Common Parameterization Errors In Count Models

Why Parameter Audits Matter

A count model can be named correctly and still be used incorrectly. A Poisson rate may be mistaken for a count-period mean, a negative binomial probability may be confused with its complement, or a binomial trial count may not match the number of opportunities for a claim. Each error changes the implied distribution, sometimes substantially.

A reliable audit begins with three questions: What values can the count take? What mean do the parameters imply? What variance do they imply? These checks do not replace a model's assumptions, but they often reveal a misread parameter before it affects a probability or a pricing calculation. The methods here are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Three Count Models and Their Parameter Meanings

Let $N$ denote a claim count over a stated observation period or exposure. The letters used for parameters are not enough to identify their meanings: state the model convention explicitly before substituting values.

Definition: In a Poisson model, $N\sim\operatorname{Poisson}(\lambda)$ has support $0,1,2,\ldots$, mean $E[N]=\lambda$, and variance $\operatorname{Var}(N)=\lambda$. If the frequency rate is $\rho$ claims per exposure unit and total exposure is $e$, the count-period parameter is $\lambda=\rho e$, under a constant-rate Poisson model.

The rate $\rho$ and the Poisson parameter $\lambda$ are numerically equal only when the exposure is one unit. A rate per policy-year must be multiplied by the number of policy-years in the observation period. If the rate changes over time, the simple constant-rate calculation requires modification; it should not be applied as if one rate held throughout.

Definition: Use the following negative binomial convention throughout: $N$ is the number of failures before $r$ successes, with independent trials having success probability $p$, where $0<p<1$. For integer $r$, the support is $0,1,2,\ldots$, and $$ E[N]=\frac{r(1-p)}{p}, \qquad \operatorname{Var}(N)=\frac{r(1-p)}{p^2}. $$ Here $r$ is the size (or target-success) parameter, and $p$ is the success probability. The same formulas apply for positive, not necessarily integer, size in the usual generalized negative binomial model, although the literal interpretation as a fixed number of successes then does not apply.

Some references instead specify a negative binomial by its mean $\mu$ and size $r$. Under the failure-count convention above, the corresponding success probability is $p=r/(r+\mu)$. This conversion is useful, but only after confirming which negative binomial convention is being used.

Definition: In a binomial model, $N\sim\operatorname{Binomial}(n,p)$ counts successes in $n$ independent trials, each with success probability $p$. Its support is $0,1,\ldots,n$, mean $E[N]=np$, and variance $\operatorname{Var}(N)=np(1-p)$. The parameter $n$ is the number of trials, not the number of successes observed.

The binomial model presumes a fixed number of eligible trials and the same success probability on each trial, with trials independent. A count greater than $n$ is impossible under this model. A Poisson or negative binomial count, by contrast, has no finite upper bound.

A Short Justification for the Negative Binomial Moments

The failure-count interpretation gives a direct explanation of the negative binomial mean and variance. For one success, let $G$ be the number of failures before that success. The probability of exactly $k$ failures is $P(G=k)=p(1-p)^k$, for $k=0,1,2,\ldots$. This is a geometric count of failures, with

E[G]=\frac{1-p}{p}, \qquad \operatorname{Var}(G)=\frac{1-p}{p^2}.

For an integer $r$, the total failures before $r$ successes can be written as the sum of $r$ independent copies of $G$, one for each success. Linearity of expectation gives the mean, and independence makes the variances add:

E[N]=\sum_{j=1}^{r}E[G_j] =r\frac{1-p}{p}, \qquad \operatorname{Var}(N)=\sum_{j=1}^{r}\operatorname{Var}(G_j) =r\frac{1-p}{p^2}.

This also explains why swapping $p$ and $1-p$ is not a harmless change of notation: it changes the failure probability on each trial and therefore changes both moments.

Worked Examples: Audit the Parameters Before Calculating

Worked Example: Convert a Poisson Rate to a Period Mean

A hypothetical book of commercial equipment is modeled with a constant Poisson claim rate of $0.08$ claims per equipment-year. A policy covers 2.5 equipment-years during the observation period. Find the Poisson parameter, mean, variance, and probability of no claims during the period.

