Why Claim Count Distribution Choice Matters
A claim count model describes the probability of observing each possible number of claims over a defined exposure period. Choosing a model affects quantities such as the probability of no claims, the chance of several claims, and the tail of the count distribution. The first question is not simply which formula best matches a calculated variance: it is what structure in the insured risk and its exposure makes the model plausible?
For example, a Poisson model can represent claims that occur as events over exposure, with multiple claims possible. A binomial model is natural when there is a fixed number of distinct opportunities, each of which can produce at most one claim. A negative binomial model can represent counts with more variability than a Poisson model, including variability arising from differences in underlying claim rates. These models are not interchangeable just because they can be adjusted to have the same mean.
Three Common Claim Count Models
The three distributions below are useful starting points for selection. Their mean and variance describe different dispersion patterns: equidispersion means variance equals the mean; underdispersion means variance is less than the mean; and overdispersion means variance is greater than the mean.
| Model | Probability model | Mean and variance | Typical structural rationale |
|---|---|---|---|
| Poisson | \(P(N=k)=e^{-\lambda}\lambda^k/k!\) | \(E[N]=\operatorname{Var}(N)=\lambda\) | Events occur independently at a stable rate over exposure; multiple claims are possible. |
| Binomial | \(P(N=k)=\binom{n}{k}p^k(1-p)^{n-k}\) | \(E[N]=np\), \(\operatorname{Var}(N)=np(1-p)\) | There are \(n\) fixed, independent opportunities, each with the same probability \(p\) of one claim. |
| Negative binomial | \(P(N=k)=\binom{k+r-1}{k}p^r(1-p)^k\) | \(E[N]=r(1-p)/p\), \(\operatorname{Var}(N)=r(1-p)/p^2\) | Counts are more variable than a Poisson count; rate heterogeneity is one possible explanation. |
For the negative binomial parameterization in the table, \(r>0\), \(0<p<1\), and \(N\) counts failures before \(r\) successes when \(r\) is an integer. The same probability model also has a standard extension to positive noninteger \(r\). Its variance can be written as \(\operatorname{Var}(N)=E[N]+E[N]^2/r\), so it exceeds its mean for finite positive \(r\). This tutorial uses the parameterization above throughout.
In practice, a Poisson count over exposure \(e\) often has parameter \(\lambda=e\theta\), where \(\theta\) is a constant claim rate per exposure unit. This requires the assumed rate to be appropriate over the period and exposure being modeled. Combining independent Poisson counts gives another Poisson count with parameter equal to the sum of their parameters. Those properties do not mean that a mixed group with differing rates must have Poisson variation.
Use Exposure Structure Before the Mean-Variance Pattern
A fixed number of genuine claim opportunities is a strong reason to consider the binomial model. Examples might include a fixed number of separately insured items, each capable of producing at most one covered event during the period. In contrast, the number of opportunities for claims may not be naturally fixed for a policy with repeated incidents, so a Poisson or another unbounded count model may be more appropriate. A policy count alone does not make a binomial model suitable.
The binomial model assumes independent opportunities with the same claim probability. If probabilities differ across opportunities, the total is generally not binomial; if outcomes are dependent, the binomial variance formula may not apply. The Poisson model also has assumptions: under its standard interpretation, events occur independently and the rate is stable per unit of exposure. A negative binomial model is a candidate when counts exhibit overdispersion, but overdispersion can also signal omitted differences in exposure or risk, dependence, or a change in the data-generating process.
Worked Example: Comparing Poisson and Binomial Zero-Claim Probabilities
A hypothetical portfolio segment has eight equal, independent policy opportunities, each with probability \(0.10\) of exactly one claim during the period. Compare the probability of zero claims under the binomial model with the probability under a Poisson model having the same mean. State which model follows the stated opportunity structure and quantify the difference.
Method and model choice. The stated fixed eight opportunities, each producing at most one claim, support \(N\sim\operatorname{Binomial}(8,0.10)\). Its mean is \(np=8(0.10)=0.8\). To compare distributions at the same mean, use a Poisson model with \(\lambda=0.8\).
Binomial calculation. Zero claims means all eight opportunities produce no claim: \(P(N=0)=(1-0.10)^8=0.9^8=0.43046721\), which is \(0.4305\) to four decimal places.
Poisson calculation. For a Poisson count with \(\lambda=0.8\), \(P(N=0)=e^{-0.8}=0.449328964\ldots\), which is \(0.4493\) to four decimal places.
