Start With the Values the Response Can Take
Choosing a response distribution begins with the outcome being modeled, not with a familiar model name. A claim count can be zero or a positive whole number; an individual claim amount is positive and measured on a continuous scale; an aggregate loss can be zero or positive. Those supports are different, so a distribution appropriate for one response may be structurally wrong for another.
The target must also match the unit of observation. A count per policy, a severity per reported claim, and an aggregate loss per policy period answer different questions. Earlier material on ratemaking targets distinguishes these outcomes and their construction. Here the focus is how their possible values guide a response-distribution choice.
The table summarizes common starting points. These are modeling choices, not automatic conclusions: the distribution must also be plausible for the data and the purpose of the model.
| Response | Possible values | Common candidate | Key feature |
|---|---|---|---|
| Claim count | Nonnegative integers | Poisson | Mean and variance are equal under the standard model |
| Claim count with extra variation | Nonnegative integers | Negative binomial | Can have variance greater than its mean |
| Positive claim amount | Positive continuous values | Gamma or inverse Gaussian | Both are continuous models on positive values, with different shapes |
| Aggregate loss per period | Zero or positive continuous values | Tweedie with power between 1 and 2 | Can represent a zero atom and a continuous positive component |
Discrete Responses: Claim Counts
A claim count is a nonnegative integer, so a continuous severity distribution cannot describe it directly. For a homogeneous Poisson count model, the number of claims \(N\) over exposure \(e\) has parameter \(\lambda e\), where \(\lambda\) is the expected count per unit exposure. The model assumes a constant rate over the exposure being modeled and independent occurrences over disjoint exposure intervals. Under this model, the mean and variance are both \(\lambda e\).
A Poisson model is a useful starting point when its assumptions are reasonable. If the observed count variation is greater than the mean, a negative binomial model is one alternative. One common parameterization uses mean \(\mu\) and size parameter \(k>0\), with variance \(\mu+\mu^2/k\). Its support remains the nonnegative integers. In this parameterization, increasing \(k\) reduces the extra variance; as \(k\) grows, the variance approaches the mean.
The negative binomial is not a remedy for every mismatch. It accommodates extra count variation relative to the mean, but it does not make a continuous outcome discrete or explain why variation is high. Data definitions, exposure differences, and important omitted risk differences should be considered when evaluating a count model.
Worked Example: Comparing Count Models at the Same Mean
A hypothetical class of policies has 2.5 policy-years of exposure. Suppose its expected claim frequency is 0.24 claims per policy-year. First calculate the probability of at least two claims under a Poisson model. Then compare it with a negative binomial model having the same mean count, \(\mu=0.6\), and size \(k=2\).
Method and exposure adjustment. The Poisson parameter for the full exposure is \(\lambda e=0.24(2.5)=0.60\). The outcome is a count, so both candidate models have nonnegative-integer support.
For the Poisson model:
The two probabilities are:
Therefore \(P(N\geq2)=1-0.5488-0.3293=0.1219\), to four decimals.
For the negative binomial, use \(P(N=0)=(k/(k+\mu))^k\). The probability of one count is the probability of zero multiplied by \(k\mu/(k+\mu)\), as follows from the ratio of consecutive probabilities in this parameterization:
Thus \(P(N\geq2)=1-0.5917-0.2731=0.1352\), to four decimals. The negative binomial assigns a larger probability to two or more claims in this example, reflecting its extra variation. The calculation is internally consistent: the probabilities of zero, one, and two or more sum to 1 under each model.
Common mistake avoided: Using 0.24 as the Poisson parameter ignores the 2.5 policy-years. The correct expected count for the modeled exposure is \(0.24(2.5)=0.60\).
Positive Continuous Responses: Claim Amounts
An individual claim severity is typically a positive amount, so its response support is continuous and restricted to values greater than zero. The Gamma and inverse Gaussian distributions are two common candidates. Both can describe right-skewed positive amounts, but their shapes differ; the choice should be assessed against the data and model purpose rather than made solely because the response is positive.