Method. Convert the rate to the total exposure-period mean first. This avoids treating a per-unit rate as the mean for the entire period. Assume the rate is constant over the stated exposure and the model is Poisson.

The total exposure is $e=2.5$ equipment-years, so

\lambda=\rho e=(0.08)(2.5)=0.20.

Thus $N\sim\operatorname{Poisson}(0.20)$, with $E[N]=0.20$ claims and $\operatorname{Var}(N)=0.20$ claims squared. The zero-claim probability is

P(N=0)=e^{-0.20}\frac{(0.20)^0}{0!} =e^{-0.20} =0.8187.

The check is consistent: the exposure is 2.5 times one equipment-year, so the period mean $0.20$ is 2.5 times the rate $0.08$. Common mistake avoided: using $0.08$ as the period mean would give $e^{-0.08}=0.9231$ for no claims, which answers a one-equipment-year question, not the stated 2.5-equipment-year question. Probabilities are rounded to four decimals.

Worked Example: Detect a Swapped Negative Binomial Probability

A hypothetical count is modeled as the number of failures before four successes, with success probability $p=0.80$ on each independent trial. Find its mean, variance, and probability of zero failures. A proposed calculation instead uses $0.20$ as $p$. Diagnose the error.

Method. Keep the specified convention in view: $p=0.80$ is the success probability, so the failure probability is $1-p=0.20$. With $r=4$,

E[N]=\frac{4(0.20)}{0.80}=1, \qquad \operatorname{Var}(N)=\frac{4(0.20)}{(0.80)^2} =\frac{0.80}{0.64}=1.25.

Zero failures means that the first four trials are all successes. Therefore

$$ P(N=0)=p^r=(0.80)^4=0.4096. $$

If $0.20$ is incorrectly substituted as the success probability, the resulting mean is $4(0.80)/0.20=16$ and the variance is $4(0.80)/(0.20)^2=80$. Those are the moments of a different model: one with success probability $0.20$, not $0.80$. The corrected values are mean $1$, variance $1.25$, and zero-count probability $0.4096$. The support remains $0,1,2,\ldots$ in either model, so support alone will not detect this swap; checking parameter meaning and moments is essential.

Common mistake avoided: treating the failure probability as the parameter $p$ even though the stated convention defines $p$ as the success probability. Probabilities are rounded to four decimals where needed.

Worked Example: Check the Binomial Trial Count

A hypothetical insurer models claims from 12 independent device-policy periods. Each period has probability $0.05$ of at least one claim, and the model counts periods with a claim. Find the support, mean, variance, and probability of no claim-bearing periods. Explain why using $n=3$ would be inappropriate if the 12 periods consist of four devices observed for three periods each.

Method. One trial is one device-policy period, since the outcome being counted is whether that period has a claim. There are 12 such trials, each with probability $p=0.05$. Under the stated independence and equal-probability assumptions, $N\sim\operatorname{Binomial}(12,0.05)$.

\text{Support: }0,1,\ldots,12, \qquad E[N]=np=(12)(0.05)=0.60,

\operatorname{Var}(N)=np(1-p) =(12)(0.05)(0.95)=0.57.

The probability of no claim-bearing periods is

P(N=0)=(1-p)^{12}=(0.95)^{12}=0.5404.

The proposed $n=3$ would count only three trials and would omit the distinction between the four devices. The number of periods is four devices times three periods per device, or 12 trials. The support provides a further check: this model can produce counts as high as 12, not merely 3. The calculated variance is below the mean, as it must be for this binomial model since $np(1-p)<np$ when $0<p<1$. Common mistake avoided: using the number of time periods as the trial count when each device-period is a separate eligible opportunity. Probabilities are rounded to four decimals.

A Practical Parameter-Audit Sequence

When a count-model parameterization is presented in a problem, perform the checks in a fixed order. The order matters: identifying the support and unit first can expose an incorrect interpretation before any longer calculation begins.

Name the count and its observation unit.
State whether $N$ counts claims per policy, successes per trial set, or failures before a target number of successes. Write down the exposure period or number of opportunities.

