Count Outcomes Versus Positive Continuous Outcomes

Start With What Each Observation Measures

A response can be an integer count, a continuous amount, or a rate calculated from a count. These may describe the same underlying activity, but they are not interchangeable modeling targets. Before selecting a Poisson or Gamma model, identify what one row represents and what the response records for that row.

For example, a service record might give the number of calls during a month, the amount paid for those calls, or calls per month of exposure. The number of calls is an integer count. The payment can be a positive amount with cents, or zero if there was no payment. Calls per month is a rate and may be fractional, even though the underlying number of calls is a count. A model’s response must match the recorded quantity, not merely the general subject of the data.

Definition: A count response records how many times an event occurs for an observational unit during a specified period or exposure. It takes values $0,1,2,\ldots$. A positive continuous response measures an amount that can vary continuously over values greater than zero. A rate is a count divided by an exposure amount; it need not be an integer and is not itself a count.

What the Poisson and Gamma Responses Require

A Poisson model is appropriate to consider when each response is a nonnegative integer count. In a Poisson regression, the count’s conditional distribution given its covariates is Poisson; its conditional mean is also its conditional variance. The model does not require every observed count to be positive: zero is part of the Poisson support.

N \mid \mathbf{x} \sim \operatorname{Poisson}(\mu), \qquad P(N=n \mid \mathbf{x})=\frac{e^{-\mu}\mu^n}{n!}, \quad n=0,1,2,\ldots, \qquad E[N\mid\mathbf{x}]=\operatorname{Var}(N\mid\mathbf{x})=\mu.

Here $\mu>0$ is the conditional mean. A Poisson model describes counts, not arbitrary nonnegative numbers. A payment of $525.40 is not a count, and a rate of 0.6 calls per month is not an observed count. A count divided by exposure can be fractional without making the underlying count fractional.

A Gamma model, by contrast, describes a positive continuous response. A common parameterization uses shape $k>0$ and scale $\theta>0$. Its mean is $k\theta$, its variance is $k\theta^2$, and its support is $y>0$. In a Gamma regression, the conditional mean is $\mu>0$, and a common dispersion convention expresses the conditional variance as $\phi\mu^2$, where $\phi>0$. The model can represent right-skewed positive amounts, but an ordinary Gamma response cannot take the value zero.

Y\mid\mathbf{x}\sim\operatorname{Gamma}(k,\theta), \qquad E[Y\mid\mathbf{x}]=k\theta, \qquad \operatorname{Var}(Y\mid\mathbf{x})=k\theta^2=\phi\mu^2, \quad \phi=\frac{1}{k}.

The Gamma distribution is continuous: the probability of any single exact value, including zero, is zero. That does not mean a Gamma model can accommodate an observed zero; a continuous distribution assigning probability zero to exactly zero cannot describe a response with a genuine point mass at zero. A dataset of positive amounts may be modeled with Gamma even if the amounts are recorded to the nearest cent, when the continuous model is a reasonable approximation. The important distinction is the nature of the response, not whether the data file uses decimal places.

Why the Poisson Mean and Variance Match

The Poisson mean-variance relationship follows directly from its probability mass function. This is useful when checking what a Poisson response model assumes; it is not a claim that every finite sample must have exactly equal sample mean and sample variance.

Compute the mean.
For $N\sim\operatorname{Poisson}(\mu)$, sum $nP(N=n)$ over positive integers. Re-indexing the sum gives $E[N]=\mu e^{-\mu}\sum_{j=0}^{\infty}\mu^j/j!=\mu$.

Compute the second factorial moment.
Using $n(n-1)/n!=1/(n-2)!$ for $n\geq2$, the same re-indexing gives $E[N(N-1)]=\mu^2$.

Obtain the variance.
Since $N^2=N(N-1)+N$, $E[N^2]=\mu^2+\mu$. Therefore $\operatorname{Var}(N)=E[N^2]-(E[N])^2=\mu^2+\mu-\mu^2=\mu$.

Worked Examples

Worked Example: Counting Roadside Service Calls

A hypothetical roadside assistance program records the number $N$ of calls for each policy-year. Suppose the model for a particular policy group is $N\sim\operatorname{Poisson}(0.8)$. Decide whether the response is a count, and calculate the probability of no calls, the probability of at least one call, and the probability of exactly two calls.

Method. Check the response’s support first. Calls per policy-year are nonnegative integers, so a Poisson response is structurally possible. The stated model has $\mu=0.8$, so its mean and variance are both 0.8 calls squared or calls, respectively, in the usual count interpretation.

For no calls, use the Poisson probability formula with $n=0$:

P(N=0)=\frac{e^{-0.8}(0.8)^0}{0!}=e^{-0.8}\approx0.4493.

The event “at least one call” is the complement of zero calls:

P(N\geq1)=1-P(N=0)=1-0.4493=0.5507.

For exactly two calls:

P(N=2)=\frac{e^{-0.8}(0.8)^2}{2!} \approx\frac{(0.449329)(0.64)}{2} \approx0.1438.

Check. The probabilities are between zero and one. Also, $P(N\geq1)$ is greater than $P(N=2)$, as it must be because the first event includes all positive counts. Rounding probabilities to four decimals gives 0.4493, 0.5507, and 0.1438.

Common mistake avoided. The mean 0.8 is not a possible observed number of calls for one policy-year. It is the model’s expected count; an individual observed count remains an integer.

Worked Example: Modeling Positive Repair Amounts

A hypothetical repair program records one positive repair amount $Y$, in dollars, for each completed repair. Consider a Gamma model with shape $k=4$ and scale $\theta=\$300$. Calculate its mean, variance, standard deviation, and probability of an exact zero repair amount under this model.

