Multiplicative Interpretation In Gamma Regression

Why Interpret Gamma Coefficients Multiplicatively?

A Gamma regression is useful when the response is a positive, right-skewed amount, such as a hypothetical claim severity or repair cost. With a log link, the model relates predictors to the logarithm of the conditional mean. Its coefficients therefore describe changes in the mean on a multiplicative scale: a predictor can raise or lower the expected amount by a factor, rather than by a fixed number of dollars.

This interpretation depends on reading the model carefully. A coefficient describes a change in the conditional mean, holding the other predictors fixed. The size of the change depends on the predictor’s units, the category used as the reference, and whether the model includes interactions.

Definition: Let $Y$ be a positive response and let $x=(x_1,\ldots,x_p)$ denote its predictors. In a Gamma regression with a log link, the conditional mean $\mu(x)=E[Y\mid x]$ satisfies $$ \log\{\mu(x)\}=\beta_0+\beta_1x_1+\cdots+\beta_px_p, \qquad \mu(x)=\exp\left(\beta_0+\beta_1x_1+\cdots+\beta_px_p\right). $$ The Gamma distribution specifies the response distribution around this mean; the log-link equation specifies how the mean depends on predictors.

For the examples, a Gamma distribution with shape $\alpha$ and scale $\theta$ has mean $\alpha\theta$. We will specify each response as $Y\mid x\sim\operatorname{Gamma}(\alpha,\theta(x))$, where $\theta(x)=\mu(x)/\alpha$. Thus, the stated shape and scale give the stated conditional mean. The coefficient interpretation below concerns that mean; it does not mean that every individual response changes by the same factor.

Continuous Predictors: A Change in Units Gives a Ratio

First consider a model with no interactions involving $x_j$, and hold every predictor other than $x_j$ fixed. If $x_j$ increases by $\Delta$, its contribution to the linear predictor increases by $\beta_j\Delta$. Exponentiating converts that change in the log mean to a ratio of conditional means.

\frac{\mu(x_j+\Delta)}{\mu(x_j)} = \frac{\exp(\eta+\beta_j\Delta)}{\exp(\eta)} = \exp(\beta_j\Delta),

Here, $\eta$ collects the intercept and all other fixed predictor contributions. For a one-unit increase, the ratio is $\exp(\beta_j)$. The corresponding percentage change in the conditional mean is $100\{\exp(\beta_j)-1\}\%$. For a change of $\Delta$ units, use $100\{\exp(\beta_j\Delta)-1\}\%$, not $100\beta_j\Delta\%$ except as a small-effect approximation.

Interpretation rule: In a log-link model without an interaction involving $x_j$, increasing $x_j$ by $\Delta$, with other predictors fixed, multiplies the conditional mean by $\exp(\beta_j\Delta)$. A positive coefficient gives a ratio above 1; a negative coefficient gives a ratio below 1.

Worked Example: A Continuous Predictor and a Categorical Adjustment

Suppose a hypothetical insurer models positive repair costs $Y$, in dollars. Let $z$ be the vehicle’s age in ten-year units, and let $I=1$ for a particular repair category and $I=0$ for the reference category. The fitted mean model is $$ \log\{\mu(z,I)\}=\log(800)+0.04879z+0.22314I. $$ Assume $Y\mid z,I\sim\operatorname{Gamma}(\alpha=4,\theta=\mu(z,I)/4)$. Interpret the age effect, and calculate the fitted mean for a vehicle with $z=2$ in the particular repair category.

Method and units. The age predictor is measured in ten-year units, so a one-unit increase in $z$ means an additional ten years. Holding repair category fixed, its mean ratio is $$ \exp(0.04879)\approx 1.0500. $$ Thus each additional ten years multiplies the conditional mean by about 1.0500, or increases it by about $100(1.0500-1)=5.00\%$. This is a conditional association in the fitted model, not by itself a causal claim.

For $z=2$, the age contribution to the log mean is $0.04879(2)=0.09758$. The category contribution is $0.22314(1)=0.22314$. Therefore the total amount above the intercept is $0.09758+0.22314=0.32072$. Equivalently, the multiplicative factors are $$ \exp(0.09758)\approx1.1025, \qquad \exp(0.22314)\approx1.2500. $$ The fitted conditional mean is $$ \mu(2,1)=800(1.1025)(1.2500)=\$1{,}102.50. $$ The distribution has shape 4 and scale $1{,}102.50/4=\$275.625$, giving mean $4(\$275.625)=\$1{,}102.50$, as required. Money is rounded to the nearest cent. The common mistake avoided here is treating $z=2$ as two years rather than twenty years, because the predictor was defined in ten-year units.

Categorical Predictors: Compare with the Reference Level

A categorical predictor is commonly represented by indicator variables. With a reference category and one coefficient for each other category, each coefficient compares that category’s conditional mean with the reference category, holding the other predictors fixed. Exponentiating the coefficient gives the mean ratio. The reference level has indicator zero, so its coefficient contribution is zero.

