Choosing Between Poisson And Gamma Responses

Why the Response Type Comes First

Poisson and Gamma regression can both describe outcomes whose means depend on explanatory variables, but they describe different kinds of outcomes. A Poisson response is a count: it takes values $0,1,2,\ldots$. A Gamma response is a positive continuous amount: it takes values greater than zero, such as a repair cost or time. Choosing between them is not simply a matter of picking whichever model has a more convenient variance formula. The response’s possible values must be consistent with the distribution.

The variance relationship is the next check. Under a Poisson model, the conditional variance equals the conditional mean. Under a Gamma model with fixed dispersion, the conditional variance is proportional to the square of the conditional mean. These assumptions have different implications as the mean changes. They help assess whether a candidate model is plausible, but they do not override an outcome’s type.

Definition: A Poisson regression models a count response $Y_i$, conditional on predictors, as $Y_i\sim\operatorname{Poisson}(\mu_i)$, where $\mu_i=E[Y_i\mid\mathbf{x}_i]>0$ and $\operatorname{Var}(Y_i\mid\mathbf{x}_i)=\mu_i$. A Gamma regression models a strictly positive continuous response $Y_i$, conditional on predictors, with mean $\mu_i>0$ and variance $\operatorname{Var}(Y_i\mid\mathbf{x}_i)=\phi\mu_i^2$, where $\phi>0$ is the dispersion parameter.

In either model, a link function relates the conditional mean to predictors. For example, a log link sets $\log(\mu_i)=\mathbf{x}_i^\mathsf{T}\boldsymbol{\beta}$. The link describes how the mean changes; it does not change the response distribution. A log link does not make a continuous amount a count or make a count positive-only.

These distributional assumptions apply conditional on the predictors. A marginal data set can show more variability than the conditional model implies because observations have different means or because the model is inadequate. Therefore, comparing a raw variance with a raw mean without accounting for predictors may be misleading. In the examples below, the stated means and variances are conditional or refer to groups with comparable predictors.

The Two Mean-Variance Patterns

For a Poisson random variable, the probability of observing count $y$ is determined by its mean $\mu$:

P(Y=y)=\frac{e^{-\mu}\mu^y}{y!},\qquad y=0,1,2,\ldots

Its mean and variance are both $\mu$. Thus, if the fitted conditional mean doubles, the model also predicts that the conditional variance doubles. A Poisson model assigns positive probability to zero and to every nonnegative integer, but none to a noninteger value.

A useful Gamma parameterization uses shape $\alpha>0$ and scale $\theta>0$. Its density is

f(y)=\frac{y^{\alpha-1}e^{-y/\theta}}{\Gamma(\alpha)\theta^\alpha},\qquad y>0.

With this parameterization, the mean is $\alpha\theta$ and the variance is $\alpha\theta^2$. A Gamma regression commonly expresses these quantities using the conditional mean $\mu$ and dispersion $\phi$, as derived below. The model assigns probability to positive continuous values, not to zero or negative values.

Why Gamma Variance Is Proportional to the Squared Mean

The Gamma mean-variance relationship follows directly from its shape-scale parameterization. Suppose the shape $\alpha$ is common across observations, while the scale can vary with the predictors. Set the scale to $\theta=\phi\mu$, where $\phi=1/\alpha$. Then the mean is $\alpha\theta=\alpha\phi\mu=\mu$. Its variance becomes

\operatorname{Var}(Y\mid\mathbf{x}) =\alpha\theta^2 =\frac{1}{\phi}(\phi\mu)^2 =\phi\mu^2.

This also gives a useful interpretation of dispersion. The conditional coefficient of variation, defined as standard deviation divided by mean, is

\frac{\sqrt{\phi\mu^2}}{\mu}=\sqrt{\phi}.

For a fixed-dispersion Gamma model, the coefficient of variation is constant as the mean changes. By contrast, the Poisson standard deviation is $\sqrt{\mu}$, so its coefficient of variation is $1/\sqrt{\mu}$. These differing patterns can help distinguish plausible models for a response of the appropriate type.

Worked Example: Selecting a Model for Event Counts

A hypothetical risk team models the number of equipment failures per site-month. For a site-month with specified predictors, the fitted Poisson mean is 2.5 failures. Calculate the model’s variance and standard deviation, and its probability of no failures. Decide whether a Gamma response model is an appropriate alternative.

