Identity-Link Gamma Regression

Why the Identity Link Needs a Feasibility Check

A Gamma regression is useful when the response is positive and right-skewed, such as a hypothetical insurance payment amount. With an identity link, the model describes the conditional mean as a linear sum of predictor effects. This makes coefficient interpretation direct: a one-unit predictor increase changes the mean by a fixed amount, all else equal.

That direct interpretation comes with an important restriction. A Gamma response has a strictly positive mean, but a linear predictor can be zero or negative. Therefore, an identity-link Gamma model is appropriate for a predictor range only when its fitted means are strictly positive throughout that range. It is not enough that the model fits the observed cases if it will also be used for extrapolation or for combinations of predictors not present in the data.

Definition: For observation $j$, an identity-link Gamma regression specifies a positive response $Y_j$ with conditional mean $\mu_j=E[Y_j\mid \mathbf{x}_j]$ and variance $\operatorname{Var}(Y_j\mid \mathbf{x}_j)=\phi\mu_j^2$, where $\phi>0$ is the dispersion parameter. The identity link is $g(\mu_j)=\mu_j$, so the linear predictor is $$ \mu_j=\eta_j=\beta_0+\beta_1x_{j1}+\cdots+\beta_px_{jp}. $$ The model requires $\mu_j>0$ for every predictor value at which it is to be used. Under the shape-scale parameterization, $Y_j\mid\mathbf{x}_j$ has Gamma shape $1/\phi$ and scale $\phi\mu_j$. These parameters give mean $\mu_j$ and variance $\phi\mu_j^2$.

Here, $\mathbf{x}_j$ contains the predictors, and the intercept is included in the linear expression. As is usual for a regression model, observations are treated as conditionally independent given their predictors. The model’s positive support and mean assumption distinguish it from a linear model that allows responses and fitted values on the entire real line.

What an Additive Effect Means

For a continuous predictor $x_k$, its coefficient $\beta_k$ is the change in the fitted conditional mean for a one-unit increase in $x_k$, with all other predictors fixed. The change is in the response’s units. For a binary predictor coded 0 or 1, its coefficient is the difference in fitted means between the two categories, holding other predictors fixed. Neither interpretation is automatically a percentage change.

For a contrast that changes several predictors, add their contributions. If the predictor vector changes from $\mathbf{x}$ to $\mathbf{x}+\Delta\mathbf{x}$, then the fitted-mean difference is the sum of each coefficient multiplied by its corresponding predictor change.

\Delta\mu=\sum_{k=1}^{p}\beta_k\Delta x_k

This result follows directly from the identity-link model, not from an approximation. The linear predictor equals the mean, so subtracting the old mean from the new mean cancels the intercept and leaves the sum of coefficient-weighted changes.

Write both fitted means.
For predictor values $\mathbf{x}$ and $\mathbf{x}+\Delta\mathbf{x}$, use the same fitted coefficient values in each linear expression.

Subtract the first mean from the second.
The intercept cancels, and each predictor contributes $\beta_k((x_k+\Delta x_k)-x_k)$.

Simplify the changes.
Since $(x_k+\Delta x_k)-x_k=\Delta x_k$, the difference is $\sum_{k=1}^{p}\beta_k\Delta x_k$.

\begin{aligned} \mu(\mathbf{x}+\Delta\mathbf{x})-\mu(\mathbf{x}) &=\left(\beta_0+\sum_{k=1}^{p}\beta_k(x_k+\Delta x_k)\right) -\left(\beta_0+\sum_{k=1}^{p}\beta_kx_k\right)\\ &=\sum_{k=1}^{p}\beta_k\Delta x_k. \end{aligned}

This derivation also clarifies the units: each coefficient is measured in response units per unit of its predictor. A positive coefficient raises the conditional mean as its predictor increases; a negative coefficient lowers it, provided the comparison stays within the model’s feasible predictor range.

Worked Example: Interpreting a Continuous-Predictor Effect

A hypothetical insurer models positive repair payments in thousands of dollars. For each claim, let $x$ be the vehicle’s age in years. The fitted identity-link Gamma model is $\mu=12+0.4x$, with $\phi=0.25$. Calculate the fitted means for vehicle ages 5 and 15, interpret the change, and give the Gamma parameters at each age. Round fitted means to two decimal places in thousands of dollars.

Method. Substitute each age into the linear mean equation. The conditional distribution has shape $1/\phi=1/0.25=4$, and its scale is $\phi\mu$.

At age 5, $\mu=12+0.4(5)=12+2=14$ thousand dollars. The Gamma shape is 4 and the scale is $0.25(14)=3.5$ thousand dollars.