For a predictor with several categories, each nonreference indicator is interpreted relative to the same reference level. To compare two nonreference categories, divide their mean ratios, or subtract their log-scale coefficients and exponentiate the difference. This comparison is still conditional on the other predictors.

Coefficient or comparison	Conditional mean ratio	Percentage change
Continuous predictor increases by $\Delta$	$\exp(\beta_j\Delta)$	$100\{\exp(\beta_j\Delta)-1\}\%$
Nonreference category $k$ versus reference, no interaction	$\exp(\beta_k)$	$100\{\exp(\beta_k)-1\}\%$
Category $a$ versus category $b$, no interaction	$\exp(\beta_a-\beta_b)$	$100\{\exp(\beta_a-\beta_b)-1\}\%$

Worked Example: Comparing Three Product Categories

A hypothetical company models positive service costs $Y$, in dollars, for three equipment plans: Standard (the reference), Premium, and Refurbished. Let $I_P$ and $I_R$ be indicators for Premium and Refurbished. The model is $$ \log\{\mu\}=\log(600)+0.26236I_P-0.16252I_R. $$ Assume $Y\mid I_P,I_R\sim\operatorname{Gamma}(\alpha=6,\theta=\mu/6)$. Calculate the fitted mean for each plan and compare Premium with Refurbished.

For Standard, both indicators are zero, so $\mu_S=\exp(\log 600)=\$600.00$. For Premium, $I_P=1$ and $I_R=0$, giving the ratio $\exp(0.26236)\approx1.3000$. Thus $\mu_P=600(1.3000)=\$780.00$, or 30.00% above Standard.

For Refurbished, $I_P=0$ and $I_R=1$. Its ratio to Standard is $\exp(-0.16252)\approx0.8500$, so $\mu_R=600(0.8500)=\$510.00$, or 15.00% below Standard.

To compare Premium directly with Refurbished, divide their means: $$ \frac{\mu_P}{\mu_R}=\frac{780}{510}=1.5294. $$ The same result follows from the coefficient difference: $\exp\{0.26236-(-0.16252)\}=\exp(0.42488)\approx1.5294$. Premium’s fitted mean is therefore about 52.94% higher than Refurbished’s. As a parameter check, the Gamma scale parameters are $780/6=\$130.00$, $510/6=\$85.00$, and $600/6=\$100.00$; multiplying each by shape 6 returns its fitted mean. Money is rounded to the nearest cent. The key trap is comparing each coefficient with zero but then reporting Premium’s comparison with Refurbished as 30.00%; 30.00% compares Premium with Standard, not with Refurbished.

Derivation: Why Exponentiation Gives the Mean Ratio

The multiplicative interpretation follows directly from the log-link equation. Suppose two predictor settings differ only in one continuous predictor $x_j$, which changes from $x_j$ to $x_j+\Delta$. Write the contribution of the intercept and all unchanged predictors as $c$. The two conditional means are $\exp(c+\beta_jx_j)$ and $\exp\{c+\beta_j(x_j+\Delta)\}$. Their ratio is

\frac{\exp\{c+\beta_j(x_j+\Delta)\}}{\exp(c+\beta_jx_j)} = \exp\{c+\beta_jx_j+\beta_j\Delta-c-\beta_jx_j\} = \exp(\beta_j\Delta).

The shared terms cancel, leaving only the effect of the predictor change. For an indicator that changes from zero to one, take $\Delta=1$ in its linear contribution: the ratio is $\exp(\beta)$. This justification requires the stated log-link mean model and a comparison that holds the other modeled predictors fixed. If the model has an interaction, the relevant coefficient contribution can also depend on another predictor; use the full linear predictor for the comparison.

Interactions Change the Effect

An interaction between a continuous predictor $x$ and a binary category indicator $G$ lets the slope for $x$ differ between categories. In the model $\log\{\mu(x,G)\}=\beta_0+\beta_xx+\beta_GG+\beta_{xG}xG$, increasing $x$ by $\Delta$ multiplies the mean by $\exp(\beta_x\Delta)$ when $G=0$, and by $\exp\{(\beta_x+\beta_{xG})\Delta\}$ when $G=1$. At a fixed value $x$, the ratio for $G=1$ versus $G=0$ is $\exp(\beta_G+\beta_{xG}x)$. Consequently, $\exp(\beta_G)$ is the category ratio specifically at $x=0$, not necessarily at every value of $x$.

Worked Example: Interpreting a Continuous-by-Category Interaction

A hypothetical repair program models positive repair amounts $Y$, in dollars. Let $x$ be the number of years above a specified baseline vehicle age, and let $G=1$ indicate a special vehicle class; $G=0$ is the reference class. The model is $$ \log\{\mu(x,G)\}=\log(1200)+0.03x+0.20G+0.01xG. $$ Assume $Y\mid x,G\sim\operatorname{Gamma}(\alpha=9,\theta=\mu(x,G)/9)$. Find the one-year mean ratio in each class, compare the classes at $x=4$, and calculate the fitted mean at $x=4,G=1$.

Reference class slope. When $G=0$, the interaction contribution is zero. A one-year increase in $x$ multiplies the mean by $\exp(0.03)\approx1.0305$, a 3.05% increase.