For a Gamma distribution with shape \(\alpha>0\) and scale \(\theta>0\), the mean is \(\alpha\theta\) and the variance is \(\alpha\theta^2\). For an inverse Gaussian distribution with mean \(\mu>0\) and shape parameter \(\lambda>0\), the mean is \(\mu\) and the variance is \(\mu^3/\lambda\). These parameterizations make the moments and units explicit. Other parameterizations exist, so a model specification should identify which is being used.
Worked Example: Matching Positive Severity Moments
A hypothetical insurer models individual claim severities in thousands of dollars. A Gamma candidate has shape \(\alpha=4\) and scale \(\theta=2.5\). An inverse Gaussian candidate has mean \(\mu=10\) and shape \(\lambda=40\). Calculate and compare their means and variances, and for the Gamma candidate calculate the probability that severity exceeds 10 thousand dollars.
Method. Both candidates have support above zero. Use the stated parameterization for each model; do not treat the inverse Gaussian shape parameter as a scale parameter.
For the Gamma candidate:
The units are thousands of dollars for the mean and thousands of dollars squared for the variance.
For the inverse Gaussian candidate:
Thus the two candidates have the same mean and variance, but this does not make their distributions identical: their shapes and tail probabilities can differ.
For a Gamma distribution with shape 4 and scale 2.5, \(X>10\) is equivalent to a unit-scale Gamma variable with shape 4 exceeding 4. The survival probability for this integer shape is \(e^{-4}\left(1+4+4^2/2!+4^3/3!\right)\). Substitution gives:
The final probability is 0.4335 to four decimals. The outcome remains a positive continuous severity; matching the first two moments alone is not a reason to ignore distributional shape.
Common mistake avoided: A variance of 25 in this calculation is in squared thousands of dollars, not dollars squared. Converting the mean but not the variance would create an inconsistent comparison.
Aggregate Losses: Zero and Positive Outcomes
Aggregate loss per policy period combines the count and the amounts of claims. If there are no claims, the aggregate is zero; if one or more claims occur, the aggregate is positive. A Tweedie distribution with power parameter between 1 and 2 is a candidate for this kind of response because it has a point mass at zero and a continuous positive component.
One way to understand this structure is through a compound Poisson-Gamma construction: let the claim count be Poisson, and, conditional on the count, let claim severities be independent and identically distributed positive Gamma variables. The aggregate is the sum of those severities, with a sum of zero when the count is zero. This construction helps explain the support; a fitted Tweedie model need not be estimated by separately fitting those count and severity pieces.
Why the Compound Aggregate Has Its Moments
The compound construction also gives a short justification for the aggregate mean and variance. Suppose \(N\) is Poisson with mean \(\lambda\), and severities \(X_1,X_2,\ldots\) are independent of \(N\), independent of one another, and have mean \(m\) and variance \(s^2\). Define \(S=\sum_{j=1}^{N}X_j\), with \(S=0\) when \(N=0\).
Given \(N\), the sum has conditional mean \(E[S\mid N]=Nm\) and conditional variance \(\operatorname{Var}(S\mid N)=Ns^2\), by independence of the severities.
\(E[S]=E[E[S\mid N]]=E[Nm]=\lambda m\), because a Poisson variable with parameter \(\lambda\) has mean \(\lambda\).
\(\operatorname{Var}(S)=E[\operatorname{Var}(S\mid N)]+\operatorname{Var}(E[S\mid N])=\lambda s^2+m^2\lambda=\lambda(s^2+m^2)\).
Since \(s^2+m^2=E[X^2]\), the aggregate variance is \(\lambda E[X^2]\). This result depends on the stated independence and Poisson-count assumptions; it should not be carried over unchanged to a different count process or dependent severities.
