Gamma Regression For Positive Skewed Outcomes

Why Model Positive, Right-Skewed Outcomes?

Many actuarial responses are amounts rather than counts: a payment per claim, a repair cost, or a positive benefit amount. Such outcomes cannot be negative, and their distributions may have a long right tail. A Gamma regression is one model for this setting. It describes a positive continuous response while allowing its conditional mean to depend on observed predictors.

The response type is the first decision. A count of claims is a nonnegative integer and may equal zero. A payment amount per observed claim is continuous and strictly positive under the modeling assumption. These are different random variables, even when both arise from the same portfolio. A Poisson model is designed for counts; a Gamma model is designed for positive continuous amounts. Choosing a model because both responses are “claim-related” ignores this essential distinction.

Definition: A Gamma regression models a response $Y_i>0$ whose conditional distribution given predictors $x_i$ is Gamma, with conditional mean $\mu_i=E[Y_i\mid x_i]$ related to those predictors through a link function: $g(\mu_i)=x_i^\mathsf{T}\beta$. Here $x_i$ is the predictor vector, $\beta$ is the coefficient vector, and $g$ is a specified one-to-one function on the allowed mean values.

A standard Gamma specification uses a common shape parameter $a>0$, while allowing the scale to vary with the conditional mean. In the shape-scale convention, the density for observation $i$ is

f(y_i\mid x_i)= \frac{1}{\Gamma(a)} \left(\frac{a}{\mu_i}\right)^a y_i^{a-1}\exp\left(-\frac{a y_i}{\mu_i}\right), \qquad y_i>0.

The function $\Gamma(a)$ is the gamma function. This parameterization has shape $a$ and scale $\mu_i/a$. In particular, the response is continuous and has support only on positive values. The regression part specifies how $\mu_i$ changes with $x_i$; the Gamma distribution describes the conditional response around that mean. A valid specification also requires the observations to be conditionally independent given their predictors, unless a dependence structure is explicitly added.

From the Gamma Distribution to a Regression Model

The density gives a direct justification for why $\mu_i$ is the conditional mean. For a Gamma variable with shape $a$ and scale $\theta$, its density has the form $y^{a-1}e^{-y/\theta}/(\Gamma(a)\theta^a)$. Multiplying by $y$ and integrating gives its expectation:

E[Y] =\frac{1}{\Gamma(a)\theta^a} \int_0^\infty y^a e^{-y/\theta}\,dy =\frac{\theta\,\Gamma(a+1)}{\Gamma(a)} =a\theta.

The second equality follows by substituting $z=y/\theta$, and the last uses $\Gamma(a+1)=a\Gamma(a)$. Setting $\theta=\mu_i/a$ therefore gives $E[Y_i\mid x_i]=a(\mu_i/a)=\mu_i$, as required.

The linear predictor is $x_i^\mathsf{T}\beta$, but the mean need not equal that predictor. The link function connects them. For example, if a model specifically states the log link, then $\log(\mu_i)=x_i^\mathsf{T}\beta$, so $\mu_i=\exp(x_i^\mathsf{T}\beta)>0$. The examples below use this stated link to calculate predictions. They do not imply that the log link is the only possible link or that it is automatically suitable for every data set.

A model specification should identify the response precisely, its units, the predictors, the link, and the distributional assumptions. For example, “payment amount per settled claim in dollars” is more informative than “claim cost.” It makes clear that the observations are positive amounts, not claim counts or aggregate costs over an exposure period.

Worked Example: Predicting a Positive Claim Amount

A hypothetical insurer models payment amount per settled claim. The response $Y_i$ is in dollars and is assumed strictly positive. The model uses a Gamma response with common shape $a=4$, conditional independence, and the log link $\log(\mu_i)=\beta_0+\beta_1 x_{i1}+\beta_2 x_{i2}$. The fitted coefficients are $\beta_0=\log(80)$, $\beta_1=\log(1.5)$, and $\beta_2=\log(1.25)$. For a claim with $x_{i1}=1$ and $x_{i2}=1$, calculate the fitted conditional mean and the corresponding Gamma scale.