The binomial probability is lower. Subtracting the unrounded values gives \(e^{-0.8}-0.9^8=0.018861754\ldots\), or \(0.0189\) to four decimal places. Thus, despite the same mean, the models assign different probabilities to zero claims. The binomial is the structurally supported model here; the Poisson result is a comparison, not an equally faithful description of the fixed opportunities.
Check. The binomial count cannot exceed eight, whereas a Poisson count has no finite upper bound. That difference is consistent with the different assumptions. The common mistake avoided is treating equal means as evidence that the models give the same count probabilities.
What the Mean and Variance Can Tell You
For independent, identically distributed observations of counts under one common model, compare the observed count variation with the model’s expected variation. A Poisson model implies variance equal to mean. A binomial model implies variance below the mean when \(0<p<1\), since \(np(1-p)<np\). The negative binomial model implies variance above the mean under the parameterization given here. These patterns are useful diagnostics, not automatic proof of a distribution.
For observed data, the sample variance is affected by sample size and by differences among the risks or periods being combined. A sample variance above the sample mean is a reason to investigate overdispersion; it does not establish that counts are negative binomial. Likewise, a sample variance below the mean is not by itself evidence for a binomial model unless the fixed-opportunity assumptions are credible. If exposures differ, counts should not be compared as if they represented identical periods and risk units without considering that difference.
Worked Example: Reading an Observed Mean-Variance Pattern
For a hypothetical set of eight equal-exposure policy groups, observed claim counts are \(0,0,1,1,2,2,4,4\). Calculate the sample mean and sample variance, then describe what the pattern suggests about a Poisson model.
Step 1: Calculate the mean. The total count is \(0+0+1+1+2+2+4+4=14\). With eight groups, the sample mean is \(14/8=1.75\).
Step 2: Calculate the sample variance. The squared deviations from \(1.75\) are \(3.0625,3.0625,0.5625,0.5625,0.0625,0.0625,5.0625,5.0625\). Their sum is \(17.5\). Using the sample variance denominator \(8-1\), the sample variance is \(17.5/7=2.5\).
Step 3: Interpret the comparison. The variance-to-mean ratio is \(2.5/1.75=1.428571\ldots\), approximately \(1.43\). The sample variance exceeds the sample mean, unlike the Poisson model’s population relationship \(\operatorname{Var}(N)=E[N]\). This pattern suggests overdispersion and makes a negative binomial model worth considering, but it does not establish the cause or prove that model is suitable.
Check. The mean is \(1.75\) and the sample variance is \(2.5\), so the direction of the comparison is overdispersion. The common mistake avoided is concluding that the negative binomial model is certainly correct from a single small sample; exposure comparability, dependence, and risk differences still matter.
When a Negative Binomial Model Is Plausible
The negative binomial is a natural candidate when counts vary more than a Poisson model allows and a plausible source of that extra variation exists. For example, policy groups may have different underlying claim rates even when each group’s exposure is measured in the same units. A Poisson mixture with a random rate produces a negative binomial distribution under a gamma mixing assumption, as established in the earlier material on deriving the negative binomial as a Poisson mixture. The mixture provides a structural interpretation: some groups have higher rates than others, and that variation increases the variance of the combined counts.
The negative binomial is not a universal fix for any large sample variance. Excess variability might instead reflect clustering, incorrect exposure measurement, inconsistent observation periods, or operational changes. A model should explain the source and shape of the counts as well as accommodate a variance-to-mean pattern. Conversely, a mean-variance ratio close to one does not prove that a Poisson model’s independence and stable-rate assumptions hold.
Worked Example: Checking a Negative Binomial Parameterization
A hypothetical risk group is modeled with a negative binomial count \(N\), where \(r=4\) and \(p=0.80\) in the failures-before-\(r\)-successes parameterization. Calculate its mean, variance, and probability of zero claims. Compare the zero-claim probability with a Poisson model of the same mean.
Step 1: Calculate the mean. \(E[N]=r(1-p)/p=4(0.20)/0.80=1\).
Step 2: Calculate the variance. \(\operatorname{Var}(N)=r(1-p)/p^2=4(0.20)/(0.80)^2=0.8/0.64=1.25\). The variance exceeds the mean, as expected for a finite-\(r\) negative binomial model.
Step 3: Calculate the zero probability. From the negative binomial probability formula, \(P(N=0)=\binom{3}{0}(0.80)^4(0.20)^0=0.4096\).