Method. First check the outcome’s possible values. Then use the Poisson mean-variance relationship and probability formula. The response is a count, so its values are nonnegative integers, including zero.

Variance and standard deviation. Under the stated Poisson model, $\mu=2.5$, so

\operatorname{Var}(Y\mid\mathbf{x})=\mu=2.5,\qquad \operatorname{SD}(Y\mid\mathbf{x})=\sqrt{2.5}\approx1.5811.

Probability of zero. Substituting $y=0$ into the Poisson probability formula gives

P(Y=0\mid\mathbf{x})=\frac{e^{-2.5}(2.5)^0}{0!}=e^{-2.5}\approx0.0821.

The model therefore predicts a variance of 2.5 and about an 8.21% chance of no failures for this site-month. A Gamma response is not an appropriate replacement: a failure count is discrete and can be zero, while a Gamma response is strictly positive and continuous. Check: the calculated standard deviation squared is approximately $1.5811^2=2.5$, as required. The common mistake is to choose Gamma just because counts are variable; that does not make a continuous amount model suitable for integer counts.

Worked Example: Assessing a Gamma Model for Repair Costs

A hypothetical repair study records positive repair costs in dollars, conditional on a set of predictors. A proposed Gamma model has conditional mean $\mu=\$800$ and dispersion $\phi=0.16$. Calculate its variance, standard deviation, coefficient of variation, and equivalent shape and scale parameters.

Method. Repair cost is a positive continuous amount, so Gamma is a candidate if its conditional variability is compatible with the Gamma pattern. Apply $\operatorname{Var}(Y\mid\mathbf{x})=\phi\mu^2$, then translate to shape and scale.

Variance and standard deviation. First, $800^2=640{,}000$. Therefore,

\operatorname{Var}(Y\mid\mathbf{x})=0.16(640{,}000)=102{,}400\text{ dollars}^2, \qquad \operatorname{SD}(Y\mid\mathbf{x})=\sqrt{102{,}400}=\$320.

Coefficient of variation. The standard deviation divided by the mean is $320/800=0.4$, or 40%. This agrees with $\sqrt{\phi}=\sqrt{0.16}=0.4$.

Shape and scale. Since $\alpha=1/\phi$, the shape is $1/0.16=6.25$. Since $\theta=\phi\mu$, the scale is $0.16(800)=\$128$. Checking the parameterization, $\alpha\theta=6.25(128)=\$800$, and $\alpha\theta^2=6.25(128^2)=102{,}400$ dollars squared. The arithmetic confirms the requested mean and variance. The common mistake is to report $\phi\mu$ as the variance; it is the Gamma scale, whereas the variance is $\phi\mu^2$.

Use Variance Patterns as a Check, Not a Substitute for Outcome Type

When the outcome type makes both candidate models seem conceivable, examine the conditional variance pattern as well as the support. For counts, a Poisson model predicts variance equal to mean. For positive amounts, a fixed-dispersion Gamma model predicts variance equal to a constant times the squared mean. A single estimated variance may be noisy, so comparisons across groups with different fitted means are generally more informative than relying on one variance estimate.

The following example uses positive amounts grouped by similar predictors. It illustrates how the variance pattern can support Gamma without suggesting that empirical summaries alone prove a model is correct.

Worked Example: Comparing Variance Patterns Across Cost Groups

A hypothetical service program records strictly positive continuous costs. In three groups of similar cases, the estimated conditional means are $\$100$, $\$200$, and $\$400$; the corresponding estimated conditional variances are $1{,}000$, $4{,}000$, and $16{,}000$ dollars squared. Compare the variance-to-mean and variance-to-squared-mean ratios, and identify which model’s variance structure is more plausible.

Method. Calculate both ratios for each group. A Poisson variance pattern requires $\operatorname{Var}(Y\mid\mathbf{x})/\mu=1$; a fixed-dispersion Gamma pattern requires $\operatorname{Var}(Y\mid\mathbf{x})/\mu^2$ to be constant.

For the first group, the ratios are $1{,}000/100=10$ and $1{,}000/100^2=0.10$. For the second group, they are $4{,}000/200=20$ and $4{,}000/200^2=0.10$. For the third group, they are $16{,}000/400=40$ and $16{,}000/400^2=0.10$.