Method. Substitute the predictor values into the linear predictor, then invert the stated log link. The response is an amount per claim, so the prediction is a mean payment amount, not a predicted number of claims.

Linear predictor. $\eta_i=\log(80)+1\log(1.5)+1\log(1.25)=\log(80\times1.5\times1.25)$. The product is $80\times1.5=120$, then $120\times1.25=150$, so $\eta_i=\log(150)$.

Conditional mean. Since $\mu_i=\exp(\eta_i)$, $\mu_i=\exp(\log(150))=150$. The fitted mean payment is therefore $150.

Distributional check. With shape $a=4$, the scale is $\theta_i=\mu_i/a=150/4=37.5$ dollars. The specified Gamma distribution has positive support, as required for this response. Its mean is $a\theta_i=4(37.5)=150$, confirming the calculation.

Common mistake avoided. Treating $\eta_i$ as the dollar prediction would report $\log(150)$, which is on the link scale, not in dollars. The correct prediction applies the inverse link.

Specify the Response Before the Predictors

The same explanatory variables can be used in models for different responses, but the response distribution must match what is observed. A record containing “three claims” is not a Gamma observation. A record containing “$600 paid for one claim” is not a Poisson count. The distinction is not merely a choice of link or a change of units: it determines the probability model and the meaning of the prediction.

For Gamma regression, observations at zero do not satisfy the stated positive-response support. If zero payments are possible because no claim occurred, a Gamma model for payment amount alone cannot represent that zero event. The analyst must define a response and modeling strategy appropriate to the data-generating process rather than silently adding a small constant or treating zero as a positive amount.

For a count model, exposure may be relevant because the expected number of events can increase with time at risk or policy-years. That does not mean an exposure offset should automatically be added to a Gamma model for an amount per claim. The unit and sampling mechanism determine what the response means. Modeling total paid amount over a period, for instance, combines the number of payments and their sizes; it is not automatically the same response as the payment amount conditional on a claim.

Worked Example: Separating Claim Frequency From Claim Severity

A hypothetical analyst has two records for each risk group: the number of claims during a year and the positive payment amounts for reported claims. One group has 12 claims over 6 policy-years, with total paid amount $3,600. The analyst wants both a claim frequency model and a model for payment amount per claim. State the appropriate response for each model, calculate the observed rate and average payment, and specify how the Gamma model could represent the amount response.

Count response and rate. The frequency response is a count: 12 claims. The observed exposure rate is $12/6=2$ claims per policy-year. A Poisson model is a natural count-model candidate; the count remains an integer, and the exposure is 6 policy-years.

Amount response. The severity response consists of the positive payment amount for each reported claim. There are 12 such amounts, and their average is total paid divided by claim count: $3600/12=300$. Thus the observed average payment is $300 per claim.

Gamma specification. Let $Y_j$ be the payment amount for claim $j$, in dollars. A possible model is $Y_j\mid x_j\sim\operatorname{Gamma}(a,\text{scale }\mu_j/a)$, with a common shape $a>0$ and a stated mean link such as $\log(\mu_j)=x_j^\mathsf{T}\beta$. Each Gamma response is one positive amount, not the annual count and not the total $3,600 for this group.

Check. The 12 payments average to $300 because $12\times300=3600$, and the frequency rate multiplied by exposure gives $2\times6=12$ claims. The two calculations recover their respective observed totals.

Common mistake avoided. Fitting Gamma regression to the value 12 would model a count as a continuous positive amount. Fitting it to $3,600 would instead change the response to an aggregate that depends on both claim frequency and claim sizes.

Categorical Predictors and Interactions

Gamma regression can use continuous predictors, categorical predictors, and interactions. As in other regression models, categorical predictors are represented with indicator variables relative to a reference category. The response distribution remains Gamma; changing the predictor structure does not turn an amount into a count.

With a log link, each predictor combination determines a linear predictor, and exponentiating it gives the conditional mean in the response’s units. An interaction adds a term for a combination of predictor levels. It allows the effect associated with one factor to differ across levels of another factor. The coefficient should be included in the linear predictor before applying the inverse link.