Match each parameter to its definition.
For Poisson, distinguish a rate per exposure unit from $\lambda$, the mean count over the period. For negative binomial, record whether $p$ is success or failure probability and what $r$ counts. For binomial, verify that $n$ is the number of eligible trials.

Check the support.
Poisson and negative binomial counts have no finite upper bound. A binomial count cannot exceed $n$. Reject any interpretation that assigns positive probability to an impossible count.

Compute the implied mean and variance.
Compare them with any stated targets or with the qualitative behavior required by the model. For example, Poisson variance equals its mean; binomial variance is $np(1-p)$, less than its mean when $0<p<1$.

Calculate probabilities only after the audit.
Use the parameterization that matches the problem statement, and retain the same exposure unit throughout the calculation.

Model	Support	Mean	Variance	High-value parameter check
Poisson $(\lambda)$	$0,1,2,\ldots$	$\lambda$	$\lambda$	Is $\lambda$ the period mean, rather than a rate per unit?
Negative binomial $(r,p)$, failures before $r$ successes	$0,1,2,\ldots$	$r(1-p)/p$	$r(1-p)/p^2$	Is $p$ the success probability in the stated convention?
Binomial $(n,p)$	$0,1,\ldots,n$	$np$	$np(1-p)$	Does $n$ equal the number of independent eligible trials?

Common Exam Traps

Using a frequency rate as a period mean. A rate of $0.08$ per exposure unit and an exposure of 2.5 units imply $\lambda=0.20$, not $0.08$.
Silently changing negative binomial convention. Some sources define a parameter as success probability; others may parameterize using a failure probability or the mean. State the convention and convert before using formulas.
Assuming the size parameter is the mean. In the failure-count convention, $r$ is the number of successes targeted, while the mean is $r(1-p)/p$. They are generally not equal.
Setting binomial $n$ equal to a convenient total. Count the actual eligible trials that can each produce the event being counted. If there are 12 such trials, a binomial model with $n=3$ has the wrong support.
Using the mean alone as a model check. Different parameter values can produce the same mean but different variances. Check both moments and the support whenever the parameters are uncertain.

Key takeaway: A valid parameterization must match the model's stated meaning, support, mean, and variance. Treat these as linked checks: a mismatch is evidence to revisit the exposure unit, the probability convention, or the number of trials before proceeding.

Check Your Understanding

For each problem, identify the parameter interpretation before calculating. Answers are rounded to four decimals for probabilities when needed.

A hypothetical portfolio has a Poisson rate of $0.04$ claims per vehicle-year and exposure of six vehicle-years. Find the period parameter and probability of zero claims. (Answer: $\lambda=0.24$, $P(N=0)=0.7866$)
A count is the number of failures before three successes, with success probability $p=0.75$. Find its mean, variance, and support. (Answer: mean $1$, variance $1.3333$, support $0,1,2,\ldots$)
A hypothetical group has 20 independent trials, each with success probability $0.03$. For the binomial success count, find the mean, variance, and probability of zero successes. (Answer: mean $0.60$, variance $0.582$, $P(N=0)=0.5438$)
A negative binomial model uses the failures-before-successes convention, with size $r=2$ and mean $\mu=6$. Find the success probability and variance. (Answer: $p=0.25$, variance $24$)
A proposed binomial model has $n=8$, but the recorded count is 9 claim-bearing trials. What does this imply about the model specification? (Answer: the count is outside the binomial support $0,\ldots,8$; the stated trial count, the counted event, or the model specification must be revisited.)

Model	Support	Mean	Variance	High-value parameter check
Poisson \((\lambda)\)	\(0,1,2,\ldots\)	\(\lambda\)	\(\lambda\)	Is \(\lambda\) the period mean, rather than a rate per unit?
Negative binomial \((r,p)\), failures before \(r\) successes	\(0,1,2,\ldots\)	\(r(1-p)/p\)	\(r(1-p)/p^2\)	Is \(p\) the success probability in the stated convention?
Binomial \((n,p)\)	\(0,1,\ldots,n\)	\(np\)	\(np(1-p)\)	Does \(n\) equal the number of independent eligible trials?

What You'll Learn