Step 4: Compare with Poisson. A Poisson model with the same mean has \(\lambda=1\), so \(P(N=0)=e^{-1}=0.367879\ldots\), or \(0.3679\) to four decimal places. The negative binomial zero probability is \(0.4096\), greater by \(0.041721\ldots\), or \(0.0417\) to four decimal places.
Check. The negative binomial has both greater variance and greater probability of zero claims in this calculation. The common mistake avoided is using \(p\) as though it were the mean; under this parameterization the mean is \(r(1-p)/p\).
A Fixed Number of Claim Opportunities
The binomial model’s upper bound is sometimes as important as its mean and variance. If each of \(n\) distinct items can generate at most one claim in the period, the total count cannot exceed \(n\). A Poisson or negative binomial model has no such bound. When the maximum possible count is material, selecting an unbounded model merely because its mean is convenient can assign probability to impossible outcomes.
Worked Example: Selecting a Model for Insured Items
A hypothetical commercial policy covers twelve separately monitored devices for one year. Each device can produce at most one covered breakdown claim during the year, and the probability of a claim for each device is \(0.05\), independently of the others. Find the mean, variance, and zero-claim probability for the appropriate count model. Compare the zero probability with a Poisson model of equal mean.
Step 1: Select the model. There are twelve fixed opportunities, each with at most one claim, and the outcomes are independent with a common probability. Thus \(N\sim\operatorname{Binomial}(12,0.05)\).
Step 2: Calculate mean and variance. \(E[N]=np=12(0.05)=0.60\). The variance is \(np(1-p)=12(0.05)(0.95)=0.57\).
Step 3: Calculate zero claims. \(P(N=0)=(1-0.05)^{12}=0.95^{12}=0.540360087\ldots\), or \(0.5404\) to four decimal places.
Step 4: Compare with Poisson. The equal-mean Poisson parameter is \(\lambda=0.60\). Its zero-claim probability is \(e^{-0.60}=0.548811636\ldots\), or \(0.5488\) to four decimal places. The Poisson probability exceeds the binomial probability by \(e^{-0.60}-0.95^{12}=0.008451549\ldots\), or \(0.0085\) to four decimal places.
Check. The binomial count is bounded above by twelve; the Poisson is not. The binomial is supported by the stated structure. The common mistake avoided is overlooking the one-claim-per-device limit and selecting a count model that allows impossible counts.
Common Selection Errors
- Choosing from the mean alone. Different distributions can share a mean and still produce different probabilities, as the zero-claim comparison shows.
- Calling every overdispersed count negative binomial. A large variance relative to the mean is a clue to investigate, not a complete model justification.
- Using a binomial model because the exposure is a number of policies. The relevant question is whether each policy represents one fixed, independent opportunity with at most one claim.
- Ignoring differences in exposure or risk. Counts from unlike exposure periods or risk groups may have extra variation that is not explained by a single homogeneous count model.
- Rounding before comparing probabilities. For example, \(0.4493-0.4305=0.0188\) subtracts rounded displays. The exact difference \(e^{-0.8}-0.9^8=0.018861754\ldots\) rounds to \(0.0189\), not \(0.0188\), at four decimal places.
Check Your Understanding
For each problem, state the model supported by the stated assumptions or interpret the dispersion pattern. Round probabilities to four decimal places.
- A hypothetical count is modeled as Poisson with parameter \(\lambda=0.7\). Calculate its mean, variance, and probability of zero claims. (Answer: mean \(0.7\), variance \(0.7\), \(P(N=0)=0.4966\))
- Each of ten independent scheduled inspections can produce at most one claim, with claim probability \(0.03\) at each inspection. Identify the model and calculate its mean and variance. (Answer: \(\operatorname{Binomial}(10,0.03)\), mean \(0.3\), variance \(0.291\))
- Six equal-exposure groups have observed counts \(0,1,1,2,3,3\). Calculate the sample mean and sample variance using denominator \(n-1\); what dispersion pattern do they show relative to a Poisson model? (Answer: mean \(1.6667\), sample variance \(1.4667\), underdispersed relative to the Poisson mean-variance relationship)
- For a negative binomial count with \(r=3\), \(p=0.75\), using the parameterization in this tutorial, calculate the mean and variance. (Answer: mean \(1\), variance \(1.3333\))
- A fixed set of six independent devices can each produce at most one claim, with probability \(0.20\) per device. Compare the binomial and equal-mean Poisson probabilities of zero claims, and give the difference to four decimal places. (Answer: binomial \(P(N=0)=0.2621\), Poisson \(P(N=0)=0.3012\); the Poisson probability is higher by \(0.0391\))