Special class slope. When $G=1$, the coefficient of $x$ is $0.03+0.01=0.04$. Its one-year mean ratio is $\exp(0.04)\approx1.0408$, a 4.08% increase. The interaction therefore changes the one-year ratio; applying $\exp(0.03)$ to both classes would ignore the interaction.

Class comparison at $x=4$. The log-scale difference for $G=1$ versus $G=0$ is $0.20+0.01(4)=0.24$. The mean ratio is $\exp(0.24)\approx1.2712$, so the special class has a fitted mean about 27.12% higher at this age value. This is not simply $\exp(0.20)$, because the interaction contributes $0.04$.

Fitted mean and check. At $x=4,G=1$, the predictor beyond the intercept is $0.03(4)+0.20+0.01(4)=0.12+0.20+0.04=0.36$. Therefore $\mu(4,1)=1200\exp(0.36)\approx1200(1.4333)=\$1{,}720.00$. The Gamma scale is approximately $1{,}720.00/9=\$191.11$; multiplying by shape 9 gives a mean of approximately $\$1{,}720.00$, with the small discrepancy due to rounding. Money is rounded to the nearest cent. The main trap is interpreting the category coefficient $0.20$ as the category effect at every $x$; it is the effect at $x=0$.

Common Interpretation Traps

Coefficient interpretation is precise only when the comparison is stated precisely: identify the response mean, the predictor change, the other predictors being held fixed, and—when relevant—the reference category. The following errors can produce a numerically neat but incorrect conclusion.

Using the coefficient as a percentage. A coefficient of $0.20$ does not mean exactly a 20% increase. The exact percentage change is $100\{\exp(0.20)-1\}\%=22.14\%$, since $\exp(0.20)\approx1.2214$. The incorrect calculation $100(0.20)=20\%$ is only an approximation.
Calling a ratio an additive dollar change. A factor of 1.25 means the mean is multiplied by 1.25, or is 25% higher. The dollar difference depends on the starting mean: a factor of 1.25 adds $50 to a $200 mean but $250 to a $1,000 mean.
Ignoring predictor units. A coefficient per one thousand dollars, ten years, or one standardized unit must be interpreted using that unit. Rescaling the predictor changes the coefficient’s numerical value but not the fitted mean comparisons when the model is expressed consistently.
Forgetting the reference level or an interaction. A categorical coefficient compares with its model reference. With an interaction, the comparison may depend on the continuous predictor’s value.
Turning association into causation. A fitted conditional mean ratio describes the model’s comparison at fixed covariate values. It does not, without additional design and assumptions, establish that changing the predictor causes the response mean to change.

Key takeaway: For a Gamma mean model with a log link, exponentiate the relevant change in the linear predictor to obtain a conditional mean ratio. Then subtract 1 and multiply by 100 for a percentage change. Include predictor units, reference categories, and any interaction terms in the comparison.

Check Your Understanding

For each problem, interpret or calculate the requested conditional mean comparison. Round percentage changes to two decimal places where needed.

A model has $\log\{\mu\}=5+0.12x$, with no interaction involving $x$. What is the mean ratio for a two-unit increase in $x$, and what percentage change does it represent? (Answer: ratio $\exp(0.24)\approx1.2712$; 27.12% increase.)
A category indicator has coefficient $-0.25$, with another category as the reference and no interaction. What is the ratio and percentage change relative to the reference? (Answer: $\exp(-0.25)\approx0.7788$; 22.12% lower.)
A three-level predictor has coefficients $0.10$ for A and $-0.20$ for B, relative to reference R. What is the fitted mean ratio for A versus B? (Answer: $\exp\{0.10-(-0.20)\}=\exp(0.30)\approx1.3499$; A is 34.99% higher than B.)
A model is $\log\{\mu(x,G)\}=4+0.05x+0.10G+0.02xG$. What is the one-unit mean ratio for $x$ when $G=1$, and what is the ratio for $G=1$ versus $G=0$ at $x=3$? (Answer: $\exp(0.07)\approx1.0725$ for the one-unit increase; $\exp(0.16)\approx1.1735$ for the category comparison at $x=3$.)
A log-link model gives coefficient $0.08$ for a predictor measured in five-year units. What is the percentage change in the fitted mean for an increase of 15 years, holding other predictors fixed? (Answer: $\exp\{0.08(3)\}=\exp(0.24)\approx1.2712$; 27.12% increase.)

Coefficient or comparison	Conditional mean ratio	Percentage change
Continuous predictor increases by \(\Delta\)	\(\exp(\beta_j\Delta)\)	\(100\{\exp(\beta_j\Delta)-1\}\%\)
Nonreference category \(k\) versus reference, no interaction	\(\exp(\beta_k)\)	\(100\{\exp(\beta_k)-1\}\%\)
Category \(a\) versus category \(b\), no interaction	\(\exp(\beta_a-\beta_b)\)	\(100\{\exp(\beta_a-\beta_b)-1\}\%\)

What You'll Learn