At age 15, $\mu=12+0.4(15)=12+6=18$ thousand dollars. The Gamma shape is 4 and the scale is $0.25(18)=4.5$ thousand dollars.

The fitted-mean difference is $18-14=4$ thousand dollars. This also equals $0.4(15-5)=4$ thousand dollars. Thus, each additional vehicle-age year is associated with an additive increase of $0.4$ thousand dollars, or $400, in the fitted mean, holding other predictors fixed. Both fitted means are strictly positive. The requested means, to two decimal places in thousands, are 14.00 and 18.00.

Common mistake avoided. Calling the coefficient a 40% increase is incorrect: the coefficient is in thousands of dollars per year, not a percentage. The calculation of the change is $0.4(10)=4$ thousand dollars, not $14(1.40^{10})$ or another compounded amount.

Categorical Predictors and Additive Contrasts

Categorical predictors are represented by indicator variables. Under reference coding, the reference category has all its indicators equal to zero. Each category coefficient then compares that category with the reference category, at the same values of other predictors. If the model includes multiple categorical predictors, the fitted mean for a combination of categories is found by adding the applicable coefficients.

For an identity link, a category effect is a fixed difference in means, not a ratio of means. A fitted mean may be positive for the reference category and still become nonpositive for a particular combination of category indicators. Check the sum for the combination being predicted rather than checking each coefficient in isolation.

Worked Example: Combining Two Categorical Effects

A hypothetical model predicts positive claim payments in thousands of dollars. Let $R=1$ indicate a rural territory and $H=1$ indicate a high deductible; both indicators are 0 otherwise. The fitted model is $\mu=4.8-0.6R+1.1H$, with $\phi=0.16$. Calculate the fitted means for all four combinations, interpret the effects, and check feasibility. Round fitted means to two decimal places in thousands.

Method. Substitute each pair of indicator values. The Gamma shape is $1/0.16=6.25$, and at each combination the scale is $0.16\mu$.

For the reference combination $R=0,H=0$, $\mu=4.8-0.6(0)+1.1(0)=4.8$. The scale is $0.16(4.8)=0.768$ thousand dollars.

For rural territory and no high deductible, $R=1,H=0$, so $\mu=4.8-0.6+0=4.2$. The scale is $0.16(4.2)=0.672$ thousand dollars.

For nonrural territory and a high deductible, $R=0,H=1$, so $\mu=4.8+0+1.1=5.9$. The scale is $0.16(5.9)=0.944$ thousand dollars.

For rural territory and a high deductible, $R=1,H=1$, so $\mu=4.8-0.6+1.1=5.3$. The scale is $0.16(5.3)=0.848$ thousand dollars. All four means are positive, so all four combinations are feasible under this model. The shape is 6.25 for each combination.

Holding the deductible indicator fixed, the rural coefficient changes the fitted mean by $-0.6$ thousand dollars relative to the nonrural category. Holding territory fixed, the high-deductible coefficient changes it by $1.1$ thousand dollars relative to the standard-deductible category. The final means, to two decimal places in thousands, are 4.80, 4.20, 5.90, and 5.30.

Common mistake avoided. Do not interpret $-0.6$ as a percentage reduction or calculate a multiplicative adjustment. It is a difference of $0.6$ thousand dollars in the fitted mean. Also, the rural and high-deductible effects are both included in the rural/high-deductible prediction.

Checking the Feasible Predictor Range

For an identity-link Gamma model, feasibility is a strict inequality: the fitted mean must exceed zero. If a model has one continuous predictor, this condition can often be solved as an interval. With several predictors, it defines a region of combinations. When predictors are restricted to a rectangular range, the smallest fitted mean occurs at a corner: choose the low end for coefficients that are positive and the high end for coefficients that are negative. This follows because a linear function changes monotonically in each predictor when the other predictors are held fixed.

Checking the corners is useful for a rectangular predictor domain, but it does not replace checking the actual domain. If predictors have restrictions that rule out some corners, the relevant minimum must be found over the combinations that are actually possible. And if the model will be used for future predictions, the intended prediction range—not merely the observed range—must be considered.

Worked Example: Finding an Infeasible Predictor Combination

A hypothetical model predicts positive repair costs in thousands of dollars using two continuous risk scores, $x$ and $z$: $\mu=2.4+0.8x-1.1z$. The intended ranges are $0\leq x\leq4$ and $0\leq z\leq3$. Let $\phi=0.20$. Determine whether the whole range is feasible, derive the boundary for $z$, and evaluate $(x,z)=(1,2.5)$ and $(1,3)$. Round fitted means to two decimal places in thousands.