Method. The response is a positive continuous amount, so it has the right type of support for a Gamma model, provided the Gamma distribution is otherwise a reasonable description. Apply the shape-scale parameterization:

E[Y]=k\theta=4(\$300)=\$1{,}200.

The variance is:

\operatorname{Var}(Y)=k\theta^2 =4(\$300)^2 =4(\$90{,}000) =360{,}000\ \text{dollars}^2.

Therefore the standard deviation is the square root of the variance:

\operatorname{SD}(Y)=\sqrt{360{,}000}=\$600.

Because the Gamma response is continuous on positive amounts, the probability of the single exact value zero is $P(Y=0)=0$. This is not a model for a repair record that can be zero when no repair occurs. Here the problem specifies completed repairs with positive recorded amounts, which avoids that zero-response issue.

Check. The coefficient of variation is $600/1{,}200=0.5$, consistent with $1/\sqrt{k}=1/2$ for this parameterization. The variance has squared-dollar units, while the standard deviation and mean have dollar units. Round monetary quantities to the cent if reporting cents; the values here are exact.

Common mistake avoided. A skewed response is not automatically Gamma. The response must be positive, and the Gamma mean and variance relationship must also be plausible for the application.

Worked Example: Separating Counts From Zero-Inclusive Payments

A hypothetical assistance program records, for each of six exposure units, both the number of service events and the total payment:

\begin{array}{c|rrrrrr} \text{Exposure unit}&1&2&3&4&5&6\\ \text{Event count}&0&0&1&2&0&1\\ \text{Total payment (\$)}&0&0&480.25&1{,}310.60&0&925.40 \end{array}

Determine which response could be considered for a Poisson model, whether the total payment can be modeled by an ordinary Gamma response, and the mean payment across all exposure units and across units with positive payments.

Method. Treat the count and payment as separate responses. The event counts are integers, so they meet the support requirement for a Poisson response. Their total is $0+0+1+2+0+1=4$, giving a sample mean of $4/6=0.6667$ events per exposure unit. This support check alone does not establish that a Poisson distribution fits; its conditional mean-variance assumption must also be appropriate.

The total payments include three exact zeros and three positive amounts with cents. Thus they cannot be modeled directly with an ordinary Gamma response, which has strictly positive support and no point mass at zero. The payment values are amounts, not counts, so they are not Poisson responses either.

The sum of all payments is:

\$480.25+\$1{,}310.60+\$925.40=\$2{,}716.25.

The mean payment over all six exposure units is $\$2{,}716.25/6=\$452.7083\ldots$, or $\$452.71$ to the nearest cent. Among the three units with positive payments, the mean is $\$2{,}716.25/3=\$905.4166\ldots$, or $\$905.42$ to the nearest cent.

Check. The mean across all units is lower than the mean among positive-payment units because the first calculation includes three zeros. A Gamma model could be considered for positive amounts only if that is the response being modeled and the distributional assumptions are reasonable; it does not thereby model the full zero-inclusive response.

Common mistake avoided. Applying a Gamma model to the three positive payments and silently treating the three zeros as if they were positive would change the response being modeled. Keep the count, the positive amount, and the zero-inclusive total distinct.

Common Traps in Response Selection

Calling every nonnegative response a count. A payment, balance, or duration may be nonnegative but is not an event count. Poisson support is restricted to integers.
Using Gamma because an amount is skewed. Skewness alone does not overcome the Gamma response’s strict positivity requirement. Genuine zeros need explicit treatment rather than being ignored or changed to small positive values without justification.
Confusing an exposure rate with a count. If 9 events occur over 12 exposure units, the rate is $9/12=0.75$, but the observed count is 9. When the response is the event count, exposure can be represented in the model; dividing first does not turn the rate into an integer response.
Assuming that integer support proves a Poisson model is suitable. Counts meet the basic response-type requirement, but the Poisson model also imposes a mean-variance relationship and other distributional assumptions. The response’s data type is necessary to check, not sufficient to establish model fit.
Mixing up an average and an individual observation. An average number of events across many units can be fractional. The count recorded for each individual unit is still an integer.

Key takeaway: Identify the response and its support before choosing a model. Poisson responses are nonnegative integer counts, including zero; ordinary Gamma responses are strictly positive continuous amounts. Fractional rates and zero-inclusive amounts require particular care: neither becomes a suitable Poisson or Gamma response just by being nonnegative.

Check Your Understanding

For each problem, identify the response type before carrying out the requested calculation or model choice.

A hypothetical monitoring system records the number $N$ of alerts per device-day, modeled as $\operatorname{Poisson}(0.7)$. Calculate $P(N\geq2)$ to four decimals. (Answer: 0.1558)
A positive continuous duration is modeled as Gamma with shape $k=5$ and scale $\theta=80$ minutes. Find its mean and variance. (Answer: Mean $400$ minutes; variance $32{,}000$ minutes squared.)
A dataset records total reimbursements per member-month, including exact zeros and positive dollar amounts with cents. Is an ordinary Poisson response or an ordinary Gamma response suitable for the reimbursement amount as recorded? State why. (Answer: Neither; the amounts are not integer counts, and the exact zeros violate the strictly positive Gamma support.)
A hypothetical call center observes 15 calls across 20 device-months. What is the observed call rate per device-month, and which quantity is the integer-count response if modeling call frequency? (Answer: Rate $15/20=0.75$ calls per device-month; the integer-count response is 15 calls across the stated exposure, not the fractional rate.)
Eight hypothetical exposure units have counts $0,0,1,1,2,2,3,3$. Calculate the sample mean and the variance using divisor $n$, then explain whether the difference alone establishes that a Poisson model is invalid. (Answer: Mean $1.5$; variance $1.25$; no—the sample quantities need not be exactly equal even under a Poisson model, so the difference alone does not establish invalidity.)

What You'll Learn