do not multiply by $1.8^2$: the standard deviation contains $\sqrt{\phi}$, and squaring the standardized inequality produces a factor of $\phi$, not $\phi^2$.

Worked Example: Inferring Dispersion and Finding Required Exposure

In a hypothetical segment, the claim count has mean 400 and variance 600. The expected frequency is 0.12 claims per exposure unit. Find the required exposure for 95% central confidence within a 10% relative tolerance. No interest rate or mortality assumption enters. Use the normal approximation with the calculated mean and variance relationship.

Step 1: Calculate the variance-to-mean ratio. $$ \phi=\frac{600}{400}=1.5. $$ The segment is overdispersed relative to a Poisson count with mean 400.

Step 2: Calculate the expected-claim threshold. For $q=0.95$, use $z_{0.975}=1.96$, and set $r=0.10$: $$ \mu_{\min}=1.5\left(\frac{1.96}{0.10}\right)^2 =1.5(19.6)^2 =1.5(384.16) =576.24. $$ The target is therefore at least 576.24 expected claims under the model.

Step 3: Convert to exposure. With frequency $\lambda=0.12$ claims per exposure unit, $$ m_{\min}=\frac{576.24}{0.12}=4{,}802 $$ exposure units. If units are indivisible, 4,802 already is a whole number, so no further rounding is needed.

Check. At 4,802 exposure units, the mean count is $0.12(4{,}802)=576.24$, exactly the derived expected-count threshold. Its modeled variance is $1.5(576.24)=864.36$, consistent with the stated ratio.

Method and mistake avoided. First infer $\phi$ from variance divided by mean, then find the adjusted count threshold, then divide by frequency. Using the stated variance of 600 directly as the variance at the target exposure would be incorrect: the variance relationship is specified by the ratio, and it scales with the mean count.

Worked Example: Comparing Overdispersion Levels at a Given Frequency

A hypothetical coverage has a claim-count model with mean $\mu$ and variance $2\mu$, and an expected frequency of 0.20 claims per exposure unit. Determine the required exposure for 90% central confidence within a 15% relative tolerance. Also compare the expected-claim threshold with the Poisson threshold. No interest rate or mortality assumption enters; use $N\approx N(\mu,2\mu)$.

Step 1: Calculate the Poisson threshold. With $z_{0.95}=1.645$ and $r=0.15$: $$ \mu_{\min,\text{Pois}}=\left(\frac{1.645}{0.15}\right)^2 =(10.9666667)^2 \approx120.2678. $$

Step 2: Adjust the threshold. Here $\phi=2$, so $$ \mu_{\min}=2(120.2677778)\approx240.5356 $$ expected claims. The final result is 240.5 expected claims to one decimal place.

Step 3: Convert to exposure. $$ m_{\min}=\frac{240.5355556}{0.20} \approx1{,}202.6778 $$ exposure units. If only whole exposure units are available, round up to 1,203.

Check. At 1,203 whole units, the expected count is $0.20(1{,}203)=240.6$, which exceeds 240.5356. At 1,202 units, the expected count is 240.4, which is below the threshold. Thus 1,203 is the smallest whole-unit requirement.

Method and mistake avoided. The count threshold doubles because the ratio is 2, while the required exposure is then obtained by dividing by frequency. Do not round the expected count threshold to a whole claim before converting exposure; rounding the exposure requirement is the relevant step when exposure units must be indivisible.

Interpreting the Adjustment

Overdispersion does not change the form of the relative-tolerance requirement. It changes the variance used to standardize the count’s deviation from its mean. At a fixed expected count, the standard deviation is $\sqrt{\phi\mu}$, so a larger $\phi$ makes the tolerance band less wide in standard-deviation units. More expected claims are needed to restore the desired confidence.

A variance-to-mean ratio is a concise way to summarize this effect for the standard. It does not, by itself, identify why overdispersion occurs or prove that a particular claim-count distribution is appropriate. The formula should be used with the mean-variance relationship specified for the model and with the normal approximation understood. The approximation is most naturally used for thresholds large enough that a continuous normal calculation is reasonable; discreteness and continuity corrections can matter for smaller counts.

Key takeaway: For a count with mean $\mu$ and variance $\phi\mu$, the normal-approximation limited-fluctuation threshold is $\phi$ times the Poisson threshold. To obtain exposure, divide the adjusted expected-count threshold by the expected claim frequency per exposure unit.

Common Exam Trap

A frequent error is to apply the variance-to-mean ratio twice. For example, at 90% confidence and 10% tolerance, the Poisson threshold is 270.6025. If $\phi=1.8$, the incorrect squared-factor calculation is $$ 270.6025(1.8)^2=876.7521. $$ The correction is $$ 270.6025(1.8)=487.0845. $$ The reason is visible in the standardization: the standard deviation is $\sqrt{\phi\mu}$, not $\phi\sqrt{\mu}$. Squaring the standardized inequality leaves one factor of $\phi$. Another practical trap is to confuse a required expected count with a whole-number count or exposure requirement; keep the model threshold unrounded until the final unit that must be integral is identified.

Check Your Understanding

Use the normal-approximation count standard. Unless stated otherwise, assume no interest rate or mortality assumption is relevant, and round final probabilities to four decimals and rates to three or four significant figures.

A count has mean $\mu$ and variance $1.2\mu$. Find the expected-claim threshold for 90% central confidence and an 8% relative tolerance. State the smallest whole-claim threshold. (Answer: 507.4 expected claims; 508 as a whole-claim threshold.)
A model has mean count 240 and variance 360. For 90% central confidence and 10% tolerance, calculate the variance-to-mean ratio and the adjusted expected-claim threshold. (Answer: $\phi=1.5$; 405.9 expected claims.)
A claim-count model has variance $2\mu$. Find the expected-claim threshold for 95% central confidence within a 12% tolerance. (Answer: 533.6 expected claims.)
A hypothetical portfolio has variance-to-mean ratio 1.25, expected frequency 0.10 per exposure unit, and a target of 95% central confidence within 5% tolerance. Calculate the adjusted expected-claim threshold and required exposure. (Answer: 1,920.8 expected claims; 19,208 exposure units.)
A hypothetical segment has mean count 480, variance 720, and frequency 0.16 per exposure unit. For 90% central confidence within 10% tolerance, determine the adjusted count threshold and the smallest whole-unit exposure requirement. (Answer: $\phi=1.5$; 405.9 expected claims; 2,537 exposure units.)

What You'll Learn