Method. Impose the Gamma-mean condition $\mu>0$ and solve the linear inequality. The Gamma shape, where the mean is feasible, is $1/0.20=5$; its scale is $0.20\mu$.

Starting with $2.4+0.8x-1.1z>0$, add $1.1z$ to both sides and subtract $2.4$: $1.1z<2.4+0.8x$. Since $1.1>0$, division gives $z<(2.4+0.8x)/1.1$. This is a strict boundary: equality gives a zero fitted mean, which is also infeasible.

At $(1,2.5)$, $\mu=2.4+0.8(1)-1.1(2.5)=2.4+0.8-2.75=0.45$ thousand dollars. This is positive. Its Gamma shape is 5, and its scale is $0.20(0.45)=0.09$ thousand dollars.

At $(1,3)$, $\mu=2.4+0.8(1)-1.1(3)=2.4+0.8-3.3=-0.10$ thousand dollars. This is not a valid Gamma mean, so no Gamma distribution with this model’s stated mean and dispersion can be assigned to that predictor combination.

The full intended rectangle is not feasible: at its corner $(0,3)$, the mean is $2.4-3.3=-0.9$. The two requested fitted means, to two decimal places in thousands, are 0.45 and -0.10; only the first is feasible.

Common mistake avoided. A mean of zero is not acceptable merely because it is “nonnegative.” A Gamma response has positive support and requires a strictly positive mean. Thus the boundary itself must be excluded, as must all predictor combinations below it.

Using the Model Responsibly

The identity link is attractive when additive differences in the response scale are meaningful. It can also produce negative fitted values, especially when a negative coefficient is extrapolated to large predictor values or when several negative contributions occur together. Such an output is not a small technical inconvenience: it is incompatible with the assumed Gamma distribution.

A sensible model assessment therefore includes the range of intended use. For a continuous predictor, solve $\mu>0$ for its permissible range. For categorical predictors, enumerate plausible combinations. For multiple continuous predictors over a rectangular domain, inspect the corners; if the domain is constrained, evaluate the minimum over the allowed combinations. If important prediction values fail the positivity condition, the identity-link specification is not suitable for those values. A different modeling specification may be needed, but coefficient interpretation should not be rescued by silently truncating negative fitted means.

Key takeaway: In identity-link Gamma regression, fitted means are linear and predictor effects are additive in the response units. The model is usable only where every fitted mean is strictly positive; check this condition for the entire intended predictor domain, not just for a typical observation.

Check Your Understanding

For each problem, identify both the requested additive interpretation and any feasibility restriction.

A hypothetical Gamma model for positive equipment-repair payments, measured in thousands of dollars, is $\mu=5.5+0.3x$, where $x$ is service age in years and $\phi=0.25$. Find the fitted means at $x=6$ and $x=11$, and state the increase. (Answer: 7.30 and 8.80 thousand dollars; increase 1.50 thousand dollars, or $1,500; both means are feasible.)
A hypothetical model has $\mu=9.0-1.2R+2.5S$, where $R=1$ indicates a rural site and $S=1$ indicates a specialized service. Find the fitted means for rural/standard and urban/specialized cases, and name the reference case. (Answer: rural/standard $=7.8$, urban/specialized $=11.5$, reference urban/standard $=9.0$; all in the stated response units.)
A Gamma identity-link model has $\mu=1.6+0.5a-0.9b$, where $a,b$ are nonnegative risk scores. At $a=2$, find the largest feasible integer value of $b$, and evaluate the fitted mean at $b=3$. (Answer: feasibility requires $b<2.8889$ approximately, so the largest feasible integer is 2; at $b=3$, $\mu=-0.10$, which is infeasible.)
A hypothetical positive-cost model is $\mu=2.2+0.4x-0.7z$ over $0\leq x\leq5$, $0\leq z\leq4$. Is the whole rectangle feasible? At $x=3$, is every allowed $z$ feasible? (Answer: no; the corner $(0,4)$ gives $-0.60$. At $x=3$, $\mu=3.4-0.7z$, so feasibility requires $z<4.8571$ approximately; every allowed $z\leq4$ is feasible.)
A fitted model is $\mu=6+0.5t-1.4G$, where $t$ is a continuous score and $G$ is a 0/1 group indicator. What is the group contrast at fixed $t$, and for which $t$ is the $G=1$ mean feasible? (Answer: the $G=1$ mean is 1.4 response units lower; $\mu=6+0.5t-1.4=4.6+0.5t>0$, so $t>-9.2$.)

What You'll Learn