The variance-to-mean ratios are not close to the Poisson value of 1 and change across the groups. In contrast, each variance-to-squared-mean ratio equals 0.10. This is consistent with a Gamma mean-variance relationship with $\phi=0.10$. Its implied coefficient of variation is $\sqrt{0.10}\approx0.3162$. Check: for the mean of $\$400$, the Gamma variance prediction is $0.10(400^2)=16{,}000$, matching the stated estimate. Because these are positive continuous costs, Gamma is also consistent with the response support. The common mistake is to select a model from the ratios alone without checking whether its possible response values fit the outcome.

A Practical Selection Sequence and Common Trap

Identify what one response measures.
Decide whether each recorded outcome is an event count or a positive continuous amount. Keep the response definition and units explicit.

Check possible values.
Poisson permits nonnegative integer counts, including zero. Gamma permits strictly positive continuous values. A model incompatible with observed response values is not made suitable by a convenient link.

Compare conditional variance structure.
For Poisson, compare conditional variance with $\mu$. For Gamma with fixed dispersion, compare conditional variance with $\mu^2$, or assess whether the coefficient of variation is reasonably stable.

Check the modeling context.
Allow for predictors and exposure when defining the conditional mean. A poor variance match can signal that the candidate model is inadequate; it does not by itself justify changing a count into a Gamma response.

A common exam trap is to see that a count response has variance greater than its mean and conclude that Gamma regression should be used. For example, suppose a count model has conditional mean 4 and an estimated conditional variance 10. The incorrect reasoning is: “Gamma allows variance larger than the mean, so use Gamma.” This ignores that the outcome remains an integer count, possibly including zero. Gamma’s positive continuous support does not fit it. The correct conclusion is that the Poisson mean-variance assumption may not describe the counts adequately; investigate the count-model specification rather than using Gamma as a generic overdispersion correction.

The reverse error is also possible. A positive continuous amount does not become a Poisson response merely because its mean and variance happen to be numerically similar. Poisson is a distribution for counts. In addition, a data set containing both zero amounts and positive continuous amounts requires care: a single Gamma response cannot assign probability to the zeros. The response definition may need to distinguish whether a zero represents no event or a zero-valued amount, but that distinction must be made from the problem context rather than by choosing a distribution mechanically.

Key takeaway: Choose the response family first from the outcome’s support: counts point to Poisson, while strictly positive continuous amounts can point to Gamma. Then check whether the conditional variance pattern—$\mu$ for Poisson or $\phi\mu^2$ for Gamma—is plausible. A variance mismatch calls the model into question; it does not change the type of outcome.

Check Your Understanding

For each problem, choose or assess a response model using both the outcome and its variance assumptions.

A hypothetical warehouse records the number of damaged parcels per shift. For one predictor profile, a Poisson model gives $\mu=3$. What variance does it imply, and is Gamma an appropriate substitute because the counts vary? (Answer: The implied variance is 3; no, the response is an integer count, not a strictly positive continuous amount.)
A positive continuous processing time has conditional mean 50 minutes and Gamma dispersion $\phi=0.04$. Find its modeled variance and standard deviation. (Answer: Variance $=0.04(50^2)=100$ minutes squared; standard deviation $=10$ minutes.)
Positive continuous amounts have conditional means 20, 40, and 80, with estimated conditional variances 80, 320, and 1,280. Which variance pattern is more consistent, and what is the implied Gamma dispersion? (Answer: Gamma; each variance divided by the squared mean is $0.20$, so $\phi=0.20$.)
A count response has fitted conditional mean 5 and estimated conditional variance 12. Is Gamma the appropriate choice solely because the variance exceeds the mean? State the model-selection conclusion. (Answer: No. The outcome is still a count, so Gamma is not justified; the Poisson variance assumption may be inadequate and the count-model specification needs assessment.)
A hypothetical dataset records positive continuous claim severities, with conditional means 300 and 600 dollars and conditional variances 4,500 and 18,000 dollars squared. Is a common-dispersion Gamma pattern supported? Give the dispersion and coefficient of variation. (Answer: Yes; both variance-to-squared-mean ratios equal $0.05$, so $\phi=0.05$ and the coefficient of variation is $\sqrt{0.05}\approx0.2236$.)

What You'll Learn