Worked Example: A Categorical Interaction for Repair Costs

A hypothetical warranty program models positive repair cost in dollars. Let $A=1$ indicate an advanced device and $R=1$ indicate a rural service region; the reference levels are standard device and urban region. The Gamma response has common shape $a=5$, and observations are conditionally independent. The fitted log-link model is $\log(\mu)=\log(400)+A\log(1.2)+R\log(0.9)+AR\log(1.25)$. Calculate the fitted mean for an advanced device in a rural region and compare it with the model’s prediction if the interaction term were omitted.

Substitute the indicator values. For an advanced rural device, $A=1$, $R=1$, and $AR=1$. Thus $\eta=\log(400)+\log(1.2)+\log(0.9)+\log(1.25)$.

Apply the inverse link. Combining the logarithms gives $\mu=\exp(\eta)=400(1.2)(0.9)(1.25)$. First, $400(1.2)=480$. Next, $480(0.9)=432$. Finally, $432(1.25)=540$. The fitted conditional mean repair cost is $540.

Omit the interaction for comparison. Without the $AR$ term, the mean is $400(1.2)(0.9)=432$ dollars. Including the interaction changes this prediction by $540-432=108$ dollars.

Distributional check. Under the stated shape $a=5$, the scale for the advanced rural group is $540/5=108$ dollars, and $5(108)=540$. The calculation is consistent with the stated mean.

Common mistake avoided. Omitting the interaction while calculating the advanced rural prediction would give $432. That is the prediction from the reduced model, not the specified model containing the $AR$ term.

Model Checks and Exam Traps

Before interpreting a Gamma regression, check that the response definition and support match the data. A continuous dollar amount may be positive and right-skewed, but the response might instead include zeros, represent a count, or be an aggregate over varying exposures. Each case changes what the model is being asked to describe. Also verify that predictions are expressed on the response scale: a linear predictor is not itself a fitted mean unless the identity link is explicitly specified.

A common trap is to select a distribution based only on the variable’s units. A dollar value can be an amount per claim, a total over a year, or a count multiplied by a fixed benefit. Those are not interchangeable response definitions. Another trap is to use the Gamma distribution for an outcome that can be zero without explaining how zero values are handled. Since a Gamma random variable has strictly positive support, it assigns no probability to exactly zero.

Observed response	Response type	Modeling implication
Number of claims in a period	Nonnegative integer count	Consider a count model such as Poisson regression
Positive payment for one claim	Strictly positive continuous amount	Gamma regression may be suitable
Total payment including periods with no claims	May include zero and depends on count and amounts	A Gamma model alone may not represent the full response

Key takeaway: A Gamma regression is a model for a strictly positive continuous response. Specify the amount being modeled, its units, the conditional Gamma assumption, the predictors, and the link. Do not substitute a claim count or an aggregate response without recognizing that it is a different modeling problem.

Check Your Understanding

For each problem, identify the response carefully before calculating or specifying a model.

A hypothetical model has $\log(\mu)=\log(250)+\log(1.2)x$, where $x=1$ indicates a certain risk feature. The response is positive payment per event in dollars. What is the fitted conditional mean? (Answer: $300)
A Gamma response has shape $a=3$ and fitted mean $\mu=180$. Under the shape-scale convention used here, what is its scale? Check the mean. (Answer: Scale $=60$; $3(60)=180$)
A group has 15 claims over 5 policy-years, with total positive payments of $4,500. Find its observed claim rate and average payment per claim, and identify which quantity is a count-model response and which is a candidate Gamma response. (Answer: Rate $=3$ claims per policy-year; average payment $=$$300 per claim; claim count is the count response and payment amount is the candidate Gamma response)
A positive repair-cost model has $\log(\mu)=\log(500)+A\log(1.1)+R\log(0.8)+AR\log(1.5)$. Find the mean when $A=1$ and $R=1$. (Answer: $660)
An analyst records the number of calls received each day, including days with zero calls, and proposes a Gamma regression because the values are right-skewed. Is the response suitable as stated, and what is the central modeling correction? (Answer: No; daily calls are nonnegative integer counts and can be zero, so use a count-response model rather than a Gamma model for strictly positive continuous amounts)

What You'll Learn