Worked Example: A Compound Aggregate with a Zero Atom
For a hypothetical policy period, suppose the claim count is Poisson with mean \(\lambda=0.4\). Conditional on a claim, the severity in thousands of dollars follows a Gamma distribution with shape 2 and scale 3. Find the aggregate mean, variance, standard deviation, and probability of zero aggregate loss. Explain why this response differs from an individual severity response.
Method. The count is Poisson and severities are positive, independent, and identically distributed. For the Gamma severity, \(m=2(3)=6\) thousand dollars and \(s^2=2(3^2)=18\) thousand-dollars squared. Its second moment is \(E[X^2]=s^2+m^2=18+36=54\).
Using the compound aggregate results:
The mean aggregate loss is 2.4 thousand dollars. Its variance is:
The standard deviation is \(\sqrt{21.6}=4.6476\) thousand dollars, rounded to four decimals. Because every severity is positive, aggregate loss is zero exactly when the count is zero:
The probability is 0.6703 to four decimals. The aggregate can therefore be zero or positive, while an individual severity in this model is always positive. The result checks against the construction: the positive aggregate mean is the expected count, 0.4, multiplied by the average severity, 6 thousand dollars.
Common mistake avoided: Treating the severity mean of 6 thousand dollars as the aggregate mean ignores the possibility of no claim and the possibility of multiple claims. The aggregate mean is \(0.4(6)=2.4\) thousand dollars.
Selection Checks and Common Traps
Before fitting or interpreting a response model, confirm that its support and assumptions line up with the target. A response with an impossible value under the candidate distribution signals a structural mismatch, not merely a poor parameter estimate.
- Count versus amount: Claim counts are integers; a Gamma severity is continuous. Do not use a severity model as though it directly described a claim count.
- Zero policies versus observed claims: A data set of individual claims ordinarily contains positive severities, while a policy-period aggregate can contain zero. Decide which outcome is modeled.
- Exposure: For a Poisson frequency rate per exposure unit, calculate the expected count using the actual exposure, not just the rate.
- Overdispersion: A negative binomial can represent variance exceeding the mean under its parameterization. That fact alone does not identify the cause of the excess variation.
- Parameterization and units: State Gamma shape and scale, or inverse Gaussian mean and shape, when reporting moments. A parameter named “shape” in one convention may not play the same role in another.
- Aggregate interpretation: The compound Poisson-Gamma view explains a zero mass and positive sums, but its assumptions must be stated before using its aggregate formulas.
Check Your Understanding
For each problem, identify the relevant response structure and carry out the requested calculation.
- A hypothetical account has 4 policy-years of exposure and a Poisson frequency rate of 0.15 claims per policy-year. Calculate the probability of zero claims, to four decimals. (Answer: Poisson count; \(P(N=0)=e^{-0.60}=0.5488\).)
- A negative binomial count model uses mean \(\mu=3\) and size \(k=2\), with variance \(\mu+\mu^2/k\). Calculate its variance and state whether it exceeds the mean. (Answer: Variance \(=3+9/2=7.5\), which exceeds the mean of 3.)
- A positive claim severity in thousands of dollars is modeled as Gamma with shape 3 and scale 2. Calculate its mean and variance, including units. (Answer: Mean \(=3(2)=6\) thousand dollars; variance \(=3(2^2)=12\) thousand-dollars squared.)
- A policy-period aggregate is modeled through a Poisson count with mean 0.2 and strictly positive severities. Calculate the probability of zero aggregate loss. (Answer: Zero aggregate occurs exactly when the count is zero, so \(e^{-0.2}=0.8187\).)
- A response records the total paid loss for each policy period, including periods with no claims. Explain why a positive-only Gamma severity model does not directly represent this response, and name a distributional candidate that can allow both zero and positive continuous values. (Answer: The response includes a point mass at zero, which a positive-only severity distribution does not allow; a Tweedie model with power between 1 and 2 is a candidate.)